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;
 273
 274         s->incompatible_features &= ~QCOW2_INCOMPAT_DIRTY;
 275
 276         ret = bdrv_flush(bs);
 277         if (ret < 0) {
 278             return ret;
 279         }
 280
 281         return qcow2_update_header(bs);
 282     }
 283     return 0;
 284 }
 285
 286 /*
 287  * Marks the image as corrupt.
 288  */
 289 int qcow2_mark_corrupt(BlockDriverState *bs)
 290 {
 291     BDRVQcowState *s = bs->opaque;
 292
 293     s->incompatible_features |= QCOW2_INCOMPAT_CORRUPT;
 294     return qcow2_update_header(bs);
 295 }
 296
 297 /*
 298  * Marks the image as consistent, i.e., unsets the corrupt bit, and flushes
 299  * before if necessary.
 300  */
 301 int qcow2_mark_consistent(BlockDriverState *bs)
 302 {
 303     BDRVQcowState *s = bs->opaque;
 304
 305     if (s->incompatible_features & QCOW2_INCOMPAT_CORRUPT) {
 306         int ret = bdrv_flush(bs);
 307         if (ret < 0) {
 308             return ret;
 309         }
 310
 311         s->incompatible_features &= ~QCOW2_INCOMPAT_CORRUPT;
 312         return qcow2_update_header(bs);
 313     }
 314     return 0;
 315 }
 316
 317 static int qcow2_check(BlockDriverState *bs, BdrvCheckResult *result,
 318                        BdrvCheckMode fix)
 319 {
 320     int ret = qcow2_check_refcounts(bs, result, fix);
 321     if (ret < 0) {
 322         return ret;
 323     }
 324
 325     if (fix && result->check_errors == 0 && result->corruptions == 0) {
 326         ret = qcow2_mark_clean(bs);
 327         if (ret < 0) {
 328             return ret;
 329         }
 330         return qcow2_mark_consistent(bs);
 331     }
 332     return ret;
 333 }
 334
 335 static int validate_table_offset(BlockDriverState *bs, uint64_t offset,
 336                                  uint64_t entries, size_t entry_len)
 337 {
 338     BDRVQcowState *s = bs->opaque;
 339     uint64_t size;
 340
 341     /* Use signed INT64_MAX as the maximum even for uint64_t header fields,
 342      * because values will be passed to qemu functions taking int64_t. */
 343     if (entries > INT64_MAX / entry_len) {
 344         return -EINVAL;
 345     }
 346
 347     size = entries * entry_len;
 348
 349     if (INT64_MAX - size < offset) {
 350         return -EINVAL;
 351     }
 352
 353     /* Tables must be cluster aligned */
 354     if (offset & (s->cluster_size - 1)) {
 355         return -EINVAL;
 356     }
 357
 358     return 0;
 359 }
 360
 361 static QemuOptsList qcow2_runtime_opts = {
 362     .name = "qcow2",
 363     .head = QTAILQ_HEAD_INITIALIZER(qcow2_runtime_opts.head),
 364     .desc = {
 365         {
 366             .name = QCOW2_OPT_LAZY_REFCOUNTS,
 367             .type = QEMU_OPT_BOOL,
 368             .help = "Postpone refcount updates",
 369         },
 370         {
 371             .name = QCOW2_OPT_DISCARD_REQUEST,
 372             .type = QEMU_OPT_BOOL,
 373             .help = "Pass guest discard requests to the layer below",
 374         },
 375         {
 376             .name = QCOW2_OPT_DISCARD_SNAPSHOT,
 377             .type = QEMU_OPT_BOOL,
 378             .help = "Generate discard requests when snapshot related space "
 379                     "is freed",
 380         },
 381         {
 382             .name = QCOW2_OPT_DISCARD_OTHER,
 383             .type = QEMU_OPT_BOOL,
 384             .help = "Generate discard requests when other clusters are freed",
 385         },
 386         {
 387             .name = QCOW2_OPT_OVERLAP,
 388             .type = QEMU_OPT_STRING,
 389             .help = "Selects which overlap checks to perform from a range of "
 390                     "templates (none, constant, cached, all)",
 391         },
 392         {
 393             .name = QCOW2_OPT_OVERLAP_MAIN_HEADER,
 394             .type = QEMU_OPT_BOOL,
 395             .help = "Check for unintended writes into the main qcow2 header",
 396         },
 397         {
 398             .name = QCOW2_OPT_OVERLAP_ACTIVE_L1,
 399             .type = QEMU_OPT_BOOL,
 400             .help = "Check for unintended writes into the active L1 table",
 401         },
 402         {
 403             .name = QCOW2_OPT_OVERLAP_ACTIVE_L2,
 404             .type = QEMU_OPT_BOOL,
 405             .help = "Check for unintended writes into an active L2 table",
 406         },
 407         {
 408             .name = QCOW2_OPT_OVERLAP_REFCOUNT_TABLE,
 409             .type = QEMU_OPT_BOOL,
 410             .help = "Check for unintended writes into the refcount table",
 411         },
 412         {
 413             .name = QCOW2_OPT_OVERLAP_REFCOUNT_BLOCK,
 414             .type = QEMU_OPT_BOOL,
 415             .help = "Check for unintended writes into a refcount block",
 416         },
 417         {
 418             .name = QCOW2_OPT_OVERLAP_SNAPSHOT_TABLE,
