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 <zlib.h> 26 27 #include "qemu-common.h" 28 #include "block_int.h" 29 #include "block/qcow2.h" 30 #include "trace.h" 31 32 int qcow2_grow_l1_table(BlockDriverState *bs, int min_size, bool exact_size) 33 { 34 BDRVQcowState *s = bs->opaque; 35 int new_l1_size, new_l1_size2, ret, i; 36 uint64_t *new_l1_table; 37 int64_t new_l1_table_offset; 38 uint8_t data[12]; 39 40 if (min_size <= s->l1_size) 41 return 0; 42 43 if (exact_size) { 44 new_l1_size = min_size; 45 } else { 46 /* Bump size up to reduce the number of times we have to grow */ 47 new_l1_size = s->l1_size; 48 if (new_l1_size == 0) { 49 new_l1_size = 1; 50 } 51 while (min_size > new_l1_size) { 52 new_l1_size = (new_l1_size * 3 + 1) / 2; 53 } 54 } 55 56 #ifdef DEBUG_ALLOC2 57 fprintf(stderr, "grow l1_table from %d to %d\n", s->l1_size, new_l1_size); 58 #endif 59 60 new_l1_size2 = sizeof(uint64_t) * new_l1_size; 61 new_l1_table = g_malloc0(align_offset(new_l1_size2, 512)); 62 memcpy(new_l1_table, s->l1_table, s->l1_size * sizeof(uint64_t)); 63 64 /* write new table (align to cluster) */ 65 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ALLOC_TABLE); 66 new_l1_table_offset = qcow2_alloc_clusters(bs, new_l1_size2); 67 if (new_l1_table_offset < 0) { 68 g_free(new_l1_table); 69 return new_l1_table_offset; 70 } 71 72 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 73 if (ret < 0) { 74 goto fail; 75 } 76 77 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_WRITE_TABLE); 78 for(i = 0; i < s->l1_size; i++) 79 new_l1_table[i] = cpu_to_be64(new_l1_table[i]); 80 ret = bdrv_pwrite_sync(bs->file, new_l1_table_offset, new_l1_table, new_l1_size2); 81 if (ret < 0) 82 goto fail; 83 for(i = 0; i < s->l1_size; i++) 84 new_l1_table[i] = be64_to_cpu(new_l1_table[i]); 85 86 /* set new table */ 87 BLKDBG_EVENT(bs->file, BLKDBG_L1_GROW_ACTIVATE_TABLE); 88 cpu_to_be32w((uint32_t*)data, new_l1_size); 89 cpu_to_be64wu((uint64_t*)(data + 4), new_l1_table_offset); 90 ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, l1_size), data,sizeof(data)); 91 if (ret < 0) { 92 goto fail; 93 } 94 g_free(s->l1_table); 95 qcow2_free_clusters(bs, s->l1_table_offset, s->l1_size * sizeof(uint64_t)); 96 s->l1_table_offset = new_l1_table_offset; 97 s->l1_table = new_l1_table; 98 s->l1_size = new_l1_size; 99 return 0; 100 fail: 101 g_free(new_l1_table); 102 qcow2_free_clusters(bs, new_l1_table_offset, new_l1_size2); 103 return ret; 104 } 105 106 /* 107 * l2_load 108 * 109 * Loads a L2 table into memory. If the table is in the cache, the cache 110 * is used; otherwise the L2 table is loaded from the image file. 111 * 112 * Returns a pointer to the L2 table on success, or NULL if the read from 113 * the image file failed. 114 */ 115 116 static int l2_load(BlockDriverState *bs, uint64_t l2_offset, 117 uint64_t **l2_table) 118 { 119 BDRVQcowState *s = bs->opaque; 120 int ret; 121 122 ret = qcow2_cache_get(bs, s->l2_table_cache, l2_offset, (void**) l2_table); 123 124 return ret; 125 } 126 127 /* 128 * Writes one sector of the L1 table to the disk (can't update single entries 129 * and we really don't want bdrv_pread to perform a read-modify-write) 130 */ 131 #define L1_ENTRIES_PER_SECTOR (512 / 8) 132 static int write_l1_entry(BlockDriverState *bs, int l1_index) 133 { 134 BDRVQcowState *s = bs->opaque; 135 uint64_t buf[L1_ENTRIES_PER_SECTOR]; 136 int l1_start_index; 137 int i, ret; 138 139 l1_start_index = l1_index & ~(L1_ENTRIES_PER_SECTOR - 1); 140 for (i = 0; i < L1_ENTRIES_PER_SECTOR; i++) { 141 buf[i] = cpu_to_be64(s->l1_table[l1_start_index + i]); 142 } 143 144 BLKDBG_EVENT(bs->file, BLKDBG_L1_UPDATE); 145 ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset + 8 * l1_start_index, 146 buf, sizeof(buf)); 147 if (ret < 0) { 148 return ret; 149 } 150 151 return 0; 152 } 153 154 /* 155 * l2_allocate 156 * 157 * Allocate a new l2 entry in the file. If l1_index points to an already 158 * used entry in the L2 table (i.e. we are doing a copy on write for the L2 159 * table) copy the contents of the old L2 table into the newly allocated one. 160 * Otherwise the new table is initialized with zeros. 