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 <zlib.h> 27 28 #include "qapi/error.h" 29 #include "qemu-common.h" 30 #include "block/block_int.h" 31 #include "block/qcow2.h" 32 #include "qemu/bswap.h" 33 #include "trace.h" 34 35 int qcow2_grow_l1_table(BlockDriverState *bs, uint64_t min_size, 36 bool exact_size) 37 { 38 BDRVQcow2State *s = bs->opaque; 39 int new_l1_size2, ret, i; 40 uint64_t *new_l1_table; 41 int64_t old_l1_table_offset, old_l1_size; 42 int64_t new_l1_table_offset, new_l1_size; 43 uint8_t data[12]; 44 45 if (min_size <= s->l1_size) 46 return 0; 47 48 /* Do a sanity check on min_size before trying to calculate new_l1_size 49 * (this prevents overflows during the while loop for the calculation of 50 * new_l1_size) */ 51 if (min_size > INT_MAX / sizeof(uint64_t)) { 52 return -EFBIG; 53 } 54 55 if (exact_size) { 56 new_l1_size = min_size; 57 } else { 58 /* Bump size up to reduce the number of times we have to grow */ 59 new_l1_size = s->l1_size; 60 if (new_l1_size == 0) { 61 new_l1_size = 1; 62 } 63 while (min_size > new_l1_size) { 64 new_l1_size = (new_l1_size * 3 + 1) / 2; 65 } 66 } 67 68 QEMU_BUILD_BUG_ON(QCOW_MAX_L1_SIZE > INT_MAX); 69 if (new_l1_size > QCOW_MAX_L1_SIZE / sizeof(uint64_t)) { 70 return -EFBIG; 71 } 72 73 #ifdef DEBUG_ALLOC2 74 fprintf(stderr, "grow l1_table from %d to %" PRId64 "\n", 75 s->l1_size, new_l1_size); 76 #endif 77 78 new_l1_size2 = sizeof(uint64_t) * new_l1_size; 79 new_l1_table = qemu_try_blockalign(bs->file->bs, 80 align_offset(new_l1_size2, 512)); 81 if (new_l1_table == NULL) { 82 return -ENOMEM; 83 } 84 memset(new_l1_table, 0, align_offset(new_l1_size2, 512)); 85 86 if (s->l1_size) { 87 memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t)); 88 } 89 90 /* write new table (align to cluster) */ 91 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE); 92 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2); 93 if (new_l1_table_offset < 0) { 94 qemu_vfree(new_l1_table); 95 return new_l1_table_offset; 96 } 97 98 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 99 if (ret < 0) { 100 goto fail; 101 } 102 103 /* the L1 position has not yet been updated, so these clusters must 104 * indeed be completely free */ 105 ret = qcow2_pre_write_overlap_check(bs, 0, new_l1_table_offset, 106 new_l1_size2); 107 if (ret < 0) { 108 goto fail; 109 } 110 111 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE); 112 for(i = 0; i < s->l1_size; i++) 113 new_l1_table[i] = cpu_to_be64(new_l1_table[i]); 114 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, 115 new_l1_table, new_l1_size2); 116 if (ret < 0) 117 goto fail; 118 for(i = 0; i < s->l1_size; i++) 119 new_l1_table[i] = be64_to_cpu(new_l1_table[i]); 120 121 /* set new table */ 122 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE); 123 stl_be_p(data, new_l1_size); 124 stq_be_p(data + 4, new_l1_table_offset); 125 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), 126 data, sizeof(data)); 127 if (ret < 0) { 128 goto fail; 129 } 130 qemu_vfree(s->l1_table); 131 old_l1_table_offset = s->l1_table_offset; 132 s->l1_table_offset = new_l1_table_offset; 133 s->l1_table = new_l1_table; 134 old_l1_size = s->l1_size; 135 s->l1_size = new_l1_size; 136 qcow2_free_clusters(bs, old_l1_table_offset, old_l1_size * sizeof(uint64_t), 137 QCOW2_DISCARD_OTHER); 138 return 0; 139 fail: 140 qemu_vfree(new_l1_table); 141 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2, 142 QCOW2_DISCARD_OTHER); 143 return ret; 144 } 145 146 /* 147 * l2_load 148 * 149 * Loads a L2 table into memory. If the table is in the cache, the cache 150 * is used; otherwise the L2 table is loaded from the image file. 151 * 152 * Returns a pointer to the L2 table on success, or NULL if the read from 153 * the image file failed. 154 */ 155 156 static int l2_load(BlockDriverState *bs, uint64_t l2_offset, 157 uint64_t **l2_table) 158 { 159 BDRVQcow2State *s = bs->opaque; 160 161 return qcow2_cache_get(bs, s->l2_table_cache, l2_offset, 162 (void **)l2_table); 163 } 164 165 /* 166 * Writes one sector of the L1 table to the disk (can't update single entries 167 * and we really don't want bdrv_pread to perform a read-modify-write) 168 */ 169 #define L1_ENTRIES_PER_SECTOR (512 / 8) 170 int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index) 171 { 172 BDRVQcow2State *s = bs->opaque; 173 uint64_t buf[L1_ENTRIES_PER_SECTOR] = { 0 }; 174 int l1_start_index; 175 int i, ret; 176 177 l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1); 178 for (i = 0; i < L1_ENTRIES_PER_SECTOR && l1_start_index + i < s->l1_size; 179 i++) 180 { 181 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]); 182 } 183 184 ret = qcow2_pre_write_overlap_check(bs, QCOW2_OL_ACTIVE_L1, 185 s->l1_table_offset + 8 * l1_start_index, sizeof(buf)); 186 if (ret < 0) { 187 return ret; 188 } 189 190 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE); 191 ret = bdrv_pwrite_sync(bs->file, 192 s->l1_table_offset + 8 * l1_start_index, 193 buf, sizeof(buf)); 194 if (ret < 0) { 195 return ret; 196 } 197 198 return 0; 199 } 200 201 /* 202 * l2_allocate 203 * 204 * Allocate a new l2 entry in the file. If l1_index points to an already 205 * used entry in the L2 table (i.e. we are doing a copy on write for the L2 206 * table) copy the contents of the old L2 table into the newly allocated one. 207 * Otherwise the new table is initialized with zeros. 208 * 209 */ 210 211 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table) 212 { 213 BDRVQcow2State *s = bs->opaque; 214 uint64_t old_l2_offset; 215 uint64_t *l2_table = NULL; 216 int64_t l2_offset; 217 int ret; 218 219 old_l2_offset = s->l1_table[l1_index]; 220 221 trace_qcow2_l2_allocate(bs, l1_index); 222 223 /* allocate a new l2 entry */ 224 225 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t)); 226 if (l2_offset < 0) { 227 ret = l2_offset; 228 goto fail; 229 } 230 231 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 232 if (ret < 0) { 233 goto fail; 234 } 235 236 /* allocate a new entry in the l2 cache */ 237 238 trace_qcow2_l2_allocate_get_empty(bs, l1_index); 239 ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table); 240 if (ret < 0) { 241 goto fail; 242 } 243 244 l2_table = *table; 245 246 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) { 247 /* if there was no old l2 table, clear the new table */ 248 memset(l2_table, 0, s->l2_size * sizeof(uint64_t)); 249 } else { 250 uint64_t* old_table; 251 252 /* if there was an old l2 table, read it from the disk */ 253 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ); 254 ret = qcow2_cache_get(bs, s->l2_table_cache, 255 old_l2_offset & L1E_OFFSET_MASK, 256 (void**) &old_table); 257 if (ret < 0) { 258 goto fail; 259 } 260 261 memcpy(l2_table, old_table, s->cluster_size); 262 263 qcow2_cache_put(bs, s->l2_table_cache, (void **) &old_table); 264 } 265 266 /* write the l2 table to the file */ 267 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE); 268 269 trace_qcow2_l2_allocate_write_l2(bs, l1_index); 270 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 271 ret = qcow2_cache_flush(bs, s->l2_table_cache); 272 if (ret < 0) { 273 goto fail; 274 } 275 276 /* update the L1 entry */ 277 trace_qcow2_l2_allocate_write_l1(bs, l1_index); 278 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED; 279 ret = qcow2_write_l1_entry(bs, l1_index); 280 if (ret < 0) { 281 goto fail; 282 } 283 284 *table = l2_table; 285 trace_qcow2_l2_allocate_done(bs, l1_index, 0); 286 return 0; 287 288 fail: 289 trace_qcow2_l2_allocate_done(bs, l1_index, ret); 290 if (l2_table != NULL) { 291 qcow2_cache_put(bs, s->l2_table_cache, (void**) table); 292 } 293 s->l1_table[l1_index] = old_l2_offset; 294 if (l2_offset > 0) { 295 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t), 296 QCOW2_DISCARD_ALWAYS); 297 } 298 return ret; 299 } 300 301 /* 302 * Checks how many clusters in a given L2 table are contiguous in the image 303 * file. As soon as one of the flags in the bitmask stop_flags changes compared 304 * to the first cluster, the search is stopped and the cluster is not counted 305 * as contiguous. (This allows it, for example, to stop at the first compressed 306 * cluster which may require a different handling) 307 */ 308 static int count_contiguous_clusters(int nb_clusters, int cluster_size, 309 uint64_t *l2_table, uint64_t stop_flags) 310 { 311 int i; 312 QCow2ClusterType first_cluster_type; 313 uint64_t mask = stop_flags | L2E_OFFSET_MASK | QCOW_OFLAG_COMPRESSED; 314 uint64_t first_entry = be64_to_cpu(l2_table[0]); 315 uint64_t offset = first_entry & mask; 316 317 if (!offset) { 318 return 0; 319 } 320 321 /* must be allocated */ 322 first_cluster_type = qcow2_get_cluster_type(first_entry); 323 assert(first_cluster_type == QCOW2_CLUSTER_NORMAL || 324 first_cluster_type == QCOW2_CLUSTER_ZERO_ALLOC); 325 326 for (i = 0; i < nb_clusters; i++) { 327 uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask; 328 if (offset + (uint64_t) i * cluster_size != l2_entry) { 329 break; 330 } 331 } 332 333 return i; 334 } 335 336 /* 337 * Checks how many consecutive unallocated clusters in a given L2 338 * table have the same cluster type. 339 */ 340 static int count_contiguous_clusters_unallocated(int nb_clusters, 341 uint64_t *l2_table, 342 QCow2ClusterType wanted_type) 343 { 344 int i; 345 346 assert(wanted_type == QCOW2_CLUSTER_ZERO_PLAIN || 347 wanted_type == QCOW2_CLUSTER_UNALLOCATED); 348 for (i = 0; i < nb_clusters; i++) { 349 uint64_t entry = be64_to_cpu(l2_table[i]); 350 QCow2ClusterType type = qcow2_get_cluster_type(entry); 351 352 if (type != wanted_type) { 353 break; 354 } 355 } 356 357 return i; 358 } 359 360 /* The crypt function is compatible with the linux cryptoloop 361 algorithm for < 4 GB images. NOTE: out_buf == in_buf is 362 supported */ 363 int qcow2_encrypt_sectors(BDRVQcow2State *s, int64_t sector_num, 364 uint8_t *out_buf, const uint8_t *in_buf, 365 int nb_sectors, bool enc, 366 Error **errp) 367 { 368 union { 369 uint64_t ll[2]; 370 uint8_t b[16]; 371 } ivec; 372 int i; 373 int ret; 374 375 for(i = 0; i < nb_sectors; i++) { 376 ivec.ll[0] = cpu_to_le64(sector_num); 377 ivec.ll[1] = 0; 378 if (qcrypto_cipher_setiv(s->cipher, 379 ivec.b, G_N_ELEMENTS(ivec.b), 380 errp) < 0) { 381 return -1; 382 } 383 if (enc) { 384 ret = qcrypto_cipher_encrypt(s->cipher, 385 in_buf, 386 out_buf, 387 512, 388 errp); 389 } else { 390 ret = qcrypto_cipher_decrypt(s->cipher, 391 in_buf, 392 out_buf, 393 512, 394 errp); 395 } 396 if (ret < 0) { 397 return -1; 398 } 399 sector_num++; 400 in_buf += 512; 401 out_buf += 512; 402 } 403 return 0; 404 } 405 406 static int coroutine_fn do_perform_cow(BlockDriverState *bs, 407 uint64_t src_cluster_offset, 408 uint64_t cluster_offset, 409 int offset_in_cluster, 410 int bytes) 411 { 412 BDRVQcow2State *s = bs->opaque; 413 QEMUIOVector qiov; 414 struct iovec iov; 415 int ret; 416 417 iov.iov_len = bytes; 418 iov.iov_base = qemu_try_blockalign(bs, iov.iov_len); 419 if (iov.iov_base == NULL) { 420 return -ENOMEM; 421 } 422 423 qemu_iovec_init_external(&qiov, &iov, 1); 424 425 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ); 426 427 if (!bs->drv) { 428 ret = -ENOMEDIUM; 429 goto out; 430 } 431 432 /* Call .bdrv_co_readv() directly instead of using the public block-layer 433 * interface. This avoids double I/O throttling and request tracking, 434 * which can lead to deadlock when block layer copy-on-read is enabled. 435 */ 436 ret = bs->drv->bdrv_co_preadv(bs, src_cluster_offset + offset_in_cluster, 437 bytes, &qiov, 0); 438 if (ret < 0) { 439 goto out; 440 } 441 442 if (bs->encrypted) { 443 Error *err = NULL; 444 int64_t sector = (src_cluster_offset + offset_in_cluster) 445 >> BDRV_SECTOR_BITS; 446 assert(s->cipher); 447 assert((offset_in_cluster & ~BDRV_SECTOR_MASK) == 0); 448 assert((bytes & ~BDRV_SECTOR_MASK) == 0); 449 if (qcow2_encrypt_sectors(s, sector, iov.iov_base, iov.iov_base, 450 bytes >> BDRV_SECTOR_BITS, true, &err) < 0) { 451 ret = -EIO; 452 error_free(err); 453 goto out; 454 } 455 } 456 457 ret = qcow2_pre_write_overlap_check(bs, 0, 458 cluster_offset + offset_in_cluster, bytes); 459 if (ret < 0) { 460 goto out; 461 } 462 463 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE); 464 ret = bdrv_co_pwritev(bs->file, cluster_offset + offset_in_cluster, 465 bytes, &qiov, 0); 466 if (ret < 0) { 467 goto out; 468 } 469 470 ret = 0; 471 out: 472 qemu_vfree(iov.iov_base); 473 return ret; 474 } 475 476 477 /* 478 * get_cluster_offset 479 * 480 * For a given offset of the virtual disk, find the cluster type and offset in 481 * the qcow2 file. The offset is stored in *cluster_offset. 482 * 483 * On entry, *bytes is the maximum number of contiguous bytes starting at 484 * offset that we are interested in. 485 * 486 * On exit, *bytes is the number of bytes starting at offset that have the same 487 * cluster type and (if applicable) are stored contiguously in the image file. 488 * Compressed clusters are always returned one by one. 489 * 490 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error 491 * cases. 492 */ 493 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset, 494 unsigned int *bytes, uint64_t *cluster_offset) 495 { 496 BDRVQcow2State *s = bs->opaque; 497 unsigned int l2_index; 498 uint64_t l1_index, l2_offset, *l2_table; 499 int l1_bits, c; 500 unsigned int offset_in_cluster; 501 uint64_t bytes_available, bytes_needed, nb_clusters; 502 QCow2ClusterType type; 503 int ret; 504 505 offset_in_cluster = offset_into_cluster(s, offset); 506 bytes_needed = (uint64_t) *bytes + offset_in_cluster; 507 508 l1_bits = s->l2_bits + s->cluster_bits; 509 510 /* compute how many bytes there are between the start of the cluster 511 * containing offset and the end of the l1 entry */ 512 bytes_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1)) 513 + offset_in_cluster; 514 515 if (bytes_needed > bytes_available) { 516 bytes_needed = bytes_available; 517 } 518 519 *cluster_offset = 0; 520 521 /* seek to the l2 offset in the l1 table */ 522 523 l1_index = offset >> l1_bits; 524 if (l1_index >= s->l1_size) { 525 type = QCOW2_CLUSTER_UNALLOCATED; 526 goto out; 527 } 528 529 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK; 530 if (!l2_offset) { 531 type = QCOW2_CLUSTER_UNALLOCATED; 532 goto out; 533 } 534 535 if (offset_into_cluster(s, l2_offset)) { 536 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64 