block/qcow2.c

   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 #include "qemu-common.h"
  25 #include "block/block_int.h"
  26 #include "qemu/module.h"
  27 #include <zlib.h>
  28 #include "qemu/aes.h"
  29 #include "block/qcow2.h"
  30 #include "qemu/error-report.h"
  31 #include "qapi/qmp/qerror.h"
  32 #include "qapi/qmp/qbool.h"
  33 #include "trace.h"
  34
  35 /*
  36   Differences with QCOW:
  37
  38   - Support for multiple incremental snapshots.
  39   - Memory management by reference counts.
  40   - Clusters which have a reference count of one have the bit
  41     QCOW_OFLAG_COPIED to optimize write performance.
  42   - Size of compressed clusters is stored in sectors to reduce bit usage
  43     in the cluster offsets.
  44   - Support for storing additional data (such as the VM state) in the
  45     snapshots.
  46   - If a backing store is used, the cluster size is not constrained
  47     (could be backported to QCOW).
  48   - L2 tables have always a size of one cluster.
  49 */
  50
  51
  52 typedef struct {
  53     uint32_t magic;
  54     uint32_t len;
  55 } QEMU_PACKED QCowExtension;
  56
  57 #define  QCOW2_EXT_MAGIC_END 0
  58 #define  QCOW2_EXT_MAGIC_BACKING_FORMAT 0xE2792ACA
  59 #define  QCOW2_EXT_MAGIC_FEATURE_TABLE 0x6803f857
  60
  61 static int qcow2_probe(const uint8_t *buf, int buf_size, const char *filename)
  62 {
  63     const QCowHeader *cow_header = (const void *)buf;
  64
  65     if (buf_size >= sizeof(QCowHeader) &&
  66         be32_to_cpu(cow_header->magic) == QCOW_MAGIC &&
  67         be32_to_cpu(cow_header->version) >= 2)
  68         return 100;
  69     else
  70         return 0;
  71 }
  72
  73
  74 /*
  75  * read qcow2 extension and fill bs
  76  * start reading from start_offset
  77  * finish reading upon magic of value 0 or when end_offset reached
  78  * unknown magic is skipped (future extension this version knows nothing about)
  79  * return 0 upon success, non-0 otherwise
  80  */
  81 static int qcow2_read_extensions(BlockDriverState *bs, uint64_t start_offset,
  82                                  uint64_t end_offset, void **p_feature_table,
  83                                  Error **errp)
  84 {
  85     BDRVQcowState *s = bs->opaque;
  86     QCowExtension ext;
  87     uint64_t offset;
  88     int ret;
  89
  90 #ifdef DEBUG_EXT
  91     printf("qcow2_read_extensions: start=%ld end=%ld\n", start_offset, end_offset);
  92 #endif
  93     offset = start_offset;
  94     while (offset < end_offset) {
  95
  96 #ifdef DEBUG_EXT
  97         /* Sanity check */
  98         if (offset > s->cluster_size)
  99             printf("qcow2_read_extension: suspicious offset %lu\n", offset);
 100
 101         printf("attempting to read extended header in offset %lu\n", offset);
 102 #endif
 103
 104         ret = bdrv_pread(bs->file, offset, &ext, sizeof(ext));
 105         if (ret < 0) {
 106             error_setg_errno(errp, -ret, "qcow2_read_extension: ERROR: "
 107                              "pread fail from offset %" PRIu64, offset);
 108             return 1;
 109         }
 110         be32_to_cpus(&ext.magic);
 111         be32_to_cpus(&ext.len);
 112         offset += sizeof(ext);
 113 #ifdef DEBUG_EXT
 114         printf("ext.magic = 0x%x\n", ext.magic);
 115 #endif
 116         if (ext.len > end_offset - offset) {
 117             error_setg(errp, "Header extension too large");
 118             return -EINVAL;
 119         }
 120
 121         switch (ext.magic) {
 122         case QCOW2_EXT_MAGIC_END:
 123             return 0;
 124
 125         case QCOW2_EXT_MAGIC_BACKING_FORMAT:
 126             if (ext.len >= sizeof(bs->backing_format)) {
 127                 error_setg(errp, "ERROR: ext_backing_format: len=%u too large"
 128                            " (>=%zu)", ext.len, sizeof(bs->backing_format));
 129                 return 2;
 130             }
 131             ret = bdrv_pread(bs->file, offset, bs->backing_format, ext.len);
 132             if (ret < 0) {
 133                 error_setg_errno(errp, -ret, "ERROR: ext_backing_format: "
 134                                  "Could not read format name");
 135                 return 3;
 136             }
 137             bs->backing_format[ext.len] = '\0';
 138 #ifdef DEBUG_EXT
 139             printf("Qcow2: Got format extension %s\n", bs->backing_format);
 140 #endif
 141             break;
 142
 143         case QCOW2_EXT_MAGIC_FEATURE_TABLE:
 144             if (p_feature_table != NULL) {
 145                 void* feature_table = g_malloc0(ext.len + 2 * sizeof(Qcow2Feature));
 146                 ret = bdrv_pread(bs->file, offset , feature_table, ext.len);
 147                 if (ret < 0) {
 148                     error_setg_errno(errp, -ret, "ERROR: ext_feature_table: "
 149                                      "Could not read table");
 150                     return ret;
 151                 }
 152
 153                 *p_feature_table = feature_table;
 154             }
 155             break;
 156
 157         default:
 158             /* unknown magic - save it in case we need to rewrite the header */
 159             {
 160                 Qcow2UnknownHeaderExtension *uext;
 161
 162                 uext = g_malloc0(sizeof(*uext)  + ext.len);
 163                 uext->magic = ext.magic;
 164                 uext->len = ext.len;
 165                 QLIST_INSERT_HEAD(&s->unknown_header_ext, uext, next);
 166
 167                 ret = bdrv_pread(bs->file, offset , uext->data, uext->len);
 168                 if (ret < 0) {
 169                     error_setg_errno(errp, -ret, "ERROR: unknown extension: "
 170                                      "Could not read data");
 171                     return ret;
 172                 }
 173             }
 174             break;
 175         }
 176
 177         offset += ((ext.len + 7) & ~7);
 178     }
 179
 180     return 0;
 181 }
 182
 183 static void cleanup_unknown_header_ext(BlockDriverState *bs)
 184 {
 185     BDRVQcowState *s = bs->opaque;
 186     Qcow2UnknownHeaderExtension *uext, *next;
 187
 188     QLIST_FOREACH_SAFE(uext, &s->unknown_header_ext, next, next) {
 189         QLIST_REMOVE(uext, next);
 190         g_free(uext);
 191     }
 192 }
 193
 194 static void GCC_FMT_ATTR(3, 4) report_unsupported(BlockDriverState *bs,
 195     Error **errp, const char *fmt, ...)
 196 {
 197     char msg[64];
 198     va_list ap;
 199
 200     va_start(ap, fmt);
 201     vsnprintf(msg, sizeof(msg), fmt, ap);
 202     va_end(ap);
 203
 204     error_set(errp, QERR_UNKNOWN_BLOCK_FORMAT_FEATURE, bs->device_name, "qcow2",
 205               msg);
 206 }
 207
 208 static void report_unsupported_feature(BlockDriverState *bs,
 209     Error **errp, Qcow2Feature *table, uint64_t mask)
 210 {
 211     while (table && table->name[0] != '\0') {
 212         if (table->type == QCOW2_FEAT_TYPE_INCOMPATIBLE) {
 213             if (mask & (1 << table->bit)) {
 214                 report_unsupported(bs, errp, "%.46s", table->name);
 215                 mask &= ~(1 << table->bit);
 216             }
 217         }
 218         table++;
 219     }
 220
 221     if (mask) {
 222         report_unsupported(bs, errp, "Unknown incompatible feature: %" PRIx64,
 223                            mask);
 224     }
 225 }
 226
 227 /*
 228  * Sets the dirty bit and flushes afterwards if necessary.
 229  *
 230  * The incompatible_features bit is only set if the image file header was
 231  * updated successfully.  Therefore it is not required to check the return
 232  * value of this function.
 233  */
 234 int qcow2_mark_dirty(BlockDriverState *bs)
 235 {
 236     BDRVQcowState *s = bs->opaque;
 237     uint64_t val;
 238     int ret;
 239
 240     assert(s->qcow_version >= 3);
 241
 242     if (s->incompatible_features & QCOW2_INCOMPAT_DIRTY) {
 243         return 0; /* already dirty */
 244     }
 245
 246     val = cpu_to_be64(s->incompatible_features | QCOW2_INCOMPAT_DIRTY);
 247     ret = bdrv_pwrite(bs->file, offsetof(QCowHeader, incompatible_features),
 248                       &val, sizeof(val));
 249     if (ret < 0) {
 250         return ret;
 251     }
 252     ret = bdrv_flush(bs->file);
 253     if (ret < 0) {
 254         return ret;
 255     }
 256
 257     /* Only treat image as dirty if the header was updated successfully */
 258     s->incompatible_features |= QCOW2_INCOMPAT_DIRTY;
 259     return 0;
 260 }
 261
 262 /*
 263  * Clears the dirty bit and flushes before if necessary.  Only call this
 264  * function when there are no pending requests, it does not guard against
 265  * concurrent requests dirtying the image.
 266  */
 267 static int qcow2_mark_clean(BlockDriverState *bs)
 268 {
 269     BDRVQcowState *s = bs->opaque;
 270
 271     if (s->incompatible_features & QCOW2_INCOMPAT_DIRTY) {
 272         int ret = bdrv_flush(bs);
 273         if (ret < 0) {
 274             return ret;
 275         }
 276
 277         s->incompatible_features &= ~QCOW2_INCOMPAT_DIRTY;
 278         return qcow2_update_header(bs);
 279     }
 280     return 0;
 281 }
 282
 283 /*
 284  * Marks the image as corrupt.
 285  */
 286 int qcow2_mark_corrupt(BlockDriverState *bs)
 287 {
 288     BDRVQcowState *s = bs->opaque;
 289
 290     s->incompatible_features |= QCOW2_INCOMPAT_CORRUPT;
 291     return qcow2_update_header(bs);
 292 }
 293
 294 /*
 295  * Marks the image as consistent, i.e., unsets the corrupt bit, and flushes
 296  * before if necessary.
