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/osdep.h" 26 #include "qapi/error.h" 27 #include "qemu-common.h" 28 #include "block/block_int.h" 29 #include "block/qcow2.h" 30 #include "qemu/range.h" 31 #include "qemu/bswap.h" 32 #include "qemu/cutils.h" 33 34 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size); 35 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs, 36 int64_t offset, int64_t length, uint64_t addend, 37 bool decrease, enum qcow2_discard_type type); 38 39 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index); 40 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index); 41 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index); 42 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index); 43 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index); 44 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index); 45 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index); 46 47 static void set_refcount_ro0(void *refcount_array, uint64_t index, 48 uint64_t value); 49 static void set_refcount_ro1(void *refcount_array, uint64_t index, 50 uint64_t value); 51 static void set_refcount_ro2(void *refcount_array, uint64_t index, 52 uint64_t value); 53 static void set_refcount_ro3(void *refcount_array, uint64_t index, 54 uint64_t value); 55 static void set_refcount_ro4(void *refcount_array, uint64_t index, 56 uint64_t value); 57 static void set_refcount_ro5(void *refcount_array, uint64_t index, 58 uint64_t value); 59 static void set_refcount_ro6(void *refcount_array, uint64_t index, 60 uint64_t value); 61 62 63 static Qcow2GetRefcountFunc *const get_refcount_funcs[] = { 64 &get_refcount_ro0, 65 &get_refcount_ro1, 66 &get_refcount_ro2, 67 &get_refcount_ro3, 68 &get_refcount_ro4, 69 &get_refcount_ro5, 70 &get_refcount_ro6 71 }; 72 73 static Qcow2SetRefcountFunc *const set_refcount_funcs[] = { 74 &set_refcount_ro0, 75 &set_refcount_ro1, 76 &set_refcount_ro2, 77 &set_refcount_ro3, 78 &set_refcount_ro4, 79 &set_refcount_ro5, 80 &set_refcount_ro6 81 }; 82 83 84 /*********************************************************/ 85 /* refcount handling */ 86 87 static void update_max_refcount_table_index(BDRVQcow2State *s) 88 { 89 unsigned i = s->refcount_table_size - 1; 90 while (i > 0 && (s->refcount_table[i] & REFT_OFFSET_MASK) == 0) { 91 i--; 92 } 93 /* Set s->max_refcount_table_index to the index of the last used entry */ 94 s->max_refcount_table_index = i; 95 } 96 97 int qcow2_refcount_init(BlockDriverState *bs) 98 { 99 BDRVQcow2State *s = bs->opaque; 100 unsigned int refcount_table_size2, i; 101 int ret; 102 103 assert(s->refcount_order >= 0 && s->refcount_order <= 6); 104 105 s->get_refcount = get_refcount_funcs[s->refcount_order]; 106 s->set_refcount = set_refcount_funcs[s->refcount_order]; 107 108 assert(s->refcount_table_size <= INT_MAX / sizeof(uint64_t)); 109 refcount_table_size2 = s->refcount_table_size * sizeof(uint64_t); 110 s->refcount_table = g_try_malloc(refcount_table_size2); 111 112 if (s->refcount_table_size > 0) { 113 if (s->refcount_table == NULL) { 114 ret = -ENOMEM; 115 goto fail; 116 } 117 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_LOAD); 118 ret = bdrv_pread(bs->file, s->refcount_table_offset, 119 s->refcount_table, refcount_table_size2); 120 if (ret < 0) { 121 goto fail; 122 } 123 for(i = 0; i < s->refcount_table_size; i++) 124 be64_to_cpus(&s->refcount_table[i]); 125 update_max_refcount_table_index(s); 126 } 127 return 0; 128 fail: 129 return ret; 130 } 131 132 void qcow2_refcount_close(BlockDriverState *bs) 133 { 134 BDRVQcow2State *s = bs->opaque; 135 g_free(s->refcount_table); 136 } 137 138 139 static uint64_t get_refcount_ro0(const void *refcount_array, uint64_t index) 140 { 141 return (((const uint8_t *)refcount_array)[index / 8] >> (index % 8)) & 0x1; 142 } 143 144 static void set_refcount_ro0(void *refcount_array, uint64_t index, 145 uint64_t value) 146 { 147 assert(!(value >> 1)); 148 ((uint8_t *)refcount_array)[index / 8] &= ~(0x1 << (index % 8)); 149 ((uint8_t *)refcount_array)[index / 8] |= value << (index % 8); 150 } 151 152 static uint64_t get_refcount_ro1(const void *refcount_array, uint64_t index) 153 { 154 return (((const uint8_t *)refcount_array)[index / 4] >> (2 * (index % 4))) 155 & 0x3; 156 } 157 158 static void set_refcount_ro1(void *refcount_array, uint64_t index, 159 uint64_t value) 160 { 161 assert(!(value >> 2)); 162 ((uint8_t *)refcount_array)[index / 4] &= ~(0x3 << (2 * (index % 4))); 163 ((uint8_t *)refcount_array)[index / 4] |= value << (2 * (index % 4)); 164 } 165 166 static uint64_t get_refcount_ro2(const void *refcount_array, uint64_t index) 167 { 168 return (((const uint8_t *)refcount_array)[index / 2] >> (4 * (index % 2))) 169 & 0xf; 170 } 171 172 static void set_refcount_ro2(void *refcount_array, uint64_t index, 173 uint64_t value) 174 { 175 assert(!(value >> 4)); 176 ((uint8_t *)refcount_array)[index / 2] &= ~(0xf << (4 * (index % 2))); 177 ((uint8_t *)refcount_array)[index / 2] |= value << (4 * (index % 2)); 178 } 179 180 static uint64_t get_refcount_ro3(const void *refcount_array, uint64_t index) 181 { 182 return ((const uint8_t *)refcount_array)[index]; 183 } 184 185 static void set_refcount_ro3(void *refcount_array, uint64_t index, 186 uint64_t value) 187 { 188 assert(!(value >> 8)); 189 ((uint8_t *)refcount_array)[index] = value; 190 } 191 192 static uint64_t get_refcount_ro4(const void *refcount_array, uint64_t index) 193 { 194 return be16_to_cpu(((const uint16_t *)refcount_array)[index]); 195 } 196 197 static void set_refcount_ro4(void *refcount_array, uint64_t index, 198 uint64_t value) 199 { 200 assert(!(value >> 16)); 201 ((uint16_t *)refcount_array)[index] = cpu_to_be16(value); 202 } 203 204 static uint64_t get_refcount_ro5(const void *refcount_array, uint64_t index) 205 { 206 return be32_to_cpu(((const uint32_t *)refcount_array)[index]); 207 } 208 209 static void set_refcount_ro5(void *refcount_array, uint64_t index, 210 uint64_t value) 211 { 212 assert(!(value >> 32)); 213 ((uint32_t *)refcount_array)[index] = cpu_to_be32(value); 214 } 215 216 static uint64_t get_refcount_ro6(const void *refcount_array, uint64_t index) 217 { 218 return be64_to_cpu(((const uint64_t *)refcount_array)[index]); 219 } 220 221 static void set_refcount_ro6(void *refcount_array, uint64_t index, 222 uint64_t value) 223 { 224 ((uint64_t *)refcount_array)[index] = cpu_to_be64(value); 225 } 226 227 228 static int load_refcount_block(BlockDriverState *bs, 229 int64_t refcount_block_offset, 230 void **refcount_block) 231 { 232 BDRVQcow2State *s = bs->opaque; 233 234 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_LOAD); 235 return qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset, 236 refcount_block); 237 } 238 239 /* 240 * Retrieves the refcount of the cluster given by its index and stores it in 241 * *refcount. Returns 0 on success and -errno on failure. 242 */ 243 int qcow2_get_refcount(BlockDriverState *bs, int64_t cluster_index, 244 uint64_t *refcount) 245 { 246 BDRVQcow2State *s = bs->opaque; 247 uint64_t refcount_table_index, block_index; 248 int64_t refcount_block_offset; 249 int ret; 250 void *refcount_block; 251 252 refcount_table_index = cluster_index >> s->refcount_block_bits; 253 if (refcount_table_index >= s->refcount_table_size) { 254 *refcount = 0; 255 return 0; 256 } 257 refcount_block_offset = 258 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK; 259 if (!refcount_block_offset) { 260 *refcount = 0; 261 return 0; 262 } 263 264 if (offset_into_cluster(s, refcount_block_offset)) { 265 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" PRIx64 266 " unaligned (reftable index: %#" PRIx64 ")", 267 refcount_block_offset, refcount_table_index); 268 return -EIO; 269 } 270 271 ret = qcow2_cache_get(bs, s->refcount_block_cache, refcount_block_offset, 272 &refcount_block); 273 if (ret < 0) { 274 return ret; 275 } 276 277 block_index = cluster_index & (s->refcount_block_size - 1); 278 *refcount = s->get_refcount(refcount_block, block_index); 279 280 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 281 282 return 0; 283 } 284 285 /* Checks if two offsets are described by the same refcount block */ 286 static int in_same_refcount_block(BDRVQcow2State *s, uint64_t offset_a, 287 uint64_t offset_b) 288 { 289 uint64_t block_a = offset_a >> (s->cluster_bits + s->refcount_block_bits); 290 uint64_t block_b = offset_b >> (s->cluster_bits + s->refcount_block_bits); 291 292 return (block_a == block_b); 293 } 294 295 /* 296 * Loads a refcount block. If it doesn't exist yet, it is allocated first 297 * (including growing the refcount table if needed). 298 * 299 * Returns 0 on success or -errno in error case 300 */ 301 static int alloc_refcount_block(BlockDriverState *bs, 302 int64_t cluster_index, void **refcount_block) 303 { 304 BDRVQcow2State *s = bs->opaque; 305 unsigned int refcount_table_index; 306 int64_t ret; 307 308 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC); 309 310 /* Find the refcount block for the given cluster */ 311 refcount_table_index = cluster_index >> s->refcount_block_bits; 312 313 if (refcount_table_index < s->refcount_table_size) { 314 315 uint64_t refcount_block_offset = 316 s->refcount_table[refcount_table_index] & REFT_OFFSET_MASK; 317 318 /* If it's already there, we're done */ 319 if (refcount_block_offset) { 320 if (offset_into_cluster(s, refcount_block_offset)) { 321 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" 322 PRIx64 " unaligned (reftable index: " 323 "%#x)", refcount_block_offset, 324 refcount_table_index); 325 return -EIO; 326 } 327 328 return load_refcount_block(bs, refcount_block_offset, 329 refcount_block); 330 } 331 } 332 333 /* 334 * If we came here, we need to allocate something. Something is at least 335 * a cluster for the new refcount block. It may also include a new refcount 336 * table if the old refcount table is too small. 337 * 338 * Note that allocating clusters here needs some special care: 339 * 340 * - We can't use the normal qcow2_alloc_clusters(), it would try to 341 * increase the refcount and very likely we would end up with an endless 342 * recursion. Instead we must place the refcount blocks in a way that 343 * they can describe them themselves. 344 * 345 * - We need to consider that at this point we are inside update_refcounts 346 * and potentially doing an initial refcount increase. This means that 347 * some clusters have already been allocated by the caller, but their 348 * refcount isn't accurate yet. If we allocate clusters for metadata, we 349 * need to return -EAGAIN to signal the caller that it needs to restart 350 * the search for free clusters. 351 * 352 * - alloc_clusters_noref and qcow2_free_clusters may load a different 353 * refcount block into the cache 354 */ 355 356 *refcount_block = NULL; 357 358 /* We write to the refcount table, so we might depend on L2 tables */ 359 ret = qcow2_cache_flush(bs, s->l2_table_cache); 360 if (ret < 0) { 361 return ret; 362 } 363 364 /* Allocate the refcount block itself and mark it as used */ 365 int64_t new_block = alloc_clusters_noref(bs, s->cluster_size); 366 if (new_block < 0) { 367 return new_block; 368 } 369 370 /* If we're allocating the block at offset 0 then something is wrong */ 371 if (new_block == 0) { 372 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid " 373 "allocation of refcount block at offset 0"); 374 return -EIO; 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; 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; 404 } 405 406 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 407 if (ret < 0) { 408 goto fail; 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; 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; 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; 439 } 440 441 s->refcount_table[refcount_table_index] = new_block; 442 /* If there's a hole in s->refcount_table then it can happen 443 * that refcount_table_index < s->max_refcount_table_index */ 444 s->max_refcount_table_index = 445 MAX(s->max_refcount_table_index, refcount_table_index); 446 447 /* The new refcount block may be where the caller intended to put its 448 * data, so let it restart the search. */ 449 return -EAGAIN; 450 } 451 452 qcow2_cache_put(bs, s->refcount_block_cache, refcount_block); 453 454 /* 455 * If we come here, we need to grow the refcount table. Again, a new 456 * refcount table needs some space and we can't simply allocate to avoid 457 * endless recursion. 