537 " unaligned (L1 index: %#" PRIx64 ")", 538 l2_offset, l1_index); 539 return -EIO; 540 } 541 542 /* load the l2 table in memory */ 543 544 ret = l2_load(bs, l2_offset, &l2_table); 545 if (ret < 0) { 546 return ret; 547 } 548 549 /* find the cluster offset for the given disk offset */ 550 551 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1); 552 *cluster_offset = be64_to_cpu(l2_table[l2_index]); 553 554 nb_clusters = size_to_clusters(s, bytes_needed); 555 /* bytes_needed <= *bytes + offset_in_cluster, both of which are unsigned 556 * integers; the minimum cluster size is 512, so this assertion is always 557 * true */ 558 assert(nb_clusters <= INT_MAX); 559 560 type = qcow2_get_cluster_type(*cluster_offset); 561 if (s->qcow_version < 3 && (type == QCOW2_CLUSTER_ZERO_PLAIN || 562 type == QCOW2_CLUSTER_ZERO_ALLOC)) { 563 qcow2_signal_corruption(bs, true, -1, -1, "Zero cluster entry found" 564 " in pre-v3 image (L2 offset: %#" PRIx64 565 ", L2 index: %#x)", l2_offset, l2_index); 566 ret = -EIO; 567 goto fail; 568 } 569 switch (type) { 570 case QCOW2_CLUSTER_COMPRESSED: 571 /* Compressed clusters can only be processed one by one */ 572 c = 1; 573 *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK; 574 break; 575 case QCOW2_CLUSTER_ZERO_PLAIN: 576 case QCOW2_CLUSTER_UNALLOCATED: 577 /* how many empty clusters ? */ 578 c = count_contiguous_clusters_unallocated(nb_clusters, 579 &l2_table[l2_index], type); 580 *cluster_offset = 0; 581 break; 582 case QCOW2_CLUSTER_ZERO_ALLOC: 583 case QCOW2_CLUSTER_NORMAL: 584 /* how many allocated clusters ? */ 585 c = count_contiguous_clusters(nb_clusters, s->cluster_size, 586 &l2_table[l2_index], QCOW_OFLAG_ZERO); 587 *cluster_offset &= L2E_OFFSET_MASK; 588 if (offset_into_cluster(s, *cluster_offset)) { 589 qcow2_signal_corruption(bs, true, -1, -1, 590 "Cluster allocation offset %#" 591 PRIx64 " unaligned (L2 offset: %#" PRIx64 592 ", L2 index: %#x)", *cluster_offset, 593 l2_offset, l2_index); 594 ret = -EIO; 595 goto fail; 596 } 597 break; 598 default: 599 abort(); 600 } 601 602 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 603 604 bytes_available = (int64_t)c * s->cluster_size; 605 606 out: 607 if (bytes_available > bytes_needed) { 608 bytes_available = bytes_needed; 609 } 610 611 /* bytes_available <= bytes_needed <= *bytes + offset_in_cluster; 612 * subtracting offset_in_cluster will therefore definitely yield something 613 * not exceeding UINT_MAX */ 614 assert(bytes_available - offset_in_cluster <= UINT_MAX); 615 *bytes = bytes_available - offset_in_cluster; 616 617 return type; 618 619 fail: 620 qcow2_cache_put(bs, s->l2_table_cache, (void **)&l2_table); 621 return ret; 622 } 623 624 /* 625 * get_cluster_table 626 * 627 * for a given disk offset, load (and allocate if needed) 628 * the l2 table. 629 * 630 * the l2 table offset in the qcow2 file and the cluster index 631 * in the l2 table are given to the caller. 632 * 633 * Returns 0 on success, -errno in failure case 634 */ 635 static int get_cluster_table(BlockDriverState *bs, uint64_t offset, 636 uint64_t **new_l2_table, 637 int *new_l2_index) 638 { 639 BDRVQcow2State *s = bs->opaque; 640 unsigned int l2_index; 641 uint64_t l1_index, l2_offset; 642 uint64_t *l2_table = NULL; 643 int ret; 644 645 /* seek to the l2 offset in the l1 table */ 646 647 l1_index = offset >> (s->l2_bits + s->cluster_bits); 648 if (l1_index >= s->l1_size) { 649 ret = qcow2_grow_l1_table(bs, l1_index + 1, false); 650 if (ret < 0) { 651 return ret; 652 } 653 } 654 655 assert(l1_index < s->l1_size); 656 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK; 657 if (offset_into_cluster(s, l2_offset)) { 658 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" PRIx64 659 " unaligned (L1 index: %#" PRIx64 ")", 660 l2_offset, l1_index); 661 return -EIO; 662 } 663 664 /* seek the l2 table of the given l2 offset */ 665 666 if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) { 667 /* load the l2 table in memory */ 668 ret = l2_load(bs, l2_offset, &l2_table); 669 if (ret < 0) { 670 return ret; 671 } 672 } else { 673 /* First allocate a new L2 table (and do COW if needed) */ 674 ret = l2_allocate(bs, l1_index, &l2_table); 675 if (ret < 0) { 676 return ret; 677 } 678 679 /* Then decrease the refcount of the old table */ 680 if (l2_offset) { 681 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t), 682 QCOW2_DISCARD_OTHER); 683 } 684 } 685 686 /* find the cluster offset for the given disk offset */ 687 688 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1); 689 690 *new_l2_table = l2_table; 691 *new_l2_index = l2_index; 692 693 return 0; 694 } 695 696 /* 697 * alloc_compressed_cluster_offset 698 * 699 * For a given offset of the disk image, return cluster offset in 700 * qcow2 file. 701 * 702 * If the offset is not found, allocate a new compressed cluster. 703 * 704 * Return the cluster offset if successful, 705 * Return 0, otherwise. 706 * 707 */ 708 709 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs, 710 uint64_t offset, 711 int compressed_size) 712 { 713 BDRVQcow2State *s = bs->opaque; 714 int l2_index, ret; 715 uint64_t *l2_table; 716 int64_t cluster_offset; 717 int nb_csectors; 718 719 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 720 if (ret < 0) { 721 return 0; 722 } 723 724 /* Compression can't overwrite anything. Fail if the cluster was already 725 * allocated. */ 726 cluster_offset = be64_to_cpu(l2_table[l2_index]); 727 if (cluster_offset & L2E_OFFSET_MASK) { 728 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 729 return 0; 730 } 731 732 cluster_offset = qcow2_alloc_bytes(bs, compressed_size); 733 if (cluster_offset < 0) { 734 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 735 return 0; 736 } 737 738 nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) - 739 (cluster_offset >> 9); 740 741 cluster_offset |= QCOW_OFLAG_COMPRESSED | 742 ((uint64_t)nb_csectors << s->csize_shift); 743 744 /* update L2 table */ 745 746 /* compressed clusters never have the copied flag */ 747 748 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED); 749 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 750 l2_table[l2_index] = cpu_to_be64(cluster_offset); 751 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 752 753 return cluster_offset; 754 } 755 756 static int perform_cow(BlockDriverState *bs, QCowL2Meta *m, Qcow2COWRegion *r) 757 { 758 BDRVQcow2State *s = bs->opaque; 759 int ret; 760 761 if (r->nb_bytes == 0) { 762 return 0; 763 } 764 765 qemu_co_mutex_unlock(&s->lock); 766 ret = do_perform_cow(bs, m->offset, m->alloc_offset, r->offset, r->nb_bytes); 767 qemu_co_mutex_lock(&s->lock); 768 769 if (ret < 0) { 770 return ret; 771 } 772 773 /* 774 * Before we update the L2 table to actually point to the new cluster, we 775 * need to be sure that the refcounts have been increased and COW was 776 * handled. 