 419             .type = QEMU_OPT_BOOL,
 420             .help = "Check for unintended writes into the snapshot table",
 421         },
 422         {
 423             .name = QCOW2_OPT_OVERLAP_INACTIVE_L1,
 424             .type = QEMU_OPT_BOOL,
 425             .help = "Check for unintended writes into an inactive L1 table",
 426         },
 427         {
 428             .name = QCOW2_OPT_OVERLAP_INACTIVE_L2,
 429             .type = QEMU_OPT_BOOL,
 430             .help = "Check for unintended writes into an inactive L2 table",
 431         },
 432         { /* end of list */ }
 433     },
 434 };
 435
 436 static const char *overlap_bool_option_names[QCOW2_OL_MAX_BITNR] = {
 437     [QCOW2_OL_MAIN_HEADER_BITNR]    = QCOW2_OPT_OVERLAP_MAIN_HEADER,
 438     [QCOW2_OL_ACTIVE_L1_BITNR]      = QCOW2_OPT_OVERLAP_ACTIVE_L1,
 439     [QCOW2_OL_ACTIVE_L2_BITNR]      = QCOW2_OPT_OVERLAP_ACTIVE_L2,
 440     [QCOW2_OL_REFCOUNT_TABLE_BITNR] = QCOW2_OPT_OVERLAP_REFCOUNT_TABLE,
 441     [QCOW2_OL_REFCOUNT_BLOCK_BITNR] = QCOW2_OPT_OVERLAP_REFCOUNT_BLOCK,
 442     [QCOW2_OL_SNAPSHOT_TABLE_BITNR] = QCOW2_OPT_OVERLAP_SNAPSHOT_TABLE,
 443     [QCOW2_OL_INACTIVE_L1_BITNR]    = QCOW2_OPT_OVERLAP_INACTIVE_L1,
 444     [QCOW2_OL_INACTIVE_L2_BITNR]    = QCOW2_OPT_OVERLAP_INACTIVE_L2,
 445 };
 446
 447 static int qcow2_open(BlockDriverState *bs, QDict *options, int flags,
 448                       Error **errp)
 449 {
 450     BDRVQcowState *s = bs->opaque;
 451     unsigned int len, i;
 452     int ret = 0;
 453     QCowHeader header;
 454     QemuOpts *opts;
 455     Error *local_err = NULL;
 456     uint64_t ext_end;
 457     uint64_t l1_vm_state_index;
 458     const char *opt_overlap_check;
 459     int overlap_check_template = 0;
 460
 461     ret = bdrv_pread(bs->file, 0, &header, sizeof(header));
 462     if (ret < 0) {
 463         error_setg_errno(errp, -ret, "Could not read qcow2 header");
 464         goto fail;
 465     }
 466     be32_to_cpus(&header.magic);
 467     be32_to_cpus(&header.version);
 468     be64_to_cpus(&header.backing_file_offset);
 469     be32_to_cpus(&header.backing_file_size);
 470     be64_to_cpus(&header.size);
 471     be32_to_cpus(&header.cluster_bits);
 472     be32_to_cpus(&header.crypt_method);
 473     be64_to_cpus(&header.l1_table_offset);
 474     be32_to_cpus(&header.l1_size);
 475     be64_to_cpus(&header.refcount_table_offset);
 476     be32_to_cpus(&header.refcount_table_clusters);
 477     be64_to_cpus(&header.snapshots_offset);
 478     be32_to_cpus(&header.nb_snapshots);
 479
 480     if (header.magic != QCOW_MAGIC) {
 481         error_setg(errp, "Image is not in qcow2 format");
 482         ret = -EINVAL;
 483         goto fail;
 484     }
 485     if (header.version < 2 || header.version > 3) {
 486         report_unsupported(bs, errp, "QCOW version %d", header.version);
 487         ret = -ENOTSUP;
 488         goto fail;
 489     }
 490
 491     s->qcow_version = header.version;
 492
 493     /* Initialise cluster size */
 494     if (header.cluster_bits < MIN_CLUSTER_BITS ||
 495         header.cluster_bits > MAX_CLUSTER_BITS) {
 496         error_setg(errp, "Unsupported cluster size: 2^%i", header.cluster_bits);
 497         ret = -EINVAL;
 498         goto fail;
 499     }
 500
 501     s->cluster_bits = header.cluster_bits;
 502     s->cluster_size = 1 << s->cluster_bits;
 503     s->cluster_sectors = 1 << (s->cluster_bits - 9);
 504
 505     /* Initialise version 3 header fields */
 506     if (header.version == 2) {
 507         header.incompatible_features    = 0;
 508         header.compatible_features      = 0;
 509         header.autoclear_features       = 0;
 510         header.refcount_order           = 4;
 511         header.header_length            = 72;
 512     } else {
 513         be64_to_cpus(&header.incompatible_features);
 514         be64_to_cpus(&header.compatible_features);
 515         be64_to_cpus(&header.autoclear_features);
 516         be32_to_cpus(&header.refcount_order);
 517         be32_to_cpus(&header.header_length);
 518
 519         if (header.header_length < 104) {
 520             error_setg(errp, "qcow2 header too short");
 521             ret = -EINVAL;
 522             goto fail;
 523         }
 524     }
 525
 526     if (header.header_length > s->cluster_size) {
 527         error_setg(errp, "qcow2 header exceeds cluster size");
 528         ret = -EINVAL;
 529         goto fail;
 530     }
 531
 532     if (header.header_length > sizeof(header)) {
 533         s->unknown_header_fields_size = header.header_length - sizeof(header);
 534         s->unknown_header_fields = g_malloc(s->unknown_header_fields_size);
 535         ret = bdrv_pread(bs->file, sizeof(header), s->unknown_header_fields,
 536                          s->unknown_header_fields_size);