458 * 459 * Therefore let's grab new refcount blocks at the end of the image, which 460 * will describe themselves and the new refcount table. This way we can 461 * reference them only in the new table and do the switch to the new 462 * refcount table at once without producing an inconsistent state in 463 * between. 464 */ 465 BLKDBG_EVENT(bs->file, BLKDBG_REFTABLE_GROW); 466 467 /* Calculate the number of refcount blocks needed so far; this will be the 468 * basis for calculating the index of the first cluster used for the 469 * self-describing refcount structures which we are about to create. 470 * 471 * Because we reached this point, there cannot be any refcount entries for 472 * cluster_index or higher indices yet. However, because new_block has been 473 * allocated to describe that cluster (and it will assume this role later 474 * on), we cannot use that index; also, new_block may actually have a higher 475 * cluster index than cluster_index, so it needs to be taken into account 476 * here (and 1 needs to be added to its value because that cluster is used). 477 */ 478 uint64_t blocks_used = DIV_ROUND_UP(MAX(cluster_index + 1, 479 (new_block >> s->cluster_bits) + 1), 480 s->refcount_block_size); 481 482 /* Create the new refcount table and blocks */ 483 uint64_t meta_offset = (blocks_used * s->refcount_block_size) * 484 s->cluster_size; 485 486 ret = qcow2_refcount_area(bs, meta_offset, 0, false, 487 refcount_table_index, new_block); 488 if (ret < 0) { 489 return ret; 490 } 491 492 ret = load_refcount_block(bs, new_block, refcount_block); 493 if (ret < 0) { 494 return ret; 495 } 496 497 /* If we were trying to do the initial refcount update for some cluster 498 * allocation, we might have used the same clusters to store newly 499 * allocated metadata. Make the caller search some new space. */ 500 return -EAGAIN; 501 502 fail: 503 if (*refcount_block != NULL) { 504 qcow2_cache_put(bs, s->refcount_block_cache, refcount_block); 505 } 506 return ret; 507 } 508 509 /* 510 * Starting at @start_offset, this function creates new self-covering refcount 511 * structures: A new refcount table and refcount blocks which cover all of 512 * themselves, and a number of @additional_clusters beyond their end. 513 * @start_offset must be at the end of the image file, that is, there must be 514 * only empty space beyond it. 515 * If @exact_size is false, the refcount table will have 50 % more entries than 516 * necessary so it will not need to grow again soon. 517 * If @new_refblock_offset is not zero, it contains the offset of a refcount 518 * block that should be entered into the new refcount table at index 519 * @new_refblock_index. 520 * 521 * Returns: The offset after the new refcount structures (i.e. where the 522 * @additional_clusters may be placed) on success, -errno on error. 523 */ 524 int64_t qcow2_refcount_area(BlockDriverState *bs, uint64_t start_offset, 525 uint64_t additional_clusters, bool exact_size, 526 int new_refblock_index, 527 uint64_t new_refblock_offset) 528 { 529 BDRVQcow2State *s = bs->opaque; 530 uint64_t total_refblock_count_u64, additional_refblock_count; 531 int total_refblock_count, table_size, area_reftable_index, table_clusters; 532 int i; 533 uint64_t table_offset, block_offset, end_offset; 534 int ret; 535 uint64_t *new_table; 536 537 assert(!(start_offset % s->cluster_size)); 538 539 qcow2_refcount_metadata_size(start_offset / s->cluster_size + 540 additional_clusters, 541 s->cluster_size, s->refcount_order, 542 !exact_size, &total_refblock_count_u64); 543 if (total_refblock_count_u64 > QCOW_MAX_REFTABLE_SIZE) { 544 return -EFBIG; 545 } 546 total_refblock_count = total_refblock_count_u64; 547 548 /* Index in the refcount table of the first refcount block to cover the area 549 * of refcount structures we are about to create; we know that 550 * @total_refblock_count can cover @start_offset, so this will definitely 551 * fit into an int. */ 552 area_reftable_index = (start_offset / s->cluster_size) / 553 s->refcount_block_size; 554 555 if (exact_size) { 556 table_size = total_refblock_count; 557 } else { 558 table_size = total_refblock_count + 559 DIV_ROUND_UP(total_refblock_count, 2); 560 } 561 /* The qcow2 file can only store the reftable size in number of clusters */ 562 table_size = ROUND_UP(table_size, s->cluster_size / sizeof(uint64_t)); 563 table_clusters = (table_size * sizeof(uint64_t)) / s->cluster_size; 564 565 if (table_size > QCOW_MAX_REFTABLE_SIZE) { 566 return -EFBIG; 567 } 568 569 new_table = g_try_new0(uint64_t, table_size); 570 571 assert(table_size > 0); 572 if (new_table == NULL) { 573 ret = -ENOMEM; 574 goto fail; 575 } 576 577 /* Fill the new refcount table */ 578 if (table_size > s->max_refcount_table_index) { 579 /* We're actually growing the reftable */ 580 memcpy(new_table, s->refcount_table, 581 (s->max_refcount_table_index + 1) * sizeof(uint64_t)); 582 } else { 583 /* Improbable case: We're shrinking the reftable. However, the caller 584 * has assured us that there is only empty space beyond @start_offset, 585 * so we can simply drop all of the refblocks that won't fit into the 586 * new reftable. */ 587 memcpy(new_table, s->refcount_table, table_size * sizeof(uint64_t)); 588 } 589 590 if (new_refblock_offset) { 591 assert(new_refblock_index < total_refblock_count); 592 new_table[new_refblock_index] = new_refblock_offset; 593 } 594 595 /* Count how many new refblocks we have to create */ 596 additional_refblock_count = 0; 597 for (i = area_reftable_index; i < total_refblock_count; i++) { 598 if (!new_table[i]) { 599 additional_refblock_count++; 600 } 601 } 602 603 table_offset = start_offset + additional_refblock_count * s->cluster_size; 604 end_offset = table_offset + table_clusters * s->cluster_size; 605 606 /* Fill the refcount blocks, and create new ones, if necessary */ 607 block_offset = start_offset; 608 for (i = area_reftable_index; i < total_refblock_count; i++) { 609 void *refblock_data; 610 uint64_t first_offset_covered; 611 612 /* Reuse an existing refblock if possible, create a new one otherwise */ 613 if (new_table[i]) { 614 ret = qcow2_cache_get(bs, s->refcount_block_cache, new_table[i], 615 &refblock_data); 616 if (ret < 0) { 617 goto fail; 618 } 619 } else { 620 ret = qcow2_cache_get_empty(bs, s->refcount_block_cache, 621 block_offset, &refblock_data); 622 if (ret < 0) { 623 goto fail; 624 } 625 memset(refblock_data, 0, s->cluster_size); 626 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, 627 refblock_data); 628 629 new_table[i] = block_offset; 630 block_offset += s->cluster_size; 631 } 632 633 /* First host offset covered by this refblock */ 634 first_offset_covered = (uint64_t)i * s->refcount_block_size * 635 s->cluster_size; 636 if (first_offset_covered < end_offset) { 637 int j, end_index; 638 639 /* Set the refcount of all of the new refcount structures to 1 */ 640 641 if (first_offset_covered < start_offset) { 642 assert(i == area_reftable_index); 643 j = (start_offset - first_offset_covered) / s->cluster_size; 644 assert(j < s->refcount_block_size); 645 } else { 646 j = 0; 647 } 648 649 end_index = MIN((end_offset - first_offset_covered) / 650 s->cluster_size, 651 s->refcount_block_size); 652 653 for (; j < end_index; j++) { 654 /* The caller guaranteed us this space would be empty */ 655 assert(s->get_refcount(refblock_data, j) == 0); 656 s->set_refcount(refblock_data, j, 1); 657 } 658 659 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, 660 refblock_data); 661 } 662 663 qcow2_cache_put(bs, s->refcount_block_cache, &refblock_data); 664 } 665 666 assert(block_offset == table_offset); 667 668 /* Write refcount blocks to disk */ 669 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_BLOCKS); 670 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 671 if (ret < 0) { 672 goto fail; 673 } 674 675 /* Write refcount table to disk */ 676 for (i = 0; i < total_refblock_count; i++) { 677 cpu_to_be64s(&new_table[i]); 678 } 679 680 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_WRITE_TABLE); 681 ret = bdrv_pwrite_sync(bs->file, table_offset, new_table, 682 table_size * sizeof(uint64_t)); 683 if (ret < 0) { 684 goto fail; 685 } 686 687 for (i = 0; i < total_refblock_count; i++) { 688 be64_to_cpus(&new_table[i]); 689 } 690 691 /* Hook up the new refcount table in the qcow2 header */ 692 struct QEMU_PACKED { 693 uint64_t d64; 694 uint32_t d32; 695 } data; 696 data.d64 = cpu_to_be64(table_offset); 697 data.d32 = cpu_to_be32(table_clusters); 698 BLKDBG_EVENT(bs->file, BLKDBG_REFBLOCK_ALLOC_SWITCH_TABLE); 699 ret = bdrv_pwrite_sync(bs->file, 700 offsetof(QCowHeader, refcount_table_offset), 701 &data, sizeof(data)); 702 if (ret < 0) { 703 goto fail; 704 } 705 706 /* And switch it in memory */ 707 uint64_t old_table_offset = s->refcount_table_offset; 708 uint64_t old_table_size = s->refcount_table_size; 709 710 g_free(s->refcount_table); 711 s->refcount_table = new_table; 712 s->refcount_table_size = table_size; 713 s->refcount_table_offset = table_offset; 714 update_max_refcount_table_index(s); 715 716 /* Free old table. */ 717 qcow2_free_clusters(bs, old_table_offset, old_table_size * sizeof(uint64_t), 718 QCOW2_DISCARD_OTHER); 719 720 return end_offset; 721 722 fail: 723 g_free(new_table); 724 return ret; 725 } 726 727 void qcow2_process_discards(BlockDriverState *bs, int ret) 728 { 729 BDRVQcow2State *s = bs->opaque; 730 Qcow2DiscardRegion *d, *next; 731 732 QTAILQ_FOREACH_SAFE(d, &s->discards, next, next) { 733 QTAILQ_REMOVE(&s->discards, d, next); 734 735 /* Discard is optional, ignore the return value */ 736 if (ret >= 0) { 737 bdrv_pdiscard(bs->file->bs, d->offset, d->bytes); 738 } 739 740 g_free(d); 741 } 742 } 743 744 static void update_refcount_discard(BlockDriverState *bs, 745 uint64_t offset, uint64_t length) 746 { 747 BDRVQcow2State *s = bs->opaque; 748 Qcow2DiscardRegion *d, *p, *next; 749 750 QTAILQ_FOREACH(d, &s->discards, next) { 751 uint64_t new_start = MIN(offset, d->offset); 752 uint64_t new_end = MAX(offset + length, d->offset + d->bytes); 753 754 if (new_end - new_start <= length + d->bytes) { 755 /* There can't be any overlap, areas ending up here have no 756 * references any more and therefore shouldn't get freed another 757 * time. */ 758 assert(d->bytes + length == new_end - new_start); 759 d->offset = new_start; 760 d->bytes = new_end - new_start; 761 goto found; 762 } 763 } 764 765 d = g_malloc(sizeof(*d)); 766 *d = (Qcow2DiscardRegion) { 767 .bs = bs, 768 .offset = offset, 769 .bytes = length, 770 }; 771 QTAILQ_INSERT_TAIL(&s->discards, d, next); 772 773 found: 774 /* Merge discard requests if they are adjacent now */ 775 QTAILQ_FOREACH_SAFE(p, &s->discards, next, next) { 776 if (p == d 777 || p->offset > d->offset + d->bytes 778 || d->offset > p->offset + p->bytes) 779 { 780 continue; 781 } 782 783 /* Still no overlap possible */ 784 assert(p->offset == d->offset + d->bytes 785 || d->offset == p->offset + p->bytes); 786 787 QTAILQ_REMOVE(&s->discards, p, next); 788 d->offset = MIN(d->offset, p->offset); 789 d->bytes += p->bytes; 790 g_free(p); 791 } 792 } 793 794 /* XXX: cache several refcount block clusters ? */ 795 /* @addend is the absolute value of the addend; if @decrease is set, @addend 796 * will be subtracted from the current refcount, otherwise it will be added */ 797 static int QEMU_WARN_UNUSED_RESULT update_refcount(BlockDriverState *bs, 798 int64_t offset, 799 int64_t length, 800 uint64_t addend, 801 bool decrease, 802 enum qcow2_discard_type type) 803 { 804 BDRVQcow2State *s = bs->opaque; 805 int64_t start, last, cluster_offset; 806 void *refcount_block = NULL; 807 int64_t old_table_index = -1; 808 int ret; 809 810 #ifdef DEBUG_ALLOC2 811 fprintf(stderr, "update_refcount: offset=%" PRId64 " size=%" PRId64 812 " addend=%s%" PRIu64 "\n", offset, length, decrease ? "-" : "", 813 addend); 814 #endif 815 if (length < 0) { 816 return -EINVAL; 817 } else if (length == 0) { 818 return 0; 819 } 820 821 if (decrease) { 822 qcow2_cache_set_dependency(bs, s->refcount_block_cache, 823 s->l2_table_cache); 824 } 825 826 start = start_of_cluster(s, offset); 827 last = start_of_cluster(s, offset + length - 1); 828 for(cluster_offset = start; cluster_offset <= last; 829 cluster_offset += s->cluster_size) 830 { 831 int block_index; 832 uint64_t refcount; 833 int64_t cluster_index = cluster_offset >> s->cluster_bits; 834 int64_t table_index = cluster_index >> s->refcount_block_bits; 835 836 /* Load the refcount block and allocate it if needed */ 837 if (table_index != old_table_index) { 838 if (refcount_block) { 839 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 840 } 841 ret = alloc_refcount_block(bs, cluster_index, &refcount_block); 842 if (ret < 0) { 843 goto fail; 844 } 845 } 846 old_table_index = table_index; 847 848 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, 849 refcount_block); 850 851 /* we can update the count and save it */ 852 block_index = cluster_index & (s->refcount_block_size - 1); 853 854 refcount = s->get_refcount(refcount_block, block_index); 855 if (decrease ? (refcount - addend > refcount) 856 : (refcount + addend < refcount || 857 refcount + addend > s->refcount_max)) 858 { 859 ret = -EINVAL; 860 goto fail; 861 } 862 if (decrease) { 863 refcount -= addend; 864 } else { 865 refcount += addend; 866 } 867 if (refcount == 0 && cluster_index < s->free_cluster_index) { 868 s->free_cluster_index = cluster_index; 869 } 870 s->set_refcount(refcount_block, block_index, refcount); 871 872 if (refcount == 0) { 873 void *table; 874 875 table = qcow2_cache_is_table_offset(bs, s->refcount_block_cache, 876 offset); 877 if (table != NULL) { 878 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 879 qcow2_cache_discard(bs, s->refcount_block_cache, table); 880 } 881 882 table = qcow2_cache_is_table_offset(bs, s->l2_table_cache, offset); 883 if (table != NULL) { 884 qcow2_cache_discard(bs, s->l2_table_cache, table); 885 } 886 887 if (s->discard_passthrough[type]) { 888 update_refcount_discard(bs, cluster_offset, s->cluster_size); 889 } 890 } 891 } 892 893 ret = 0; 894 fail: 895 if (!s->cache_discards) { 896 qcow2_process_discards(bs, ret); 897 } 898 899 /* Write last changed block to disk */ 900 if (refcount_block) { 901 qcow2_cache_put(bs, s->refcount_block_cache, &refcount_block); 902 } 903 904 /* 905 * Try do undo any updates if an error is returned (This may succeed in 906 * some cases like ENOSPC for allocating a new refcount block) 907 */ 908 if (ret < 0) { 909 int dummy; 910 dummy = update_refcount(bs, offset, cluster_offset - offset, addend, 911 !decrease, QCOW2_DISCARD_NEVER); 912 (void)dummy; 913 } 914 915 return ret; 916 } 917 918 /* 919 * Increases or decreases the refcount of a given cluster. 920 * 921 * @addend is the absolute value of the addend; if @decrease is set, @addend 922 * will be subtracted from the current refcount, otherwise it will be added. 923 * 924 * On success 0 is returned; on failure -errno is returned. 925 */ 926 int qcow2_update_cluster_refcount(BlockDriverState *bs, 927 int64_t cluster_index, 928 uint64_t addend, bool decrease, 929 enum qcow2_discard_type type) 930 { 931 BDRVQcow2State *s = bs->opaque; 932 int ret; 933 934 ret = update_refcount(bs, cluster_index << s->cluster_bits, 1, addend, 935 decrease, type); 936 if (ret < 0) { 937 return ret; 938 } 939 940 return 0; 941 } 942 943 944 945 /*********************************************************/ 946 /* cluster allocation functions */ 947 948 949 950 /* return < 0 if error */ 951 static int64_t alloc_clusters_noref(BlockDriverState *bs, uint64_t size) 952 { 953 BDRVQcow2State *s = bs->opaque; 954 uint64_t i, nb_clusters, refcount; 955 int ret; 956 957 /* We can't allocate clusters if they may still be queued for discard. */ 958 if (s->cache_discards) { 959 qcow2_process_discards(bs, 0); 960 } 961 962 nb_clusters = size_to_clusters(s, size); 963 retry: 964 for(i = 0; i < nb_clusters; i++) { 965 uint64_t next_cluster_index = s->free_cluster_index++; 966 ret = qcow2_get_refcount(bs, next_cluster_index, &refcount); 967 968 if (ret < 0) { 969 return ret; 970 } else if (refcount != 0) { 971 goto retry; 972 } 973 } 974 975 /* Make sure that all offsets in the "allocated" range are representable 976 * in an int64_t */ 977 if (s->free_cluster_index > 0 && 978 s->free_cluster_index - 1 > (INT64_MAX >> s->cluster_bits)) 979 { 980 return -EFBIG; 981 } 982 983 #ifdef DEBUG_ALLOC2 984 fprintf(stderr, "alloc_clusters: size=%" PRId64 " -> %" PRId64 "\n", 985 size, 986 (s->free_cluster_index - nb_clusters) << s->cluster_bits); 987 #endif 988 return (s->free_cluster_index - nb_clusters) << s->cluster_bits; 989 } 990 991 int64_t qcow2_alloc_clusters(BlockDriverState *bs, uint64_t size) 992 { 993 int64_t offset; 994 int ret; 995 996 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC); 997 do { 998 offset = alloc_clusters_noref(bs, size); 999 if (offset < 0) { 1000 return offset; 1001 } 1002 1003 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER); 1004 } while (ret == -EAGAIN); 1005 1006 if (ret < 0) { 1007 return ret; 1008 } 1009 1010 return offset; 1011 } 1012 1013 int64_t qcow2_alloc_clusters_at(BlockDriverState *bs, uint64_t offset, 1014 int64_t nb_clusters) 1015 { 1016 BDRVQcow2State *s = bs->opaque; 1017 uint64_t cluster_index, refcount; 1018 uint64_t i; 1019 int ret; 1020 1021 assert(nb_clusters >= 0); 1022 if (nb_clusters == 0) { 1023 return 0; 1024 } 1025 1026 do { 1027 /* Check how many clusters there are free */ 1028 cluster_index = offset >> s->cluster_bits; 1029 for(i = 0; i < nb_clusters; i++) { 1030 ret = qcow2_get_refcount(bs, cluster_index++, &refcount); 1031 if (ret < 0) { 1032 return ret; 1033 } else if (refcount != 0) { 1034 break; 1035 } 1036 } 1037 1038 /* And then allocate them */ 1039 ret = update_refcount(bs, offset, i << s->cluster_bits, 1, false, 1040 QCOW2_DISCARD_NEVER); 1041 } while (ret == -EAGAIN); 1042 1043 if (ret < 0) { 1044 return ret; 1045 } 1046 1047 return i; 1048 } 1049 1050 /* only used to allocate compressed sectors. We try to allocate 1051 contiguous sectors. size must be <= cluster_size */ 1052 int64_t qcow2_alloc_bytes(BlockDriverState *bs, int size) 1053 { 1054 BDRVQcow2State *s = bs->opaque; 1055 int64_t offset; 1056 size_t free_in_cluster; 1057 int ret; 1058 1059 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_ALLOC_BYTES); 1060 assert(size > 0 && size <= s->cluster_size); 1061 assert(!s->free_byte_offset || offset_into_cluster(s, s->free_byte_offset)); 1062 1063 offset = s->free_byte_offset; 1064 1065 if (offset) { 1066 uint64_t refcount; 1067 ret = qcow2_get_refcount(bs, offset >> s->cluster_bits, &refcount); 1068 if (ret < 0) { 1069 return ret; 1070 } 1071 1072 if (refcount == s->refcount_max) { 1073 offset = 0; 1074 } 1075 } 1076 1077 free_in_cluster = s->cluster_size - offset_into_cluster(s, offset); 1078 do { 1079 if (!offset || free_in_cluster < size) { 1080 int64_t new_cluster = alloc_clusters_noref(bs, s->cluster_size); 1081 if (new_cluster < 0) { 1082 return new_cluster; 1083 } 1084 1085 if (new_cluster == 0) { 1086 qcow2_signal_corruption(bs, true, -1, -1, "Preventing invalid " 1087 "allocation of compressed cluster " 1088 "at offset 0"); 1089 return -EIO; 1090 } 1091 1092 if (!offset || ROUND_UP(offset, s->cluster_size) != new_cluster) { 1093 offset = new_cluster; 1094 free_in_cluster = s->cluster_size; 1095 } else { 1096 free_in_cluster += s->cluster_size; 1097 } 1098 } 1099 1100 assert(offset); 1101 ret = update_refcount(bs, offset, size, 1, false, QCOW2_DISCARD_NEVER); 1102 if (ret < 0) { 1103 offset = 0; 1104 } 1105 } while (ret == -EAGAIN); 1106 if (ret < 0) { 1107 return ret; 1108 } 1109 1110 /* The cluster refcount was incremented; refcount blocks must be flushed 1111 * before the caller's L2 table updates. */ 1112 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache); 1113 1114 s->free_byte_offset = offset + size; 1115 if (!offset_into_cluster(s, s->free_byte_offset)) { 1116 s->free_byte_offset = 0; 1117 } 1118 1119 return offset; 1120 } 1121 1122 void qcow2_free_clusters(BlockDriverState *bs, 1123 int64_t offset, int64_t size, 1124 enum qcow2_discard_type type) 1125 { 1126 int ret; 1127 1128 BLKDBG_EVENT(bs->file, BLKDBG_CLUSTER_FREE); 1129 ret = update_refcount(bs, offset, size, 1, true, type); 1130 if (ret < 0) { 1131 fprintf(stderr, "qcow2_free_clusters failed: %s\n", strerror(-ret)); 1132 /* TODO Remember the clusters to free them later and avoid leaking */ 1133 } 1134 } 1135 1136 /* 1137 * Free a cluster using its L2 entry (handles clusters of all types, e.g. 1138 * normal cluster, compressed cluster, etc.) 1139 */ 1140 void qcow2_free_any_clusters(BlockDriverState *bs, uint64_t l2_entry, 1141 int nb_clusters, enum qcow2_discard_type type) 1142 { 1143 BDRVQcow2State *s = bs->opaque; 1144 1145 switch (qcow2_get_cluster_type(l2_entry)) { 1146 case QCOW2_CLUSTER_COMPRESSED: 1147 { 1148 int nb_csectors; 1149 nb_csectors = ((l2_entry >> s->csize_shift) & 1150 s->csize_mask) + 1; 1151 qcow2_free_clusters(bs, 1152 (l2_entry & s->cluster_offset_mask) & ~511, 1153 nb_csectors * 512, type); 1154 } 1155 break; 1156 case QCOW2_CLUSTER_NORMAL: 1157 case QCOW2_CLUSTER_ZERO_ALLOC: 1158 if (offset_into_cluster(s, l2_entry & L2E_OFFSET_MASK)) { 1159 qcow2_signal_corruption(bs, false, -1, -1, 1160 "Cannot free unaligned cluster %#llx", 1161 l2_entry & L2E_OFFSET_MASK); 1162 } else { 1163 qcow2_free_clusters(bs, l2_entry & L2E_OFFSET_MASK, 1164 nb_clusters << s->cluster_bits, type); 1165 } 1166 break; 1167 case QCOW2_CLUSTER_ZERO_PLAIN: 1168 case QCOW2_CLUSTER_UNALLOCATED: 1169 break; 1170 default: 1171 abort(); 1172 } 1173 } 1174 1175 1176 1177 /*********************************************************/ 1178 /* snapshots and image creation */ 1179 1180 1181 1182 /* update the refcounts of snapshots and the copied flag */ 1183 int qcow2_update_snapshot_refcount(BlockDriverState *bs, 1184 int64_t l1_table_offset, int l1_size, int addend) 1185 { 1186 BDRVQcow2State *s = bs->opaque; 1187 uint64_t *l1_table, *l2_table, l2_offset, entry, l1_size2, refcount; 1188 bool l1_allocated = false; 1189 int64_t old_entry, old_l2_offset; 1190 int i, j, l1_modified = 0, nb_csectors; 1191 int ret; 1192 1193 assert(addend >= -1 && addend <= 1); 1194 1195 l2_table = NULL; 1196 l1_table = NULL; 1197 l1_size2 = l1_size * sizeof(uint64_t); 1198 1199 s->cache_discards = true; 1200 1201 /* WARNING: qcow2_snapshot_goto relies on this function not using the 1202 * l1_table_offset when it is the current s->l1_table_offset! Be careful 1203 * when changing this! */ 1204 if (l1_table_offset != s->l1_table_offset) { 1205 l1_table = g_try_malloc0(align_offset(l1_size2, 512)); 1206 if (l1_size2 && l1_table == NULL) { 1207 ret = -ENOMEM; 1208 goto fail; 1209 } 1210 l1_allocated = true; 1211 1212 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2); 1213 if (ret < 0) { 1214 goto fail; 1215 } 1216 1217 for (i = 0; i < l1_size; i++) { 1218 be64_to_cpus(&l1_table[i]); 1219 } 1220 } else { 1221 assert(l1_size == s->l1_size); 1222 l1_table = s->l1_table; 1223 l1_allocated = false; 1224 } 1225 1226 for (i = 0; i < l1_size; i++) { 1227 l2_offset = l1_table[i]; 1228 if (l2_offset) { 1229 old_l2_offset = l2_offset; 1230 l2_offset &= L1E_OFFSET_MASK; 1231 1232 if (offset_into_cluster(s, l2_offset)) { 1233 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" 1234 PRIx64 " unaligned (L1 index: %#x)", 1235 l2_offset, i); 1236 ret = -EIO; 1237 goto fail; 1238 } 1239 1240 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, 1241 (void**) &l2_table); 1242 if (ret < 0) { 1243 goto fail; 1244 } 1245 1246 for (j = 0; j < s->l2_size; j++) { 1247 uint64_t cluster_index; 1248 uint64_t offset; 1249 1250 entry = be64_to_cpu(l2_table[j]); 1251 old_entry = entry; 1252 entry &= ~QCOW_OFLAG_COPIED; 1253 offset = entry & L2E_OFFSET_MASK; 1254 1255 switch (qcow2_get_cluster_type(entry)) { 1256 case QCOW2_CLUSTER_COMPRESSED: 1257 nb_csectors = ((entry >> s->csize_shift) & 1258 s->csize_mask) + 1; 1259 if (addend != 