161 * 162 */ 163 164 static int l2_allocate(BlockDriverState *bs, int l1_index, uint64_t **table) 165 { 166 BDRVQcowState *s = bs->opaque; 167 uint64_t old_l2_offset; 168 uint64_t *l2_table; 169 int64_t l2_offset; 170 int ret; 171 172 old_l2_offset = s->l1_table[l1_index]; 173 174 trace_qcow2_l2_allocate(bs, l1_index); 175 176 /* allocate a new l2 entry */ 177 178 l2_offset = qcow2_alloc_clusters(bs, s->l2_size * sizeof(uint64_t)); 179 if (l2_offset < 0) { 180 return l2_offset; 181 } 182 183 ret = qcow2_cache_flush(bs, s->refcount_block_cache); 184 if (ret < 0) { 185 goto fail; 186 } 187 188 /* allocate a new entry in the l2 cache */ 189 190 trace_qcow2_l2_allocate_get_empty(bs, l1_index); 191 ret = qcow2_cache_get_empty(bs, s->l2_table_cache, l2_offset, (void**) table); 192 if (ret < 0) { 193 return ret; 194 } 195 196 l2_table = *table; 197 198 if ((old_l2_offset & L1E_OFFSET_MASK) == 0) { 199 /* if there was no old l2 table, clear the new table */ 200 memset(l2_table, 0, s->l2_size * sizeof(uint64_t)); 201 } else { 202 uint64_t* old_table; 203 204 /* if there was an old l2 table, read it from the disk */ 205 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_COW_READ); 206 ret = qcow2_cache_get(bs, s->l2_table_cache, 207 old_l2_offset & L1E_OFFSET_MASK, 208 (void**) &old_table); 209 if (ret < 0) { 210 goto fail; 211 } 212 213 memcpy(l2_table, old_table, s->cluster_size); 214 215 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &old_table); 216 if (ret < 0) { 217 goto fail; 218 } 219 } 220 221 /* write the l2 table to the file */ 222 BLKDBG_EVENT(bs->file, BLKDBG_L2_ALLOC_WRITE); 223 224 trace_qcow2_l2_allocate_write_l2(bs, l1_index); 225 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table); 226 ret = qcow2_cache_flush(bs, s->l2_table_cache); 227 if (ret < 0) { 228 goto fail; 229 } 230 231 /* update the L1 entry */ 232 trace_qcow2_l2_allocate_write_l1(bs, l1_index); 233 s->l1_table[l1_index] = l2_offset | QCOW_OFLAG_COPIED; 234 ret = write_l1_entry(bs, l1_index); 235 if (ret < 0) { 236 goto fail; 237 } 238 239 *table = l2_table; 240 trace_qcow2_l2_allocate_done(bs, l1_index, 0); 241 return 0; 242 243 fail: 244 trace_qcow2_l2_allocate_done(bs, l1_index, ret); 245 qcow2_cache_put(bs, s->l2_table_cache, (void**) table); 246 s->l1_table[l1_index] = old_l2_offset; 247 return ret; 248 } 249 250 /* 251 * Checks how many clusters in a given L2 table are contiguous in the image 252 * file. As soon as one of the flags in the bitmask stop_flags changes compared 253 * to the first cluster, the search is stopped and the cluster is not counted 254 * as contiguous. (This allows it, for example, to stop at the first compressed 255 * cluster which may require a different handling) 256 */ 257 static int count_contiguous_clusters(uint64_t nb_clusters, int cluster_size, 258 uint64_t *l2_table, uint64_t start, uint64_t stop_flags) 259 { 260 int i; 261 uint64_t mask = stop_flags | L2E_OFFSET_MASK; 262 uint64_t offset = be64_to_cpu(l2_table[0]) & mask; 263 264 if (!offset) 265 return 0; 266 267 for (i = start; i < start + nb_clusters; i++) { 268 uint64_t l2_entry = be64_to_cpu(l2_table[i]) & mask; 269 if (offset + (uint64_t) i * cluster_size != l2_entry) { 270 break; 271 } 272 } 273 274 return (i - start); 275 } 276 277 static int count_contiguous_free_clusters(uint64_t nb_clusters, uint64_t *l2_table) 278 { 279 int i; 280 281 for (i = 0; i < nb_clusters; i++) { 282 int type = qcow2_get_cluster_type(be64_to_cpu(l2_table[i])); 283 284 if (type != QCOW2_CLUSTER_UNALLOCATED) { 285 break; 286 } 287 } 288 289 return i; 290 } 291 292 /* The crypt function is compatible with the linux cryptoloop 293 algorithm for < 4 GB images. NOTE: out_buf == in_buf is 294 supported */ 295 void qcow2_encrypt_sectors(BDRVQcowState *s, int64_t sector_num, 296 uint8_t *out_buf, const uint8_t *in_buf, 297 int nb_sectors, int enc, 298 const AES_KEY *key) 299 { 300 union { 301 uint64_t ll[2]; 302 uint8_t b[16]; 303 } ivec; 304 int i; 305 306 for(i = 0; i < nb_sectors; i++) { 307 ivec.ll[0] = cpu_to_le64(sector_num); 308 ivec.ll[1] = 0; 309 AES_cbc_encrypt(in_buf, out_buf, 512, key, 310 ivec.b, enc); 311 sector_num++; 312 in_buf += 512; 313 out_buf += 512; 314 } 315 } 316 317 static int coroutine_fn copy_sectors(BlockDriverState *bs, 318 uint64_t start_sect, 319 uint64_t cluster_offset, 320 int n_start, int n_end) 321 { 322 BDRVQcowState *s = bs->opaque; 323 QEMUIOVector qiov; 324 struct iovec iov; 325 int n, ret; 326 327 /* 328 * If this is the last cluster and it is only partially used, we must only 329 * copy until the end of the image, or bdrv_check_request will fail for the 330 * bdrv_read/write calls below. 