 537         if (ret < 0) {
 538             error_setg_errno(errp, -ret, "Could not read unknown qcow2 header "
 539                              "fields");
 540             goto fail;
 541         }
 542     }
 543
 544     if (header.backing_file_offset > s->cluster_size) {
 545         error_setg(errp, "Invalid backing file offset");
 546         ret = -EINVAL;
 547         goto fail;
 548     }
 549
 550     if (header.backing_file_offset) {
 551         ext_end = header.backing_file_offset;
 552     } else {
 553         ext_end = 1 << header.cluster_bits;
 554     }
 555
 556     /* Handle feature bits */
 557     s->incompatible_features    = header.incompatible_features;
 558     s->compatible_features      = header.compatible_features;
 559     s->autoclear_features       = header.autoclear_features;
 560
 561     if (s->incompatible_features & ~QCOW2_INCOMPAT_MASK) {
 562         void *feature_table = NULL;
 563         qcow2_read_extensions(bs, header.header_length, ext_end,
 564                               &feature_table, NULL);
 565         report_unsupported_feature(bs, errp, feature_table,
 566                                    s->incompatible_features &
 567                                    ~QCOW2_INCOMPAT_MASK);
 568         ret = -ENOTSUP;
 569         g_free(feature_table);
 570         goto fail;
 571     }
 572
 573     if (s->incompatible_features & QCOW2_INCOMPAT_CORRUPT) {
 574         /* Corrupt images may not be written to unless they are being repaired
 575          */
 576         if ((flags & BDRV_O_RDWR) && !(flags & BDRV_O_CHECK)) {
 577             error_setg(errp, "qcow2: Image is corrupt; cannot be opened "
 578                        "read/write");
 579             ret = -EACCES;
 580             goto fail;
 581         }
 582     }
 583
 584     /* Check support for various header values */
 585     if (header.refcount_order != 4) {
 586         report_unsupported(bs, errp, "%d bit reference counts",
 587                            1 << header.refcount_order);
 588         ret = -ENOTSUP;
 589         goto fail;
 590     }
 591     s->refcount_order = header.refcount_order;
 592
 593     if (header.crypt_method > QCOW_CRYPT_AES) {
 594         error_setg(errp, "Unsupported encryption method: %i",
 595                    header.crypt_method);
 596         ret = -EINVAL;
 597         goto fail;
 598     }
 599     s->crypt_method_header = header.crypt_method;
 600     if (s->crypt_method_header) {
 601         bs->encrypted = 1;
 602     }
 603
 604     s->l2_bits = s->cluster_bits - 3; /* L2 is always one cluster */
 605     s->l2_size = 1 << s->l2_bits;
 606     bs->total_sectors = header.size / 512;
 607     s->csize_shift = (62 - (s->cluster_bits - 8));
 608     s->csize_mask = (1 << (s->cluster_bits - 8)) - 1;
 609     s->cluster_offset_mask = (1LL << s->csize_shift) - 1;
 610
 611     s->refcount_table_offset = header.refcount_table_offset;
 612     s->refcount_table_size =
 613         header.refcount_table_clusters << (s->cluster_bits - 3);
 614
 615     if (header.refcount_table_clusters > qcow2_max_refcount_clusters(s)) {
 616         error_setg(errp, "Reference count table too large");
 617         ret = -EINVAL;
 618         goto fail;
 619     }
 620
 621     ret = validate_table_offset(bs, s->refcount_table_offset,
 622                                 s->refcount_table_size, sizeof(uint64_t));
 623     if (ret < 0) {
 624         error_setg(errp, "Invalid reference count table offset");
 625         goto fail;
 626     }
 627
 628     /* Snapshot table offset/length */
 629     if (header.nb_snapshots > QCOW_MAX_SNAPSHOTS) {
 630         error_setg(errp, "Too many snapshots");
 631         ret = -EINVAL;
 632         goto fail;
 633     }
 634
 635     ret = validate_table_offset(bs, header.snapshots_offset,
 636                                 header.nb_snapshots,
 637                                 sizeof(QCowSnapshotHeader));
 638     if (ret < 0) {
 639         error_setg(errp, "Invalid snapshot table offset");
 640         goto fail;
 641     }
 642
 643     /* read the level 1 table */
 644     if (header.l1_size > QCOW_MAX_L1_SIZE) {
 645         error_setg(errp, "Active L1 table too large");
 646         ret = -EFBIG;
 647         goto fail;
 648     }
 649     s->l1_size = header.l1_size;
 650
 651     l1_vm_state_index = size_to_l1(s, header.size);
 652     if (l1_vm_state_index > INT_MAX) {
 653         error_setg(errp, "Image is too big");
 654         ret = -EFBIG;
 655         goto fail;
 656     }
 657     s->l1_vm_state_index = l1_vm_state_index;
 658
 659     /* the L1 table must contain at least enough entries to put
 660        header.size bytes */
 661     if (s->l1_size < s->l1_vm_state_index) {
 662         error_setg(errp, "L1 table is too small");