 297  */
 298 int qcow2_mark_consistent(BlockDriverState *bs)
 299 {
 300     BDRVQcowState *s = bs->opaque;
 301
 302     if (s->incompatible_features & QCOW2_INCOMPAT_CORRUPT) {
 303         int ret = bdrv_flush(bs);
 304         if (ret < 0) {
 305             return ret;
 306         }
 307
 308         s->incompatible_features &= ~QCOW2_INCOMPAT_CORRUPT;
 309         return qcow2_update_header(bs);
 310     }
 311     return 0;
 312 }
 313
 314 static int qcow2_check(BlockDriverState *bs, BdrvCheckResult *result,
 315                        BdrvCheckMode fix)
 316 {
 317     int ret = qcow2_check_refcounts(bs, result, fix);
 318     if (ret < 0) {
 319         return ret;
 320     }
 321
 322     if (fix && result->check_errors == 0 && result->corruptions == 0) {
 323         ret = qcow2_mark_clean(bs);
 324         if (ret < 0) {
 325             return ret;
 326         }
 327         return qcow2_mark_consistent(bs);
 328     }
 329     return ret;
 330 }
 331
 332 static int validate_table_offset(BlockDriverState *bs, uint64_t offset,
 333                                  uint64_t entries, size_t entry_len)
 334 {
 335     BDRVQcowState *s = bs->opaque;
 336     uint64_t size;
 337
 338     /* Use signed INT64_MAX as the maximum even for uint64_t header fields,
 339      * because values will be passed to qemu functions taking int64_t. */
 340     if (entries > INT64_MAX / entry_len) {
 341         return -EINVAL;
 342     }
 343
 344     size = entries * entry_len;
 345
 346     if (INT64_MAX - size < offset) {
 347         return -EINVAL;
 348     }
 349
 350     /* Tables must be cluster aligned */
 351     if (offset & (s->cluster_size - 1)) {
 352         return -EINVAL;
 353     }
 354
 355     return 0;
 356 }
 357
 358 static QemuOptsList qcow2_runtime_opts = {
 359     .name = "qcow2",
 360     .head = QTAILQ_HEAD_INITIALIZER(qcow2_runtime_opts.head),
 361     .desc = {
 362         {
 363             .name = QCOW2_OPT_LAZY_REFCOUNTS,
 364             .type = QEMU_OPT_BOOL,
 365             .help = "Postpone refcount updates",
 366         },
 367         {
 368             .name = QCOW2_OPT_DISCARD_REQUEST,
 369             .type = QEMU_OPT_BOOL,
 370             .help = "Pass guest discard requests to the layer below",
 371         },
 372         {
 373             .name = QCOW2_OPT_DISCARD_SNAPSHOT,
 374             .type = QEMU_OPT_BOOL,
 375             .help = "Generate discard requests when snapshot related space "
 376                     "is freed",
 377         },
 378         {
 379             .name = QCOW2_OPT_DISCARD_OTHER,
 380             .type = QEMU_OPT_BOOL,
 381             .help = "Generate discard requests when other clusters are freed",
 382         },
 383         {
 384             .name = QCOW2_OPT_OVERLAP,
 385             .type = QEMU_OPT_STRING,
 386             .help = "Selects which overlap checks to perform from a range of "
 387                     "templates (none, constant, cached, all)",
 388         },
 389         {
 390             .name = QCOW2_OPT_OVERLAP_MAIN_HEADER,
 391             .type = QEMU_OPT_BOOL,
 392             .help = "Check for unintended writes into the main qcow2 header",
 393         },
 394         {
 395             .name = QCOW2_OPT_OVERLAP_ACTIVE_L1,
 396             .type = QEMU_OPT_BOOL,
 397             .help = "Check for unintended writes into the active L1 table",
 398         },
 399         {
 400             .name = QCOW2_OPT_OVERLAP_ACTIVE_L2,
 401             .type = QEMU_OPT_BOOL,
 402             .help = "Check for unintended writes into an active L2 table",
 403         },
 404         {
 405             .name = QCOW2_OPT_OVERLAP_REFCOUNT_TABLE,
 406             .type = QEMU_OPT_BOOL,
 407             .help = "Check for unintended writes into the refcount table",
 408         },
 409         {
 410             .name = QCOW2_OPT_OVERLAP_REFCOUNT_BLOCK,
 411             .type = QEMU_OPT_BOOL,
 412             .help = "Check for unintended writes into a refcount block",
 413         },
 414         {
 415             .name = QCOW2_OPT_OVERLAP_SNAPSHOT_TABLE,
 416             .type = QEMU_OPT_BOOL,
 417             .help = "Check for unintended writes into the snapshot table",
 418         },
 419         {
 420             .name = QCOW2_OPT_OVERLAP_INACTIVE_L1,
 421             .type = QEMU_OPT_BOOL,
 422             .help = "Check for unintended writes into an inactive L1 table",
 423         },
 424         {
 425             .name = QCOW2_OPT_OVERLAP_INACTIVE_L2,
 426             .type = QEMU_OPT_BOOL,
 427             .help = "Check for unintended writes into an inactive L2 table",
 428         },
 429         { /* end of list */ }
 430     },
 431 };
 432
 433 static const char *overlap_bool_option_names[QCOW2_OL_MAX_BITNR] = {
 434     [QCOW2_OL_MAIN_HEADER_BITNR]    = QCOW2_OPT_OVERLAP_MAIN_HEADER,
 435     [QCOW2_OL_ACTIVE_L1_BITNR]      = QCOW2_OPT_OVERLAP_ACTIVE_L1,
 436     [QCOW2_OL_ACTIVE_L2_BITNR]      = QCOW2_OPT_OVERLAP_ACTIVE_L2,
 437     [QCOW2_OL_REFCOUNT_TABLE_BITNR] = QCOW2_OPT_OVERLAP_REFCOUNT_TABLE,
 438     [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = QCOW2_OPT_OVERLAP_REFCOUNT_BLOCK,
 439     [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = QCOW2_OPT_OVERLAP_SNAPSHOT_TABLE,
 440     [QCOW2_OL_INACTIVE_L1_BITNR]    = QCOW2_OPT_OVERLAP_INACTIVE_L1,
 441     [QCOW2_OL_INACTIVE_L2_BITNR]    = QCOW2_OPT_OVERLAP_INACTIVE_L2,
 442 };
 443
 444 static int qcow2_open(BlockDriverState *bs, QDict *options, int flags,
 445                       Error **errp)
 446 {
 447     BDRVQcowState *s = bs->opaque;
 448     unsigned int len, i;
 449     int ret = 0;
 450     QCowHeader header;
 451     QemuOpts *opts;
 452     Error *local_err = NULL;
 453     uint64_t ext_end;
 454     uint64_t l1_vm_state_index;
 455     const char *opt_overlap_check;
 456     int overlap_check_template = 0;
 457
 458     ret = bdrv_pread(bs->file, 0, &header, sizeof(header));
 459     if (ret < 0) {
 460         error_setg_errno(errp, -ret, "Could not read qcow2 header");
 461         goto fail;
 462     }
 463     be32_to_cpus(&header.magic);
 464     be32_to_cpus(&header.version);
 465     be64_to_cpus(&header.backing_file_offset);
 466     be32_to_cpus(&header.backing_file_size);
 467     be64_to_cpus(&header.size);
 468     be32_to_cpus(&header.cluster_bits);
 469     be32_to_cpus(&header.crypt_method);
 470     be64_to_cpus(&header.l1_table_offset);
 471     be32_to_cpus(&header.l1_size);
 472     be64_to_cpus(&header.refcount_table_offset);
 473     be32_to_cpus(&header.refcount_table_clusters);
 474     be64_to_cpus(&header.snapshots_offset);
 475     be32_to_cpus(&header.nb_snapshots);
 476
 477     if (header.magic != QCOW_MAGIC) {
 478         error_setg(errp, "Image is not in qcow2 format");
 479         ret = -EINVAL;
 480         goto fail;
 481     }
 482     if (header.version < 2 || header.version > 3) {
 483         report_unsupported(bs, errp, "QCOW version %d", header.version);
 484         ret = -ENOTSUP;
 485         goto fail;
 486     }
 487
 488     s->qcow_version = header.version;
 489
 490     /* Initialise cluster size */
 491     if (header.cluster_bits < MIN_CLUSTER_BITS ||
 492         header.cluster_bits > MAX_CLUSTER_BITS) {
 493         error_setg(errp, "Unsupported cluster size: 2^%i", header.cluster_bits);
 494         ret = -EINVAL;
 495         goto fail;
 496     }
 497
 498     s->cluster_bits = header.cluster_bits;
 499     s->cluster_size = 1 << s->cluster_bits;
 500     s->cluster_sectors = 1 << (s->cluster_bits - 9);
 501
 502     /* Initialise version 3 header fields */
 503     if (header.version == 2) {
 504         header.incompatible_features    = 0;
 505         header.compatible_features      = 0;
 506         header.autoclear_features       = 0;
 507         header.refcount_order           = 4;
 508         header.header_length            = 72;
 509     } else {
 510         be64_to_cpus(&header.incompatible_features);
 511         be64_to_cpus(&header.compatible_features);
 512         be64_to_cpus(&header.autoclear_features);
 513         be32_to_cpus(&header.refcount_order);
 514         be32_to_cpus(&header.header_length);
 515
 516         if (header.header_length < 104) {
 517             error_setg(errp, "qcow2 header too short");
 518             ret = -EINVAL;
 519             goto fail;
 520         }
 521     }
 522
 523     if (header.header_length > s->cluster_size) {
 524         error_setg(errp, "qcow2 header exceeds cluster size");
 525         ret = -EINVAL;
 526         goto fail;
 527     }
 528
 529     if (header.header_length > sizeof(header)) {
 530         s->unknown_header_fields_size = header.header_length - sizeof(header);
 531         s->unknown_header_fields = g_malloc(s->unknown_header_fields_size);
 532         ret = bdrv_pread(bs->file, sizeof(header), s->unknown_header_fields,
 533                          s->unknown_header_fields_size);
 534         if (ret < 0) {
 535             error_setg_errno(errp, -ret, "Could not read unknown qcow2 header "
 536                              "fields");
 537             goto fail;
 538         }
 539     }
 540
 541     if (header.backing_file_offset > s->cluster_size) {
 542         error_setg(errp, "Invalid backing file offset");
 543         ret = -EINVAL;
 544         goto fail;
 545     }
 546
 547     if (header.backing_file_offset) {
 548         ext_end = header.backing_file_offset;
 549     } else {
 550         ext_end = 1 << header.cluster_bits;
 551     }
 552
 553     /* Handle feature bits */
 554     s->incompatible_features    = header.incompatible_features;
 555     s->compatible_features      = header.compatible_features;
 556     s->autoclear_features       = header.autoclear_features;
 557
 558     if (s->incompatible_features & ~QCOW2_INCOMPAT_MASK) {
 559         void *feature_table = NULL;
 560         qcow2_read_extensions(bs, header.header_length, ext_end,
 561                               &feature_table, NULL);
 562         report_unsupported_feature(bs, errp, feature_table,
 563                                    s->incompatible_features &
 564                                    ~QCOW2_INCOMPAT_MASK);
 565         ret = -ENOTSUP;
 566         g_free(feature_table);
 567         goto fail;
 568     }
 569
 570     if (s->incompatible_features & QCOW2_INCOMPAT_CORRUPT) {
 571         /* Corrupt images may not be written to unless they are being repaired
 572          */
 573         if ((flags & BDRV_O_RDWR) && !(flags & BDRV_O_CHECK)) {
 574             error_setg(errp, "qcow2: Image is corrupt; cannot be opened "
 575                        "read/write");
 576             ret = -EACCES;
 577             goto fail;
 578         }
 579     }
 580
 581     /* Check support for various header values */
 582     if (header.refcount_order != 4) {
 583         report_unsupported(bs, errp, "%d bit reference counts",
 584                            1 << header.refcount_order);
 585         ret = -ENOTSUP;
 586         goto fail;
 587     }
 588     s->refcount_order = header.refcount_order;
 589
 590     if (header.crypt_method > QCOW_CRYPT_AES) {
 591         error_setg(errp, "Unsupported encryption method: %i",
 592                    header.crypt_method);
 593         ret = -EINVAL;
 594         goto fail;
 595     }
 596     s->crypt_method_header = header.crypt_method;
 597     if (s->crypt_method_header) {
 598         bs->encrypted = 1;
 599     }
 600
 601     s->l2_bits = s->cluster_bits - 3; /* L2 is always one cluster */
 602     s->l2_size = 1 << s->l2_bits;
 603     bs->total_sectors = header.size / 512;
 604     s->csize_shift = (62 - (s->cluster_bits - 8));
 605     s->csize_mask = (1 << (s->cluster_bits - 8)) - 1;
 606     s->cluster_offset_mask = (1LL << s->csize_shift) - 1;
 607
 608     s->refcount_table_offset = header.refcount_table_offset;
 609     s->refcount_table_size =
 610         header.refcount_table_clusters << (s->cluster_bits - 3);