0) { 1260 ret = update_refcount(bs, 1261 (entry & s->cluster_offset_mask) & ~511, 1262 nb_csectors * 512, abs(addend), addend < 0, 1263 QCOW2_DISCARD_SNAPSHOT); 1264 if (ret < 0) { 1265 goto fail; 1266 } 1267 } 1268 /* compressed clusters are never modified */ 1269 refcount = 2; 1270 break; 1271 1272 case QCOW2_CLUSTER_NORMAL: 1273 case QCOW2_CLUSTER_ZERO_ALLOC: 1274 if (offset_into_cluster(s, offset)) { 1275 qcow2_signal_corruption(bs, true, -1, -1, "Cluster " 1276 "allocation offset %#" PRIx64 1277 " unaligned (L2 offset: %#" 1278 PRIx64 ", L2 index: %#x)", 1279 offset, l2_offset, j); 1280 ret = -EIO; 1281 goto fail; 1282 } 1283 1284 cluster_index = offset >> s->cluster_bits; 1285 assert(cluster_index); 1286 if (addend != 0) { 1287 ret = qcow2_update_cluster_refcount(bs, 1288 cluster_index, abs(addend), addend < 0, 1289 QCOW2_DISCARD_SNAPSHOT); 1290 if (ret < 0) { 1291 goto fail; 1292 } 1293 } 1294 1295 ret = qcow2_get_refcount(bs, cluster_index, &refcount); 1296 if (ret < 0) { 1297 goto fail; 1298 } 1299 break; 1300 1301 case QCOW2_CLUSTER_ZERO_PLAIN: 1302 case QCOW2_CLUSTER_UNALLOCATED: 1303 refcount = 0; 1304 break; 1305 1306 default: 1307 abort(); 1308 } 1309 1310 if (refcount == 1) { 1311 entry |= QCOW_OFLAG_COPIED; 1312 } 1313 if (entry != old_entry) { 1314 if (addend > 0) { 1315 qcow2_cache_set_dependency(bs, s->l2_table_cache, 1316 s->refcount_block_cache); 1317 } 1318 l2_table[j] = cpu_to_be64(entry); 1319 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, 1320 l2_table); 1321 } 1322 } 1323 1324 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1325 1326 if (addend != 0) { 1327 ret = qcow2_update_cluster_refcount(bs, l2_offset >> 1328 s->cluster_bits, 1329 abs(addend), addend < 0, 1330 QCOW2_DISCARD_SNAPSHOT); 1331 if (ret < 0) { 1332 goto fail; 1333 } 1334 } 1335 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits, 1336 &refcount); 1337 if (ret < 0) { 1338 goto fail; 1339 } else if (refcount == 1) { 1340 l2_offset |= QCOW_OFLAG_COPIED; 1341 } 1342 if (l2_offset != old_l2_offset) { 1343 l1_table[i] = l2_offset; 1344 l1_modified = 1; 1345 } 1346 } 1347 } 1348 1349 ret = bdrv_flush(bs); 1350 fail: 1351 if (l2_table) { 1352 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 1353 } 1354 1355 s->cache_discards = false; 1356 qcow2_process_discards(bs, ret); 1357 1358 /* Update L1 only if it isn't deleted anyway (addend = -1) */ 1359 if (ret == 0 && addend >= 0 && l1_modified) { 1360 for (i = 0; i < l1_size; i++) { 1361 cpu_to_be64s(&l1_table[i]); 1362 } 1363 1364 ret = bdrv_pwrite_sync(bs->file, l1_table_offset, 1365 l1_table, l1_size2); 1366 1367 for (i = 0; i < l1_size; i++) { 1368 be64_to_cpus(&l1_table[i]); 1369 } 1370 } 1371 if (l1_allocated) 1372 g_free(l1_table); 1373 return ret; 1374 } 1375 1376 1377 1378 1379 /*********************************************************/ 1380 /* refcount checking functions */ 1381 1382 1383 static uint64_t refcount_array_byte_size(BDRVQcow2State *s, uint64_t entries) 1384 { 1385 /* This assertion holds because there is no way we can address more than 1386 * 2^(64 - 9) clusters at once (with cluster size 512 = 2^9, and because 1387 * offsets have to be representable in bytes); due to every cluster 1388 * corresponding to one refcount entry, we are well below that limit */ 1389 assert(entries < (UINT64_C(1) << (64 - 9))); 1390 1391 /* Thanks to the assertion this will not overflow, because 1392 * s->refcount_order < 7. 1393 * (note: x << s->refcount_order == x * s->refcount_bits) */ 1394 return DIV_ROUND_UP(entries << s->refcount_order, 8); 1395 } 1396 1397 /** 1398 * Reallocates *array so that it can hold new_size entries. *size must contain 1399 * the current number of entries in *array. If the reallocation fails, *array 1400 * and *size will not be modified and -errno will be returned. If the 1401 * reallocation is successful, *array will be set to the new buffer, *size 1402 * will be set to new_size and 0 will be returned. The size of the reallocated 1403 * refcount array buffer will be aligned to a cluster boundary, and the newly 1404 * allocated area will be zeroed. 1405 */ 1406 static int realloc_refcount_array(BDRVQcow2State *s, void **array, 1407 int64_t *size, int64_t new_size) 1408 { 1409 int64_t old_byte_size, new_byte_size; 1410 void *new_ptr; 1411 1412 /* Round to clusters so the array can be directly written to disk */ 1413 old_byte_size = size_to_clusters(s, refcount_array_byte_size(s, *size)) 1414 * s->cluster_size; 1415 new_byte_size = size_to_clusters(s, refcount_array_byte_size(s, new_size)) 1416 * s->cluster_size; 1417 1418 if (new_byte_size == old_byte_size) { 1419 *size = new_size; 1420 return 0; 1421 } 1422 1423 assert(new_byte_size > 0); 1424 1425 if (new_byte_size > SIZE_MAX) { 1426 return -ENOMEM; 1427 } 1428 1429 new_ptr = g_try_realloc(*array, new_byte_size); 1430 if (!new_ptr) { 1431 return -ENOMEM; 1432 } 1433 1434 if (new_byte_size > old_byte_size) { 1435 memset((char *)new_ptr + old_byte_size, 0, 1436 new_byte_size - old_byte_size); 1437 } 1438 1439 *array = new_ptr; 1440 *size = new_size; 1441 1442 return 0; 1443 } 1444 1445 /* 1446 * Increases the refcount for a range of clusters in a given refcount table. 1447 * This is used to construct a temporary refcount table out of L1 and L2 tables 1448 * which can be compared to the refcount table saved in the image. 1449 * 1450 * Modifies the number of errors in res. 1451 */ 1452 int qcow2_inc_refcounts_imrt(BlockDriverState *bs, BdrvCheckResult *res, 1453 void **refcount_table, 1454 int64_t *refcount_table_size, 1455 int64_t offset, int64_t size) 1456 { 1457 BDRVQcow2State *s = bs->opaque; 1458 uint64_t start, last, cluster_offset, k, refcount; 1459 int ret; 1460 1461 if (size <= 0) { 1462 return 0; 1463 } 1464 1465 start = start_of_cluster(s, offset); 1466 last = start_of_cluster(s, offset + size - 1); 1467 for(cluster_offset = start; cluster_offset <= last; 1468 cluster_offset += s->cluster_size) { 1469 k = cluster_offset >> s->cluster_bits; 1470 if (k >= *refcount_table_size) { 1471 ret = realloc_refcount_array(s, refcount_table, 1472 refcount_table_size, k + 1); 1473 if (ret < 0) { 1474 res->check_errors++; 1475 return ret; 1476 } 1477 } 1478 1479 refcount = s->get_refcount(*refcount_table, k); 1480 if (refcount == s->refcount_max) { 1481 fprintf(stderr, "ERROR: overflow cluster offset=0x%" PRIx64 1482 "\n", cluster_offset); 1483 fprintf(stderr, "Use qemu-img amend to increase the refcount entry " 1484 "width or qemu-img convert to create a clean copy if the " 1485 "image cannot be opened for writing\n"); 1486 res->corruptions++; 1487 continue; 1488 } 1489 s->set_refcount(*refcount_table, k, refcount + 1); 1490 } 1491 1492 return 0; 1493 } 1494 1495 /* Flags for check_refcounts_l1() and check_refcounts_l2() */ 1496 enum { 1497 CHECK_FRAG_INFO = 0x2, /* update BlockFragInfo counters */ 1498 }; 1499 1500 /* 1501 * Increases the refcount in the given refcount table for the all clusters 1502 * referenced in the L2 table. While doing so, performs some checks on L2 1503 * entries. 1504 * 1505 * Returns the number of errors found by the checks or -errno if an internal 1506 * error occurred. 1507 */ 1508 static int check_refcounts_l2(BlockDriverState *bs, BdrvCheckResult *res, 1509 void **refcount_table, 1510 int64_t *refcount_table_size, int64_t l2_offset, 1511 int flags) 1512 { 1513 BDRVQcow2State *s = bs->opaque; 1514 uint64_t *l2_table, l2_entry; 1515 uint64_t next_contiguous_offset = 0; 1516 int i, l2_size, nb_csectors, ret; 1517 1518 /* Read L2 table from disk */ 1519 l2_size = s->l2_size * sizeof(uint64_t); 1520 l2_table = g_malloc(l2_size); 1521 1522 ret = bdrv_pread(bs->file, l2_offset, l2_table, l2_size); 1523 if (ret < 0) { 1524 fprintf(stderr, "ERROR: I/O error in check_refcounts_l2\n"); 1525 res->check_errors++; 1526 goto fail; 1527 } 1528 1529 /* Do the actual checks */ 1530 for(i = 0; i < s->l2_size; i++) { 1531 l2_entry = be64_to_cpu(l2_table[i]); 1532 1533 switch (qcow2_get_cluster_type(l2_entry)) { 1534 case QCOW2_CLUSTER_COMPRESSED: 1535 /* Compressed clusters don't have QCOW_OFLAG_COPIED */ 1536 if (l2_entry & QCOW_OFLAG_COPIED) { 1537 fprintf(stderr, "ERROR: cluster %" PRId64 ": " 1538 "copied flag must never be set for compressed " 1539 "clusters\n", l2_entry >> s->cluster_bits); 1540 l2_entry &= ~QCOW_OFLAG_COPIED; 1541 res->corruptions++; 1542 } 1543 1544 /* Mark cluster as used */ 1545 nb_csectors = ((l2_entry >> s->csize_shift) & 1546 s->csize_mask) + 1; 1547 l2_entry &= s->cluster_offset_mask; 1548 ret = qcow2_inc_refcounts_imrt(bs, res, 1549 refcount_table, refcount_table_size, 1550 l2_entry & ~511, nb_csectors * 512); 1551 if (ret < 0) { 1552 goto fail; 1553 } 1554 1555 if (flags & CHECK_FRAG_INFO) { 1556 res->bfi.allocated_clusters++; 1557 res->bfi.compressed_clusters++; 1558 1559 /* Compressed clusters are fragmented by nature. Since they 1560 * take up sub-sector space but we only have sector granularity 1561 * I/O we need to re-read the same sectors even for adjacent 1562 * compressed clusters. 1563 */ 1564 res->bfi.fragmented_clusters++; 1565 } 1566 break; 1567 1568 case QCOW2_CLUSTER_ZERO_ALLOC: 1569 case QCOW2_CLUSTER_NORMAL: 1570 { 1571 uint64_t offset = l2_entry & L2E_OFFSET_MASK; 1572 1573 if (flags & CHECK_FRAG_INFO) { 1574 res->bfi.allocated_clusters++; 1575 if (next_contiguous_offset && 1576 offset != next_contiguous_offset) { 1577 res->bfi.fragmented_clusters++; 1578 } 1579 next_contiguous_offset = offset + s->cluster_size; 1580 } 1581 1582 /* Mark cluster as used */ 1583 ret = qcow2_inc_refcounts_imrt(bs, res, 1584 refcount_table, refcount_table_size, 1585 offset, s->cluster_size); 1586 if (ret < 0) { 1587 goto fail; 1588 } 1589 1590 /* Correct offsets are cluster aligned */ 1591 if (offset_into_cluster(s, offset)) { 1592 fprintf(stderr, "ERROR offset=%" PRIx64 ": Cluster is not " 1593 "properly aligned; L2 entry corrupted.\n", offset); 1594 res->corruptions++; 1595 } 1596 break; 1597 } 1598 1599 case QCOW2_CLUSTER_ZERO_PLAIN: 1600 case QCOW2_CLUSTER_UNALLOCATED: 1601 break; 1602 1603 default: 1604 abort(); 1605 } 1606 } 1607 1608 g_free(l2_table); 1609 return 0; 1610 1611 fail: 1612 g_free(l2_table); 1613 return ret; 1614 } 1615 1616 /* 1617 * Increases the refcount for the L1 table, its L2 tables and all referenced 1618 * clusters in the given refcount table. While doing so, performs some checks 1619 * on L1 and L2 entries. 1620 * 1621 * Returns the number of errors found by the checks or -errno if an internal 1622 * error occurred. 1623 */ 1624 static int check_refcounts_l1(BlockDriverState *bs, 1625 BdrvCheckResult *res, 1626 void **refcount_table, 1627 int64_t *refcount_table_size, 1628 int64_t l1_table_offset, int l1_size, 1629 int flags) 1630 { 1631 BDRVQcow2State *s = bs->opaque; 1632 uint64_t *l1_table = NULL, l2_offset, l1_size2; 1633 int i, ret; 1634 1635 l1_size2 = l1_size * sizeof(uint64_t); 1636 1637 /* Mark L1 table as used */ 1638 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, refcount_table_size, 1639 l1_table_offset, l1_size2); 1640 if (ret < 0) { 1641 goto fail; 1642 } 1643 1644 /* Read L1 table entries from disk */ 1645 if (l1_size2 > 0) { 1646 l1_table = g_try_malloc(l1_size2); 1647 if (l1_table == NULL) { 1648 ret = -ENOMEM; 1649 res->check_errors++; 1650 goto fail; 1651 } 1652 ret = bdrv_pread(bs->file, l1_table_offset, l1_table, l1_size2); 1653 if (ret < 0) { 1654 fprintf(stderr, "ERROR: I/O error in check_refcounts_l1\n"); 1655 res->check_errors++; 1656 goto fail; 1657 } 1658 for(i = 0;i < l1_size; i++) 1659 be64_to_cpus(&l1_table[i]); 1660 } 1661 1662 /* Do the actual checks */ 1663 for(i = 0; i < l1_size; i++) { 1664 l2_offset = l1_table[i]; 1665 if (l2_offset) { 1666 /* Mark L2 table as used */ 1667 l2_offset &= L1E_OFFSET_MASK; 1668 ret = qcow2_inc_refcounts_imrt(bs, res, 1669 refcount_table, refcount_table_size, 1670 l2_offset, s->cluster_size); 1671 if (ret < 0) { 1672 goto fail; 1673 } 1674 1675 /* L2 tables are cluster aligned */ 1676 if (offset_into_cluster(s, l2_offset)) { 1677 fprintf(stderr, "ERROR l2_offset=%" PRIx64 ": Table is not " 1678 "cluster aligned; L1 entry corrupted\n", l2_offset); 1679 res->corruptions++; 1680 } 1681 1682 /* Process and check L2 entries */ 1683 ret = check_refcounts_l2(bs, res, refcount_table, 1684 refcount_table_size, l2_offset, flags); 1685 if (ret < 0) { 1686 goto fail; 1687 } 1688 } 1689 } 1690 g_free(l1_table); 1691 return 0; 1692 1693 fail: 1694 g_free(l1_table); 1695 return ret; 1696 } 1697 1698 /* 1699 * Checks the OFLAG_COPIED flag for all L1 and L2 entries. 