331 */ 332 if (start_sect + n_end > bs->total_sectors) { 333 n_end = bs->total_sectors - start_sect; 334 } 335 336 n = n_end - n_start; 337 if (n <= 0) { 338 return 0; 339 } 340 341 iov.iov_len = n * BDRV_SECTOR_SIZE; 342 iov.iov_base = qemu_blockalign(bs, iov.iov_len); 343 344 qemu_iovec_init_external(&qiov, &iov, 1); 345 346 BLKDBG_EVENT(bs->file, BLKDBG_COW_READ); 347 348 /* Call .bdrv_co_readv() directly instead of using the public block-layer 349 * interface. This avoids double I/O throttling and request tracking, 350 * which can lead to deadlock when block layer copy-on-read is enabled. 351 */ 352 ret = bs->drv->bdrv_co_readv(bs, start_sect + n_start, n, &qiov); 353 if (ret < 0) { 354 goto out; 355 } 356 357 if (s->crypt_method) { 358 qcow2_encrypt_sectors(s, start_sect + n_start, 359 iov.iov_base, iov.iov_base, n, 1, 360 &s->aes_encrypt_key); 361 } 362 363 BLKDBG_EVENT(bs->file, BLKDBG_COW_WRITE); 364 ret = bdrv_co_writev(bs->file, (cluster_offset >> 9) + n_start, n, &qiov); 365 if (ret < 0) { 366 goto out; 367 } 368 369 ret = 0; 370 out: 371 qemu_vfree(iov.iov_base); 372 return ret; 373 } 374 375 376 /* 377 * get_cluster_offset 378 * 379 * For a given offset of the disk image, find the cluster offset in 380 * qcow2 file. The offset is stored in *cluster_offset. 381 * 382 * on entry, *num is the number of contiguous sectors we'd like to 383 * access following offset. 384 * 385 * on exit, *num is the number of contiguous sectors we can read. 386 * 387 * Returns the cluster type (QCOW2_CLUSTER_*) on success, -errno in error 388 * cases. 389 */ 390 int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset, 391 int *num, uint64_t *cluster_offset) 392 { 393 BDRVQcowState *s = bs->opaque; 394 unsigned int l1_index, l2_index; 395 uint64_t l2_offset, *l2_table; 396 int l1_bits, c; 397 unsigned int index_in_cluster, nb_clusters; 398 uint64_t nb_available, nb_needed; 399 int ret; 400 401 index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1); 402 nb_needed = *num + index_in_cluster; 403 404 l1_bits = s->l2_bits + s->cluster_bits; 405 406 /* compute how many bytes there are between the offset and 407 * the end of the l1 entry 408 */ 409 410 nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1)); 411 412 /* compute the number of available sectors */ 413 414 nb_available = (nb_available >> 9) + index_in_cluster; 415 416 if (nb_needed > nb_available) { 417 nb_needed = nb_available; 418 } 419 420 *cluster_offset = 0; 421 422 /* seek the the l2 offset in the l1 table */ 423 424 l1_index = offset >> l1_bits; 425 if (l1_index >= s->l1_size) { 426 ret = QCOW2_CLUSTER_UNALLOCATED; 427 goto out; 428 } 429 430 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK; 431 if (!l2_offset) { 432 ret = QCOW2_CLUSTER_UNALLOCATED; 433 goto out; 434 } 435 436 /* load the l2 table in memory */ 437 438 ret = l2_load(bs, l2_offset, &l2_table); 439 if (ret < 0) { 440 return ret; 441 } 442 443 /* find the cluster offset for the given disk offset */ 444 445 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1); 446 *cluster_offset = be64_to_cpu(l2_table[l2_index]); 447 nb_clusters = size_to_clusters(s, nb_needed << 9); 448 449 ret = qcow2_get_cluster_type(*cluster_offset); 450 switch (ret) { 451 case QCOW2_CLUSTER_COMPRESSED: 452 /* Compressed clusters can only be processed one by one */ 453 c = 1; 454 *cluster_offset &= L2E_COMPRESSED_OFFSET_SIZE_MASK; 455 break; 456 case QCOW2_CLUSTER_ZERO: 457 c = count_contiguous_clusters(nb_clusters, s->cluster_size, 458 &l2_table[l2_index], 0, 459 QCOW_OFLAG_COMPRESSED | QCOW_OFLAG_ZERO); 460 *cluster_offset = 0; 461 break; 462 case QCOW2_CLUSTER_UNALLOCATED: 463 /* how many empty clusters ? */ 464 c = count_contiguous_free_clusters(nb_clusters, &l2_table[l2_index]); 465 *cluster_offset = 0; 466 break; 467 case