 611
 612     if (header.refcount_table_clusters > qcow2_max_refcount_clusters(s)) {
 613         error_setg(errp, "Reference count table too large");
 614         ret = -EINVAL;
 615         goto fail;
 616     }
 617
 618     ret = validate_table_offset(bs, s->refcount_table_offset,
 619                                 s->refcount_table_size, sizeof(uint64_t));
 620     if (ret < 0) {
 621         error_setg(errp, "Invalid reference count table offset");
 622         goto fail;
 623     }
 624
 625     /* Snapshot table offset/length */
 626     if (header.nb_snapshots > QCOW_MAX_SNAPSHOTS) {
 627         error_setg(errp, "Too many snapshots");
 628         ret = -EINVAL;
 629         goto fail;
 630     }
 631
 632     ret = validate_table_offset(bs, header.snapshots_offset,
 633                                 header.nb_snapshots,
 634                                 sizeof(QCowSnapshotHeader));
 635     if (ret < 0) {
 636         error_setg(errp, "Invalid snapshot table offset");
 637         goto fail;
 638     }
 639
 640     /* read the level 1 table */
 641     if (header.l1_size > QCOW_MAX_L1_SIZE) {
 642         error_setg(errp, "Active L1 table too large");
 643         ret = -EFBIG;
 644         goto fail;
 645     }
 646     s->l1_size = header.l1_size;
 647
 648     l1_vm_state_index = size_to_l1(s, header.size);
 649     if (l1_vm_state_index > INT_MAX) {
 650         error_setg(errp, "Image is too big");
 651         ret = -EFBIG;
 652         goto fail;
 653     }
 654     s->l1_vm_state_index = l1_vm_state_index;
 655
 656     /* the L1 table must contain at least enough entries to put
 657        header.size bytes */
 658     if (s->l1_size < s->l1_vm_state_index) {
 659         error_setg(errp, "L1 table is too small");
 660         ret = -EINVAL;
 661         goto fail;
 662     }
 663
 664     ret = validate_table_offset(bs, header.l1_table_offset,
 665                                 header.l1_size, sizeof(uint64_t));
 666     if (ret < 0) {
 667         error_setg(errp, "Invalid L1 table offset");
 668         goto fail;
 669     }
 670     s->l1_table_offset = header.l1_table_offset;
 671
 672
 673     if (s->l1_size > 0) {
 674         s->l1_table = g_malloc0(
 675             align_offset(s->l1_size * sizeof(uint64_t), 512));
 676         ret = bdrv_pread(bs->file, s->l1_table_offset, s->l1_table,
 677                          s->l1_size * sizeof(uint64_t));
 678         if (ret < 0) {
 679             error_setg_errno(errp, -ret, "Could not read L1 table");
 680             goto fail;
 681         }
 682         for(i = 0;i < s->l1_size; i++) {
 683             be64_to_cpus(&s->l1_table[i]);
 684         }
 685     }
 686
 687     /* alloc L2 table/refcount block cache */
 688     s->l2_table_cache = qcow2_cache_create(bs, L2_CACHE_SIZE);
 689     s->refcount_block_cache = qcow2_cache_create(bs, REFCOUNT_CACHE_SIZE);
 690
 691     s->cluster_cache = g_malloc(s->cluster_size);
 692     /* one more sector for decompressed data alignment */
 693     s->cluster_data = qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size
 694                                   + 512);
 695     s->cluster_cache_offset = -1;
 696     s->flags = flags;
 697
 698     ret = qcow2_refcount_init(bs);
 699     if (ret != 0) {
 700         error_setg_errno(errp, -ret, "Could not initialize refcount handling");
 701         goto fail;
 702     }
 703
 704     QLIST_INIT(&s->cluster_allocs);
 705     QTAILQ_INIT(&s->discards);
 706
 707     /* read qcow2 extensions */
 708     if (qcow2_read_extensions(bs, header.header_length, ext_end, NULL,
 709         &local_err)) {
 710         error_propagate(errp, local_err);
 711         ret = -EINVAL;
 712         goto fail;
 713     }
 714
 715     /* read the backing file name */
 716     if (header.backing_file_offset != 0) {
 717         len = header.backing_file_size;
 718         if (len > MIN(1023, s->cluster_size - header.backing_file_offset)) {
 719             error_setg(errp, "Backing file name too long");
 720             ret = -EINVAL;
 721             goto fail;
 722         }
 723         ret = bdrv_pread(bs->file, header.backing_file_offset,
 724                          bs->backing_file, len);
 725         if (ret < 0) {
 726             error_setg_errno(errp, -ret, "Could not read backing file name");
 727             goto fail;
 728         }
 729         bs->backing_file[len] = '\0';
 730     }
 731
 732     /* Internal snapshots */
 733     s->snapshots_offset = header.snapshots_offset;
 734     s->nb_snapshots = header.nb_snapshots;
 735
 736     ret = qcow2_read_snapshots(bs);
 737     if (ret < 0) {
 738         error_setg_errno(errp, -ret, "Could not read snapshots");
 739         goto fail;
 740     }
 741
 742     /* Clear unknown autoclear feature bits */
 743     if (!bs->read_only && !(flags & BDRV_O_INCOMING) && s->autoclear_features) {
 744         s->autoclear_features = 0;
 745         ret = qcow2_update_header(bs);
 746         if (ret < 0) {
 747             error_setg_errno(errp, -ret, "Could not update qcow2 header");
 748             goto fail;
 749         }
 750     }
 751
 752     /* Initialise locks */
 753     qemu_co_mutex_init(&s->lock);
 754
 755     /* Repair image if dirty */
 756     if (!(flags & (BDRV_O_CHECK | BDRV_O_INCOMING)) && !bs->read_only &&
 757         (s->incompatible_features & QCOW2_INCOMPAT_DIRTY)) {
 758         BdrvCheckResult result = {0};
 759
 760         ret = qcow2_check(bs, &result, BDRV_FIX_ERRORS);
 761         if (ret < 0) {
 762             error_setg_errno(errp, -ret, "Could not repair dirty image");
 763             goto fail;
 764         }
 765     }
 766
 767     /* Enable lazy_refcounts according to image and command line options */
 768     opts = qemu_opts_create(&qcow2_runtime_opts, NULL, 0, &error_abort);
 769     qemu_opts_absorb_qdict(opts, options, &local_err);
 770     if (local_err) {
 771         error_propagate(errp, local_err);
 772         ret = -EINVAL;
 773         goto fail;
 774     }
 775
 776     s->use_lazy_refcounts = qemu_opt_get_bool(opts, QCOW2_OPT_LAZY_REFCOUNTS,
 777         (s->compatible_features & QCOW2_COMPAT_LAZY_REFCOUNTS));
 778
 779     s->discard_passthrough[QCOW2_DISCARD_NEVER] = false;
 780     s->discard_passthrough[QCOW2_DISCARD_ALWAYS] = true;
 781     s->discard_passthrough[QCOW2_DISCARD_REQUEST] =
 782         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_REQUEST,
 783                           flags & BDRV_O_UNMAP);
 784     s->discard_passthrough[QCOW2_DISCARD_SNAPSHOT] =
 785         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_SNAPSHOT, true);
 786     s->discard_passthrough[QCOW2_DISCARD_OTHER] =
 787         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_OTHER, false);
 788
 789     opt_overlap_check = qemu_opt_get(opts, "overlap-check") ?: "cached";
 790     if (!strcmp(opt_overlap_check, "none")) {
 791         overlap_check_template = 0;
 792     } else if (!strcmp(opt_overlap_check, "constant")) {
 793         overlap_check_template = QCOW2_OL_CONSTANT;
 794     } else if (!strcmp(opt_overlap_check, "cached")) {
 795         overlap_check_template = QCOW2_OL_CACHED;
 796     } else if (!strcmp(opt_overlap_check, "all")) {
 797         overlap_check_template = QCOW2_OL_ALL;
 798     } else {
 799         error_setg(errp, "Unsupported value '%s' for qcow2 option "
 800                    "'overlap-check'. Allowed are either of the following: "
 801                    "none, constant, cached, all", opt_overlap_check);
 802         qemu_opts_del(opts);
 803         ret = -EINVAL;
 804         goto fail;
 805     }
 806
 807     s->overlap_check = 0;
 808     for (i = 0; i < QCOW2_OL_MAX_BITNR; i++) {
 809         /* overlap-check defines a template bitmask, but every flag may be
 810          * overwritten through the associated boolean option */
 811         s->overlap_check |=
 812             qemu_opt_get_bool(opts, overlap_bool_option_names[i],
 813                               overlap_check_template & (1 << i)) << i;
 814     }
 815
 816     qemu_opts_del(opts);
 817
 818     if (s->use_lazy_refcounts && s->qcow_version < 3) {
 819         error_setg(errp, "Lazy refcounts require a qcow2 image with at least "
 820                    "qemu 1.1 compatibility level");
 821         ret = -EINVAL;
 822         goto fail;
 823     }
 824
 825 #ifdef DEBUG_ALLOC
 826     {
 827         BdrvCheckResult result = {0};
 828         qcow2_check_refcounts(bs, &result, 0);
 829     }
 830 #endif
 831     return ret;
 832
 833  fail:
 834     g_free(s->unknown_header_fields);
 835     cleanup_unknown_header_ext(bs);
 836     qcow2_free_snapshots(bs);
 837     qcow2_refcount_close(bs);
 838     g_free(s->l1_table);
 839     /* else pre-write overlap checks in cache_destroy may crash */
 840     s->l1_table = NULL;
 841     if (s->l2_table_cache) {
 842         qcow2_cache_destroy(bs, s->l2_table_cache);
 843     }
 844     if (s->refcount_block_cache) {
 845         qcow2_cache_destroy(bs, s->refcount_block_cache);
 846     }
 847     g_free(s->cluster_cache);
 848     qemu_vfree(s->cluster_data);
 849     return ret;
 850 }
 851
 852 static int qcow2_refresh_limits(BlockDriverState *bs)
 853 {
 854     BDRVQcowState *s = bs->opaque;
 855
 856     bs->bl.write_zeroes_alignment = s->cluster_sectors;
 857
 858     return 0;
 859 }
 860
 861 static int qcow2_set_key(BlockDriverState *bs, const char *key)
 862 {
 863     BDRVQcowState *s = bs->opaque;
 864     uint8_t keybuf[16];
 865     int len, i;
 866
 867     memset(keybuf, 0, 16);
 868     len = strlen(key);
 869     if (len > 16)
 870         len = 16;
 871     /* XXX: we could compress the chars to 7 bits to increase
 872        entropy */
 873     for(i = 0;i < len;i++) {
 874         keybuf[i] = key[i];
 875     }
 876     s->crypt_method = s->crypt_method_header;
 877
 878     if (AES_set_encrypt_key(keybuf, 128, &s->aes_encrypt_key) != 0)
 879         return -1;
 880     if (AES_set_decrypt_key(keybuf, 128, &s->aes_decrypt_key) != 0)
 881         return -1;
 882 #if 0
 883     /* test */
 884     {
 885         uint8_t in[16];
 886         uint8_t out[16];
 887         uint8_t tmp[16];
 888         for(i=0;i<16;i++)
 889             in[i] = i;
 890         AES_encrypt(in, tmp, &s->aes_encrypt_key);
 891         AES_decrypt(tmp, out, &s->aes_decrypt_key);
 892         for(i = 0; i < 16; i++)
 893             printf(" %02x", tmp[i]);
 894         printf("\n");
 895         for(i = 0; i < 16; i++)
 896             printf(" %02x", out[i]);
 897         printf("\n");
 898     }
 899 #endif
 900     return 0;
 901 }
 902
 903 /* We have nothing to do for QCOW2 reopen, stubs just return
 904  * success */
 905 static int qcow2_reopen_prepare(BDRVReopenState *state,
 906                                 BlockReopenQueue *queue, Error **errp)
 907 {
 908     return 0;
 909 }
 910
 911 static int64_t coroutine_fn qcow2_co_get_block_status(BlockDriverState *bs,
 912         int64_t sector_num, int nb_sectors, int *pnum)
 913 {
 914     BDRVQcowState *s = bs->opaque;
 915     uint64_t cluster_offset;
 916     int index_in_cluster, ret;
 917     int64_t status = 0;
 918
 919     *pnum = nb_sectors;
 920     qemu_co_mutex_lock(&s->lock);
 921     ret = qcow2_get_cluster_offset(bs, sector_num << 9, pnum, &cluster_offset);
 922     qemu_co_mutex_unlock(&s->lock);
 923     if (ret < 0) {
 924         return ret;
 925     }
 926
 927     if (cluster_offset != 0 && ret != QCOW2_CLUSTER_COMPRESSED &&
 928         !s->crypt_method) {
 929         index_in_cluster = sector_num & (s->cluster_sectors - 1);
 930         cluster_offset |= (index_in_cluster << BDRV_SECTOR_BITS);
 931         status |= BDRV_BLOCK_OFFSET_VALID | cluster_offset;
 932     }
 933     if (ret == QCOW2_CLUSTER_ZERO) {
 934         status |= BDRV_BLOCK_ZERO;
 935     } else if (ret != QCOW2_CLUSTER_UNALLOCATED) {
 936         status |= BDRV_BLOCK_DATA;