777 */ 778 qcow2_cache_depends_on_flush(s->l2_table_cache); 779 780 return 0; 781 } 782 783 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m) 784 { 785 BDRVQcow2State *s = bs->opaque; 786 int i, j = 0, l2_index, ret; 787 uint64_t *old_cluster, *l2_table; 788 uint64_t cluster_offset = m->alloc_offset; 789 790 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters); 791 assert(m->nb_clusters > 0); 792 793 old_cluster = g_try_new(uint64_t, m->nb_clusters); 794 if (old_cluster == NULL) { 795 ret = -ENOMEM; 796 goto err; 797 } 798 799 /* copy content of unmodified sectors */ 800 ret = perform_cow(bs, m, &m->cow_start); 801 if (ret < 0) { 802 goto err; 803 } 804 805 ret = perform_cow(bs, m, &m->cow_end); 806 if (ret < 0) { 807 goto err; 808 } 809 810 /* Update L2 table. */ 811 if (s->use_lazy_refcounts) { 812 qcow2_mark_dirty(bs); 813 } 814 if (qcow2_need_accurate_refcounts(s)) { 815 qcow2_cache_set_dependency(bs, s->l2_table_cache, 816 s->refcount_block_cache); 817 } 818 819 ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index); 820 if (ret < 0) { 821 goto err; 822 } 823 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 824 825 assert(l2_index + m->nb_clusters <= s->l2_size); 826 for (i = 0; i < m->nb_clusters; i++) { 827 /* if two concurrent writes happen to the same unallocated cluster 828 * each write allocates separate cluster and writes data concurrently. 829 * The first one to complete updates l2 table with pointer to its 830 * cluster the second one has to do RMW (which is done above by 831 * perform_cow()), update l2 table with its cluster pointer and free 832 * old cluster. This is what this loop does */ 833 if (l2_table[l2_index + i] != 0) { 834 old_cluster[j++] = l2_table[l2_index + i]; 835 } 836 837 l2_table[l2_index + i] = cpu_to_be64((cluster_offset + 838 (i << s->cluster_bits)) | QCOW_OFLAG_COPIED); 839 } 840 841 842 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 843 844 /* 845 * If this was a COW, we need to decrease the refcount of the old cluster. 846 * 847 * Don't discard clusters that reach a refcount of 0 (e.g. compressed 848 * clusters), the next write will reuse them anyway. 849 */ 850 if (!m->keep_old_clusters && j != 0) { 851 for (i = 0; i < j; i++) { 852 qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1, 853 QCOW2_DISCARD_NEVER); 854 } 855 } 856 857 ret = 0; 858 err: 859 g_free(old_cluster); 860 return ret; 861 } 862 863 /* 864 * Returns the number of contiguous clusters that can be used for an allocating 865 * write, but require COW to be performed (this includes yet unallocated space, 866 * which must copy from the backing file) 867 */ 868 static int count_cow_clusters(BDRVQcow2State *s, int nb_clusters, 869 uint64_t *l2_table, int l2_index) 870 { 871 int i; 872 873 for (i = 0; i < nb_clusters; i++) { 874 uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]); 875 QCow2ClusterType cluster_type = qcow2_get_cluster_type(l2_entry); 876 877 switch(cluster_type) { 878 case QCOW2_CLUSTER_NORMAL: 879 if (l2_entry & QCOW_OFLAG_COPIED) { 880 goto out; 881 } 882 break; 883 case QCOW2_CLUSTER_UNALLOCATED: 884 case QCOW2_CLUSTER_COMPRESSED: 885 case QCOW2_CLUSTER_ZERO_PLAIN: 886 case QCOW2_CLUSTER_ZERO_ALLOC: 887 break; 888 default: 889 abort(); 890 } 891 } 892 893 out: 894 assert(i <= nb_clusters); 895 return i; 896 } 897 898 /* 899 * Check if there already is an AIO write request in flight which allocates 900 * the same cluster. In this case we need to wait until the previous 901 * request has completed and updated the L2 table accordingly. 902 * 903 * Returns: 904 * 0 if there was no dependency. *cur_bytes indicates the number of 905 * bytes from guest_offset that can be read before the next 906 * dependency must be processed (or the request is complete) 907 * 908 * -EAGAIN if we had to wait for another request, previously gathered 909 * information on cluster allocation may be invalid now. The caller 910 * must start over anyway, so consider *cur_bytes undefined. 911 */ 912 static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset, 913 uint64_t *cur_bytes, QCowL2Meta **m) 914 { 915 BDRVQcow2State *s = bs->opaque; 916 QCowL2Meta *old_alloc; 917 uint64_t bytes = *cur_bytes; 918 919 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) { 920 921 uint64_t start = guest_offset; 922 uint64_t end = start + bytes; 923 uint64_t old_start = l2meta_cow_start(old_alloc); 924 uint64_t old_end = l2meta_cow_end(old_alloc); 925 926 if (end <= old_start || start >= old_end) { 927 /* No intersection */ 928 } else { 929 if (start < old_start) { 930 /* Stop at the start of a running allocation */ 931 bytes = old_start - start; 932 } else { 933 bytes = 0; 934 } 935 936 /* Stop if already an l2meta exists. After yielding, it wouldn't 937 * be valid any more, so we'd have to clean up the old L2Metas 938 * and deal with requests depending on them before starting to 939 * gather new ones. Not worth the trouble. */ 940 if (bytes == 0 && *m) { 941 *cur_bytes = 0; 942 return 0; 943 } 944 945 if (bytes == 0) { 946 /* Wait for the dependency to complete. We need to recheck 947 * the free/allocated clusters when we continue. */ 948 qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock); 949 return -EAGAIN; 950 } 951 } 952 } 953 954 /* Make sure that existing clusters and new allocations are only used up to 955 * the next dependency if we shortened the request above */ 956 *cur_bytes = bytes; 957 958 return 0; 959 } 960 961 /* 962 * Checks how many already allocated clusters that don't require a copy on 963 * write there are at the given guest_offset (up to *bytes). If 964 * *host_offset is not zero, only physically contiguous clusters beginning at 965 * this host offset are counted. 966 * 967 * Note that guest_offset may not be cluster aligned. In this case, the 968 * returned *host_offset points to exact byte referenced by guest_offset and 969 * therefore isn't cluster aligned as well. 970 * 971 * Returns: 972 * 0: if no allocated clusters are available at the given offset. 973 * *bytes is normally unchanged. It is set to 0 if the cluster 974 * is allocated and doesn't need COW, but doesn't have the right 975 * physical offset. 976 * 977 * 1: if allocated clusters that don't require a COW are available at 978 * the requested offset. *bytes may have decreased and describes 979 * the length of the area that can be written to. 980 * 981 * -errno: in error cases 982 */ 983 static int handle_copied(BlockDriverState *bs, uint64_t guest_offset, 984 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m) 985 { 986 BDRVQcow2State *s = bs->opaque; 987 int l2_index; 988 uint64_t cluster_offset; 989 uint64_t *l2_table; 990 uint64_t nb_clusters; 991 unsigned int keep_clusters; 992 int ret; 993 994 trace_qcow2_handle_copied(qemu_coroutine_self(), guest_offset, *host_offset, 995 *bytes); 996 997 assert(*host_offset == 0 || offset_into_cluster(s, guest_offset) 998 == offset_into_cluster(s, *host_offset)); 999 1000 /* 1001 * Calculate the number of clusters to look for. We stop at L2 table 1002 * boundaries to keep things simple. 