 663         ret = -EINVAL;
 664         goto fail;
 665     }
 666
 667     ret = validate_table_offset(bs, header.l1_table_offset,
 668                                 header.l1_size, sizeof(uint64_t));
 669     if (ret < 0) {
 670         error_setg(errp, "Invalid L1 table offset");
 671         goto fail;
 672     }
 673     s->l1_table_offset = header.l1_table_offset;
 674
 675
 676     if (s->l1_size > 0) {
 677         s->l1_table = g_malloc0(
 678             align_offset(s->l1_size * sizeof(uint64_t), 512));
 679         ret = bdrv_pread(bs->file, s->l1_table_offset, s->l1_table,
 680                          s->l1_size * sizeof(uint64_t));
 681         if (ret < 0) {
 682             error_setg_errno(errp, -ret, "Could not read L1 table");
 683             goto fail;
 684         }
 685         for(i = 0;i < s->l1_size; i++) {
 686             be64_to_cpus(&s->l1_table[i]);
 687         }
 688     }
 689
 690     /* alloc L2 table/refcount block cache */
 691     s->l2_table_cache = qcow2_cache_create(bs, L2_CACHE_SIZE);
 692     s->refcount_block_cache = qcow2_cache_create(bs, REFCOUNT_CACHE_SIZE);
 693
 694     s->cluster_cache = g_malloc(s->cluster_size);
 695     /* one more sector for decompressed data alignment */
 696     s->cluster_data = qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size
 697                                   + 512);
 698     s->cluster_cache_offset = -1;
 699     s->flags = flags;
 700
 701     ret = qcow2_refcount_init(bs);
 702     if (ret != 0) {
 703         error_setg_errno(errp, -ret, "Could not initialize refcount handling");
 704         goto fail;
 705     }
 706
 707     QLIST_INIT(&s->cluster_allocs);
 708     QTAILQ_INIT(&s->discards);
 709
 710     /* read qcow2 extensions */
 711     if (qcow2_read_extensions(bs, header.header_length, ext_end, NULL,
 712         &local_err)) {
 713         error_propagate(errp, local_err);
 714         ret = -EINVAL;
 715         goto fail;
 716     }
 717
 718     /* read the backing file name */
 719     if (header.backing_file_offset != 0) {
 720         len = header.backing_file_size;
 721         if (len > MIN(1023, s->cluster_size - header.backing_file_offset)) {
 722             error_setg(errp, "Backing file name too long");
 723             ret = -EINVAL;
 724             goto fail;
 725         }
 726         ret = bdrv_pread(bs->file, header.backing_file_offset,
 727                          bs->backing_file, len);
 728         if (ret < 0) {
 729             error_setg_errno(errp, -ret, "Could not read backing file name");
 730             goto fail;
 731         }
 732         bs->backing_file[len] = '\0';
 733     }
 734
 735     /* Internal snapshots */
 736     s->snapshots_offset = header.snapshots_offset;
 737     s->nb_snapshots = header.nb_snapshots;
 738
 739     ret = qcow2_read_snapshots(bs);
 740     if (ret < 0) {
 741         error_setg_errno(errp, -ret, "Could not read snapshots");
 742         goto fail;
 743     }
 744
 745     /* Clear unknown autoclear feature bits */
 746     if (!bs->read_only && !(flags & BDRV_O_INCOMING) && s->autoclear_features) {
 747         s->autoclear_features = 0;
 748         ret = qcow2_update_header(bs);
 749         if (ret < 0) {
 750             error_setg_errno(errp, -ret, "Could not update qcow2 header");
 751             goto fail;
 752         }
 753     }
 754
 755     /* Initialise locks */
 756     qemu_co_mutex_init(&s->lock);
 757
 758     /* Repair image if dirty */
 759     if (!(flags & (BDRV_O_CHECK | BDRV_O_INCOMING)) && !bs->read_only &&
 760         (s->incompatible_features & QCOW2_INCOMPAT_DIRTY)) {
 761         BdrvCheckResult result = {0};
 762
 763         ret = qcow2_check(bs, &result, BDRV_FIX_ERRORS);
 764         if (ret < 0) {
 765             error_setg_errno(errp, -ret, "Could not repair dirty image");
 766             goto fail;
 767         }
 768     }
 769
 770     /* Enable lazy_refcounts according to image and command line options */
 771     opts = qemu_opts_create(&qcow2_runtime_opts, NULL, 0, &error_abort);
 772     qemu_opts_absorb_qdict(opts, options, &local_err);
 773     if (local_err) {
 774         error_propagate(errp, local_err);
 775         ret = -EINVAL;
 776         goto fail;
 777     }
 778
 779     s->use_lazy_refcounts = qemu_opt_get_bool(opts, QCOW2_OPT_LAZY_REFCOUNTS,
 780         (s->compatible_features & QCOW2_COMPAT_LAZY_REFCOUNTS));
 781
 782     s->discard_passthrough[QCOW2_DISCARD_NEVER] = false;
 783     s->discard_passthrough[QCOW2_DISCARD_ALWAYS] = true;
 784     s->discard_passthrough[QCOW2_DISCARD_REQUEST] =
 785         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_REQUEST,
 786                           flags & BDRV_O_UNMAP);
 787     s->discard_passthrough[QCOW2_DISCARD_SNAPSHOT] =