 937     }
 938     return status;
 939 }
 940
 941 /* handle reading after the end of the backing file */
 942 int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
 943                   int64_t sector_num, int nb_sectors)
 944 {
 945     int n1;
 946     if ((sector_num + nb_sectors) <= bs->total_sectors)
 947         return nb_sectors;
 948     if (sector_num >= bs->total_sectors)
 949         n1 = 0;
 950     else
 951         n1 = bs->total_sectors - sector_num;
 952
 953     qemu_iovec_memset(qiov, 512 * n1, 0, 512 * (nb_sectors - n1));
 954
 955     return n1;
 956 }
 957
 958 static coroutine_fn int qcow2_co_readv(BlockDriverState *bs, int64_t sector_num,
 959                           int remaining_sectors, QEMUIOVector *qiov)
 960 {
 961     BDRVQcowState *s = bs->opaque;
 962     int index_in_cluster, n1;
 963     int ret;
 964     int cur_nr_sectors; /* number of sectors in current iteration */
 965     uint64_t cluster_offset = 0;
 966     uint64_t bytes_done = 0;
 967     QEMUIOVector hd_qiov;
 968     uint8_t *cluster_data = NULL;
 969
 970     qemu_iovec_init(&hd_qiov, qiov->niov);
 971
 972     qemu_co_mutex_lock(&s->lock);
 973
 974     while (remaining_sectors != 0) {
 975
 976         /* prepare next request */
 977         cur_nr_sectors = remaining_sectors;
 978         if (s->crypt_method) {
 979             cur_nr_sectors = MIN(cur_nr_sectors,
 980                 QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
 981         }
 982
 983         ret = qcow2_get_cluster_offset(bs, sector_num << 9,
 984             &cur_nr_sectors, &cluster_offset);
 985         if (ret < 0) {
 986             goto fail;
 987         }
 988
 989         index_in_cluster = sector_num & (s->cluster_sectors - 1);
 990
 991         qemu_iovec_reset(&hd_qiov);
 992         qemu_iovec_concat(&hd_qiov, qiov, bytes_done,
 993             cur_nr_sectors * 512);
 994
 995         switch (ret) {
 996         case QCOW2_CLUSTER_UNALLOCATED:
 997
 998             if (bs->backing_hd) {
 999                 /* read from the base image */
1000                 n1 = qcow2_backing_read1(bs->backing_hd, &hd_qiov,
1001                     sector_num, cur_nr_sectors);
1002                 if (n1 > 0) {
1003                     BLKDBG_EVENT(bs->file, BLKDBG_READ_BACKING_AIO);
1004                     qemu_co_mutex_unlock(&s->lock);
1005                     ret = bdrv_co_readv(bs->backing_hd, sector_num,
1006                                         n1, &hd_qiov);
1007                     qemu_co_mutex_lock(&s->lock);
1008                     if (ret < 0) {
1009                         goto fail;
1010                     }
1011                 }
1012             } else {
1013                 /* Note: in this case, no need to wait */
1014                 qemu_iovec_memset(&hd_qiov, 0, 0, 512 * cur_nr_sectors);
1015             }
1016             break;
1017
1018         case QCOW2_CLUSTER_ZERO:
1019             qemu_iovec_memset(&hd_qiov, 0, 0, 512 * cur_nr_sectors);
1020             break;
1021
1022         case QCOW2_CLUSTER_COMPRESSED:
1023             /* add AIO support for compressed blocks ? */
1024             ret = qcow2_decompress_cluster(bs, cluster_offset);
1025             if (ret < 0) {
1026                 goto fail;
1027             }
1028
1029             qemu_iovec_from_buf(&hd_qiov, 0,
1030                 s->cluster_cache + index_in_cluster * 512,
1031                 512 * cur_nr_sectors);
1032             break;
1033
1034         case QCOW2_CLUSTER_NORMAL:
1035             if ((cluster_offset & 511) != 0) {
1036                 ret = -EIO;
1037                 goto fail;
1038             }
1039
1040             if (s->crypt_method) {
1041                 /*
1042                  * For encrypted images, read everything into a temporary
1043                  * contiguous buffer on which the AES functions can work.
1044                  */
1045                 if (!cluster_data) {
1046                     cluster_data =
1047                         qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
1048                 }
1049
1050                 assert(cur_nr_sectors <=
1051                     QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
1052                 qemu_iovec_reset(&hd_qiov);
1053                 qemu_iovec_add(&hd_qiov, cluster_data,
1054                     512 * cur_nr_sectors);
1055             }
1056
1057             BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1058             qemu_co_mutex_unlock(&s->lock);
1059             ret = bdrv_co_readv(bs->file,
1060                                 (cluster_offset >> 9) + index_in_cluster,
1061                                 cur_nr_sectors, &hd_qiov);
1062             qemu_co_mutex_lock(&s->lock);
1063             if (ret < 0) {
1064                 goto fail;
1065             }
1066             if (s->crypt_method) {
1067                 qcow2_encrypt_sectors(s, sector_num,  cluster_data,
1068                     cluster_data, cur_nr_sectors, 0, &s->aes_decrypt_key);
1069                 qemu_iovec_from_buf(qiov, bytes_done,
1070                     cluster_data, 512 * cur_nr_sectors);
1071             }
1072             break;
1073
1074         default:
1075             g_assert_not_reached();
1076             ret = -EIO;
1077             goto fail;
1078         }
1079
1080         remaining_sectors -= cur_nr_sectors;
1081         sector_num += cur_nr_sectors;
1082         bytes_done += cur_nr_sectors * 512;
1083     }
1084     ret = 0;
1085
1086 fail:
1087     qemu_co_mutex_unlock(&s->lock);
1088
1089     qemu_iovec_destroy(&hd_qiov);
1090     qemu_vfree(cluster_data);
1091
1092     return ret;
1093 }
1094
1095 static coroutine_fn int qcow2_co_writev(BlockDriverState *bs,
1096                            int64_t sector_num,
1097                            int remaining_sectors,
1098                            QEMUIOVector *qiov)
1099 {
1100     BDRVQcowState *s = bs->opaque;
1101     int index_in_cluster;
1102     int ret;
1103     int cur_nr_sectors; /* number of sectors in current iteration */
1104     uint64_t cluster_offset;
1105     QEMUIOVector hd_qiov;
1106     uint64_t bytes_done = 0;
1107     uint8_t *cluster_data = NULL;
1108     QCowL2Meta *l2meta = NULL;
1109
1110     trace_qcow2_writev_start_req(qemu_coroutine_self(), sector_num,
1111                                  remaining_sectors);
1112
1113     qemu_iovec_init(&hd_qiov, qiov->niov);
1114
1115     s->cluster_cache_offset = -1; /* disable compressed cache */
1116
1117     qemu_co_mutex_lock(&s->lock);
1118
1119     while (remaining_sectors != 0) {
1120
1121         l2meta = NULL;
1122
1123         trace_qcow2_writev_start_part(qemu_coroutine_self());
1124         index_in_cluster = sector_num & (s->cluster_sectors - 1);
1125         cur_nr_sectors = remaining_sectors;
1126         if (s->crypt_method &&
1127             cur_nr_sectors >
1128             QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors - index_in_cluster) {
1129             cur_nr_sectors =
1130                 QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors - index_in_cluster;
1131         }
1132
1133         ret = qcow2_alloc_cluster_offset(bs, sector_num << 9,
1134             &cur_nr_sectors, &cluster_offset, &l2meta);
1135         if (ret < 0) {
1136             goto fail;
1137         }
1138
1139         assert((cluster_offset & 511) == 0);
1140
1141         qemu_iovec_reset(&hd_qiov);
1142         qemu_iovec_concat(&hd_qiov, qiov, bytes_done,
1143             cur_nr_sectors * 512);
1144
1145         if (s->crypt_method) {
1146             if (!cluster_data) {
1147                 cluster_data = qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS *
1148                                                  s->cluster_size);
1149             }
1150
1151             assert(hd_qiov.size <=
1152                    QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
1153             qemu_iovec_to_buf(&hd_qiov, 0, cluster_data, hd_qiov.size);
1154
1155             qcow2_encrypt_sectors(s, sector_num, cluster_data,
1156                 cluster_data, cur_nr_sectors, 1, &s->aes_encrypt_key);
1157
1158             qemu_iovec_reset(&hd_qiov);
1159             qemu_iovec_add(&hd_qiov, cluster_data,
1160                 cur_nr_sectors * 512);
1161         }
1162
1163         ret = qcow2_pre_write_overlap_check(bs, 0,
1164                 cluster_offset + index_in_cluster * BDRV_SECTOR_SIZE,
1165                 cur_nr_sectors * BDRV_SECTOR_SIZE);
1166         if (ret < 0) {
1167             goto fail;
1168         }
1169
1170         qemu_co_mutex_unlock(&s->lock);
1171         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
1172         trace_qcow2_writev_data(qemu_coroutine_self(),
1173                                 (cluster_offset >> 9) + index_in_cluster);
1174         ret = bdrv_co_writev(bs->file,
1175                              (cluster_offset >> 9) + index_in_cluster,
1176                              cur_nr_sectors, &hd_qiov);
1177         qemu_co_mutex_lock(&s->lock);
1178         if (ret < 0) {
1179             goto fail;
1180         }
1181
1182         while (l2meta != NULL) {
1183             QCowL2Meta *next;
1184
1185             ret = qcow2_alloc_cluster_link_l2(bs, l2meta);
1186             if (ret < 0) {
1187                 goto fail;
1188             }
1189
1190             /* Take the request off the list of running requests */
1191             if (l2meta->nb_clusters != 0) {
1192                 QLIST_REMOVE(l2meta, next_in_flight);
1193             }
1194
1195             qemu_co_queue_restart_all(&l2meta->dependent_requests);
1196
1197             next = l2meta->next;
1198             g_free(l2meta);
1199             l2meta = next;
1200         }
1201
1202         remaining_sectors -= cur_nr_sectors;
1203         sector_num += cur_nr_sectors;
1204         bytes_done += cur_nr_sectors * 512;
1205         trace_qcow2_writev_done_part(qemu_coroutine_self(), cur_nr_sectors);
1206     }
1207     ret = 0;
1208
1209 fail:
1210     qemu_co_mutex_unlock(&s->lock);
1211
1212     while (l2meta != NULL) {
1213         QCowL2Meta *next;
1214
1215         if (l2meta->nb_clusters != 0) {
1216             QLIST_REMOVE(l2meta, next_in_flight);
1217         }
1218         qemu_co_queue_restart_all(&l2meta->dependent_requests);
1219
1220         next = l2meta->next;
1221         g_free(l2meta);
1222         l2meta = next;
1223     }
1224
1225     qemu_iovec_destroy(&hd_qiov);
1226     qemu_vfree(cluster_data);
1227     trace_qcow2_writev_done_req(qemu_coroutine_self(), ret);
1228
1229     return ret;
1230 }
1231
1232 static void qcow2_close(BlockDriverState *bs)
1233 {
1234     BDRVQcowState *s = bs->opaque;
1235     g_free(s->l1_table);
1236     /* else pre-write overlap checks in cache_destroy may crash */
1237     s->l1_table = NULL;
1238
1239     if (!(bs->open_flags & BDRV_O_INCOMING)) {
1240         qcow2_cache_flush(bs, s->l2_table_cache);
1241         qcow2_cache_flush(bs, s->refcount_block_cache);
1242
1243         qcow2_mark_clean(bs);
1244     }
1245
1246     qcow2_cache_destroy(bs, s->l2_table_cache);
1247     qcow2_cache_destroy(bs, s->refcount_block_cache);
1248
1249     g_free(s->unknown_header_fields);
1250     cleanup_unknown_header_ext(bs);
1251
1252     g_free(s->cluster_cache);
1253     qemu_vfree(s->cluster_data);
1254     qcow2_refcount_close(bs);
1255     qcow2_free_snapshots(bs);
1256 }
1257
1258 static void qcow2_invalidate_cache(BlockDriverState *bs, Error **errp)
1259 {
1260     BDRVQcowState *s = bs->opaque;
1261     int flags = s->flags;
1262     AES_KEY aes_encrypt_key;
1263     AES_KEY aes_decrypt_key;
1264     uint32_t crypt_method = 0;
1265     QDict *options;
1266     Error *local_err = NULL;
1267     int ret;
1268
1269     /*
1270      * Backing files are read-only which makes all of their metadata immutable,
1271      * that means we don't have to worry about reopening them here.