QCOW2_CLUSTER_NORMAL: 468 /* how many allocated clusters ? */ 469 c = count_contiguous_clusters(nb_clusters, s->cluster_size, 470 &l2_table[l2_index], 0, 471 QCOW_OFLAG_COMPRESSED | QCOW_OFLAG_ZERO); 472 *cluster_offset &= L2E_OFFSET_MASK; 473 break; 474 } 475 476 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 477 478 nb_available = (c * s->cluster_sectors); 479 480 out: 481 if (nb_available > nb_needed) 482 nb_available = nb_needed; 483 484 *num = nb_available - index_in_cluster; 485 486 return ret; 487 } 488 489 /* 490 * get_cluster_table 491 * 492 * for a given disk offset, load (and allocate if needed) 493 * the l2 table. 494 * 495 * the l2 table offset in the qcow2 file and the cluster index 496 * in the l2 table are given to the caller. 497 * 498 * Returns 0 on success, -errno in failure case 499 */ 500 static int get_cluster_table(BlockDriverState *bs, uint64_t offset, 501 uint64_t **new_l2_table, 502 int *new_l2_index) 503 { 504 BDRVQcowState *s = bs->opaque; 505 unsigned int l1_index, l2_index; 506 uint64_t l2_offset; 507 uint64_t *l2_table = NULL; 508 int ret; 509 510 /* seek the the l2 offset in the l1 table */ 511 512 l1_index = offset >> (s->l2_bits + s->cluster_bits); 513 if (l1_index >= s->l1_size) { 514 ret = qcow2_grow_l1_table(bs, l1_index + 1, false); 515 if (ret < 0) { 516 return ret; 517 } 518 } 519 520 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK; 521 522 /* seek the l2 table of the given l2 offset */ 523 524 if (s->l1_table[l1_index] & QCOW_OFLAG_COPIED) { 525 /* load the l2 table in memory */ 526 ret = l2_load(bs, l2_offset, &l2_table); 527 if (ret < 0) { 528 return ret; 529 } 530 } else { 531 /* First allocate a new L2 table (and do COW if needed) */ 532 ret = l2_allocate(bs, l1_index, &l2_table); 533 if (ret < 0) { 534 return ret; 535 } 536 537 /* Then decrease the refcount of the old table */ 538 if (l2_offset) { 539 qcow2_free_clusters(bs, l2_offset, s->l2_size * sizeof(uint64_t)); 540 } 541 l2_offset = s->l1_table[l1_index] & L1E_OFFSET_MASK; 542 } 543 544 /* find the cluster offset for the given disk offset */ 545 546 l2_index = (offset >> s->cluster_bits) & (s->l2_size - 1); 547 548 *new_l2_table = l2_table; 549 *new_l2_index = l2_index; 550 551 return 0; 552 } 553 554 /* 555 * alloc_compressed_cluster_offset 556 * 557 * For a given offset of the disk image, return cluster offset in 558 * qcow2 file. 559 * 560 * If the offset is not found, allocate a new compressed cluster. 561 * 562 * Return the cluster offset if successful, 563 * Return 0, otherwise. 564 * 565 */ 566 567 uint64_t qcow2_alloc_compressed_cluster_offset(BlockDriverState *bs, 568 uint64_t offset, 569 int compressed_size) 570 { 571 BDRVQcowState *s = bs->opaque; 572 int l2_index, ret; 573 uint64_t *l2_table; 574 int64_t cluster_offset; 575 int nb_csectors; 576 577 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 578 if (ret < 0) { 579 return 0; 580 } 581 582 /* Compression can't overwrite anything. Fail if the cluster was already 583 * allocated. */ 584 cluster_offset = be64_to_cpu(l2_table[l2_index]); 585 if (cluster_offset & L2E_OFFSET_MASK) { 586 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 587 return 0; 588 } 589 590 cluster_offset = qcow2_alloc_bytes(bs, compressed_size); 591 if (cluster_offset < 0) { 592 qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 593 return 0; 594 } 595 596 nb_csectors = ((cluster_offset + compressed_size - 1) >> 9) - 597 (cluster_offset >> 9); 598 599 cluster_offset |= QCOW_OFLAG_COMPRESSED | 600 ((uint64_t)nb_csectors << s->csize_shift); 601 602 /* update L2 table */ 603 604 /* compressed clusters never have the copied flag */ 605 606 BLKDBG_EVENT(bs->file, BLKDBG_L2_UPDATE_COMPRESSED); 607 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table); 608 l2_table[l2_index] = cpu_to_be64(cluster_offset); 609 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 610 if (ret < 0) { 611 return 0; 612 } 613 614 return cluster_offset; 615 } 616 617 int qcow2_alloc_cluster_link_l2(BlockDriverState *bs, QCowL2Meta *m) 618 { 619 BDRVQcowState *s = bs->opaque; 620 int i, j = 0, l2_index, ret; 621 uint64_t *old_cluster, start_sect, *l2_table; 622 uint64_t cluster_offset = m->alloc_offset; 623 bool cow = false; 624 625 trace_qcow2_cluster_link_l2(qemu_coroutine_self(), m->nb_clusters); 626 627 if (m->nb_clusters == 0) 628 return 0; 629 630 old_cluster = g_malloc(m->nb_clusters * sizeof(uint64_t)); 631 632 /* copy content of unmodified sectors */ 633 start_sect = (m->offset & ~(s->cluster_size - 1)) >> 9; 634 if (m->n_start) { 635 cow = true; 636 qemu_co_mutex_unlock(&s->lock); 637 ret = copy_sectors(bs, start_sect, cluster_offset, 0, m->n_start); 638 qemu_co_mutex_lock(&s->lock); 639 if (ret < 0) 640 goto err; 641 } 642 643 if (m->nb_available & (s->cluster_sectors - 1)) { 644 uint64_t end = m->nb_available & ~(uint64_t)(s->cluster_sectors - 1); 645 cow = true; 646 qemu_co_mutex_unlock(&s->lock); 647 ret = copy_sectors(bs, start_sect + end, cluster_offset + (end << 9), 648 m->nb_available - end, s->cluster_sectors); 649 qemu_co_mutex_lock(&s->lock); 650 if (ret < 0) 651 goto err; 652 } 653 654 /* 655 * Update L2 table. 656 * 657 * Before we update the L2 table to actually point to the new cluster, we 658 * need to be sure that the refcounts have been increased and COW was 659 * handled. 660 */ 661 if (cow) { 662 qcow2_cache_depends_on_flush(s->l2_table_cache); 663 } 664 665 qcow2_cache_set_dependency(bs, s->l2_table_cache, s->refcount_block_cache); 666 ret = get_cluster_table(bs, m->offset, &l2_table, &l2_index); 667 if (ret < 0) { 668 goto err; 669 } 670 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table); 671 672 for (i = 0; i < m->nb_clusters; i++) { 673 /* if two concurrent writes happen to the same unallocated cluster 674 * each write allocates separate cluster and writes data concurrently. 675 * The first one to complete updates l2 table with pointer to its 676 * cluster the second one has to do RMW (which is done above by 677 * copy_sectors()), update l2 table with its cluster pointer and free 678 * old cluster. This is what this loop does */ 679 if(l2_table[l2_index + i] != 0) 680 old_cluster[j++] = l2_table[l2_index + i]; 681 682 l2_table[l2_index + i] = cpu_to_be64((cluster_offset + 683 (i << s->cluster_bits)) | QCOW_OFLAG_COPIED); 684 } 685 686 687 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 688 if (ret < 0) { 689 goto err; 690 } 691 692 /* 693 * If this was a COW, we need to decrease the refcount of the old cluster. 694 * Also flush bs->file to get the right order for L2 and refcount update. 695 */ 696 if (j != 0) { 697 for (i = 0; i < j; i++) { 698 qcow2_free_any_clusters(bs, be64_to_cpu(old_cluster[i]), 1); 699 } 700 } 701 702 ret = 0; 703 err: 704 g_free(old_cluster); 705 return ret; 706 } 707 708 /* 709 * Returns the number of contiguous clusters that can be used for an allocating 710 * write, but require COW to be performed (this includes yet unallocated space, 711 * which must copy from the backing file) 712 */ 713 static int count_cow_clusters(BDRVQcowState *s, int nb_clusters, 714 uint64_t *l2_table, int l2_index) 715 { 716 int i; 717 718 for (i = 0; i < nb_clusters; i++) { 719 uint64_t l2_entry = be64_to_cpu(l2_table[l2_index + i]); 720 int cluster_type = qcow2_get_cluster_type(l2_entry); 721 722 switch(cluster_type) { 723 case QCOW2_CLUSTER_NORMAL: 724 if (l2_entry & QCOW_OFLAG_COPIED) { 725 goto out; 726 } 727 break; 728 case QCOW2_CLUSTER_UNALLOCATED: 729 case QCOW2_CLUSTER_COMPRESSED: 730 case QCOW2_CLUSTER_ZERO: 731 break; 732 default: 733 abort(); 734 } 735 } 736 737 out: 738 assert(i <= nb_clusters); 739 return i; 740 } 741 742 /* 743 * Allocates new clusters for the given guest_offset. 744 * 745 * At most *nb_clusters are allocated, and on return *nb_clusters is updated to 746 * contain the number of clusters that have been allocated and are contiguous 747 * in the image file. 748 * 749 * If *host_offset is non-zero, it specifies the offset in the image file at 750 * which the new clusters must start. *nb_clusters can be 0 on return in this 751 * case if the cluster at host_offset is already in use. If *host_offset is 752 * zero, the clusters can be allocated anywhere in the image file. 753 * 754 * *host_offset is updated to contain the offset into the image file at which 755 * the first allocated cluster starts. 