 788         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_SNAPSHOT, true);
 789     s->discard_passthrough[QCOW2_DISCARD_OTHER] =
 790         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_OTHER, false);
 791
 792     opt_overlap_check = qemu_opt_get(opts, "overlap-check") ?: "cached";
 793     if (!strcmp(opt_overlap_check, "none")) {
 794         overlap_check_template = 0;
 795     } else if (!strcmp(opt_overlap_check, "constant")) {
 796         overlap_check_template = QCOW2_OL_CONSTANT;
 797     } else if (!strcmp(opt_overlap_check, "cached")) {
 798         overlap_check_template = QCOW2_OL_CACHED;
 799     } else if (!strcmp(opt_overlap_check, "all")) {
 800         overlap_check_template = QCOW2_OL_ALL;
 801     } else {
 802         error_setg(errp, "Unsupported value '%s' for qcow2 option "
 803                    "'overlap-check'. Allowed are either of the following: "
 804                    "none, constant, cached, all", opt_overlap_check);
 805         qemu_opts_del(opts);
 806         ret = -EINVAL;
 807         goto fail;
 808     }
 809
 810     s->overlap_check = 0;
 811     for (i = 0; i < QCOW2_OL_MAX_BITNR; i++) {
 812         /* overlap-check defines a template bitmask, but every flag may be
 813          * overwritten through the associated boolean option */
 814         s->overlap_check |=
 815             qemu_opt_get_bool(opts, overlap_bool_option_names[i],
 816                               overlap_check_template & (1 << i)) << i;
 817     }
 818
 819     qemu_opts_del(opts);
 820
 821     if (s->use_lazy_refcounts && s->qcow_version < 3) {
 822         error_setg(errp, "Lazy refcounts require a qcow2 image with at least "
 823                    "qemu 1.1 compatibility level");
 824         ret = -EINVAL;
 825         goto fail;
 826     }
 827
 828 #ifdef DEBUG_ALLOC
 829     {
 830         BdrvCheckResult result = {0};
 831         qcow2_check_refcounts(bs, &result, 0);
 832     }
 833 #endif
 834     return ret;
 835
 836  fail:
 837     g_free(s->unknown_header_fields);
 838     cleanup_unknown_header_ext(bs);
 839     qcow2_free_snapshots(bs);
 840     qcow2_refcount_close(bs);
 841     g_free(s->l1_table);
 842     /* else pre-write overlap checks in cache_destroy may crash */
 843     s->l1_table = NULL;
 844     if (s->l2_table_cache) {
 845         qcow2_cache_destroy(bs, s->l2_table_cache);
 846     }
 847     if (s->refcount_block_cache) {
 848         qcow2_cache_destroy(bs, s->refcount_block_cache);
 849     }
 850     g_free(s->cluster_cache);
 851     qemu_vfree(s->cluster_data);
 852     return ret;
 853 }
 854
 855 static int qcow2_refresh_limits(BlockDriverState *bs)
 856 {
 857     BDRVQcowState *s = bs->opaque;
 858
 859     bs->bl.write_zeroes_alignment = s->cluster_sectors;
 860
 861     return 0;
 862 }
 863
 864 static int qcow2_set_key(BlockDriverState *bs, const char *key)
 865 {
 866     BDRVQcowState *s = bs->opaque;
 867     uint8_t keybuf[16];
 868     int len, i;
 869
 870     memset(keybuf, 0, 16);
 871     len = strlen(key);
 872     if (len > 16)
 873         len = 16;
 874     /* XXX: we could compress the chars to 7 bits to increase
 875        entropy */
 876     for(i = 0;i < len;i++) {
 877         keybuf[i] = key[i];
 878     }
 879     s->crypt_method = s->crypt_method_header;
 880
 881     if (AES_set_encrypt_key(keybuf, 128, &s->aes_encrypt_key) != 0)
 882         return -1;
 883     if (AES_set_decrypt_key(keybuf, 128, &s->aes_decrypt_key) != 0)
 884         return -1;
 885 #if 0
 886     /* test */
 887     {
 888         uint8_t in[16];
 889         uint8_t out[16];
 890         uint8_t tmp[16];
 891         for(i=0;i<16;i++)
 892             in[i] = i;
 893         AES_encrypt(in, tmp, &s->aes_encrypt_key);
 894         AES_decrypt(tmp, out, &s->aes_decrypt_key);
 895         for(i = 0; i < 16; i++)
 896             printf(" %02x", tmp[i]);
 897         printf("\n");
 898         for(i = 0; i < 16; i++)
 899             printf(" %02x", out[i]);
 900         printf("\n");
 901     }
 902 #endif
 903     return 0;
 904 }
 905
 906 /* We have no actual commit/abort logic for qcow2, but we need to write out any
 907  * unwritten data if we reopen read-only. */
 908 static int qcow2_reopen_prepare(BDRVReopenState *state,
 909                                 BlockReopenQueue *queue, Error **errp)
 910 {
 911     int ret;
 912
 913     if ((state->flags & BDRV_O_RDWR) == 0) {
 914         ret = bdrv_flush(state->bs);
 915         if (ret < 0) {
 916             return ret;
 917         }
 918
 919         ret = qcow2_mark_clean(state->bs);
 920         if (ret < 0) {
 921             return ret;
 922         }
 923     }
 924
 925     return 0;
 926 }
 927
 928 static int64_t coroutine_fn qcow2_co_get_block_status(BlockDriverState *bs,
 929         int64_t sector_num, int nb_sectors, int *pnum)