1272      */
1273
1274     if (s->crypt_method) {
1275         crypt_method = s->crypt_method;
1276         memcpy(&aes_encrypt_key, &s->aes_encrypt_key, sizeof(aes_encrypt_key));
1277         memcpy(&aes_decrypt_key, &s->aes_decrypt_key, sizeof(aes_decrypt_key));
1278     }
1279
1280     qcow2_close(bs);
1281
1282     bdrv_invalidate_cache(bs->file, &local_err);
1283     if (local_err) {
1284         error_propagate(errp, local_err);
1285         return;
1286     }
1287
1288     memset(s, 0, sizeof(BDRVQcowState));
1289     options = qdict_clone_shallow(bs->options);
1290
1291     ret = qcow2_open(bs, options, flags, &local_err);
1292     if (local_err) {
1293         error_setg(errp, "Could not reopen qcow2 layer: %s",
1294                    error_get_pretty(local_err));
1295         error_free(local_err);
1296         return;
1297     } else if (ret < 0) {
1298         error_setg_errno(errp, -ret, "Could not reopen qcow2 layer");
1299         return;
1300     }
1301
1302     QDECREF(options);
1303
1304     if (crypt_method) {
1305         s->crypt_method = crypt_method;
1306         memcpy(&s->aes_encrypt_key, &aes_encrypt_key, sizeof(aes_encrypt_key));
1307         memcpy(&s->aes_decrypt_key, &aes_decrypt_key, sizeof(aes_decrypt_key));
1308     }
1309 }
1310
1311 static size_t header_ext_add(char *buf, uint32_t magic, const void *s,
1312     size_t len, size_t buflen)
1313 {
1314     QCowExtension *ext_backing_fmt = (QCowExtension*) buf;
1315     size_t ext_len = sizeof(QCowExtension) + ((len + 7) & ~7);
1316
1317     if (buflen < ext_len) {
1318         return -ENOSPC;
1319     }
1320
1321     *ext_backing_fmt = (QCowExtension) {
1322         .magic  = cpu_to_be32(magic),
1323         .len    = cpu_to_be32(len),
1324     };
1325     memcpy(buf + sizeof(QCowExtension), s, len);
1326
1327     return ext_len;
1328 }
1329
1330 /*
1331  * Updates the qcow2 header, including the variable length parts of it, i.e.
1332  * the backing file name and all extensions. qcow2 was not designed to allow
1333  * such changes, so if we run out of space (we can only use the first cluster)
1334  * this function may fail.
1335  *
1336  * Returns 0 on success, -errno in error cases.
1337  */
1338 int qcow2_update_header(BlockDriverState *bs)
1339 {
1340     BDRVQcowState *s = bs->opaque;
1341     QCowHeader *header;
1342     char *buf;
1343     size_t buflen = s->cluster_size;
1344     int ret;
1345     uint64_t total_size;
1346     uint32_t refcount_table_clusters;
1347     size_t header_length;
1348     Qcow2UnknownHeaderExtension *uext;
1349
1350     buf = qemu_blockalign(bs, buflen);
1351
1352     /* Header structure */
1353     header = (QCowHeader*) buf;
1354
1355     if (buflen < sizeof(*header)) {
1356         ret = -ENOSPC;
1357         goto fail;
1358     }
1359
1360     header_length = sizeof(*header) + s->unknown_header_fields_size;
1361     total_size = bs->total_sectors * BDRV_SECTOR_SIZE;
1362     refcount_table_clusters = s->refcount_table_size >> (s->cluster_bits - 3);
1363
1364     *header = (QCowHeader) {
1365         /* Version 2 fields */
1366         .magic                  = cpu_to_be32(QCOW_MAGIC),
1367         .version                = cpu_to_be32(s->qcow_version),
1368         .backing_file_offset    = 0,
1369         .backing_file_size      = 0,
1370         .cluster_bits           = cpu_to_be32(s->cluster_bits),
1371         .size                   = cpu_to_be64(total_size),
1372         .crypt_method           = cpu_to_be32(s->crypt_method_header),
1373         .l1_size                = cpu_to_be32(s->l1_size),
1374         .l1_table_offset        = cpu_to_be64(s->l1_table_offset),
1375         .refcount_table_offset  = cpu_to_be64(s->refcount_table_offset),
1376         .refcount_table_clusters = cpu_to_be32(refcount_table_clusters),
1377         .nb_snapshots           = cpu_to_be32(s->nb_snapshots),
1378         .snapshots_offset       = cpu_to_be64(s->snapshots_offset),
1379
1380         /* Version 3 fields */
1381         .incompatible_features  = cpu_to_be64(s->incompatible_features),
1382         .compatible_features    = cpu_to_be64(s->compatible_features),
1383         .autoclear_features     = cpu_to_be64(s->autoclear_features),
1384         .refcount_order         = cpu_to_be32(s->refcount_order),
1385         .header_length          = cpu_to_be32(header_length),
1386     };
1387
1388     /* For older versions, write a shorter header */
1389     switch (s->qcow_version) {
1390     case 2:
1391         ret = offsetof(QCowHeader, incompatible_features);
1392         break;
1393     case 3:
1394         ret = sizeof(*header);
1395         break;
1396     default:
1397         ret = -EINVAL;
1398         goto fail;
1399     }
1400
1401     buf += ret;
1402     buflen -= ret;
1403     memset(buf, 0, buflen);
1404
1405     /* Preserve any unknown field in the header */
1406     if (s->unknown_header_fields_size) {
1407         if (buflen < s->unknown_header_fields_size) {
1408             ret = -ENOSPC;
1409             goto fail;
1410         }
1411
1412         memcpy(buf, s->unknown_header_fields, s->unknown_header_fields_size);
1413         buf += s->unknown_header_fields_size;
1414         buflen -= s->unknown_header_fields_size;
1415     }
1416
1417     /* Backing file format header extension */
1418     if (*bs->backing_format) {
1419         ret = header_ext_add(buf, QCOW2_EXT_MAGIC_BACKING_FORMAT,
1420                              bs->backing_format, strlen(bs->backing_format),
1421                              buflen);
1422         if (ret < 0) {
1423             goto fail;
1424         }
1425
1426         buf += ret;
1427         buflen -= ret;
1428     }
1429
1430     /* Feature table */
1431     Qcow2Feature features[] = {
1432         {
1433             .type = QCOW2_FEAT_TYPE_INCOMPATIBLE,
1434             .bit  = QCOW2_INCOMPAT_DIRTY_BITNR,
1435             .name = "dirty bit",
1436         },
1437         {
1438             .type = QCOW2_FEAT_TYPE_INCOMPATIBLE,
1439             .bit  = QCOW2_INCOMPAT_CORRUPT_BITNR,
1440             .name = "corrupt bit",
1441         },
1442         {
1443             .type = QCOW2_FEAT_TYPE_COMPATIBLE,
1444             .bit  = QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR,
1445             .name = "lazy refcounts",
1446         },
1447     };
1448
1449     ret = header_ext_add(buf, QCOW2_EXT_MAGIC_FEATURE_TABLE,
1450                          features, sizeof(features), buflen);
1451     if (ret < 0) {
1452         goto fail;
1453     }
1454     buf += ret;
1455     buflen -= ret;
1456
1457     /* Keep unknown header extensions */
1458     QLIST_FOREACH(uext, &s->unknown_header_ext, next) {
1459         ret = header_ext_add(buf, uext->magic, uext->data, uext->len, buflen);
1460         if (ret < 0) {
1461             goto fail;
1462         }
1463
1464         buf += ret;
1465         buflen -= ret;
1466     }
1467
1468     /* End of header extensions */
1469     ret = header_ext_add(buf, QCOW2_EXT_MAGIC_END, NULL, 0, buflen);
1470     if (ret < 0) {
1471         goto fail;
1472     }
1473
1474     buf += ret;
1475     buflen -= ret;
1476
1477     /* Backing file name */
1478     if (*bs->backing_file) {
1479         size_t backing_file_len = strlen(bs->backing_file);
1480
1481         if (buflen < backing_file_len) {
1482             ret = -ENOSPC;
1483             goto fail;
1484         }
1485
1486         /* Using strncpy is ok here, since buf is not NUL-terminated. */
1487         strncpy(buf, bs->backing_file, buflen);
1488
1489         header->backing_file_offset = cpu_to_be64(buf - ((char*) header));
1490         header->backing_file_size   = cpu_to_be32(backing_file_len);
1491     }
1492
1493     /* Write the new header */
1494     ret = bdrv_pwrite(bs->file, 0, header, s->cluster_size);
1495     if (ret < 0) {
1496         goto fail;
1497     }
1498
1499     ret = 0;
1500 fail:
1501     qemu_vfree(header);
1502     return ret;
1503 }
1504
1505 static int qcow2_change_backing_file(BlockDriverState *bs,
1506     const char *backing_file, const char *backing_fmt)
1507 {
1508     pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: "");
1509     pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: "");
1510
1511     return qcow2_update_header(bs);
1512 }
1513
1514 static int preallocate(BlockDriverState *bs)
1515 {
1516     uint64_t nb_sectors;
1517     uint64_t offset;
1518     uint64_t host_offset = 0;
1519     int num;
1520     int ret;
1521     QCowL2Meta *meta;
1522
1523     nb_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
1524     offset = 0;
1525
1526     while (nb_sectors) {
1527         num = MIN(nb_sectors, INT_MAX >> BDRV_SECTOR_BITS);
1528         ret = qcow2_alloc_cluster_offset(bs, offset, &num,
1529                                          &host_offset, &meta);
1530         if (ret < 0) {
1531             return ret;
1532         }
1533
1534         while (meta) {
1535             QCowL2Meta *next = meta->next;
1536
1537             ret = qcow2_alloc_cluster_link_l2(bs, meta);
1538             if (ret < 0) {
1539                 qcow2_free_any_clusters(bs, meta->alloc_offset,
1540                                         meta->nb_clusters, QCOW2_DISCARD_NEVER);
1541                 return ret;
1542             }
1543
1544             /* There are no dependent requests, but we need to remove our
1545              * request from the list of in-flight requests */
1546             QLIST_REMOVE(meta, next_in_flight);
1547
1548             g_free(meta);
1549             meta = next;
1550         }
1551
1552         /* TODO Preallocate data if requested */
1553
1554         nb_sectors -= num;
1555         offset += num << BDRV_SECTOR_BITS;
1556     }
1557
1558     /*
1559      * It is expected that the image file is large enough to actually contain
1560      * all of the allocated clusters (otherwise we get failing reads after
1561      * EOF). Extend the image to the last allocated sector.