756 * 757 * Return 0 on success and -errno in error cases. -EAGAIN means that the 758 * function has been waiting for another request and the allocation must be 759 * restarted, but the whole request should not be failed. 760 */ 761 static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset, 762 uint64_t *host_offset, unsigned int *nb_clusters) 763 { 764 BDRVQcowState *s = bs->opaque; 765 QCowL2Meta *old_alloc; 766 767 trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset, 768 *host_offset, *nb_clusters); 769 770 /* 771 * Check if there already is an AIO write request in flight which allocates 772 * the same cluster. In this case we need to wait until the previous 773 * request has completed and updated the L2 table accordingly. 774 */ 775 QLIST_FOREACH(old_alloc, &s->cluster_allocs, next_in_flight) { 776 777 uint64_t start = guest_offset >> s->cluster_bits; 778 uint64_t end = start + *nb_clusters; 779 uint64_t old_start = old_alloc->offset >> s->cluster_bits; 780 uint64_t old_end = old_start + old_alloc->nb_clusters; 781 782 if (end < old_start || start > old_end) { 783 /* No intersection */ 784 } else { 785 if (start < old_start) { 786 /* Stop at the start of a running allocation */ 787 *nb_clusters = old_start - start; 788 } else { 789 *nb_clusters = 0; 790 } 791 792 if (*nb_clusters == 0) { 793 /* Wait for the dependency to complete. We need to recheck 794 * the free/allocated clusters when we continue. */ 795 qemu_co_mutex_unlock(&s->lock); 796 qemu_co_queue_wait(&old_alloc->dependent_requests); 797 qemu_co_mutex_lock(&s->lock); 798 return -EAGAIN; 799 } 800 } 801 } 802 803 if (!*nb_clusters) { 804 abort(); 805 } 806 807 /* Allocate new clusters */ 808 trace_qcow2_cluster_alloc_phys(qemu_coroutine_self()); 809 if (*host_offset == 0) { 810 int64_t cluster_offset = 811 qcow2_alloc_clusters(bs, *nb_clusters * s->cluster_size); 812 if (cluster_offset < 0) { 813 return cluster_offset; 814 } 815 *host_offset = cluster_offset; 816 return 0; 817 } else { 818 int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters); 819 if (ret < 0) { 820 return ret; 821 } 822 *nb_clusters = ret; 823 return 0; 824 } 825 } 826 827 /* 828 * alloc_cluster_offset 829 * 830 * For a given offset on the virtual disk, find the cluster offset in qcow2 831 * file. If the offset is not found, allocate a new cluster. 832 * 833 * If the cluster was already allocated, m->nb_clusters is set to 0 and 834 * other fields in m are meaningless. 835 * 836 * If the cluster is newly allocated, m->nb_clusters is set to the number of 837 * contiguous clusters that have been allocated. In this case, the other 838 * fields of m are valid and contain information about the first allocated 839 * cluster. 840 * 841 * If the request conflicts with another write request in flight, the coroutine 842 * is queued and will be reentered when the dependency has completed. 843 * 844 * Return 0 on success and -errno in error cases 845 */ 846 int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset, 847 int n_start, int n_end, int *num, QCowL2Meta *m) 848 { 849 BDRVQcowState *s = bs->opaque; 850 int l2_index, ret, sectors; 851 uint64_t *l2_table; 852 unsigned int nb_clusters, keep_clusters; 853 uint64_t cluster_offset; 854 855 trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset, 856 n_start, n_end); 857 858 /* Find L2 entry for the first involved cluster */ 859 again: 860 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 861 if (ret < 0) { 862 return ret; 863 } 864 865 /* 866 * Calculate the number of clusters to look for. We stop at L2 table 867 * boundaries to keep things simple. 868 */ 869 nb_clusters = MIN(size_to_clusters(s, n_end << BDRV_SECTOR_BITS), 870 s->l2_size - l2_index); 871 872 cluster_offset = be64_to_cpu(l2_table[l2_index]); 873 874 /* 875 * Check how many clusters are already allocated and don't need COW, and how 876 * many need a new allocation. 