1700 * 1701 * This function does not print an error message nor does it increment 1702 * check_errors if qcow2_get_refcount fails (this is because such an error will 1703 * have been already detected and sufficiently signaled by the calling function 1704 * (qcow2_check_refcounts) by the time this function is called). 1705 */ 1706 static int check_oflag_copied(BlockDriverState *bs, BdrvCheckResult *res, 1707 BdrvCheckMode fix) 1708 { 1709 BDRVQcow2State *s = bs->opaque; 1710 uint64_t *l2_table = qemu_blockalign(bs, s->cluster_size); 1711 int ret; 1712 uint64_t refcount; 1713 int i, j; 1714 1715 for (i = 0; i < s->l1_size; i++) { 1716 uint64_t l1_entry = s->l1_table[i]; 1717 uint64_t l2_offset = l1_entry & L1E_OFFSET_MASK; 1718 bool l2_dirty = false; 1719 1720 if (!l2_offset) { 1721 continue; 1722 } 1723 1724 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits, 1725 &refcount); 1726 if (ret < 0) { 1727 /* don't print message nor increment check_errors */ 1728 continue; 1729 } 1730 if ((refcount == 1) != ((l1_entry & QCOW_OFLAG_COPIED) != 0)) { 1731 fprintf(stderr, "%s OFLAG_COPIED L2 cluster: l1_index=%d " 1732 "l1_entry=%" PRIx64 " refcount=%" PRIu64 "\n", 1733 fix & BDRV_FIX_ERRORS ? "Repairing" : 1734 "ERROR", 1735 i, l1_entry, refcount); 1736 if (fix & BDRV_FIX_ERRORS) { 1737 s->l1_table[i] = refcount == 1 1738 ? l1_entry | QCOW_OFLAG_COPIED 1739 : l1_entry & ~QCOW_OFLAG_COPIED; 1740 ret = qcow2_write_l1_entry(bs, i); 1741 if (ret < 0) { 1742 res->check_errors++; 1743 goto fail; 1744 } 1745 res->corruptions_fixed++; 1746 } else { 1747 res->corruptions++; 1748 } 1749 } 1750 1751 ret = bdrv_pread(bs->file, l2_offset, l2_table, 1752 s->l2_size * sizeof(uint64_t)); 1753 if (ret < 0) { 1754 fprintf(stderr, "ERROR: Could not read L2 table: %s\n", 1755 strerror(-ret)); 1756 res->check_errors++; 1757 goto fail; 1758 } 1759 1760 for (j = 0; j < s->l2_size; j++) { 1761 uint64_t l2_entry = be64_to_cpu(l2_table[j]); 1762 uint64_t data_offset = l2_entry & L2E_OFFSET_MASK; 1763 QCow2ClusterType cluster_type = qcow2_get_cluster_type(l2_entry); 1764 1765 if (cluster_type == QCOW2_CLUSTER_NORMAL || 1766 cluster_type == QCOW2_CLUSTER_ZERO_ALLOC) { 1767 ret = qcow2_get_refcount(bs, 1768 data_offset >> s->cluster_bits, 1769 &refcount); 1770 if (ret < 0) { 1771 /* don't print message nor increment check_errors */ 1772 continue; 1773 } 1774 if ((refcount == 1) != ((l2_entry & QCOW_OFLAG_COPIED) != 0)) { 1775 fprintf(stderr, "%s OFLAG_COPIED data cluster: " 1776 "l2_entry=%" PRIx64 " refcount=%" PRIu64 "\n", 1777 fix & BDRV_FIX_ERRORS ? "Repairing" : 1778 "ERROR", 1779 l2_entry, refcount); 1780 if (fix & BDRV_FIX_ERRORS) { 1781 l2_table[j] = cpu_to_be64(refcount == 1 1782 ? l2_entry | QCOW_OFLAG_COPIED 1783 : l2_entry & ~QCOW_OFLAG_COPIED); 1784 l2_dirty = true; 1785 res->corruptions_fixed++; 1786 } else { 1787 res->corruptions++; 1788 } 1789 } 1790 } 1791 } 1792 1793 if (l2_dirty) { 1794 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L2, 1795 l2_offset, s->cluster_size); 1796 if (ret < 0) { 1797 fprintf(stderr, "ERROR: Could not write L2 table; metadata " 1798 "overlap check failed: %s\n", strerror(-ret)); 1799 res->check_errors++; 1800 goto fail; 1801 } 1802 1803 ret = bdrv_pwrite(bs->file, l2_offset, l2_table, 1804 s->cluster_size); 1805 if (ret < 0) { 1806 fprintf(stderr, "ERROR: Could not write L2 table: %s\n", 1807 strerror(-ret)); 1808 res->check_errors++; 1809 goto fail; 1810 } 1811 } 1812 } 1813 1814 ret = 0; 1815 1816 fail: 1817 qemu_vfree(l2_table); 1818 return ret; 1819 } 1820 1821 /* 1822 * Checks consistency of refblocks and accounts for each refblock in 1823 * *refcount_table. 1824 */ 1825 static int check_refblocks(BlockDriverState *bs, BdrvCheckResult *res, 1826 BdrvCheckMode fix, bool *rebuild, 1827 void **refcount_table, int64_t *nb_clusters) 1828 { 1829 BDRVQcow2State *s = bs->opaque; 1830 int64_t i, size; 1831 int ret; 1832 1833 for(i = 0; i < s->refcount_table_size; i++) { 1834 uint64_t offset, cluster; 1835 offset = s->refcount_table[i]; 1836 cluster = offset >> s->cluster_bits; 1837 1838 /* Refcount blocks are cluster aligned */ 1839 if (offset_into_cluster(s, offset)) { 1840 fprintf(stderr, "ERROR refcount block %" PRId64 " is not " 1841 "cluster aligned; refcount table entry corrupted\n", i); 1842 res->corruptions++; 1843 *rebuild = true; 1844 continue; 1845 } 1846 1847 if (cluster >= *nb_clusters) { 1848 fprintf(stderr, "%s refcount block %" PRId64 " is outside image\n", 1849 fix & BDRV_FIX_ERRORS ? "Repairing" : "ERROR", i); 1850 1851 if (fix & BDRV_FIX_ERRORS) { 1852 int64_t new_nb_clusters; 1853 Error *local_err = NULL; 1854 1855 if (offset > INT64_MAX - s->cluster_size) { 1856 ret = -EINVAL; 1857 goto resize_fail; 1858 } 1859 1860 ret = bdrv_truncate(bs->file, offset + s->cluster_size, 1861 PREALLOC_MODE_OFF, &local_err); 1862 if (ret < 0) { 1863 error_report_err(local_err); 1864 goto resize_fail; 1865 } 1866 size = bdrv_getlength(bs->file->bs); 1867 if (size < 0) { 1868 ret = size; 1869 goto resize_fail; 1870 } 1871 1872 new_nb_clusters = size_to_clusters(s, size); 1873 assert(new_nb_clusters >= *nb_clusters); 1874 1875 ret = realloc_refcount_array(s, refcount_table, 1876 nb_clusters, new_nb_clusters); 1877 if (ret < 0) { 1878 res->check_errors++; 1879 return ret; 1880 } 1881 1882 if (cluster >= *nb_clusters) { 1883 ret = -EINVAL; 1884 goto resize_fail; 1885 } 1886 1887 res->corruptions_fixed++; 1888 ret = qcow2_inc_refcounts_imrt(bs, res, 1889 refcount_table, nb_clusters, 1890 offset, s->cluster_size); 1891 if (ret < 0) { 1892 return ret; 1893 } 1894 /* No need to check whether the refcount is now greater than 1: 1895 * This area was just allocated and zeroed, so it can only be 1896 * exactly 1 after qcow2_inc_refcounts_imrt() */ 1897 continue; 1898 1899 resize_fail: 1900 res->corruptions++; 1901 *rebuild = true; 1902 fprintf(stderr, "ERROR could not resize image: %s\n", 1903 strerror(-ret)); 1904 } else { 1905 res->corruptions++; 1906 } 1907 continue; 1908 } 1909 1910 if (offset != 0) { 1911 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters, 1912 offset, s->cluster_size); 1913 if (ret < 0) { 1914 return ret; 1915 } 1916 if (s->get_refcount(*refcount_table, cluster) != 1) { 1917 fprintf(stderr, "ERROR refcount block %" PRId64 1918 " refcount=%" PRIu64 "\n", i, 1919 s->get_refcount(*refcount_table, cluster)); 1920 res->corruptions++; 1921 *rebuild = true; 1922 } 1923 } 1924 } 1925 1926 return 0; 1927 } 1928 1929 /* 1930 * Calculates an in-memory refcount table. 1931 */ 1932 static int calculate_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 1933 BdrvCheckMode fix, bool *rebuild, 1934 void **refcount_table, int64_t *nb_clusters) 1935 { 1936 BDRVQcow2State *s = bs->opaque; 1937 int64_t i; 1938 QCowSnapshot *sn; 1939 int ret; 1940 1941 if (!*refcount_table) { 1942 int64_t old_size = 0; 1943 ret = realloc_refcount_array(s, refcount_table, 1944 &old_size, *nb_clusters); 1945 if (ret < 0) { 1946 res->check_errors++; 1947 return ret; 1948 } 1949 } 1950 1951 /* header */ 1952 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters, 1953 0, s->cluster_size); 1954 if (ret < 0) { 1955 return ret; 1956 } 1957 1958 /* current L1 table */ 1959 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters, 1960 s->l1_table_offset, s->l1_size, CHECK_FRAG_INFO); 1961 if (ret < 0) { 1962 return ret; 1963 } 1964 1965 /* snapshots */ 1966 for (i = 0; i < s->nb_snapshots; i++) { 1967 sn = s->snapshots + i; 1968 ret = check_refcounts_l1(bs, res, refcount_table, nb_clusters, 1969 sn->l1_table_offset, sn->l1_size, 0); 1970 if (ret < 0) { 1971 return ret; 1972 } 1973 } 1974 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters, 1975 s->snapshots_offset, s->snapshots_size); 1976 if (ret < 0) { 1977 return ret; 1978 } 1979 1980 /* refcount data */ 1981 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters, 1982 s->refcount_table_offset, 1983 s->refcount_table_size * sizeof(uint64_t)); 1984 if (ret < 0) { 1985 return ret; 1986 } 1987 1988 /* encryption */ 1989 if (s->crypto_header.length) { 1990 ret = qcow2_inc_refcounts_imrt(bs, res, refcount_table, nb_clusters, 1991 s->crypto_header.offset, 1992 s->crypto_header.length); 1993 if (ret < 0) { 1994 return ret; 1995 } 1996 } 1997 1998 /* bitmaps */ 1999 ret = qcow2_check_bitmaps_refcounts(bs, res, refcount_table, nb_clusters); 2000 if (ret < 0) { 2001 return ret; 2002 } 2003 2004 return check_refblocks(bs, res, fix, rebuild, refcount_table, nb_clusters); 2005 } 2006 2007 /* 2008 * Compares the actual reference count for each cluster in the image against the 2009 * refcount as reported by the refcount structures on-disk. 2010 */ 2011 static void compare_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 2012 BdrvCheckMode fix, bool *rebuild, 2013 int64_t *highest_cluster, 2014 void *refcount_table, int64_t nb_clusters) 2015 { 2016 BDRVQcow2State *s = bs->opaque; 2017 int64_t i; 2018 uint64_t refcount1, refcount2; 2019 int ret; 2020 2021 for (i = 0, *highest_cluster = 0; i < nb_clusters; i++) { 2022 ret = qcow2_get_refcount(bs, i, &refcount1); 2023 if (ret < 0) { 2024 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n", 2025 i, strerror(-ret)); 2026 res->check_errors++; 2027 continue; 2028 } 2029 2030 refcount2 = s->get_refcount(refcount_table, i); 2031 2032 if (refcount1 > 0 || refcount2 > 0) { 2033 *highest_cluster = i; 2034 } 2035 2036 if (refcount1 != refcount2) { 2037 /* Check if we're allowed to fix the mismatch */ 2038 int *num_fixed = NULL; 2039 if (refcount1 == 0) { 2040 *rebuild = true; 2041 } else if (refcount1 > refcount2 && (fix & BDRV_FIX_LEAKS)) { 2042 num_fixed = &res->leaks_fixed; 2043 } else if (refcount1 < refcount2 && (fix & BDRV_FIX_ERRORS)) { 2044 num_fixed = &res->corruptions_fixed; 2045 } 2046 2047 fprintf(stderr, "%s cluster %" PRId64 " refcount=%" PRIu64 2048 " reference=%" PRIu64 "\n", 2049 num_fixed != NULL ? "Repairing" : 2050 refcount1 < refcount2 ? "ERROR" : 2051 "Leaked", 2052 i, refcount1, refcount2); 2053 2054 if (num_fixed) { 2055 ret = update_refcount(bs, i << s->cluster_bits, 1, 2056 refcount_diff(refcount1, refcount2), 2057 refcount1 > refcount2, 2058 QCOW2_DISCARD_ALWAYS); 2059 if (ret >= 0) { 2060 (*num_fixed)++; 2061 continue; 2062 } 2063 } 2064 2065 /* And if we couldn't, print an error */ 2066 if (refcount1 < refcount2) { 2067 res->corruptions++; 2068 } else { 2069 res->leaks++; 2070 } 2071 } 2072 } 2073 } 2074 2075 /* 2076 * Allocates clusters using an in-memory refcount table (IMRT) in contrast to 2077 * the on-disk refcount structures. 2078 * 2079 * On input, *first_free_cluster tells where to start looking, and need not 2080 * actually be a free cluster; the returned offset will not be before that 2081 * cluster. On output, *first_free_cluster points to the first gap found, even 2082 * if that gap was too small to be used as the returned offset. 2083 * 2084 * Note that *first_free_cluster is a cluster index whereas the return value is 2085 * an offset. 