1562      */
1563     if (host_offset != 0) {
1564         uint8_t buf[BDRV_SECTOR_SIZE];
1565         memset(buf, 0, BDRV_SECTOR_SIZE);
1566         ret = bdrv_write(bs->file, (host_offset >> BDRV_SECTOR_BITS) + num - 1,
1567                          buf, 1);
1568         if (ret < 0) {
1569             return ret;
1570         }
1571     }
1572
1573     return 0;
1574 }
1575
1576 static int qcow2_create2(const char *filename, int64_t total_size,
1577                          const char *backing_file, const char *backing_format,
1578                          int flags, size_t cluster_size, int prealloc,
1579                          QEMUOptionParameter *options, int version,
1580                          Error **errp)
1581 {
1582     /* Calculate cluster_bits */
1583     int cluster_bits;
1584     cluster_bits = ffs(cluster_size) - 1;
1585     if (cluster_bits < MIN_CLUSTER_BITS || cluster_bits > MAX_CLUSTER_BITS ||
1586         (1 << cluster_bits) != cluster_size)
1587     {
1588         error_setg(errp, "Cluster size must be a power of two between %d and "
1589                    "%dk", 1 << MIN_CLUSTER_BITS, 1 << (MAX_CLUSTER_BITS - 10));
1590         return -EINVAL;
1591     }
1592
1593     /*
1594      * Open the image file and write a minimal qcow2 header.
1595      *
1596      * We keep things simple and start with a zero-sized image. We also
1597      * do without refcount blocks or a L1 table for now. We'll fix the
1598      * inconsistency later.
1599      *
1600      * We do need a refcount table because growing the refcount table means
1601      * allocating two new refcount blocks - the seconds of which would be at
1602      * 2 GB for 64k clusters, and we don't want to have a 2 GB initial file
1603      * size for any qcow2 image.
1604      */
1605     BlockDriverState* bs;
1606     QCowHeader *header;
1607     uint64_t* refcount_table;
1608     Error *local_err = NULL;
1609     int ret;
1610
1611     ret = bdrv_create_file(filename, options, &local_err);
1612     if (ret < 0) {
1613         error_propagate(errp, local_err);
1614         return ret;
1615     }
1616
1617     bs = NULL;
1618     ret = bdrv_open(&bs, filename, NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1619                     NULL, &local_err);
1620     if (ret < 0) {
1621         error_propagate(errp, local_err);
1622         return ret;
1623     }
1624
1625     /* Write the header */
1626     QEMU_BUILD_BUG_ON((1 << MIN_CLUSTER_BITS) < sizeof(*header));
1627     header = g_malloc0(cluster_size);
1628     *header = (QCowHeader) {
1629         .magic                      = cpu_to_be32(QCOW_MAGIC),
1630         .version                    = cpu_to_be32(version),
1631         .cluster_bits               = cpu_to_be32(cluster_bits),
1632         .size                       = cpu_to_be64(0),
1633         .l1_table_offset            = cpu_to_be64(0),
1634         .l1_size                    = cpu_to_be32(0),
1635         .refcount_table_offset      = cpu_to_be64(cluster_size),
1636         .refcount_table_clusters    = cpu_to_be32(1),
1637         .refcount_order             = cpu_to_be32(3 + REFCOUNT_SHIFT),
1638         .header_length              = cpu_to_be32(sizeof(*header)),
1639     };
1640
1641     if (flags & BLOCK_FLAG_ENCRYPT) {
1642         header->crypt_method = cpu_to_be32(QCOW_CRYPT_AES);
1643     } else {
1644         header->crypt_method = cpu_to_be32(QCOW_CRYPT_NONE);
1645     }
1646
1647     if (flags & BLOCK_FLAG_LAZY_REFCOUNTS) {
1648         header->compatible_features |=
1649             cpu_to_be64(QCOW2_COMPAT_LAZY_REFCOUNTS);
1650     }
1651
1652     ret = bdrv_pwrite(bs, 0, header, cluster_size);
1653     g_free(header);
1654     if (ret < 0) {
1655         error_setg_errno(errp, -ret, "Could not write qcow2 header");
1656         goto out;
1657     }
1658
1659     /* Write a refcount table with one refcount block */
1660     refcount_table = g_malloc0(2 * cluster_size);
1661     refcount_table[0] = cpu_to_be64(2 * cluster_size);
1662     ret = bdrv_pwrite(bs, cluster_size, refcount_table, 2 * cluster_size);
1663     g_free(refcount_table);
1664
1665     if (ret < 0) {
1666         error_setg_errno(errp, -ret, "Could not write refcount table");
1667         goto out;
1668     }
1669
1670     bdrv_unref(bs);
1671     bs = NULL;
1672
1673     /*
1674      * And now open the image and make it consistent first (i.e. increase the
1675      * refcount of the cluster that is occupied by the header and the refcount
1676      * table)
1677      */
1678     BlockDriver* drv = bdrv_find_format("qcow2");
1679     assert(drv != NULL);
1680     ret = bdrv_open(&bs, filename, NULL, NULL,
1681         BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_NO_FLUSH, drv, &local_err);
1682     if (ret < 0) {
1683         error_propagate(errp, local_err);
1684         goto out;
1685     }
1686
1687     ret = qcow2_alloc_clusters(bs, 3 * cluster_size);
1688     if (ret < 0) {
1689         error_setg_errno(errp, -ret, "Could not allocate clusters for qcow2 "
1690                          "header and refcount table");
1691         goto out;
1692
1693     } else if (ret != 0) {
1694         error_report("Huh, first cluster in empty image is already in use?");
1695         abort();
1696     }
1697
1698     /* Okay, now that we have a valid image, let's give it the right size */
1699     ret = bdrv_truncate(bs, total_size * BDRV_SECTOR_SIZE);
1700     if (ret < 0) {
1701         error_setg_errno(errp, -ret, "Could not resize image");
1702         goto out;
1703     }
1704
1705     /* Want a backing file? There you go.*/
1706     if (backing_file) {
1707         ret = bdrv_change_backing_file(bs, backing_file, backing_format);
1708         if (ret < 0) {
1709             error_setg_errno(errp, -ret, "Could not assign backing file '%s' "
1710                              "with format '%s'", backing_file, backing_format);
1711             goto out;
1712         }
1713     }
1714
1715     /* And if we're supposed to preallocate metadata, do that now */
1716     if (prealloc) {
1717         BDRVQcowState *s = bs->opaque;
1718         qemu_co_mutex_lock(&s->lock);
1719         ret = preallocate(bs);
1720         qemu_co_mutex_unlock(&s->lock);
1721         if (ret < 0) {
1722             error_setg_errno(errp, -ret, "Could not preallocate metadata");
1723             goto out;
1724         }
1725     }
1726
1727     bdrv_unref(bs);
1728     bs = NULL;
1729
1730     /* Reopen the image without BDRV_O_NO_FLUSH to flush it before returning */
1731     ret = bdrv_open(&bs, filename, NULL, NULL,
1732                     BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_NO_BACKING,
1733                     drv, &local_err);
1734     if (local_err) {
1735         error_propagate(errp, local_err);
1736         goto out;
1737     }
1738
1739     ret = 0;
1740 out:
1741     if (bs) {
1742         bdrv_unref(bs);
1743     }
1744     return ret;
1745 }
1746
1747 static int qcow2_create(const char *filename, QEMUOptionParameter *options,
1748                         Error **errp)
1749 {
1750     const char *backing_file = NULL;
1751     const char *backing_fmt = NULL;
1752     uint64_t sectors = 0;
1753     int flags = 0;
1754     size_t cluster_size = DEFAULT_CLUSTER_SIZE;
1755     int prealloc = 0;
1756     int version = 3;
1757     Error *local_err = NULL;
1758     int ret;
1759
1760     /* Read out options */
1761     while (options && options->name) {
1762         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
1763             sectors = options->value.n / 512;
1764         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
1765             backing_file = options->value.s;
1766         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) {
1767             backing_fmt = options->value.s;
1768         } else if (!strcmp(options->name, BLOCK_OPT_ENCRYPT)) {
1769             flags |= options->value.n ? BLOCK_FLAG_ENCRYPT : 0;
1770         } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) {
1771             if (options->value.n) {
1772                 cluster_size = options->value.n;
1773             }
1774         } else if (!strcmp(options->name, BLOCK_OPT_PREALLOC)) {
1775             if (!options->value.s || !strcmp(options->value.s, "off")) {
1776                 prealloc = 0;
1777             } else if (!strcmp(options->value.s, "metadata")) {
1778                 prealloc = 1;
1779             } else {
1780                 error_setg(errp, "Invalid preallocation mode: '%s'",
1781                            options->value.s);
1782                 return -EINVAL;
1783             }
1784         } else if (!strcmp(options->name, BLOCK_OPT_COMPAT_LEVEL)) {
1785             if (!options->value.s) {
1786                 /* keep the default */
1787             } else if (!strcmp(options->value.s, "0.10")) {
1788                 version = 2;
1789             } else if (!strcmp(options->value.s, "1.1")) {
1790                 version = 3;
1791             } else {
1792                 error_setg(errp, "Invalid compatibility level: '%s'",
1793                            options->value.s);
1794                 return -EINVAL;
1795             }
1796         } else if (!strcmp(options->name, BLOCK_OPT_LAZY_REFCOUNTS)) {
1797             flags |= options->value.n ? BLOCK_FLAG_LAZY_REFCOUNTS : 0;
1798         }
1799         options++;
1800     }
1801
1802     if (backing_file && prealloc) {
1803         error_setg(errp, "Backing file and preallocation cannot be used at "
1804                    "the same time");
1805         return -EINVAL;
1806     }
1807
1808     if (version < 3 && (flags & BLOCK_FLAG_LAZY_REFCOUNTS)) {
1809         error_setg(errp, "Lazy refcounts only supported with compatibility "
1810                    "level 1.1 and above (use compat=1.1 or greater)");
1811         return -EINVAL;
1812     }
1813
1814     ret = qcow2_create2(filename, sectors, backing_file, backing_fmt, flags,
1815                         cluster_size, prealloc, options, version, &local_err);
1816     if (local_err) {
1817         error_propagate(errp, local_err);
1818     }