877 */ 878 if (qcow2_get_cluster_type(cluster_offset) == QCOW2_CLUSTER_NORMAL 879 && (cluster_offset & QCOW_OFLAG_COPIED)) 880 { 881 /* We keep all QCOW_OFLAG_COPIED clusters */ 882 keep_clusters = 883 count_contiguous_clusters(nb_clusters, s->cluster_size, 884 &l2_table[l2_index], 0, 885 QCOW_OFLAG_COPIED | QCOW_OFLAG_ZERO); 886 assert(keep_clusters <= nb_clusters); 887 nb_clusters -= keep_clusters; 888 } else { 889 keep_clusters = 0; 890 cluster_offset = 0; 891 } 892 893 if (nb_clusters > 0) { 894 /* For the moment, overwrite compressed clusters one by one */ 895 uint64_t entry = be64_to_cpu(l2_table[l2_index + keep_clusters]); 896 if (entry & QCOW_OFLAG_COMPRESSED) { 897 nb_clusters = 1; 898 } else { 899 nb_clusters = count_cow_clusters(s, nb_clusters, l2_table, 900 l2_index + keep_clusters); 901 } 902 } 903 904 cluster_offset &= L2E_OFFSET_MASK; 905 906 /* 907 * The L2 table isn't used any more after this. As long as the cache works 908 * synchronously, it's important to release it before calling 909 * do_alloc_cluster_offset, which may yield if we need to wait for another 910 * request to complete. If we still had the reference, we could use up the 911 * whole cache with sleeping requests. 912 */ 913 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 914 if (ret < 0) { 915 return ret; 916 } 917 918 /* If there is something left to allocate, do that now */ 919 *m = (QCowL2Meta) { 920 .cluster_offset = cluster_offset, 921 .nb_clusters = 0, 922 }; 923 qemu_co_queue_init(&m->dependent_requests); 924 925 if (nb_clusters > 0) { 926 uint64_t alloc_offset; 927 uint64_t alloc_cluster_offset; 928 uint64_t keep_bytes = keep_clusters * s->cluster_size; 929 930 /* Calculate start and size of allocation */ 931 alloc_offset = offset + keep_bytes; 932 933 if (keep_clusters == 0) { 934 alloc_cluster_offset = 0; 935 } else { 936 alloc_cluster_offset = cluster_offset + keep_bytes; 937 } 938 939 /* Allocate, if necessary at a given offset in the image file */ 940 ret = do_alloc_cluster_offset(bs, alloc_offset, &alloc_cluster_offset, 941 &nb_clusters); 942 if (ret == -EAGAIN) { 943 goto again; 944 } else if (ret < 0) { 945 goto fail; 946 } 947 948 /* save info needed for meta data update */ 949 if (nb_clusters > 0) { 950 int requested_sectors = n_end - keep_clusters * s->cluster_sectors; 951 int avail_sectors = (keep_clusters + nb_clusters) 952 << (s->cluster_bits - BDRV_SECTOR_BITS); 953 954 *m = (QCowL2Meta) { 955 .cluster_offset = keep_clusters == 0 ? 956 alloc_cluster_offset : cluster_offset, 957 .alloc_offset = alloc_cluster_offset, 958 .offset = alloc_offset, 959 .n_start = keep_clusters == 0 ? n_start : 0, 960 .nb_clusters = nb_clusters, 961 .nb_available = MIN(requested_sectors, avail_sectors), 962 }; 963 qemu_co_queue_init(&m->dependent_requests); 964 QLIST_INSERT_HEAD(&s->cluster_allocs, m, next_in_flight); 965 } 966 } 967 968 /* Some cleanup work */ 969 sectors = (keep_clusters + nb_clusters) << (s->cluster_bits - 9); 970 if (sectors > n_end) { 971 sectors = n_end; 972 } 973 974 assert(sectors > n_start); 975 *num = sectors - n_start; 976 977 return 0; 978 979 fail: 980 if (m->nb_clusters > 0) { 981 QLIST_REMOVE(m, next_in_flight); 982 } 983 return ret; 984 } 985 986 static int decompress_buffer(uint8_t *out_buf, int out_buf_size, 987 const uint8_t *buf, int buf_size) 988 { 989 z_stream strm1, *strm = &strm1; 990 int ret, out_len; 991 992 memset(strm, 0, sizeof(*strm)); 993 994 strm->next_in = (uint8_t *)buf; 995 strm->avail_in = buf_size; 996 strm->next_out = out_buf; 997 strm->avail_out = out_buf_size; 998 999 ret = inflateInit2(strm, -12); 1000 if (ret != Z_OK) 1001 return -1; 1002 ret = inflate(strm, Z_FINISH); 1003 out_len = strm->next_out - out_buf; 1004 if ((ret != Z_STREAM_END && ret != Z_BUF_ERROR) || 1005 out_len != out_buf_size) { 1006 inflateEnd(strm); 1007 return -1; 1008 } 1009 inflateEnd(strm); 1010 return 0; 1011 } 1012 1013 int qcow2_decompress_cluster(BlockDriverState *bs, uint64_t cluster_offset) 1014 { 1015 BDRVQcowState *s = bs->opaque; 1016 int ret, csize, nb_csectors, sector_offset; 1017 uint64_t coffset; 1018 1019 coffset = cluster_offset & s->cluster_offset_mask; 1020 if (s->cluster_cache_offset != coffset) { 1021 nb_csectors = ((cluster_offset >> s->csize_shift) & s->csize_mask) + 1; 1022 sector_offset = coffset & 511; 1023 csize = nb_csectors * 512 - sector_offset; 1024 BLKDBG_EVENT(bs->file, BLKDBG_READ_COMPRESSED); 1025 ret = bdrv_read(bs->file, coffset >> 9, s->cluster_data, nb_csectors); 1026 if (ret < 0) { 1027 return ret; 1028 } 1029 if (decompress_buffer(s->cluster_cache, s->cluster_size, 1030 s->cluster_data + sector_offset, csize) < 0) { 1031 return -EIO; 1032 } 1033 s->cluster_cache_offset = coffset; 1034 } 1035 return 0; 1036 } 1037 1038 /* 1039 * This discards as many clusters of nb_clusters as possible at once (i.e. 1040 * all clusters in the same L2 table) and returns the number of discarded 1041 * clusters. 