 930 {
 931     BDRVQcowState *s = bs->opaque;
 932     uint64_t cluster_offset;
 933     int index_in_cluster, ret;
 934     int64_t status = 0;
 935
 936     *pnum = nb_sectors;
 937     qemu_co_mutex_lock(&s->lock);
 938     ret = qcow2_get_cluster_offset(bs, sector_num << 9, pnum, &cluster_offset);
 939     qemu_co_mutex_unlock(&s->lock);
 940     if (ret < 0) {
 941         return ret;
 942     }
 943
 944     if (cluster_offset != 0 && ret != QCOW2_CLUSTER_COMPRESSED &&
 945         !s->crypt_method) {
 946         index_in_cluster = sector_num & (s->cluster_sectors - 1);
 947         cluster_offset |= (index_in_cluster << BDRV_SECTOR_BITS);
 948         status |= BDRV_BLOCK_OFFSET_VALID | cluster_offset;
 949     }
 950     if (ret == QCOW2_CLUSTER_ZERO) {
 951         status |= BDRV_BLOCK_ZERO;
 952     } else if (ret != QCOW2_CLUSTER_UNALLOCATED) {
 953         status |= BDRV_BLOCK_DATA;
 954     }
 955     return status;
 956 }
 957
 958 /* handle reading after the end of the backing file */
 959 int qcow2_backing_read1(BlockDriverState *bs, QEMUIOVector *qiov,
 960                   int64_t sector_num, int nb_sectors)
 961 {
 962     int n1;
 963     if ((sector_num + nb_sectors) <= bs->total_sectors)
 964         return nb_sectors;
 965     if (sector_num >= bs->total_sectors)
 966         n1 = 0;
 967     else
 968         n1 = bs->total_sectors - sector_num;
 969
 970     qemu_iovec_memset(qiov, 512 * n1, 0, 512 * (nb_sectors - n1));
 971
 972     return n1;
 973 }
 974
 975 static coroutine_fn int qcow2_co_readv(BlockDriverState *bs, int64_t sector_num,
 976                           int remaining_sectors, QEMUIOVector *qiov)
 977 {
 978     BDRVQcowState *s = bs->opaque;
 979     int index_in_cluster, n1;
 980     int ret;
 981     int cur_nr_sectors; /* number of sectors in current iteration */
 982     uint64_t cluster_offset = 0;
 983     uint64_t bytes_done = 0;
 984     QEMUIOVector hd_qiov;
 985     uint8_t *cluster_data = NULL;
 986
 987     qemu_iovec_init(&hd_qiov, qiov->niov);
 988
 989     qemu_co_mutex_lock(&s->lock);
 990
 991     while (remaining_sectors != 0) {
 992
 993         /* prepare next request */
 994         cur_nr_sectors = remaining_sectors;
 995         if (s->crypt_method) {
 996             cur_nr_sectors = MIN(cur_nr_sectors,
 997                 QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
 998         }
 999
1000         ret = qcow2_get_cluster_offset(bs, sector_num << 9,
1001             &cur_nr_sectors, &cluster_offset);
1002         if (ret < 0) {
1003             goto fail;
1004         }
1005
1006         index_in_cluster = sector_num & (s->cluster_sectors - 1);
1007
1008         qemu_iovec_reset(&hd_qiov);
1009         qemu_iovec_concat(&hd_qiov, qiov, bytes_done,
1010             cur_nr_sectors * 512);
1011
1012         switch (ret) {
1013         case QCOW2_CLUSTER_UNALLOCATED:
1014
1015             if (bs->backing_hd) {
1016                 /* read from the base image */
1017                 n1 = qcow2_backing_read1(bs->backing_hd, &hd_qiov,
1018                     sector_num, cur_nr_sectors);
1019                 if (n1 > 0) {
1020                     BLKDBG_EVENT(bs->file, BLKDBG_READ_BACKING_AIO);
1021                     qemu_co_mutex_unlock(&s->lock);
1022                     ret = bdrv_co_readv(bs->backing_hd, sector_num,
1023                                         n1, &hd_qiov);
1024                     qemu_co_mutex_lock(&s->lock);
1025                     if (ret < 0) {
1026                         goto fail;
1027                     }
1028                 }
1029             } else {
1030                 /* Note: in this case, no need to wait */
1031                 qemu_iovec_memset(&hd_qiov, 0, 0, 512 * cur_nr_sectors);
1032             }
1033             break;
1034
1035         case QCOW2_CLUSTER_ZERO:
1036             qemu_iovec_memset(&hd_qiov, 0, 0, 512 * cur_nr_sectors);
1037             break;
1038
1039         case QCOW2_CLUSTER_COMPRESSED:
1040             /* add AIO support for compressed blocks ? */
1041             ret = qcow2_decompress_cluster(bs, cluster_offset);
1042             if (ret < 0) {
1043                 goto fail;
1044             }
1045
1046             qemu_iovec_from_buf(&hd_qiov, 0,
1047                 s->cluster_cache + index_in_cluster * 512,
1048                 512 * cur_nr_sectors);
1049             break;
1050
1051         case QCOW2_CLUSTER_NORMAL:
1052             if ((cluster_offset & 511) != 0) {
1053                 ret = -EIO;
1054                 goto fail;
1055             }
1056
1057             if (s->crypt_method) {
1058                 /*
1059                  * For encrypted images, read everything into a temporary
1060                  * contiguous buffer on which the AES functions can work.