1003 */ 1004 nb_clusters = 1005 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes); 1006 1007 l2_index = offset_to_l2_index(s, guest_offset); 1008 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1009 assert(nb_clusters <= INT_MAX); 1010 1011 /* Find L2 entry for the first involved cluster */ 1012 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index); 1013 if (ret < 0) { 1014 return ret; 1015 } 1016 1017 cluster_offset = be64_to_cpu(l2_table[l2_index]); 1018 1019 /* Check how many clusters are already allocated and don't need COW */ 1020 if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL 1021 && (cluster_offset & QCOW_OFLAG_COPIED)) 1022 { 1023 /* If a specific host_offset is required, check it */ 1024 bool offset_matches = 1025 (cluster_offset & L2E_OFFSET_MASK) == *host_offset; 1026 1027 if (offset_into_cluster(s, cluster_offset & L2E_OFFSET_MASK)) { 1028 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset " 1029 "%#llx unaligned (guest offset: %#" PRIx64 1030 ")", cluster_offset & L2E_OFFSET_MASK, 1031 guest_offset); 1032 ret = -EIO; 1033 goto out; 1034 } 1035 1036 if (*host_offset != 0 && !offset_matches) { 1037 *bytes = 0; 1038 ret = 0; 1039 goto out; 1040 } 1041 1042 /* We keep all QCOW_OFLAG_COPIED clusters */ 1043 keep_clusters = 1044 count_contiguous_clusters(nb_clusters, s->cluster_size, 1045 &l2_table[l2_index], 1046 QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO); 1047 assert(keep_clusters <= nb_clusters); 1048 1049 *bytes = MIN(*bytes, 1050 keep_clusters * s->cluster_size 1051 - offset_into_cluster(s, guest_offset)); 1052 1053 ret = 1; 1054 } else { 1055 ret = 0; 1056 } 1057 1058 /* Cleanup */ 1059 out: 1060 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1061 1062 /* Only return a host offset if we actually made progress. Otherwise we 1063 * would make requirements for handle_alloc() that it can't fulfill */ 1064 if (ret > 0) { 1065 *host_offset = (cluster_offset & L2E_OFFSET_MASK) 1066 + offset_into_cluster(s, guest_offset); 1067 } 1068 1069 return ret; 1070 } 1071 1072 /* 1073 * Allocates new clusters for the given guest_offset. 1074 * 1075 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to 1076 * contain the number of clusters that have been allocated and are contiguous 1077 * in the image file. 1078 * 1079 * If *host_offset is non-zero, it specifies the offset in the image file at 1080 * which the new clusters must start. *nb_clusters can be 0 on return in this 1081 * case if the cluster at host_offset is already in use. If *host_offset is 1082 * zero, the clusters can be allocated anywhere in the image file. 1083 * 1084 * *host_offset is updated to contain the offset into the image file at which 1085 * the first allocated cluster starts. 1086 * 1087 * Return 0 on success and -errno in error cases. -EAGAIN means that the 1088 * function has been waiting for another request and the allocation must be 1089 * restarted, but the whole request should not be failed. 1090 */ 1091 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset, 1092 uint64_t *host_offset, uint64_t *nb_clusters) 1093 { 1094 BDRVQcow2State *s = bs->opaque; 1095 1096 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset, 1097 *host_offset, *nb_clusters); 1098 1099 /* Allocate new clusters */ 1100 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self()); 1101 if (*host_offset == 0) { 1102 int64_t cluster_offset = 1103 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size); 1104 if (cluster_offset < 0) { 1105 return cluster_offset; 1106 } 1107 *host_offset = cluster_offset; 1108 return 0; 1109 } else { 1110 int64_t ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters); 1111 if (ret < 0) { 1112 return ret; 1113 } 1114 *nb_clusters = ret; 1115 return 0; 1116 } 1117 } 1118 1119 /* 1120 * Allocates new clusters for an area that either is yet unallocated or needs a 1121 * copy on write. If *host_offset is non-zero, clusters are only allocated if 1122 * the new allocation can match the specified host offset. 1123 * 1124 * Note that guest_offset may not be cluster aligned. In this case, the 1125 * returned *host_offset points to exact byte referenced by guest_offset and 1126 * therefore isn't cluster aligned as well. 1127 * 1128 * Returns: 1129 * 0: if no clusters could be allocated. *bytes is set to 0, 1130 * *host_offset is left unchanged. 1131 * 1132 * 1: if new clusters were allocated. *bytes may be decreased if the 1133 * new allocation doesn't cover all of the requested area. 1134 * *host_offset is updated to contain the host offset of the first 1135 * newly allocated cluster. 1136 * 1137 * -errno: in error cases 1138 */ 1139 static int handle_alloc(BlockDriverState *bs, uint64_t guest_offset, 1140 uint64_t *host_offset, uint64_t *bytes, QCowL2Meta **m) 1141 { 1142 BDRVQcow2State *s = bs->opaque; 1143 int l2_index; 1144 uint64_t *l2_table; 1145 uint64_t entry; 1146 uint64_t nb_clusters; 1147 int ret; 1148 bool keep_old_clusters = false; 1149 1150 uint64_t alloc_cluster_offset = 0; 1151 1152 trace_qcow2_handle_alloc(qemu_coroutine_self(), guest_offset, *host_offset, 1153 *bytes); 1154 assert(*bytes > 0); 1155 1156 /* 1157 * Calculate the number of clusters to look for. We stop at L2 table 1158 * boundaries to keep things simple. 1159 */ 1160 nb_clusters = 1161 size_to_clusters(s, offset_into_cluster(s, guest_offset) + *bytes); 1162 1163 l2_index = offset_to_l2_index(s, guest_offset); 1164 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1165 assert(nb_clusters <= INT_MAX); 1166 1167 /* Find L2 entry for the first involved cluster */ 1168 ret = get_cluster_table(bs, guest_offset, &l2_table, &l2_index); 1169 if (ret < 0) { 1170 return ret; 1171 } 1172 1173 entry = be64_to_cpu(l2_table[l2_index]); 1174 1175 /* For the moment, overwrite compressed clusters one by one */ 1176 if (entry & QCOW_OFLAG_COMPRESSED) { 1177 nb_clusters = 1; 1178 } else { 1179 nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, l2_index); 1180 } 1181 1182 /* This function is only called when there were no non-COW clusters, so if 1183 * we can't find any unallocated or COW clusters either, something is 1184 * wrong with our code. */ 1185 assert(nb_clusters > 0); 1186 1187 if (qcow2_get_cluster_type(entry) == QCOW2_CLUSTER_ZERO_ALLOC && 1188 (entry & QCOW_OFLAG_COPIED) && 1189 (!*host_offset || 1190 start_of_cluster(s, *host_offset) == (entry & L2E_OFFSET_MASK))) 1191 { 1192 /* Try to reuse preallocated zero clusters; contiguous normal clusters 1193 * would be fine, too, but count_cow_clusters() above has limited 1194 * nb_clusters already to a range of COW clusters */ 1195 int preallocated_nb_clusters = 1196 count_contiguous_clusters(nb_clusters, s->cluster_size, 1197 &l2_table[l2_index], QCOW_OFLAG_COPIED); 1198 assert(preallocated_nb_clusters > 0); 1199 1200 nb_clusters = preallocated_nb_clusters; 1201 alloc_cluster_offset = entry & L2E_OFFSET_MASK; 1202 1203 /* We want to reuse these clusters, so qcow2_alloc_cluster_link_l2() 1204 * should not free them. */ 1205 keep_old_clusters = true; 1206 } 1207 1208 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1209 1210 if (!alloc_cluster_offset) { 1211 /* Allocate, if necessary at a given offset in the image file */ 1212 alloc_cluster_offset = start_of_cluster(s, *host_offset); 1213 ret = do_alloc_cluster_offset(bs, guest_offset, &alloc_cluster_offset, 1214 &nb_clusters); 1215 if (ret < 0) { 1216 goto fail; 1217 } 1218 1219 /* Can't extend contiguous allocation */ 1220 if (nb_clusters == 0) { 1221 *bytes = 0; 1222 return 0; 1223 } 1224 1225 /* !