1819     return ret;
1820 }
1821
1822 static coroutine_fn int qcow2_co_write_zeroes(BlockDriverState *bs,
1823     int64_t sector_num, int nb_sectors, BdrvRequestFlags flags)
1824 {
1825     int ret;
1826     BDRVQcowState *s = bs->opaque;
1827
1828     /* Emulate misaligned zero writes */
1829     if (sector_num % s->cluster_sectors || nb_sectors % s->cluster_sectors) {
1830         return -ENOTSUP;
1831     }
1832
1833     /* Whatever is left can use real zero clusters */
1834     qemu_co_mutex_lock(&s->lock);
1835     ret = qcow2_zero_clusters(bs, sector_num << BDRV_SECTOR_BITS,
1836         nb_sectors);
1837     qemu_co_mutex_unlock(&s->lock);
1838
1839     return ret;
1840 }
1841
1842 static coroutine_fn int qcow2_co_discard(BlockDriverState *bs,
1843     int64_t sector_num, int nb_sectors)
1844 {
1845     int ret;
1846     BDRVQcowState *s = bs->opaque;
1847
1848     qemu_co_mutex_lock(&s->lock);
1849     ret = qcow2_discard_clusters(bs, sector_num << BDRV_SECTOR_BITS,
1850         nb_sectors, QCOW2_DISCARD_REQUEST);
1851     qemu_co_mutex_unlock(&s->lock);
1852     return ret;
1853 }
1854
1855 static int qcow2_truncate(BlockDriverState *bs, int64_t offset)
1856 {
1857     BDRVQcowState *s = bs->opaque;
1858     int64_t new_l1_size;
1859     int ret;
1860
1861     if (offset & 511) {
1862         error_report("The new size must be a multiple of 512");
1863         return -EINVAL;
1864     }
1865
1866     /* cannot proceed if image has snapshots */
1867     if (s->nb_snapshots) {
1868         error_report("Can't resize an image which has snapshots");
1869         return -ENOTSUP;
1870     }
1871
1872     /* shrinking is currently not supported */
1873     if (offset < bs->total_sectors * 512) {
1874         error_report("qcow2 doesn't support shrinking images yet");
1875         return -ENOTSUP;
1876     }
1877
1878     new_l1_size = size_to_l1(s, offset);
1879     ret = qcow2_grow_l1_table(bs, new_l1_size, true);
1880     if (ret < 0) {
1881         return ret;
1882     }
1883
1884     /* write updated header.size */
1885     offset = cpu_to_be64(offset);
1886     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, size),
1887                            &offset, sizeof(uint64_t));
1888     if (ret < 0) {
1889         return ret;
1890     }
1891
1892     s->l1_vm_state_index = new_l1_size;
1893     return 0;
1894 }
1895
1896 /* XXX: put compressed sectors first, then all the cluster aligned
1897    tables to avoid losing bytes in alignment */
1898 static int qcow2_write_compressed(BlockDriverState *bs, int64_t sector_num,
1899                                   const uint8_t *buf, int nb_sectors)
1900 {
1901     BDRVQcowState *s = bs->opaque;
1902     z_stream strm;
1903     int ret, out_len;
1904     uint8_t *out_buf;
1905     uint64_t cluster_offset;
1906
1907     if (nb_sectors == 0) {
1908         /* align end of file to a sector boundary to ease reading with
1909            sector based I/Os */
1910         cluster_offset = bdrv_getlength(bs->file);
1911         cluster_offset = (cluster_offset + 511) & ~511;
1912         bdrv_truncate(bs->file, cluster_offset);
1913         return 0;
1914     }
1915
1916     if (nb_sectors != s->cluster_sectors) {
1917         ret = -EINVAL;
1918
1919         /* Zero-pad last write if image size is not cluster aligned */
1920         if (sector_num + nb_sectors == bs->total_sectors &&
1921             nb_sectors < s->cluster_sectors) {
1922             uint8_t *pad_buf = qemu_blockalign(bs, s->cluster_size);
1923             memset(pad_buf, 0, s->cluster_size);
1924             memcpy(pad_buf, buf, nb_sectors * BDRV_SECTOR_SIZE);
1925             ret = qcow2_write_compressed(bs, sector_num,
1926                                          pad_buf, s->cluster_sectors);
1927             qemu_vfree(pad_buf);
1928         }
1929         return ret;
1930     }
1931
1932     out_buf = g_malloc(s->cluster_size + (s->cluster_size / 1000) + 128);
1933
1934     /* best compression, small window, no zlib header */
1935     memset(&strm, 0, sizeof(strm));
1936     ret = deflateInit2(&strm, Z_DEFAULT_COMPRESSION,
1937                        Z_DEFLATED, -12,
1938                        9, Z_DEFAULT_STRATEGY);
1939     if (ret != 0) {
1940         ret = -EINVAL;
1941         goto fail;
1942     }
1943
1944     strm.avail_in = s->cluster_size;
1945     strm.next_in = (uint8_t *)buf;
1946     strm.avail_out = s->cluster_size;
1947     strm.next_out = out_buf;
1948
1949     ret = deflate(&strm, Z_FINISH);
1950     if (ret != Z_STREAM_END && ret != Z_OK) {
1951         deflateEnd(&strm);
1952         ret = -EINVAL;
1953         goto fail;
1954     }
1955     out_len = strm.next_out - out_buf;
1956
1957     deflateEnd(&strm);
1958
1959     if (ret != Z_STREAM_END || out_len >= s->cluster_size) {
1960         /* could not compress: write normal cluster */
1961         ret = bdrv_write(bs, sector_num, buf, s->cluster_sectors);
1962         if (ret < 0) {
1963             goto fail;
1964         }
1965     } else {
1966         cluster_offset = qcow2_alloc_compressed_cluster_offset(bs,
1967             sector_num << 9, out_len);
1968         if (!cluster_offset) {
1969             ret = -EIO;
1970             goto fail;
1971         }
1972         cluster_offset &= s->cluster_offset_mask;
1973
1974         ret = qcow2_pre_write_overlap_check(bs, 0, cluster_offset, out_len);
1975         if (ret < 0) {
1976             goto fail;
1977         }
1978
1979         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_COMPRESSED);
1980         ret = bdrv_pwrite(bs->file, cluster_offset, out_buf, out_len);
1981         if (ret < 0) {
1982             goto fail;
1983         }
1984     }
1985
1986     ret = 0;
1987 fail:
1988     g_free(out_buf);
1989     return ret;
1990 }
1991
1992 static coroutine_fn int qcow2_co_flush_to_os(BlockDriverState *bs)
1993 {
1994     BDRVQcowState *s = bs->opaque;
1995     int ret;
1996
1997     qemu_co_mutex_lock(&s->lock);
1998     ret = qcow2_cache_flush(bs, s->l2_table_cache);
1999     if (ret < 0) {
2000         qemu_co_mutex_unlock(&s->lock);
2001         return ret;
2002     }
2003
2004     if (qcow2_need_accurate_refcounts(s)) {
2005         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
2006         if (ret < 0) {
2007             qemu_co_mutex_unlock(&s->lock);
2008             return ret;
2009         }
2010     }
2011     qemu_co_mutex_unlock(&s->lock);
2012
2013     return 0;
2014 }
2015
2016 static int qcow2_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2017 {
2018     BDRVQcowState *s = bs->opaque;
2019     bdi->unallocated_blocks_are_zero = true;
2020     bdi->can_write_zeroes_with_unmap = (s->qcow_version >= 3);
2021     bdi->cluster_size = s->cluster_size;
2022     bdi->vm_state_offset = qcow2_vm_state_offset(s);
2023     return 0;
2024 }
2025
2026 static ImageInfoSpecific *qcow2_get_specific_info(BlockDriverState *bs)
2027 {
2028     BDRVQcowState *s = bs->opaque;
2029     ImageInfoSpecific *spec_info = g_new(ImageInfoSpecific, 1);
2030
2031     *spec_info = (ImageInfoSpecific){
2032         .kind  = IMAGE_INFO_SPECIFIC_KIND_QCOW2,
2033         {
2034             .qcow2 = g_new(ImageInfoSpecificQCow2, 1),
2035         },
2036     };
2037     if (s->qcow_version == 2) {
2038         *spec_info->qcow2 = (ImageInfoSpecificQCow2){
2039             .compat = g_strdup("0.10"),
2040         };
2041     } else if (s->qcow_version == 3) {
2042         *spec_info->qcow2 = (ImageInfoSpecificQCow2){
2043             .compat             = g_strdup("1.1"),
2044             .lazy_refcounts     = s->compatible_features &
2045                                   QCOW2_COMPAT_LAZY_REFCOUNTS,
2046             .has_lazy_refcounts = true,
2047         };
2048     }
2049
2050     return spec_info;
2051 }
2052
2053 #if 0
2054 static void dump_refcounts(BlockDriverState *bs)
2055 {
2056     BDRVQcowState *s = bs->opaque;
2057     int64_t nb_clusters, k, k1, size;
2058     int refcount;
2059
2060     size = bdrv_getlength(bs->file);
2061     nb_clusters = size_to_clusters(s, size);
2062     for(k = 0; k < nb_clusters;) {
2063         k1 = k;
2064         refcount = get_refcount(bs, k);
2065         k++;
2066         while (k < nb_clusters && get_refcount(bs, k) == refcount)
2067             k++;
2068         printf("%" PRId64 ": refcount=%d nb=%" PRId64 "\n", k, refcount,
2069                k - k1);
2070     }
2071 }
2072 #endif
2073
2074 static int qcow2_save_vmstate(BlockDriverState *bs, QEMUIOVector *qiov,
2075                               int64_t pos)
2076 {
2077     BDRVQcowState *s = bs->opaque;
2078     int64_t total_sectors = bs->total_sectors;
2079     int growable = bs->growable;
2080     bool zero_beyond_eof = bs->zero_beyond_eof;
2081     int ret;
2082
2083     BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_SAVE);
2084     bs->growable = 1;
2085     bs->zero_beyond_eof = false;
2086     ret = bdrv_pwritev(bs, qcow2_vm_state_offset(s) + pos, qiov);
2087     bs->growable = growable;
2088     bs->zero_beyond_eof = zero_beyond_eof;
2089
2090     /* bdrv_co_do_writev will have increased the total_sectors value to include
2091      * the VM state - the VM state is however not an actual part of the block
2092      * device, therefore, we need to restore the old value. */
2093     bs->total_sectors = total_sectors;
2094
2095     return ret;
2096 }
2097
2098 static int qcow2_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2099                               int64_t pos, int size)
2100 {
2101     BDRVQcowState *s = bs->opaque;
2102     int growable = bs->growable;
2103     bool zero_beyond_eof = bs->zero_beyond_eof;
2104     int ret;
2105
2106     BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_LOAD);
2107     bs->growable = 1;
2108     bs->zero_beyond_eof = false;
2109     ret = bdrv_pread(bs, qcow2_vm_state_offset(s) + pos, buf, size);
2110     bs->growable = growable;
2111     bs->zero_beyond_eof = zero_beyond_eof;
2112
2113     return ret;
2114 }
2115
2116 /*
2117  * Downgrades an image's version. To achieve this, any incompatible features
2118  * have to be removed.