2086 */ 2087 static int64_t alloc_clusters_imrt(BlockDriverState *bs, 2088 int cluster_count, 2089 void **refcount_table, 2090 int64_t *imrt_nb_clusters, 2091 int64_t *first_free_cluster) 2092 { 2093 BDRVQcow2State *s = bs->opaque; 2094 int64_t cluster = *first_free_cluster, i; 2095 bool first_gap = true; 2096 int contiguous_free_clusters; 2097 int ret; 2098 2099 /* Starting at *first_free_cluster, find a range of at least cluster_count 2100 * continuously free clusters */ 2101 for (contiguous_free_clusters = 0; 2102 cluster < *imrt_nb_clusters && 2103 contiguous_free_clusters < cluster_count; 2104 cluster++) 2105 { 2106 if (!s->get_refcount(*refcount_table, cluster)) { 2107 contiguous_free_clusters++; 2108 if (first_gap) { 2109 /* If this is the first free cluster found, update 2110 * *first_free_cluster accordingly */ 2111 *first_free_cluster = cluster; 2112 first_gap = false; 2113 } 2114 } else if (contiguous_free_clusters) { 2115 contiguous_free_clusters = 0; 2116 } 2117 } 2118 2119 /* If contiguous_free_clusters is greater than zero, it contains the number 2120 * of continuously free clusters until the current cluster; the first free 2121 * cluster in the current "gap" is therefore 2122 * cluster - contiguous_free_clusters */ 2123 2124 /* If no such range could be found, grow the in-memory refcount table 2125 * accordingly to append free clusters at the end of the image */ 2126 if (contiguous_free_clusters < cluster_count) { 2127 /* contiguous_free_clusters clusters are already empty at the image end; 2128 * we need cluster_count clusters; therefore, we have to allocate 2129 * cluster_count - contiguous_free_clusters new clusters at the end of 2130 * the image (which is the current value of cluster; note that cluster 2131 * may exceed old_imrt_nb_clusters if *first_free_cluster pointed beyond 2132 * the image end) */ 2133 ret = realloc_refcount_array(s, refcount_table, imrt_nb_clusters, 2134 cluster + cluster_count 2135 - contiguous_free_clusters); 2136 if (ret < 0) { 2137 return ret; 2138 } 2139 } 2140 2141 /* Go back to the first free cluster */ 2142 cluster -= contiguous_free_clusters; 2143 for (i = 0; i < cluster_count; i++) { 2144 s->set_refcount(*refcount_table, cluster + i, 1); 2145 } 2146 2147 return cluster << s->cluster_bits; 2148 } 2149 2150 /* 2151 * Creates a new refcount structure based solely on the in-memory information 2152 * given through *refcount_table. All necessary allocations will be reflected 2153 * in that array. 2154 * 2155 * On success, the old refcount structure is leaked (it will be covered by the 2156 * new refcount structure). 2157 */ 2158 static int rebuild_refcount_structure(BlockDriverState *bs, 2159 BdrvCheckResult *res, 2160 void **refcount_table, 2161 int64_t *nb_clusters) 2162 { 2163 BDRVQcow2State *s = bs->opaque; 2164 int64_t first_free_cluster = 0, reftable_offset = -1, cluster = 0; 2165 int64_t refblock_offset, refblock_start, refblock_index; 2166 uint32_t reftable_size = 0; 2167 uint64_t *on_disk_reftable = NULL; 2168 void *on_disk_refblock; 2169 int ret = 0; 2170 struct { 2171 uint64_t reftable_offset; 2172 uint32_t reftable_clusters; 2173 } QEMU_PACKED reftable_offset_and_clusters; 2174 2175 qcow2_cache_empty(bs, s->refcount_block_cache); 2176 2177 write_refblocks: 2178 for (; cluster < *nb_clusters; cluster++) { 2179 if (!s->get_refcount(*refcount_table, cluster)) { 2180 continue; 2181 } 2182 2183 refblock_index = cluster >> s->refcount_block_bits; 2184 refblock_start = refblock_index << s->refcount_block_bits; 2185 2186 /* Don't allocate a cluster in a refblock already written to disk */ 2187 if (first_free_cluster < refblock_start) { 2188 first_free_cluster = refblock_start; 2189 } 2190 refblock_offset = alloc_clusters_imrt(bs, 1, refcount_table, 2191 nb_clusters, &first_free_cluster); 2192 if (refblock_offset < 0) { 2193 fprintf(stderr, "ERROR allocating refblock: %s\n", 2194 strerror(-refblock_offset)); 2195 res->check_errors++; 2196 ret = refblock_offset; 2197 goto fail; 2198 } 2199 2200 if (reftable_size <= refblock_index) { 2201 uint32_t old_reftable_size = reftable_size; 2202 uint64_t *new_on_disk_reftable; 2203 2204 reftable_size = ROUND_UP((refblock_index + 1) * sizeof(uint64_t), 2205 s->cluster_size) / sizeof(uint64_t); 2206 new_on_disk_reftable = g_try_realloc(on_disk_reftable, 2207 reftable_size * 2208 sizeof(uint64_t)); 2209 if (!new_on_disk_reftable) { 2210 res->check_errors++; 2211 ret = -ENOMEM; 2212 goto fail; 2213 } 2214 on_disk_reftable = new_on_disk_reftable; 2215 2216 memset(on_disk_reftable + old_reftable_size, 0, 2217 (reftable_size - old_reftable_size) * sizeof(uint64_t)); 2218 2219 /* The offset we have for the reftable is now no longer valid; 2220 * this will leak that range, but we can easily fix that by running 2221 * a leak-fixing check after this rebuild operation */ 2222 reftable_offset = -1; 2223 } else { 2224 assert(on_disk_reftable); 2225 } 2226 on_disk_reftable[refblock_index] = refblock_offset; 2227 2228 /* If this is apparently the last refblock (for now), try to squeeze the 2229 * reftable in */ 2230 if (refblock_index == (*nb_clusters - 1) >> s->refcount_block_bits && 2231 reftable_offset < 0) 2232 { 2233 uint64_t reftable_clusters = size_to_clusters(s, reftable_size * 2234 sizeof(uint64_t)); 2235 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters, 2236 refcount_table, nb_clusters, 2237 &first_free_cluster); 2238 if (reftable_offset < 0) { 2239 fprintf(stderr, "ERROR allocating reftable: %s\n", 2240 strerror(-reftable_offset)); 2241 res->check_errors++; 2242 ret = reftable_offset; 2243 goto fail; 2244 } 2245 } 2246 2247 ret = qcow2_pre_write_overlap_check(bs, 0, refblock_offset, 2248 s->cluster_size); 2249 if (ret < 0) { 2250 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret)); 2251 goto fail; 2252 } 2253 2254 /* The size of *refcount_table is always cluster-aligned, therefore the 2255 * write operation will not overflow */ 2256 on_disk_refblock = (void *)((char *) *refcount_table + 2257 refblock_index * s->cluster_size); 2258 2259 ret = bdrv_write(bs->file, refblock_offset / BDRV_SECTOR_SIZE, 2260 on_disk_refblock, s->cluster_sectors); 2261 if (ret < 0) { 2262 fprintf(stderr, "ERROR writing refblock: %s\n", strerror(-ret)); 2263 goto fail; 2264 } 2265 2266 /* Go to the end of this refblock */ 2267 cluster = refblock_start + s->refcount_block_size - 1; 2268 } 2269 2270 if (reftable_offset < 0) { 2271 uint64_t post_refblock_start, reftable_clusters; 2272 2273 post_refblock_start = ROUND_UP(*nb_clusters, s->refcount_block_size); 2274 reftable_clusters = size_to_clusters(s, 2275 reftable_size * sizeof(uint64_t)); 2276 /* Not pretty but simple */ 2277 if (first_free_cluster < post_refblock_start) { 2278 first_free_cluster = post_refblock_start; 2279 } 2280 reftable_offset = alloc_clusters_imrt(bs, reftable_clusters, 2281 refcount_table, nb_clusters, 2282 &first_free_cluster); 2283 if (reftable_offset < 0) { 2284 fprintf(stderr, "ERROR allocating reftable: %s\n", 2285 strerror(-reftable_offset)); 2286 res->check_errors++; 2287 ret = reftable_offset; 2288 goto fail; 2289 } 2290 2291 goto write_refblocks; 2292 } 2293 2294 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) { 2295 cpu_to_be64s(&on_disk_reftable[refblock_index]); 2296 } 2297 2298 ret = qcow2_pre_write_overlap_check(bs, 0, reftable_offset, 2299 reftable_size * sizeof(uint64_t)); 2300 if (ret < 0) { 2301 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret)); 2302 goto fail; 2303 } 2304 2305 assert(reftable_size < INT_MAX / sizeof(uint64_t)); 2306 ret = bdrv_pwrite(bs->file, reftable_offset, on_disk_reftable, 2307 reftable_size * sizeof(uint64_t)); 2308 if (ret < 0) { 2309 fprintf(stderr, "ERROR writing reftable: %s\n", strerror(-ret)); 2310 goto fail; 2311 } 2312 2313 /* Enter new reftable into the image header */ 2314 reftable_offset_and_clusters.reftable_offset = cpu_to_be64(reftable_offset); 2315 reftable_offset_and_clusters.reftable_clusters = 2316 cpu_to_be32(size_to_clusters(s, reftable_size * sizeof(uint64_t))); 2317 ret = bdrv_pwrite_sync(bs->file, 2318 offsetof(QCowHeader, refcount_table_offset), 2319 &reftable_offset_and_clusters, 2320 sizeof(reftable_offset_and_clusters)); 2321 if (ret < 0) { 2322 fprintf(stderr, "ERROR setting reftable: %s\n", strerror(-ret)); 2323 goto fail; 2324 } 2325 2326 for (refblock_index = 0; refblock_index < reftable_size; refblock_index++) { 2327 be64_to_cpus(&on_disk_reftable[refblock_index]); 2328 } 2329 s->refcount_table = on_disk_reftable; 2330 s->refcount_table_offset = reftable_offset; 2331 s->refcount_table_size = reftable_size; 2332 update_max_refcount_table_index(s); 2333 2334 return 0; 2335 2336 fail: 2337 g_free(on_disk_reftable); 2338 return ret; 2339 } 2340 2341 /* 2342 * Checks an image for refcount consistency. 2343 * 2344 * Returns 0 if no errors are found, the number of errors in case the image is 2345 * detected as corrupted, and -errno when an internal error occurred. 2346 */ 2347 int qcow2_check_refcounts(BlockDriverState *bs, BdrvCheckResult *res, 2348 BdrvCheckMode fix) 2349 { 2350 BDRVQcow2State *s = bs->opaque; 2351 BdrvCheckResult pre_compare_res; 2352 int64_t size, highest_cluster, nb_clusters; 2353 void *refcount_table = NULL; 2354 bool rebuild = false; 2355 int ret; 2356 2357 size = bdrv_getlength(bs->file->bs); 2358 if (size < 0) { 2359 res->check_errors++; 2360 return size; 2361 } 2362 2363 nb_clusters = size_to_clusters(s, size); 2364 if (nb_clusters > INT_MAX) { 2365 res->check_errors++; 2366 return -EFBIG; 2367 } 2368 2369 res->bfi.total_clusters = 2370 size_to_clusters(s, bs->total_sectors * BDRV_SECTOR_SIZE); 2371 2372 ret = calculate_refcounts(bs, res, fix, &rebuild, &refcount_table, 2373 &nb_clusters); 2374 if (ret < 0) { 2375 goto fail; 2376 } 2377 2378 /* In case we don't need to rebuild the refcount structure (but want to fix 2379 * something), this function is immediately called again, in which case the 2380 * result should be ignored */ 2381 pre_compare_res = *res; 2382 compare_refcounts(bs, res, 0, &rebuild, &highest_cluster, refcount_table, 2383 nb_clusters); 2384 2385 if (rebuild && (fix & BDRV_FIX_ERRORS)) { 2386 BdrvCheckResult old_res = *res; 2387 int fresh_leaks = 0; 2388 2389 fprintf(stderr, "Rebuilding refcount structure\n"); 2390 ret = rebuild_refcount_structure(bs, res, &refcount_table, 2391 &nb_clusters); 2392 if (ret < 0) { 2393 goto fail; 2394 } 2395 2396 res->corruptions = 0; 2397 res->leaks = 0; 2398 2399 /* Because the old reftable has been exchanged for a new one the 2400 * references have to be recalculated */ 2401 rebuild = false; 2402 memset(refcount_table, 0, refcount_array_byte_size(s, nb_clusters)); 2403 ret = calculate_refcounts(bs, res, 0, &rebuild, &refcount_table, 2404 &nb_clusters); 2405 if (ret < 0) { 2406 goto fail; 2407 } 2408 2409 if (fix & BDRV_FIX_LEAKS) { 2410 /* The old refcount structures are now leaked, fix it; the result 2411 * can be ignored, aside from leaks which were introduced by 2412 * rebuild_refcount_structure() that could not be fixed */ 2413 BdrvCheckResult saved_res = *res; 2414 *res = (BdrvCheckResult){ 0 }; 2415 2416 compare_refcounts(bs, res, BDRV_FIX_LEAKS, &rebuild, 2417 &highest_cluster, refcount_table, nb_clusters); 2418 if (rebuild) { 2419 fprintf(stderr, "ERROR rebuilt refcount structure is still " 2420 "broken\n"); 2421 } 2422 2423 /* Any leaks accounted for here were introduced by 2424 * rebuild_refcount_structure() because that function has created a 2425 * new refcount structure from scratch */ 2426 fresh_leaks = res->leaks; 2427 *res = saved_res; 2428 } 2429 2430 if (res->corruptions < old_res.corruptions) { 2431 res->corruptions_fixed += old_res.corruptions - res->corruptions; 2432 } 2433 if (res->leaks < old_res.leaks) { 2434 res->leaks_fixed += old_res.leaks - res->leaks; 2435 } 2436 res->leaks += fresh_leaks; 2437 } else if (fix) { 2438 if (rebuild) { 2439 fprintf(stderr, "ERROR need to rebuild refcount structures\n"); 2440 res->check_errors++; 2441 ret = -EIO; 2442 goto fail; 2443 } 2444 2445 if (res->leaks || res->corruptions) { 2446 *res = pre_compare_res; 2447 compare_refcounts(bs, res, fix, &rebuild, &highest_cluster, 2448 refcount_table, nb_clusters); 2449 } 2450 } 2451 2452 /* check OFLAG_COPIED */ 2453 ret = check_oflag_copied(bs, res, fix); 2454 if (ret < 0) { 2455 goto fail; 2456 } 2457 2458 res->image_end_offset = (highest_cluster + 1) * s->cluster_size; 2459 ret = 0; 2460 2461 fail: 2462 g_free(refcount_table); 2463 2464 return ret; 2465 } 2466 2467 #define overlaps_with(ofs, sz) \ 2468 ranges_overlap(offset, size, ofs, sz) 2469 2470 /* 2471 * Checks if the given offset into the image file is actually free to use by 2472 * looking for overlaps with important metadata sections (L1/L2 tables etc.), 2473 * i.e. a sanity check without relying on the refcount tables. 2474 * 2475 * The ign parameter specifies what checks not to perform (being a bitmask of 2476 * QCow2MetadataOverlap values), i.e., what sections to ignore. 