*host_offset would overwrite the image header and is reserved for 1226 * "no host offset preferred". If 0 was a valid host offset, it'd 1227 * trigger the following overlap check; do that now to avoid having an 1228 * invalid value in *host_offset. */ 1229 if (!alloc_cluster_offset) { 1230 ret = qcow2_pre_write_overlap_check(bs, 0, alloc_cluster_offset, 1231 nb_clusters * s->cluster_size); 1232 assert(ret < 0); 1233 goto fail; 1234 } 1235 } 1236 1237 /* 1238 * Save info needed for meta data update. 1239 * 1240 * requested_bytes: Number of bytes from the start of the first 1241 * newly allocated cluster to the end of the (possibly shortened 1242 * before) write request. 1243 * 1244 * avail_bytes: Number of bytes from the start of the first 1245 * newly allocated to the end of the last newly allocated cluster. 1246 * 1247 * nb_bytes: The number of bytes from the start of the first 1248 * newly allocated cluster to the end of the area that the write 1249 * request actually writes to (excluding COW at the end) 1250 */ 1251 uint64_t requested_bytes = *bytes + offset_into_cluster(s, guest_offset); 1252 int avail_bytes = MIN(INT_MAX, nb_clusters << s->cluster_bits); 1253 int nb_bytes = MIN(requested_bytes, avail_bytes); 1254 QCowL2Meta *old_m = *m; 1255 1256 *m = g_malloc0(sizeof(**m)); 1257 1258 **m = (QCowL2Meta) { 1259 .next = old_m, 1260 1261 .alloc_offset = alloc_cluster_offset, 1262 .offset = start_of_cluster(s, guest_offset), 1263 .nb_clusters = nb_clusters, 1264 1265 .keep_old_clusters = keep_old_clusters, 1266 1267 .cow_start = { 1268 .offset = 0, 1269 .nb_bytes = offset_into_cluster(s, guest_offset), 1270 }, 1271 .cow_end = { 1272 .offset = nb_bytes, 1273 .nb_bytes = avail_bytes - nb_bytes, 1274 }, 1275 }; 1276 qemu_co_queue_init(&(*m)->dependent_requests); 1277 QLIST_INSERT_HEAD(&s->cluster_allocs, *m, next_in_flight); 1278 1279 *host_offset = alloc_cluster_offset + offset_into_cluster(s, guest_offset); 1280 *bytes = MIN(*bytes, nb_bytes - offset_into_cluster(s, guest_offset)); 1281 assert(*bytes != 0); 1282 1283 return 1; 1284 1285 fail: 1286 if (*m && (*m)->nb_clusters > 0) { 1287 QLIST_REMOVE(*m, next_in_flight); 1288 } 1289 return ret; 1290 } 1291 1292 /* 1293 * alloc_cluster_offset 1294 * 1295 * For a given offset on the virtual disk, find the cluster offset in qcow2 1296 * file. If the offset is not found, allocate a new cluster. 1297 * 1298 * If the cluster was already allocated, m->nb_clusters is set to 0 and 1299 * other fields in m are meaningless. 1300 * 1301 * If the cluster is newly allocated, m->nb_clusters is set to the number of 1302 * contiguous clusters that have been allocated. In this case, the other 1303 * fields of m are valid and contain information about the first allocated 1304 * cluster. 1305 * 1306 * If the request conflicts with another write request in flight, the coroutine 1307 * is queued and will be reentered when the dependency has completed. 1308 * 1309 * Return 0 on success and -errno in error cases 1310 */ 1311 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset, 1312 unsigned int *bytes, uint64_t *host_offset, 1313 QCowL2Meta **m) 1314 { 1315 BDRVQcow2State *s = bs->opaque; 1316 uint64_t start, remaining; 1317 uint64_t cluster_offset; 1318 uint64_t cur_bytes; 1319 int ret; 1320 1321 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, *bytes); 1322 1323 again: 1324 start = offset; 1325 remaining = *bytes; 1326 cluster_offset = 0; 1327 *host_offset = 0; 1328 cur_bytes = 0; 1329 *m = NULL; 1330 1331 while (true) { 1332 1333 if (!*host_offset) { 1334 *host_offset = start_of_cluster(s, cluster_offset); 1335 } 1336 1337 assert(remaining >= cur_bytes); 1338 1339 start += cur_bytes; 1340 remaining -= cur_bytes; 1341 cluster_offset += cur_bytes; 1342 1343 if (remaining == 0) { 1344 break; 1345 } 1346 1347 cur_bytes = remaining; 1348 1349 /* 1350 * Now start gathering as many contiguous clusters as possible: 1351 * 1352 * 1. Check for overlaps with in-flight allocations 1353 * 1354 * a) Overlap not in the first cluster -> shorten this request and 1355 * let the caller handle the rest in its next loop iteration. 1356 * 1357 * b) Real overlaps of two requests. Yield and restart the search 1358 * for contiguous clusters (the situation could have changed 1359 * while we were sleeping) 1360 * 1361 * c) TODO: Request starts in the same cluster as the in-flight 1362 * allocation ends. Shorten the COW of the in-fight allocation, 1363 * set cluster_offset to write to the same cluster and set up 1364 * the right synchronisation between the in-flight request and 1365 * the new one. 1366 */ 1367 ret = handle_dependencies(bs, start, &cur_bytes, m); 1368 if (ret == -EAGAIN) { 1369 /* Currently handle_dependencies() doesn't yield if we already had 1370 * an allocation. If it did, we would have to clean up the L2Meta 1371 * structs before starting over. */ 1372 assert(*m == NULL); 1373 goto again; 1374 } else if (ret < 0) { 1375 return ret; 1376 } else if (cur_bytes == 0) { 1377 break; 1378 } else { 1379 /* handle_dependencies() may have decreased cur_bytes (shortened 1380 * the allocations below) so that the next dependency is processed 1381 * correctly during the next loop iteration. */ 1382 } 1383 1384 /* 1385 * 2. Count contiguous COPIED clusters. 1386 */ 1387 ret = handle_copied(bs, start, &cluster_offset, &cur_bytes, m); 1388 if (ret < 0) { 1389 return ret; 1390 } else if (ret) { 1391 continue; 1392 } else if (cur_bytes == 0) { 1393 break; 1394 } 1395 1396 /* 1397 * 3. If the request still hasn't completed, allocate new clusters, 1398 * considering any cluster_offset of steps 1c or 2. 1399 */ 1400 ret = handle_alloc(bs, start, &cluster_offset, &cur_bytes, m); 1401 if (ret < 0) { 1402 return ret; 1403 } else if (ret) { 1404 continue; 1405 } else { 1406 assert(cur_bytes == 0); 1407 break; 1408 } 1409 } 1410 1411 *bytes -= remaining; 1412 assert(*bytes > 0); 1413 assert(*host_offset != 0); 1414 1415 return 0; 1416 } 1417 1418 static int decompress_buffer(uint8_t *out_buf, int out_buf_size, 1419 const uint8_t *buf, int buf_size) 1420 { 1421 z_stream strm1, *strm = &strm1; 1422 int ret, out_len; 1423 1424 memset(strm, 0, sizeof(*strm)); 1425 1426 strm->next_in = (uint8_t *)buf; 1427 strm->avail_in = buf_size; 1428 strm->next_out = out_buf; 1429 strm->avail_out = out_buf_size; 1430 1431 ret = inflateInit2(strm, -12); 1432 if (ret != Z_OK) 1433 return -1; 1434 ret = inflate(strm, Z_FINISH); 1435 out_len = strm->next_out - out_buf; 1436 if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) || 1437 out_len != out_buf_size) { 1438 inflateEnd(strm); 1439 return -1; 1440 } 1441 inflateEnd(strm); 1442 return 0; 1443 } 1444 1445 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset) 1446 { 1447 BDRVQcow2State *s = bs->opaque; 1448 int ret, csize, nb_csectors, sector_offset; 1449 uint64_t coffset; 1450 1451 coffset = cluster_offset & s->cluster_offset_mask; 1452 if (s->cluster_cache_offset != coffset) { 1453 nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1; 1454 sector_offset = coffset & 511; 1455 csize = nb_csectors * 512 - sector_offset; 1456 BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED); 1457 ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, 1458 nb_csectors); 1459 if (ret < 0) { 1460 return ret; 1461 } 1462 if (decompress_buffer(s->cluster_cache, s->cluster_size, 1463 s->cluster_data + sector_offset, csize) < 0) { 1464 return -EIO; 1465 } 1466 s->cluster_cache_offset = coffset; 1467 } 1468 return 0; 1469 } 1470 1471 /* 1472 * This discards as many clusters of nb_clusters as possible at once (i.e. 1473 * all clusters in the same L2 table) and returns the number of discarded 1474 * clusters. 