2119  */
2120 static int qcow2_downgrade(BlockDriverState *bs, int target_version)
2121 {
2122     BDRVQcowState *s = bs->opaque;
2123     int current_version = s->qcow_version;
2124     int ret;
2125
2126     if (target_version == current_version) {
2127         return 0;
2128     } else if (target_version > current_version) {
2129         return -EINVAL;
2130     } else if (target_version != 2) {
2131         return -EINVAL;
2132     }
2133
2134     if (s->refcount_order != 4) {
2135         /* we would have to convert the image to a refcount_order == 4 image
2136          * here; however, since qemu (at the time of writing this) does not
2137          * support anything different than 4 anyway, there is no point in doing
2138          * so right now; however, we should error out (if qemu supports this in
2139          * the future and this code has not been adapted) */
2140         error_report("qcow2_downgrade: Image refcount orders other than 4 are "
2141                      "currently not supported.");
2142         return -ENOTSUP;
2143     }
2144
2145     /* clear incompatible features */
2146     if (s->incompatible_features & QCOW2_INCOMPAT_DIRTY) {
2147         ret = qcow2_mark_clean(bs);
2148         if (ret < 0) {
2149             return ret;
2150         }
2151     }
2152
2153     /* with QCOW2_INCOMPAT_CORRUPT, it is pretty much impossible to get here in
2154      * the first place; if that happens nonetheless, returning -ENOTSUP is the
2155      * best thing to do anyway */
2156
2157     if (s->incompatible_features) {
2158         return -ENOTSUP;
2159     }
2160
2161     /* since we can ignore compatible features, we can set them to 0 as well */
2162     s->compatible_features = 0;
2163     /* if lazy refcounts have been used, they have already been fixed through
2164      * clearing the dirty flag */
2165
2166     /* clearing autoclear features is trivial */
2167     s->autoclear_features = 0;
2168
2169     ret = qcow2_expand_zero_clusters(bs);
2170     if (ret < 0) {
2171         return ret;
2172     }
2173
2174     s->qcow_version = target_version;
2175     ret = qcow2_update_header(bs);
2176     if (ret < 0) {
2177         s->qcow_version = current_version;
2178         return ret;
2179     }
2180     return 0;
2181 }
2182
2183 static int qcow2_amend_options(BlockDriverState *bs,
2184                                QEMUOptionParameter *options)
2185 {
2186     BDRVQcowState *s = bs->opaque;
2187     int old_version = s->qcow_version, new_version = old_version;
2188     uint64_t new_size = 0;
2189     const char *backing_file = NULL, *backing_format = NULL;
2190     bool lazy_refcounts = s->use_lazy_refcounts;
2191     int ret;
2192     int i;
2193
2194     for (i = 0; options[i].name; i++)
2195     {
2196         if (!options[i].assigned) {
2197             /* only change explicitly defined options */
2198             continue;
2199         }
2200
2201         if (!strcmp(options[i].name, "compat")) {
2202             if (!options[i].value.s) {
2203                 /* preserve default */
2204             } else if (!strcmp(options[i].value.s, "0.10")) {
2205                 new_version = 2;
2206             } else if (!strcmp(options[i].value.s, "1.1")) {
2207                 new_version = 3;
2208             } else {
2209                 fprintf(stderr, "Unknown compatibility level %s.\n",
2210                         options[i].value.s);
2211                 return -EINVAL;
2212             }
2213         } else if (!strcmp(options[i].name, "preallocation")) {
2214             fprintf(stderr, "Cannot change preallocation mode.\n");
2215             return -ENOTSUP;
2216         } else if (!strcmp(options[i].name, "size")) {
2217             new_size = options[i].value.n;
2218         } else if (!strcmp(options[i].name, "backing_file")) {
2219             backing_file = options[i].value.s;
2220         } else if (!strcmp(options[i].name, "backing_fmt")) {
2221             backing_format = options[i].value.s;
2222         } else if (!strcmp(options[i].name, "encryption")) {
2223             if ((options[i].value.n != !!s->crypt_method)) {
2224                 fprintf(stderr, "Changing the encryption flag is not "
2225                         "supported.\n");
2226                 return -ENOTSUP;
2227             }
2228         } else if (!strcmp(options[i].name, "cluster_size")) {
2229             if (options[i].value.n != s->cluster_size) {
2230                 fprintf(stderr, "Changing the cluster size is not "
2231                         "supported.\n");
2232                 return -ENOTSUP;
2233             }
2234         } else if (!strcmp(options[i].name, "lazy_refcounts")) {
2235             lazy_refcounts = options[i].value.n;
2236         } else {
2237             /* if this assertion fails, this probably means a new option was
2238              * added without having it covered here */
2239             assert(false);
2240         }
2241     }
2242
2243     if (new_version != old_version) {
2244         if (new_version > old_version) {
2245             /* Upgrade */
2246             s->qcow_version = new_version;
2247             ret = qcow2_update_header(bs);
2248             if (ret < 0) {
2249                 s->qcow_version = old_version;
2250                 return ret;
2251             }
2252         } else {
2253             ret = qcow2_downgrade(bs, new_version);
2254             if (ret < 0) {
2255                 return ret;
2256             }
2257         }
2258     }
2259
2260     if (backing_file || backing_format) {
2261         ret = qcow2_change_backing_file(bs, backing_file ?: bs->backing_file,
2262                                         backing_format ?: bs->backing_format);
2263         if (ret < 0) {
2264             return ret;
2265         }
2266     }
2267
2268     if (s->use_lazy_refcounts != lazy_refcounts) {
2269         if (lazy_refcounts) {
2270             if (s->qcow_version < 3) {
2271                 fprintf(stderr, "Lazy refcounts only supported with compatibility "
2272                         "level 1.1 and above (use compat=1.1 or greater)\n");
2273                 return -EINVAL;
2274             }
2275             s->compatible_features |= QCOW2_COMPAT_LAZY_REFCOUNTS;
2276             ret = qcow2_update_header(bs);
2277             if (ret < 0) {
2278                 s->compatible_features &= ~QCOW2_COMPAT_LAZY_REFCOUNTS;
2279                 return ret;
2280             }
2281             s->use_lazy_refcounts = true;
2282         } else {
2283             /* make image clean first */
2284             ret = qcow2_mark_clean(bs);
2285             if (ret < 0) {
2286                 return ret;
2287             }
2288             /* now disallow lazy refcounts */
2289             s->compatible_features &= ~QCOW2_COMPAT_LAZY_REFCOUNTS;
2290             ret = qcow2_update_header(bs);
2291             if (ret < 0) {
2292                 s->compatible_features |= QCOW2_COMPAT_LAZY_REFCOUNTS;
2293                 return ret;
2294             }
2295             s->use_lazy_refcounts = false;
2296         }
2297     }
2298
2299     if (new_size) {
2300         ret = bdrv_truncate(bs, new_size);
2301         if (ret < 0) {
2302             return ret;
2303         }
2304     }
2305
2306     return 0;
2307 }
2308
2309 static QEMUOptionParameter qcow2_create_options[] = {
2310     {
2311         .name = BLOCK_OPT_SIZE,
2312         .type = OPT_SIZE,
2313         .help = "Virtual disk size"
2314     },
2315     {
2316         .name = BLOCK_OPT_COMPAT_LEVEL,
2317         .type = OPT_STRING,
2318         .help = "Compatibility level (0.10 or 1.1)"
2319     },
2320     {
2321         .name = BLOCK_OPT_BACKING_FILE,
2322         .type = OPT_STRING,
2323         .help = "File name of a base image"
2324     },
2325     {
2326         .name = BLOCK_OPT_BACKING_FMT,
2327         .type = OPT_STRING,
2328         .help = "Image format of the base image"
2329     },
2330     {
2331         .name = BLOCK_OPT_ENCRYPT,
2332         .type = OPT_FLAG,
2333         .help = "Encrypt the image"
2334     },
2335     {
2336         .name = BLOCK_OPT_CLUSTER_SIZE,
2337         .type = OPT_SIZE,
2338         .help = "qcow2 cluster size",
2339         .value = { .n = DEFAULT_CLUSTER_SIZE },
2340     },
2341     {
2342         .name = BLOCK_OPT_PREALLOC,
2343         .type = OPT_STRING,
2344         .help = "Preallocation mode (allowed values: off, metadata)"
2345     },
2346     {
2347         .name = BLOCK_OPT_LAZY_REFCOUNTS,
2348         .type = OPT_FLAG,
2349         .help = "Postpone refcount updates",
2350     },
2351     { NULL }
2352 };
2353
2354 static BlockDriver bdrv_qcow2 = {
2355     .format_name        = "qcow2",
2356     .instance_size      = sizeof(BDRVQcowState),
2357     .bdrv_probe         = qcow2_probe,
2358     .bdrv_open          = qcow2_open,
2359     .bdrv_close         = qcow2_close,
2360     .bdrv_reopen_prepare  = qcow2_reopen_prepare,
2361     .bdrv_create        = qcow2_create,
2362     .bdrv_has_zero_init = bdrv_has_zero_init_1,
2363     .bdrv_co_get_block_status = qcow2_co_get_block_status,
2364     .bdrv_set_key       = qcow2_set_key,
2365
2366     .bdrv_co_readv          = qcow2_co_readv,
2367     .bdrv_co_writev         = qcow2_co_writev,
2368     .bdrv_co_flush_to_os    = qcow2_co_flush_to_os,
2369
2370     .bdrv_co_write_zeroes   = qcow2_co_write_zeroes,
2371     .bdrv_co_discard        = qcow2_co_discard,
2372     .bdrv_truncate          = qcow2_truncate,
2373     .bdrv_write_compressed  = qcow2_write_compressed,
2374
2375     .bdrv_snapshot_create   = qcow2_snapshot_create,
2376     .bdrv_snapshot_goto     = qcow2_snapshot_goto,
2377     .bdrv_snapshot_delete   = qcow2_snapshot_delete,
2378     .bdrv_snapshot_list     = qcow2_snapshot_list,
2379     .bdrv_snapshot_load_tmp     = qcow2_snapshot_load_tmp,
2380     .bdrv_get_info      = qcow2_get_info,
2381     .bdrv_get_specific_info = qcow2_get_specific_info,
2382
2383     .bdrv_save_vmstate    = qcow2_save_vmstate,
2384     .bdrv_load_vmstate    = qcow2_load_vmstate,
2385
2386     .bdrv_change_backing_file   = qcow2_change_backing_file,
2387
2388     .bdrv_refresh_limits        = qcow2_refresh_limits,
2389     .bdrv_invalidate_cache      = qcow2_invalidate_cache,
2390
2391     .create_options = qcow2_create_options,
2392     .bdrv_check = qcow2_check,
2393     .bdrv_amend_options = qcow2_amend_options,
2394 };
2395
2396 static void bdrv_qcow2_init(void)
2397 {
2398     bdrv_register(&bdrv_qcow2);
2399 }
2400
2401 block_init(bdrv_qcow2_init);