2477 * 2478 * Returns: 2479 * - 0 if writing to this offset will not affect the mentioned metadata 2480 * - a positive QCow2MetadataOverlap value indicating one overlapping section 2481 * - a negative value (-errno) indicating an error while performing a check, 2482 * e.g. when bdrv_read failed on QCOW2_OL_INACTIVE_L2 2483 */ 2484 int qcow2_check_metadata_overlap(BlockDriverState *bs, int ign, int64_t offset, 2485 int64_t size) 2486 { 2487 BDRVQcow2State *s = bs->opaque; 2488 int chk = s->overlap_check & ~ign; 2489 int i, j; 2490 2491 if (!size) { 2492 return 0; 2493 } 2494 2495 if (chk & QCOW2_OL_MAIN_HEADER) { 2496 if (offset < s->cluster_size) { 2497 return QCOW2_OL_MAIN_HEADER; 2498 } 2499 } 2500 2501 /* align range to test to cluster boundaries */ 2502 size = align_offset(offset_into_cluster(s, offset) + size, s->cluster_size); 2503 offset = start_of_cluster(s, offset); 2504 2505 if ((chk & QCOW2_OL_ACTIVE_L1) && s->l1_size) { 2506 if (overlaps_with(s->l1_table_offset, s->l1_size * sizeof(uint64_t))) { 2507 return QCOW2_OL_ACTIVE_L1; 2508 } 2509 } 2510 2511 if ((chk & QCOW2_OL_REFCOUNT_TABLE) && s->refcount_table_size) { 2512 if (overlaps_with(s->refcount_table_offset, 2513 s->refcount_table_size * sizeof(uint64_t))) { 2514 return QCOW2_OL_REFCOUNT_TABLE; 2515 } 2516 } 2517 2518 if ((chk & QCOW2_OL_SNAPSHOT_TABLE) && s->snapshots_size) { 2519 if (overlaps_with(s->snapshots_offset, s->snapshots_size)) { 2520 return QCOW2_OL_SNAPSHOT_TABLE; 2521 } 2522 } 2523 2524 if ((chk & QCOW2_OL_INACTIVE_L1) && s->snapshots) { 2525 for (i = 0; i < s->nb_snapshots; i++) { 2526 if (s->snapshots[i].l1_size && 2527 overlaps_with(s->snapshots[i].l1_table_offset, 2528 s->snapshots[i].l1_size * sizeof(uint64_t))) { 2529 return QCOW2_OL_INACTIVE_L1; 2530 } 2531 } 2532 } 2533 2534 if ((chk & QCOW2_OL_ACTIVE_L2) && s->l1_table) { 2535 for (i = 0; i < s->l1_size; i++) { 2536 if ((s->l1_table[i] & L1E_OFFSET_MASK) && 2537 overlaps_with(s->l1_table[i] & L1E_OFFSET_MASK, 2538 s->cluster_size)) { 2539 return QCOW2_OL_ACTIVE_L2; 2540 } 2541 } 2542 } 2543 2544 if ((chk & QCOW2_OL_REFCOUNT_BLOCK) && s->refcount_table) { 2545 unsigned last_entry = s->max_refcount_table_index; 2546 assert(last_entry < s->refcount_table_size); 2547 assert(last_entry + 1 == s->refcount_table_size || 2548 (s->refcount_table[last_entry + 1] & REFT_OFFSET_MASK) == 0); 2549 for (i = 0; i <= last_entry; i++) { 2550 if ((s->refcount_table[i] & REFT_OFFSET_MASK) && 2551 overlaps_with(s->refcount_table[i] & REFT_OFFSET_MASK, 2552 s->cluster_size)) { 2553 return QCOW2_OL_REFCOUNT_BLOCK; 2554 } 2555 } 2556 } 2557 2558 if ((chk & QCOW2_OL_INACTIVE_L2) && s->snapshots) { 2559 for (i = 0; i < s->nb_snapshots; i++) { 2560 uint64_t l1_ofs = s->snapshots[i].l1_table_offset; 2561 uint32_t l1_sz = s->snapshots[i].l1_size; 2562 uint64_t l1_sz2 = l1_sz * sizeof(uint64_t); 2563 uint64_t *l1 = g_try_malloc(l1_sz2); 2564 int ret; 2565 2566 if (l1_sz2 && l1 == NULL) { 2567 return -ENOMEM; 2568 } 2569 2570 ret = bdrv_pread(bs->file, l1_ofs, l1, l1_sz2); 2571 if (ret < 0) { 2572 g_free(l1); 2573 return ret; 2574 } 2575 2576 for (j = 0; j < l1_sz; j++) { 2577 uint64_t l2_ofs = be64_to_cpu(l1[j]) & L1E_OFFSET_MASK; 2578 if (l2_ofs && overlaps_with(l2_ofs, s->cluster_size)) { 2579 g_free(l1); 2580 return QCOW2_OL_INACTIVE_L2; 2581 } 2582 } 2583 2584 g_free(l1); 2585 } 2586 } 2587 2588 return 0; 2589 } 2590 2591 static const char *metadata_ol_names[] = { 2592 [QCOW2_OL_MAIN_HEADER_BITNR] = "qcow2_header", 2593 [QCOW2_OL_ACTIVE_L1_BITNR] = "active L1 table", 2594 [QCOW2_OL_ACTIVE_L2_BITNR] = "active L2 table", 2595 [QCOW2_OL_REFCOUNT_TABLE_BITNR] = "refcount table", 2596 [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = "refcount block", 2597 [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = "snapshot table", 2598 [QCOW2_OL_INACTIVE_L1_BITNR] = "inactive L1 table", 2599 [QCOW2_OL_INACTIVE_L2_BITNR] = "inactive L2 table", 2600 }; 2601 2602 /* 2603 * First performs a check for metadata overlaps (through 2604 * qcow2_check_metadata_overlap); if that fails with a negative value (error 2605 * while performing a check), that value is returned. If an impending overlap 2606 * is detected, the BDS will be made unusable, the qcow2 file marked corrupt 2607 * and -EIO returned. 2608 * 2609 * Returns 0 if there were neither overlaps nor errors while checking for 2610 * overlaps; or a negative value (-errno) on error. 2611 */ 2612 int qcow2_pre_write_overlap_check(BlockDriverState *bs, int ign, int64_t offset, 2613 int64_t size) 2614 { 2615 int ret = qcow2_check_metadata_overlap(bs, ign, offset, size); 2616 2617 if (ret < 0) { 2618 return ret; 2619 } else if (ret > 0) { 2620 int metadata_ol_bitnr = ctz32(ret); 2621 assert(metadata_ol_bitnr < QCOW2_OL_MAX_BITNR); 2622 2623 qcow2_signal_corruption(bs, true, offset, size, "Preventing invalid " 2624 "write on metadata (overlaps with %s)", 2625 metadata_ol_names[metadata_ol_bitnr]); 2626 return -EIO; 2627 } 2628 2629 return 0; 2630 } 2631 2632 /* A pointer to a function of this type is given to walk_over_reftable(). That 2633 * function will create refblocks and pass them to a RefblockFinishOp once they 2634 * are completed (@refblock). @refblock_empty is set if the refblock is 2635 * completely empty. 2636 * 2637 * Along with the refblock, a corresponding reftable entry is passed, in the 2638 * reftable @reftable (which may be reallocated) at @reftable_index. 2639 * 2640 * @allocated should be set to true if a new cluster has been allocated. 2641 */ 2642 typedef int (RefblockFinishOp)(BlockDriverState *bs, uint64_t **reftable, 2643 uint64_t reftable_index, uint64_t *reftable_size, 2644 void *refblock, bool refblock_empty, 2645 bool *allocated, Error **errp); 2646 2647 /** 2648 * This "operation" for walk_over_reftable() allocates the refblock on disk (if 2649 * it is not empty) and inserts its offset into the new reftable. The size of 2650 * this new reftable is increased as required. 2651 */ 2652 static int alloc_refblock(BlockDriverState *bs, uint64_t **reftable, 2653 uint64_t reftable_index, uint64_t *reftable_size, 2654 void *refblock, bool refblock_empty, bool *allocated, 2655 Error **errp) 2656 { 2657 BDRVQcow2State *s = bs->opaque; 2658 int64_t offset; 2659 2660 if (!refblock_empty && reftable_index >= *reftable_size) { 2661 uint64_t *new_reftable; 2662 uint64_t new_reftable_size; 2663 2664 new_reftable_size = ROUND_UP(reftable_index + 1, 2665 s->cluster_size / sizeof(uint64_t)); 2666 if (new_reftable_size > QCOW_MAX_REFTABLE_SIZE / sizeof(uint64_t)) { 2667 error_setg(errp, 2668 "This operation would make the refcount table grow " 2669 "beyond the maximum size supported by QEMU, aborting"); 2670 return -ENOTSUP; 2671 } 2672 2673 new_reftable = g_try_realloc(*reftable, new_reftable_size * 2674 sizeof(uint64_t)); 2675 if (!new_reftable) { 2676 error_setg(errp, "Failed to increase reftable buffer size"); 2677 return -ENOMEM; 2678 } 2679 2680 memset(new_reftable + *reftable_size, 0, 2681 (new_reftable_size - *reftable_size) * sizeof(uint64_t)); 2682 2683 *reftable = new_reftable; 2684 *reftable_size = new_reftable_size; 2685 } 2686 2687 if (!refblock_empty && !(*reftable)[reftable_index]) { 2688 offset = qcow2_alloc_clusters(bs, s->cluster_size); 2689 if (offset < 0) { 2690 error_setg_errno(errp, -offset, "Failed to allocate refblock"); 2691 return offset; 2692 } 2693 (*reftable)[reftable_index] = offset; 2694 *allocated = true; 2695 } 2696 2697 return 0; 2698 } 2699 2700 /** 2701 * This "operation" for walk_over_reftable() writes the refblock to disk at the 2702 * offset specified by the new reftable's entry. It does not modify the new 2703 * reftable or change any refcounts. 2704 */ 2705 static int flush_refblock(BlockDriverState *bs, uint64_t **reftable, 2706 uint64_t reftable_index, uint64_t *reftable_size, 2707 void *refblock, bool refblock_empty, bool *allocated, 2708 Error **errp) 2709 { 2710 BDRVQcow2State *s = bs->opaque; 2711 int64_t offset; 2712 int ret; 2713 2714 if (reftable_index < *reftable_size && (*reftable)[reftable_index]) { 2715 offset = (*reftable)[reftable_index]; 2716 2717 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size); 2718 if (ret < 0) { 2719 error_setg_errno(errp, -ret, "Overlap check failed"); 2720 return ret; 2721 } 2722 2723 ret = bdrv_pwrite(bs->file, offset, refblock, s->cluster_size); 2724 if (ret < 0) { 2725 error_setg_errno(errp, -ret, "Failed to write refblock"); 2726 return ret; 2727 } 2728 } else { 2729 assert(refblock_empty); 2730 } 2731 2732 return 0; 2733 } 2734 2735 /** 2736 * This function walks over the existing reftable and every referenced refblock; 2737 * if @new_set_refcount is non-NULL, it is called for every refcount entry to 2738 * create an equal new entry in the passed @new_refblock. Once that 2739 * @new_refblock is completely filled, @operation will be called. 2740 * 2741 * @status_cb and @cb_opaque are used for the amend operation's status callback. 2742 * @index is the index of the walk_over_reftable() calls and @total is the total 2743 * number of walk_over_reftable() calls per amend operation. Both are used for 2744 * calculating the parameters for the status callback. 2745 * 2746 * @allocated is set to true if a new cluster has been allocated. 2747 */ 2748 static int walk_over_reftable(BlockDriverState *bs, uint64_t **new_reftable, 2749 uint64_t *new_reftable_index, 2750 uint64_t *new_reftable_size, 2751 void *new_refblock, int new_refblock_size, 2752 int new_refcount_bits, 2753 RefblockFinishOp *operation, bool *allocated, 2754 Qcow2SetRefcountFunc *new_set_refcount, 2755 BlockDriverAmendStatusCB *status_cb, 2756 void *cb_opaque, int index, int total, 2757 Error **errp) 2758 { 2759 BDRVQcow2State *s = bs->opaque; 2760 uint64_t reftable_index; 2761 bool new_refblock_empty = true; 2762 int refblock_index; 2763 int new_refblock_index = 0; 2764 int ret; 2765 2766 for (reftable_index = 0; reftable_index < s->refcount_table_size; 2767 reftable_index++) 2768 { 2769 uint64_t refblock_offset = s->refcount_table[reftable_index] 2770 & REFT_OFFSET_MASK; 2771 2772 status_cb(bs, (uint64_t)index * s->refcount_table_size + reftable_index, 2773 (uint64_t)total * s->refcount_table_size, cb_opaque); 2774 2775 if (refblock_offset) { 2776 void *refblock; 2777 2778 if (offset_into_cluster(s, refblock_offset)) { 2779 qcow2_signal_corruption(bs, true, -1, -1, "Refblock offset %#" 2780 PRIx64 " unaligned (reftable index: %#" 2781 PRIx64 ")", refblock_offset, 2782 reftable_index); 2783 error_setg(errp, 2784 "Image is corrupt (unaligned refblock offset)"); 2785 return -EIO; 2786 } 2787 2788 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offset, 2789 &refblock); 2790 if (ret < 0) { 2791 error_setg_errno(errp, -ret, "Failed to retrieve refblock"); 2792 return ret; 2793 } 2794 2795 for (refblock_index = 0; refblock_index < s->refcount_block_size; 2796 refblock_index++) 2797 { 2798 uint64_t refcount; 2799 2800 if (new_refblock_index >= new_refblock_size) { 2801 /* new_refblock is now complete */ 2802 ret = operation(bs, new_reftable, *new_reftable_index, 2803 new_reftable_size, new_refblock, 2804 new_refblock_empty, allocated, errp); 2805 if (ret < 0) { 2806 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 2807 return ret; 2808 } 2809 2810 (*new_reftable_index)++; 2811 new_refblock_index = 0; 2812 new_refblock_empty = true; 2813 } 2814 2815 refcount = s->get_refcount(refblock, refblock_index); 2816 if (new_refcount_bits < 64 && refcount >> new_refcount_bits) { 2817 uint64_t offset; 2818 2819 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 2820 2821 offset = ((reftable_index << s->refcount_block_bits) 2822 + refblock_index) << s->cluster_bits; 2823 2824 error_setg(errp, "Cannot decrease refcount entry width to " 2825 "%i bits: Cluster at offset %#" PRIx64 " has a " 2826 "refcount of %" PRIu64, new_refcount_bits, 2827 offset, refcount); 2828 return -EINVAL; 2829 } 2830 2831 if (new_set_refcount) { 2832 new_set_refcount(new_refblock, new_refblock_index++, 2833 refcount); 2834 } else { 2835 new_refblock_index++; 2836 } 2837 new_refblock_empty = new_refblock_empty && refcount == 0; 2838 } 2839 2840 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 2841 } else { 2842 /* No refblock means every refcount is 0 */ 2843 for (refblock_index = 0; refblock_index < s->refcount_block_size; 2844 refblock_index++) 2845 { 2846 if (new_refblock_index >= new_refblock_size) { 2847 /* new_refblock is now complete */ 2848 ret = operation(bs, new_reftable, *new_reftable_index, 2849 new_reftable_size, new_refblock, 2850 new_refblock_empty, allocated, errp); 2851 if (ret < 0) { 2852 return ret; 2853 } 2854 2855 (*new_reftable_index)++; 2856 new_refblock_index = 0; 2857 new_refblock_empty = true; 2858 } 2859 2860 if (new_set_refcount) { 2861 new_set_refcount(new_refblock, new_refblock_index++, 0); 2862 } else { 2863 new_refblock_index++; 2864 } 2865 } 2866 } 2867 } 2868 2869 if (new_refblock_index > 0) { 2870 /* Complete the potentially existing partially filled final refblock */ 2871 if (new_set_refcount) { 2872 for (; new_refblock_index < new_refblock_size; 2873 new_refblock_index++) 2874 { 2875 new_set_refcount(new_refblock, new_refblock_index, 0); 2876 } 2877 } 2878 2879 ret = operation(bs, new_reftable, *new_reftable_index, 2880 new_reftable_size, new_refblock, new_refblock_empty, 2881 allocated, errp); 2882 if (ret < 0) { 2883 return ret; 2884 } 2885 2886 (*new_reftable_index)++; 2887 } 2888 2889 status_cb(bs, (uint64_t)(index + 1) * s->refcount_table_size, 2890 (uint64_t)total * s->refcount_table_size, cb_opaque); 2891 2892 return 0; 2893 } 2894 2895 int qcow2_change_refcount_order(BlockDriverState *bs, int refcount_order, 2896 BlockDriverAmendStatusCB *status_cb, 2897 void *cb_opaque, Error **errp) 2898 { 2899 BDRVQcow2State *s = bs->opaque; 2900 Qcow2GetRefcountFunc *new_get_refcount; 2901 Qcow2SetRefcountFunc *new_set_refcount; 2902 void *new_refblock = qemu_blockalign(bs->file->bs, s->cluster_size); 2903 uint64_t *new_reftable = NULL, new_reftable_size = 0; 2904 uint64_t *old_reftable, old_reftable_size, old_reftable_offset; 2905 uint64_t new_reftable_index = 0; 2906 uint64_t i; 2907 int64_t new_reftable_offset = 0, allocated_reftable_size = 0; 2908 int new_refblock_size, new_refcount_bits = 1 << refcount_order; 2909 int old_refcount_order; 2910 int walk_index = 0; 2911 int ret; 2912 bool new_allocation; 2913 2914 assert(s->qcow_version >= 3); 2915 assert(refcount_order >= 0 && refcount_order <= 6); 2916 2917 /* see qcow2_open() */ 2918 new_refblock_size = 1 << (s->cluster_bits - (refcount_order - 3)); 2919 2920 new_get_refcount = get_refcount_funcs[refcount_order]; 2921 new_set_refcount = set_refcount_funcs[refcount_order]; 2922 2923 2924 do { 2925 int total_walks; 2926 2927 new_allocation = false; 2928 2929 /* At least we have to do this walk and the one which writes the 2930 * refblocks; also, at least we have to do this loop here at least 2931 * twice (normally), first to do the allocations, and second to 2932 * determine that everything is correctly allocated, this then makes 2933 * three walks in total */ 2934 total_walks = MAX(walk_index + 2, 3); 2935 2936 /* First, allocate the structures so they are present in the refcount 2937 * structures */ 2938 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index, 2939 &new_reftable_size, NULL, new_refblock_size, 2940 new_refcount_bits, &alloc_refblock, 2941 &new_allocation, NULL, status_cb, cb_opaque, 2942 walk_index++, total_walks, errp); 2943 if (ret < 0) { 2944 goto done; 2945 } 2946 2947 new_reftable_index = 0; 2948 2949 if (new_allocation) { 2950 if (new_reftable_offset) { 2951 qcow2_free_clusters(bs, new_reftable_offset, 2952 allocated_reftable_size * sizeof(uint64_t), 2953 QCOW2_DISCARD_NEVER); 2954 } 2955 2956 new_reftable_offset = qcow2_alloc_clusters(bs, new_reftable_size * 2957 sizeof(uint64_t)); 2958 if (new_reftable_offset < 0) { 2959 error_setg_errno(errp, -new_reftable_offset, 2960 "Failed to allocate the new reftable"); 2961 ret = new_reftable_offset; 2962 goto done; 2963 } 2964 allocated_reftable_size = new_reftable_size; 2965 } 2966 } while (new_allocation); 2967 2968 /* Second, write the new refblocks */ 2969 ret = walk_over_reftable(bs, &new_reftable, &new_reftable_index, 2970 &new_reftable_size, new_refblock, 2971 new_refblock_size, new_refcount_bits, 2972 &flush_refblock, &new_allocation, new_set_refcount, 2973 status_cb, cb_opaque, walk_index, walk_index + 1, 2974 errp); 2975 if (ret < 0) { 2976 goto done; 2977 } 2978 assert(!new_allocation); 2979 2980 2981 /* Write the new reftable */ 2982 ret = qcow2_pre_write_overlap_check(bs, 0, new_reftable_offset, 2983 new_reftable_size * sizeof(uint64_t)); 2984 if (ret < 0) { 2985 error_setg_errno(errp, -ret, "Overlap check failed"); 2986 goto done; 2987 } 2988 2989 for (i = 0; i < new_reftable_size; i++) { 2990 cpu_to_be64s(&new_reftable[i]); 2991 } 2992 2993 ret = bdrv_pwrite(bs->file, new_reftable_offset, new_reftable, 2994 new_reftable_size * sizeof(uint64_t)); 2995 2996 for (i = 0; i < new_reftable_size; i++) { 2997 be64_to_cpus(&new_reftable[i]); 2998 } 2999 3000 if (ret < 0) { 3001 error_setg_errno(errp, -ret, "Failed to write the new reftable"); 3002 goto done; 3003 } 3004 3005 3006 /* Empty the refcount cache */ 3007 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 3008 if (ret < 0) { 3009 error_setg_errno(errp, -ret, "Failed to flush the refblock cache"); 3010 goto done; 3011 } 3012 3013 /* Update the image header to point to the new reftable; this only updates 3014 * the fields which are relevant to qcow2_update_header(); other fields 3015 * such as s->refcount_table or s->refcount_bits stay stale for now 3016 * (because we have to restore everything if qcow2_update_header() fails) */ 3017 old_refcount_order = s->refcount_order; 3018 old_reftable_size = s->refcount_table_size; 3019 old_reftable_offset = s->refcount_table_offset; 3020 3021 s->refcount_order = refcount_order; 3022 s->refcount_table_size = new_reftable_size; 3023 s->refcount_table_offset = new_reftable_offset; 3024 3025 ret = qcow2_update_header(bs); 3026 if (ret < 0) { 3027 s->refcount_order = old_refcount_order; 3028 s->refcount_table_size = old_reftable_size; 3029 s->refcount_table_offset = old_reftable_offset; 3030 error_setg_errno(errp, -ret, "Failed to update the qcow2 header"); 3031 goto done; 3032 } 3033 3034 /* Now update the rest of the in-memory information */ 3035 old_reftable = s->refcount_table; 3036 s->refcount_table = new_reftable; 3037 update_max_refcount_table_index(s); 3038 3039 s->refcount_bits = 1 << refcount_order; 3040 s->refcount_max = UINT64_C(1) << (s->refcount_bits - 1); 3041 s->refcount_max += s->refcount_max - 1; 3042 3043 s->refcount_block_bits = s->cluster_bits - (refcount_order - 3); 3044 s->refcount_block_size = 1 << s->refcount_block_bits; 3045 3046 s->get_refcount = new_get_refcount; 3047 s->set_refcount = new_set_refcount; 3048 3049 /* For cleaning up all old refblocks and the old reftable below the "done" 3050 * label */ 3051 new_reftable = old_reftable; 3052 new_reftable_size = old_reftable_size; 3053 new_reftable_offset = old_reftable_offset; 3054 3055 done: 3056 if (new_reftable) { 3057 /* On success, new_reftable actually points to the old reftable (and 3058 * new_reftable_size is the old reftable's size); but that is just 3059 * fine */ 3060 for (i = 0; i < new_reftable_size; i++) { 3061 uint64_t offset = new_reftable[i] & REFT_OFFSET_MASK; 3062 if (offset) { 3063 qcow2_free_clusters(bs, offset, s->cluster_size, 3064 QCOW2_DISCARD_OTHER); 3065 } 3066 } 3067 g_free(new_reftable); 3068 3069 if (new_reftable_offset > 0) { 3070 qcow2_free_clusters(bs, new_reftable_offset, 3071 new_reftable_size * sizeof(uint64_t), 3072 QCOW2_DISCARD_OTHER); 3073 } 3074 } 3075 3076 qemu_vfree(new_refblock); 3077 return ret; 3078 } 3079 3080 static int qcow2_discard_refcount_block(BlockDriverState *bs, 3081 uint64_t discard_block_offs) 3082 { 3083 BDRVQcow2State *s = bs->opaque; 3084 uint64_t refblock_offs = get_refblock_offset(s, discard_block_offs); 3085 uint64_t cluster_index = discard_block_offs >> s->cluster_bits; 3086 uint32_t block_index = cluster_index & (s->refcount_block_size - 1); 3087 void *refblock; 3088 int ret; 3089 3090 assert(discard_block_offs != 0); 3091 3092 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs, 3093 &refblock); 3094 if (ret < 0) { 3095 return ret; 3096 } 3097 3098 if (s->get_refcount(refblock, block_index) != 1) { 3099 qcow2_signal_corruption(bs, true, -1, -1, "Invalid refcount:" 3100 " refblock offset %#" PRIx64 3101 ", reftable index %u" 3102 ", block offset %#" PRIx64 3103 ", refcount %#" PRIx64, 3104 refblock_offs, 3105 offset_to_reftable_index(s, discard_block_offs), 3106 discard_block_offs, 3107 s->get_refcount(refblock, block_index)); 3108 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 3109 return -EINVAL; 3110 } 3111 s->set_refcount(refblock, block_index, 0); 3112 3113 qcow2_cache_entry_mark_dirty(bs, s->refcount_block_cache, refblock); 3114 3115 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 3116 3117 if (cluster_index < s->free_cluster_index) { 3118 s->free_cluster_index = cluster_index; 3119 } 3120 3121 refblock = qcow2_cache_is_table_offset(bs, s->refcount_block_cache, 3122 discard_block_offs); 3123 if (refblock) { 3124 /* discard refblock from the cache if refblock is cached */ 3125 qcow2_cache_discard(bs, s->refcount_block_cache, refblock); 3126 } 3127 update_refcount_discard(bs, discard_block_offs, s->cluster_size); 3128 3129 return 0; 3130 } 3131 3132 int qcow2_shrink_reftable(BlockDriverState *bs) 3133 { 3134 BDRVQcow2State *s = bs->opaque; 3135 uint64_t *reftable_tmp = 3136 g_malloc(s->refcount_table_size * sizeof(uint64_t)); 3137 int i, ret; 3138 3139 for (i = 0; i < s->refcount_table_size; i++) { 3140 int64_t refblock_offs = s->refcount_table[i] & REFT_OFFSET_MASK; 3141 void *refblock; 3142 bool unused_block; 3143 3144 if (refblock_offs == 0) { 3145 reftable_tmp[i] = 0; 3146 continue; 3147 } 3148 ret = qcow2_cache_get(bs, s->refcount_block_cache, refblock_offs, 3149 &refblock); 3150 if (ret < 0) { 3151 goto out; 3152 } 3153 3154 /* the refblock has own reference */ 3155 if (i == offset_to_reftable_index(s, refblock_offs)) { 3156 uint64_t block_index = (refblock_offs >> s->cluster_bits) & 3157 (s->refcount_block_size - 1); 3158 uint64_t refcount = s->get_refcount(refblock, block_index); 3159 3160 s->set_refcount(refblock, block_index, 0); 3161 3162 unused_block = buffer_is_zero(refblock, s->cluster_size); 3163 3164 s->set_refcount(refblock, block_index, refcount); 3165 } else { 3166 unused_block = buffer_is_zero(refblock, s->cluster_size); 3167 } 3168 qcow2_cache_put(bs, s->refcount_block_cache, &refblock); 3169 3170 reftable_tmp[i] = unused_block ? 0 : cpu_to_be64(s->refcount_table[i]); 3171 } 3172 3173 ret = bdrv_pwrite_sync(bs->file, s->refcount_table_offset, reftable_tmp, 3174 s->refcount_table_size * sizeof(uint64_t)); 3175 /* 3176 * If the write in the reftable failed the image may contain a partially 3177 * overwritten reftable. In this case it would be better to clear the 3178 * reftable in memory to avoid possible image corruption. 3179 */ 3180 for (i = 0; i < s->refcount_table_size; i++) { 3181 if (s->refcount_table[i] && !reftable_tmp[i]) { 3182 if (ret == 0) { 3183 ret = qcow2_discard_refcount_block(bs, s->refcount_table[i] & 3184 REFT_OFFSET_MASK); 3185 } 3186 s->refcount_table[i] = 0; 3187 } 3188 } 3189 3190 if (!s->cache_discards) { 3191 qcow2_process_discards(bs, ret); 3192 } 3193 3194 out: 3195 g_free(reftable_tmp); 3196 return ret; 3197 } 3198 3199 int64_t qcow2_get_last_cluster(BlockDriverState *bs, int64_t size) 3200 { 3201 BDRVQcow2State *s = bs->opaque; 3202 int64_t i; 3203 3204 for (i = size_to_clusters(s, size) - 1; i >= 0; i--) { 3205 uint64_t refcount; 3206 int ret = qcow2_get_refcount(bs, i, &refcount); 3207 if (ret < 0) { 3208 fprintf(stderr, "Can't get refcount for cluster %" PRId64 ": %s\n", 3209 i, strerror(-ret)); 3210 return ret; 3211 } 3212 if (refcount > 0) { 3213 return i; 3214 } 3215 } 3216 qcow2_signal_corruption(bs, true, -1, -1, 3217 "There are no references in the refcount table."); 3218 return -EIO; 3219 } 3220