1475 */ 1476 static int discard_single_l2(BlockDriverState *bs, uint64_t offset, 1477 uint64_t nb_clusters, enum qcow2_discard_type type, 1478 bool full_discard) 1479 { 1480 BDRVQcow2State *s = bs->opaque; 1481 uint64_t *l2_table; 1482 int l2_index; 1483 int ret; 1484 int i; 1485 1486 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 1487 if (ret < 0) { 1488 return ret; 1489 } 1490 1491 /* Limit nb_clusters to one L2 table */ 1492 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1493 assert(nb_clusters <= INT_MAX); 1494 1495 for (i = 0; i < nb_clusters; i++) { 1496 uint64_t old_l2_entry; 1497 1498 old_l2_entry = be64_to_cpu(l2_table[l2_index + i]); 1499 1500 /* 1501 * If full_discard is false, make sure that a discarded area reads back 1502 * as zeroes for v3 images (we cannot do it for v2 without actually 1503 * writing a zero-filled buffer). We can skip the operation if the 1504 * cluster is already marked as zero, or if it's unallocated and we 1505 * don't have a backing file. 1506 * 1507 * TODO We might want to use bdrv_get_block_status(bs) here, but we're 1508 * holding s->lock, so that doesn't work today. 1509 * 1510 * If full_discard is true, the sector should not read back as zeroes, 1511 * but rather fall through to the backing file. 1512 */ 1513 switch (qcow2_get_cluster_type(old_l2_entry)) { 1514 case QCOW2_CLUSTER_UNALLOCATED: 1515 if (full_discard || !bs->backing) { 1516 continue; 1517 } 1518 break; 1519 1520 case QCOW2_CLUSTER_ZERO_PLAIN: 1521 if (!full_discard) { 1522 continue; 1523 } 1524 break; 1525 1526 case QCOW2_CLUSTER_ZERO_ALLOC: 1527 case QCOW2_CLUSTER_NORMAL: 1528 case QCOW2_CLUSTER_COMPRESSED: 1529 break; 1530 1531 default: 1532 abort(); 1533 } 1534 1535 /* First remove L2 entries */ 1536 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 1537 if (!full_discard && s->qcow_version >= 3) { 1538 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO); 1539 } else { 1540 l2_table[l2_index + i] = cpu_to_be64(0); 1541 } 1542 1543 /* Then decrease the refcount */ 1544 qcow2_free_any_clusters(bs, old_l2_entry, 1, type); 1545 } 1546 1547 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1548 1549 return nb_clusters; 1550 } 1551 1552 int qcow2_cluster_discard(BlockDriverState *bs, uint64_t offset, 1553 uint64_t bytes, enum qcow2_discard_type type, 1554 bool full_discard) 1555 { 1556 BDRVQcow2State *s = bs->opaque; 1557 uint64_t end_offset = offset + bytes; 1558 uint64_t nb_clusters; 1559 int64_t cleared; 1560 int ret; 1561 1562 /* Caller must pass aligned values, except at image end */ 1563 assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); 1564 assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) || 1565 end_offset == bs->total_sectors << BDRV_SECTOR_BITS); 1566 1567 nb_clusters = size_to_clusters(s, bytes); 1568 1569 s->cache_discards = true; 1570 1571 /* Each L2 table is handled by its own loop iteration */ 1572 while (nb_clusters > 0) { 1573 cleared = discard_single_l2(bs, offset, nb_clusters, type, 1574 full_discard); 1575 if (cleared < 0) { 1576 ret = cleared; 1577 goto fail; 1578 } 1579 1580 nb_clusters -= cleared; 1581 offset += (cleared * s->cluster_size); 1582 } 1583 1584 ret = 0; 1585 fail: 1586 s->cache_discards = false; 1587 qcow2_process_discards(bs, ret); 1588 1589 return ret; 1590 } 1591 1592 /* 1593 * This zeroes as many clusters of nb_clusters as possible at once (i.e. 1594 * all clusters in the same L2 table) and returns the number of zeroed 1595 * clusters. 1596 */ 1597 static int zero_single_l2(BlockDriverState *bs, uint64_t offset, 1598 uint64_t nb_clusters, int flags) 1599 { 1600 BDRVQcow2State *s = bs->opaque; 1601 uint64_t *l2_table; 1602 int l2_index; 1603 int ret; 1604 int i; 1605 bool unmap = !!(flags & BDRV_REQ_MAY_UNMAP); 1606 1607 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 1608 if (ret < 0) { 1609 return ret; 1610 } 1611 1612 /* Limit nb_clusters to one L2 table */ 1613 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1614 assert(nb_clusters <= INT_MAX); 1615 1616 for (i = 0; i < nb_clusters; i++) { 1617 uint64_t old_offset; 1618 QCow2ClusterType cluster_type; 1619 1620 old_offset = be64_to_cpu(l2_table[l2_index + i]); 1621 1622 /* 1623 * Minimize L2 changes if the cluster already reads back as 1624 * zeroes with correct allocation. 1625 */ 1626 cluster_type = qcow2_get_cluster_type(old_offset); 1627 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN || 1628 (cluster_type == QCOW2_CLUSTER_ZERO_ALLOC && !unmap)) { 1629 continue; 1630 } 1631 1632 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 1633 if (cluster_type == QCOW2_CLUSTER_COMPRESSED || unmap) { 1634 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO); 1635 qcow2_free_any_clusters(bs, old_offset, 1, QCOW2_DISCARD_REQUEST); 1636 } else { 1637 l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO); 1638 } 1639 } 1640 1641 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1642 1643 return nb_clusters; 1644 } 1645 1646 int qcow2_cluster_zeroize(BlockDriverState *bs, uint64_t offset, 1647 uint64_t bytes, int flags) 1648 { 1649 BDRVQcow2State *s = bs->opaque; 1650 uint64_t end_offset = offset + bytes; 1651 uint64_t nb_clusters; 1652 int64_t cleared; 1653 int ret; 1654 1655 /* Caller must pass aligned values, except at image end */ 1656 assert(QEMU_IS_ALIGNED(offset, s->cluster_size)); 1657 assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size) || 1658 end_offset == bs->total_sectors << BDRV_SECTOR_BITS); 1659 1660 /* The zero flag is only supported by version 3 and newer */ 1661 if (s->qcow_version < 3) { 1662 return -ENOTSUP; 1663 } 1664 1665 /* Each L2 table is handled by its own loop iteration */ 1666 nb_clusters = size_to_clusters(s, bytes); 1667 1668 s->cache_discards = true; 1669 1670 while (nb_clusters > 0) { 1671 cleared = zero_single_l2(bs, offset, nb_clusters, flags); 1672 if (cleared < 0) { 1673 ret = cleared; 1674 goto fail; 1675 } 1676 1677 nb_clusters -= cleared; 1678 offset += (cleared * s->cluster_size); 1679 } 1680 1681 ret = 0; 1682 fail: 1683 s->cache_discards = false; 1684 qcow2_process_discards(bs, ret); 1685 1686 return ret; 1687 } 1688 1689 /* 1690 * Expands all zero clusters in a specific L1 table (or deallocates them, for 1691 * non-backed non-pre-allocated zero clusters). 1692 * 1693 * l1_entries and *visited_l1_entries are used to keep track of progress for 1694 * status_cb(). l1_entries contains the total number of L1 entries and 1695 * *visited_l1_entries counts all visited L1 entries. 