1042 */ 1043 static int discard_single_l2(BlockDriverState *bs, uint64_t offset, 1044 unsigned int nb_clusters) 1045 { 1046 BDRVQcowState *s = bs->opaque; 1047 uint64_t *l2_table; 1048 int l2_index; 1049 int ret; 1050 int i; 1051 1052 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 1053 if (ret < 0) { 1054 return ret; 1055 } 1056 1057 /* Limit nb_clusters to one L2 table */ 1058 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1059 1060 for (i = 0; i < nb_clusters; i++) { 1061 uint64_t old_offset; 1062 1063 old_offset = be64_to_cpu(l2_table[l2_index + i]); 1064 if ((old_offset & L2E_OFFSET_MASK) == 0) { 1065 continue; 1066 } 1067 1068 /* First remove L2 entries */ 1069 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table); 1070 l2_table[l2_index + i] = cpu_to_be64(0); 1071 1072 /* Then decrease the refcount */ 1073 qcow2_free_any_clusters(bs, old_offset, 1); 1074 } 1075 1076 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 1077 if (ret < 0) { 1078 return ret; 1079 } 1080 1081 return nb_clusters; 1082 } 1083 1084 int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset, 1085 int nb_sectors) 1086 { 1087 BDRVQcowState *s = bs->opaque; 1088 uint64_t end_offset; 1089 unsigned int nb_clusters; 1090 int ret; 1091 1092 end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS); 1093 1094 /* Round start up and end down */ 1095 offset = align_offset(offset, s->cluster_size); 1096 end_offset &= ~(s->cluster_size - 1); 1097 1098 if (offset > end_offset) { 1099 return 0; 1100 } 1101 1102 nb_clusters = size_to_clusters(s, end_offset - offset); 1103 1104 /* Each L2 table is handled by its own loop iteration */ 1105 while (nb_clusters > 0) { 1106 ret = discard_single_l2(bs, offset, nb_clusters); 1107 if (ret < 0) { 1108 return ret; 1109 } 1110 1111 nb_clusters -= ret; 1112 offset += (ret * s->cluster_size); 1113 } 1114 1115 return 0; 1116 } 1117 1118 /* 1119 * This zeroes as many clusters of nb_clusters as possible at once (i.e. 1120 * all clusters in the same L2 table) and returns the number of zeroed 1121 * clusters. 1122 */ 1123 static int zero_single_l2(BlockDriverState *bs, uint64_t offset, 1124 unsigned int nb_clusters) 1125 { 1126 BDRVQcowState *s = bs->opaque; 1127 uint64_t *l2_table; 1128 int l2_index; 1129 int ret; 1130 int i; 1131 1132 ret = get_cluster_table(bs, offset, &l2_table, &l2_index); 1133 if (ret < 0) { 1134 return ret; 1135 } 1136 1137 /* Limit nb_clusters to one L2 table */ 1138 nb_clusters = MIN(nb_clusters, s->l2_size - l2_index); 1139 1140 for (i = 0; i < nb_clusters; i++) { 1141 uint64_t old_offset; 1142 1143 old_offset = be64_to_cpu(l2_table[l2_index + i]); 1144 1145 /* Update L2 entries */ 1146 qcow2_cache_entry_mark_dirty(s->l2_table_cache, l2_table); 1147 if (old_offset & QCOW_OFLAG_COMPRESSED) { 1148 l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO); 1149 qcow2_free_any_clusters(bs, old_offset, 1); 1150 } else { 1151 l2_table[l2_index + i] |= cpu_to_be64(QCOW_OFLAG_ZERO); 1152 } 1153 } 1154 1155 ret = qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table); 1156 if (ret < 0) { 1157 return ret; 1158 } 1159 1160 return nb_clusters; 1161 } 1162 1163 int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors) 1164 { 1165 BDRVQcowState *s = bs->opaque; 1166 unsigned int nb_clusters; 1167 int ret; 1168 1169 /* The zero flag is only supported by version 3 and newer */ 1170 if (s->qcow_version < 3) { 1171 return -ENOTSUP; 1172 } 1173 1174 /* Each L2 table is handled by its own loop iteration */ 1175 nb_clusters = size_to_clusters(s, nb_sectors << BDRV_SECTOR_BITS); 1176 1177 while (nb_clusters > 0) { 1178 ret = zero_single_l2(bs, offset, nb_clusters); 1179 if (ret < 0) { 1180 return ret; 1181 } 1182 1183 nb_clusters -= ret; 1184 offset += (ret * s->cluster_size); 1185 } 1186 1187 return 0; 1188 } 1189