1061                  */
1062                 if (!cluster_data) {
1063                     cluster_data =
1064                         qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
1065                 }
1066
1067                 assert(cur_nr_sectors <=
1068                     QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors);
1069                 qemu_iovec_reset(&hd_qiov);
1070                 qemu_iovec_add(&hd_qiov, cluster_data,
1071                     512 * cur_nr_sectors);
1072             }
1073
1074             BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
1075             qemu_co_mutex_unlock(&s->lock);
1076             ret = bdrv_co_readv(bs->file,
1077                                 (cluster_offset >> 9) + index_in_cluster,
1078                                 cur_nr_sectors, &hd_qiov);
1079             qemu_co_mutex_lock(&s->lock);
1080             if (ret < 0) {
1081                 goto fail;
1082             }
1083             if (s->crypt_method) {
1084                 qcow2_encrypt_sectors(s, sector_num,  cluster_data,
1085                     cluster_data, cur_nr_sectors, 0, &s->aes_decrypt_key);
1086                 qemu_iovec_from_buf(qiov, bytes_done,
1087                     cluster_data, 512 * cur_nr_sectors);
1088             }
1089             break;
1090
1091         default:
1092             g_assert_not_reached();
1093             ret = -EIO;
1094             goto fail;
1095         }
1096
1097         remaining_sectors -= cur_nr_sectors;
1098         sector_num += cur_nr_sectors;
1099         bytes_done += cur_nr_sectors * 512;
1100     }
1101     ret = 0;
1102
1103 fail:
1104     qemu_co_mutex_unlock(&s->lock);
1105
1106     qemu_iovec_destroy(&hd_qiov);
1107     qemu_vfree(cluster_data);
1108
1109     return ret;
1110 }
1111
1112 static coroutine_fn int qcow2_co_writev(BlockDriverState *bs,
1113                            int64_t sector_num,
1114                            int remaining_sectors,
1115                            QEMUIOVector *qiov)
1116 {
1117     BDRVQcowState *s = bs->opaque;
1118     int index_in_cluster;
1119     int ret;
1120     int cur_nr_sectors; /* number of sectors in current iteration */
1121     uint64_t cluster_offset;
1122     QEMUIOVector hd_qiov;
1123     uint64_t bytes_done = 0;
1124     uint8_t *cluster_data = NULL;
1125     QCowL2Meta *l2meta = NULL;
1126
1127     trace_qcow2_writev_start_req(qemu_coroutine_self(), sector_num,
1128                                  remaining_sectors);
1129
1130     qemu_iovec_init(&hd_qiov, qiov->niov);
1131
1132     s->cluster_cache_offset = -1; /* disable compressed cache */
1133
1134     qemu_co_mutex_lock(&s->lock);
1135
1136     while (remaining_sectors != 0) {
1137
1138         l2meta = NULL;
1139
1140         trace_qcow2_writev_start_part(qemu_coroutine_self());
1141         index_in_cluster = sector_num & (s->cluster_sectors - 1);
1142         cur_nr_sectors = remaining_sectors;
1143         if (s->crypt_method &&
1144             cur_nr_sectors >
1145             QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors - index_in_cluster) {
1146             cur_nr_sectors =
1147                 QCOW_MAX_CRYPT_CLUSTERS * s->cluster_sectors - index_in_cluster;
1148         }
1149
1150         ret = qcow2_alloc_cluster_offset(bs, sector_num << 9,
1151             &cur_nr_sectors, &cluster_offset, &l2meta);
1152         if (ret < 0) {
1153             goto fail;
1154         }
1155
1156         assert((cluster_offset & 511) == 0);
1157
1158         qemu_iovec_reset(&hd_qiov);
1159         qemu_iovec_concat(&hd_qiov, qiov, bytes_done,
1160             cur_nr_sectors * 512);
1161
1162         if (s->crypt_method) {
1163             if (!cluster_data) {
1164                 cluster_data = qemu_blockalign(bs, QCOW_MAX_CRYPT_CLUSTERS *
1165                                                  s->cluster_size);
1166             }
1167
1168             assert(hd_qiov.size <=
1169                    QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
1170             qemu_iovec_to_buf(&hd_qiov, 0, cluster_data, hd_qiov.size);
1171
1172             qcow2_encrypt_sectors(s, sector_num, cluster_data,
1173                 cluster_data, cur_nr_sectors, 1, &s->aes_encrypt_key);
1174
1175             qemu_iovec_reset(&hd_qiov);
1176             qemu_iovec_add(&hd_qiov, cluster_data,
1177                 cur_nr_sectors * 512);
1178         }
1179
1180         ret = qcow2_pre_write_overlap_check(bs, 0,
1181                 cluster_offset + index_in_cluster * BDRV_SECTOR_SIZE,
1182                 cur_nr_sectors * BDRV_SECTOR_SIZE);
1183         if (ret < 0) {
1184             goto fail;
1185         }
1186
1187         qemu_co_mutex_unlock(&s->lock);
1188         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
1189         trace_qcow2_writev_data(qemu_coroutine_self(),
1190                                 (cluster_offset >> 9) + index_in_cluster);
1191         ret = bdrv_co_writev(bs->file,
1192                              (cluster_offset >> 9) + index_in_cluster,
1193                              cur_nr_sectors, &hd_qiov);