1696 */ 1697 static int expand_zero_clusters_in_l1(BlockDriverState *bs, uint64_t *l1_table, 1698 int l1_size, int64_t *visited_l1_entries, 1699 int64_t l1_entries, 1700 BlockDriverAmendStatusCB *status_cb, 1701 void *cb_opaque) 1702 { 1703 BDRVQcow2State *s = bs->opaque; 1704 bool is_active_l1 = (l1_table == s->l1_table); 1705 uint64_t *l2_table = NULL; 1706 int ret; 1707 int i, j; 1708 1709 if (!is_active_l1) { 1710 /* inactive L2 tables require a buffer to be stored in when loading 1711 * them from disk */ 1712 l2_table = qemu_try_blockalign(bs->file->bs, s->cluster_size); 1713 if (l2_table == NULL) { 1714 return -ENOMEM; 1715 } 1716 } 1717 1718 for (i = 0; i < l1_size; i++) { 1719 uint64_t l2_offset = l1_table[i] & L1E_OFFSET_MASK; 1720 bool l2_dirty = false; 1721 uint64_t l2_refcount; 1722 1723 if (!l2_offset) { 1724 /* unallocated */ 1725 (*visited_l1_entries)++; 1726 if (status_cb) { 1727 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque); 1728 } 1729 continue; 1730 } 1731 1732 if (offset_into_cluster(s, l2_offset)) { 1733 qcow2_signal_corruption(bs, true, -1, -1, "L2 table offset %#" 1734 PRIx64 " unaligned (L1 index: %#x)", 1735 l2_offset, i); 1736 ret = -EIO; 1737 goto fail; 1738 } 1739 1740 if (is_active_l1) { 1741 /* get active L2 tables from cache */ 1742 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, 1743 (void **)&l2_table); 1744 } else { 1745 /* load inactive L2 tables from disk */ 1746 ret = bdrv_read(bs->file, l2_offset / BDRV_SECTOR_SIZE, 1747 (void *)l2_table, s->cluster_sectors); 1748 } 1749 if (ret < 0) { 1750 goto fail; 1751 } 1752 1753 ret = qcow2_get_refcount(bs, l2_offset >> s->cluster_bits, 1754 &l2_refcount); 1755 if (ret < 0) { 1756 goto fail; 1757 } 1758 1759 for (j = 0; j < s->l2_size; j++) { 1760 uint64_t l2_entry = be64_to_cpu(l2_table[j]); 1761 int64_t offset = l2_entry & L2E_OFFSET_MASK; 1762 QCow2ClusterType cluster_type = qcow2_get_cluster_type(l2_entry); 1763 1764 if (cluster_type != QCOW2_CLUSTER_ZERO_PLAIN && 1765 cluster_type != QCOW2_CLUSTER_ZERO_ALLOC) { 1766 continue; 1767 } 1768 1769 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) { 1770 if (!bs->backing) { 1771 /* not backed; therefore we can simply deallocate the 1772 * cluster */ 1773 l2_table[j] = 0; 1774 l2_dirty = true; 1775 continue; 1776 } 1777 1778 offset = qcow2_alloc_clusters(bs, s->cluster_size); 1779 if (offset < 0) { 1780 ret = offset; 1781 goto fail; 1782 } 1783 1784 if (l2_refcount > 1) { 1785 /* For shared L2 tables, set the refcount accordingly (it is 1786 * already 1 and needs to be l2_refcount) */ 1787 ret = qcow2_update_cluster_refcount(bs, 1788 offset >> s->cluster_bits, 1789 refcount_diff(1, l2_refcount), false, 1790 QCOW2_DISCARD_OTHER); 1791 if (ret < 0) { 1792 qcow2_free_clusters(bs, offset, s->cluster_size, 1793 QCOW2_DISCARD_OTHER); 1794 goto fail; 1795 } 1796 } 1797 } 1798 1799 if (offset_into_cluster(s, offset)) { 1800 qcow2_signal_corruption(bs, true, -1, -1, "Data cluster offset " 1801 "%#" PRIx64 " unaligned (L2 offset: %#" 1802 PRIx64 ", L2 index: %#x)", offset, 1803 l2_offset, j); 1804 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) { 1805 qcow2_free_clusters(bs, offset, s->cluster_size, 1806 QCOW2_DISCARD_ALWAYS); 1807 } 1808 ret = -EIO; 1809 goto fail; 1810 } 1811 1812 ret = qcow2_pre_write_overlap_check(bs, 0, offset, s->cluster_size); 1813 if (ret < 0) { 1814 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) { 1815 qcow2_free_clusters(bs, offset, s->cluster_size, 1816 QCOW2_DISCARD_ALWAYS); 1817 } 1818 goto fail; 1819 } 1820 1821 ret = bdrv_pwrite_zeroes(bs->file, offset, s->cluster_size, 0); 1822 if (ret < 0) { 1823 if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN) { 1824 qcow2_free_clusters(bs, offset, s->cluster_size, 1825 QCOW2_DISCARD_ALWAYS); 1826 } 1827 goto fail; 1828 } 1829 1830 if (l2_refcount == 1) { 1831 l2_table[j] = cpu_to_be64(offset | QCOW_OFLAG_COPIED); 1832 } else { 1833 l2_table[j] = cpu_to_be64(offset); 1834 } 1835 l2_dirty = true; 1836 } 1837 1838 if (is_active_l1) { 1839 if (l2_dirty) { 1840 qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table); 1841 qcow2_cache_depends_on_flush(s->l2_table_cache); 1842 } 1843 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1844 } else { 1845 if (l2_dirty) { 1846 ret = qcow2_pre_write_overlap_check(bs, 1847 QCOW2_OL_INACTIVE_L2 | QCOW2_OL_ACTIVE_L2, l2_offset, 1848 s->cluster_size); 1849 if (ret < 0) { 1850 goto fail; 1851 } 1852 1853 ret = bdrv_write(bs->file, l2_offset / BDRV_SECTOR_SIZE, 1854 (void *)l2_table, s->cluster_sectors); 1855 if (ret < 0) { 1856 goto fail; 1857 } 1858 } 1859 } 1860 1861 (*visited_l1_entries)++; 1862 if (status_cb) { 1863 status_cb(bs, *visited_l1_entries, l1_entries, cb_opaque); 1864 } 1865 } 1866 1867 ret = 0; 1868 1869 fail: 1870 if (l2_table) { 1871 if (!is_active_l1) { 1872 qemu_vfree(l2_table); 1873 } else { 1874 qcow2_cache_put(bs, s->l2_table_cache, (void **) &l2_table); 1875 } 1876 } 1877 return ret; 1878 } 1879 1880 /* 1881 * For backed images, expands all zero clusters on the image. For non-backed 1882 * images, deallocates all non-pre-allocated zero clusters (and claims the 1883 * allocation for pre-allocated ones). This is important for downgrading to a 1884 * qcow2 version which doesn't yet support metadata zero clusters. 1885 */ 1886 int qcow2_expand_zero_clusters(BlockDriverState *bs, 1887 BlockDriverAmendStatusCB *status_cb, 1888 void *cb_opaque) 1889 { 1890 BDRVQcow2State *s = bs->opaque; 1891 uint64_t *l1_table = NULL; 1892 int64_t l1_entries = 0, visited_l1_entries = 0; 1893 int ret; 1894 int i, j; 1895 1896 if (status_cb) { 1897 l1_entries = s->l1_size; 1898 for (i = 0; i < s->nb_snapshots; i++) { 1899 l1_entries += s->snapshots[i].l1_size; 1900 } 1901 } 1902 1903 ret = expand_zero_clusters_in_l1(bs, s->l1_table, s->l1_size, 1904 &visited_l1_entries, l1_entries, 1905 status_cb, cb_opaque); 1906 if (ret < 0) { 1907 goto fail; 1908 } 1909 1910 /* Inactive L1 tables may point to active L2 tables - therefore it is 1911 * necessary to flush the L2 table cache before trying to access the L2 1912 * tables pointed to by inactive L1 entries (else we might try to expand 1913 * zero clusters that have already been expanded); furthermore, it is also 1914 * necessary to empty the L2 table cache, since it may contain tables which 1915 * are now going to be modified directly on disk, bypassing the cache. 1916 * qcow2_cache_empty() does both for us. */ 1917 ret = qcow2_cache_empty(bs, s->l2_table_cache); 1918 if (ret < 0) { 1919 goto fail; 1920 } 1921 1922 for (i = 0; i < s->nb_snapshots; i++) { 1923 int l1_sectors = DIV_ROUND_UP(s->snapshots[i].l1_size * 1924 sizeof(uint64_t), BDRV_SECTOR_SIZE); 1925 1926 l1_table = g_realloc(l1_table, l1_sectors * BDRV_SECTOR_SIZE); 1927 1928 ret = bdrv_read(bs->file, 1929 s->snapshots[i].l1_table_offset / BDRV_SECTOR_SIZE, 1930 (void *)l1_table, l1_sectors); 1931 if (ret < 0) { 1932 goto fail; 1933 } 1934 1935 for (j = 0; j < s->snapshots[i].l1_size; j++) { 1936 be64_to_cpus(&l1_table[j]); 1937 } 1938 1939 ret = expand_zero_clusters_in_l1(bs, l1_table, s->snapshots[i].l1_size, 1940 &visited_l1_entries, l1_entries, 1941 status_cb, cb_opaque); 1942 if (ret < 0) { 1943 goto fail; 1944 } 1945 } 1946 1947 ret = 0; 1948 1949 fail: 1950 g_free(l1_table); 1951 return ret; 1952 } 1953