1194         qemu_co_mutex_lock(&s->lock);
1195         if (ret < 0) {
1196             goto fail;
1197         }
1198
1199         while (l2meta != NULL) {
1200             QCowL2Meta *next;
1201
1202             ret = qcow2_alloc_cluster_link_l2(bs, l2meta);
1203             if (ret < 0) {
1204                 goto fail;
1205             }
1206
1207             /* Take the request off the list of running requests */
1208             if (l2meta->nb_clusters != 0) {
1209                 QLIST_REMOVE(l2meta, next_in_flight);
1210             }
1211
1212             qemu_co_queue_restart_all(&l2meta->dependent_requests);
1213
1214             next = l2meta->next;
1215             g_free(l2meta);
1216             l2meta = next;
1217         }
1218
1219         remaining_sectors -= cur_nr_sectors;
1220         sector_num += cur_nr_sectors;
1221         bytes_done += cur_nr_sectors * 512;
1222         trace_qcow2_writev_done_part(qemu_coroutine_self(), cur_nr_sectors);
1223     }
1224     ret = 0;
1225
1226 fail:
1227     qemu_co_mutex_unlock(&s->lock);
1228
1229     while (l2meta != NULL) {
1230         QCowL2Meta *next;
1231
1232         if (l2meta->nb_clusters != 0) {
1233             QLIST_REMOVE(l2meta, next_in_flight);
1234         }
1235         qemu_co_queue_restart_all(&l2meta->dependent_requests);
1236
1237         next = l2meta->next;
1238         g_free(l2meta);
1239         l2meta = next;
1240     }
1241
1242     qemu_iovec_destroy(&hd_qiov);
1243     qemu_vfree(cluster_data);
1244     trace_qcow2_writev_done_req(qemu_coroutine_self(), ret);
1245
1246     return ret;
1247 }
1248
1249 static void qcow2_close(BlockDriverState *bs)
1250 {
1251     BDRVQcowState *s = bs->opaque;
1252     g_free(s->l1_table);
1253     /* else pre-write overlap checks in cache_destroy may crash */
1254     s->l1_table = NULL;
1255
1256     if (!(bs->open_flags & BDRV_O_INCOMING)) {
1257         qcow2_cache_flush(bs, s->l2_table_cache);
1258         qcow2_cache_flush(bs, s->refcount_block_cache);
1259
1260         qcow2_mark_clean(bs);
1261     }
1262
1263     qcow2_cache_destroy(bs, s->l2_table_cache);
1264     qcow2_cache_destroy(bs, s->refcount_block_cache);
1265
1266     g_free(s->unknown_header_fields);
1267     cleanup_unknown_header_ext(bs);
1268
1269     g_free(s->cluster_cache);
1270     qemu_vfree(s->cluster_data);
1271     qcow2_refcount_close(bs);
1272     qcow2_free_snapshots(bs);
1273 }
1274
1275 static void qcow2_invalidate_cache(BlockDriverState *bs, Error **errp)
1276 {
1277     BDRVQcowState *s = bs->opaque;
1278     int flags = s->flags;
1279     AES_KEY aes_encrypt_key;
1280     AES_KEY aes_decrypt_key;
1281     uint32_t crypt_method = 0;
1282     QDict *options;
1283     Error *local_err = NULL;
1284     int ret;
1285
1286     /*
1287      * Backing files are read-only which makes all of their metadata immutable,
1288      * that means we don't have to worry about reopening them here.
1289      */
1290
1291     if (s->crypt_method) {
1292         crypt_method = s->crypt_method;
1293         memcpy(&aes_encrypt_key, &s->aes_encrypt_key, sizeof(aes_encrypt_key));
1294         memcpy(&aes_decrypt_key, &s->aes_decrypt_key, sizeof(aes_decrypt_key));
1295     }
1296
1297     qcow2_close(bs);
1298
1299     bdrv_invalidate_cache(bs->file, &local_err);
1300     if (local_err) {
1301         error_propagate(errp, local_err);
1302         return;
1303     }
1304
1305     memset(s, 0, sizeof(BDRVQcowState));
1306     options = qdict_clone_shallow(bs->options);
1307
1308     ret = qcow2_open(bs, options, flags, &local_err);
1309     if (local_err) {
1310         error_setg(errp, "Could not reopen qcow2 layer: %s",
1311                    error_get_pretty(local_err));
1312         error_free(local_err);
1313         return;
1314     } else if (ret < 0) {
1315         error_setg_errno(errp, -ret, "Could not reopen qcow2 layer");
1316         return;
1317     }
1318
1319     QDECREF(options);
1320
1321     if (crypt_method) {
1322         s->crypt_method = crypt_method;
1323         memcpy(&s->aes_encrypt_key, &aes_encrypt_key, sizeof(aes_encrypt_key));
1324         memcpy(&s->aes_decrypt_key, &aes_decrypt_key, sizeof(aes_decrypt_key));
1325     }
1326 }
1327
1328 static size_t header_ext_add(char *buf, uint32_t magic, const void *s,
1329     size_t len, size_t buflen)
1330 {
1331     QCowExtension *ext_backing_fmt = (QCowExtension*) buf;
1332     size_t ext_len = sizeof(QCowExtension) + ((len + 7) & ~7);
1333
1334     if (buflen < ext_len) {
1335         return -ENOSPC;
1336     }
1337
1338     *ext_backing_fmt = (QCowExtension) {
1339         .magic  = cpu_to_be32(magic),
1340         .len    = cpu_to_be32(len),
1341     };
1342     memcpy(buf + sizeof(QCowExtension), s, len);
1343
1344     return ext_len;
1345 }
1346
1347 /*
1348  * Updates the qcow2 header, including the variable length parts of it, i.e.
1349  * the backing file name and all extensions. qcow2 was not designed to allow
1350  * such changes, so if we run out of space (we can only use the first cluster)
1351  * this function may fail.
1352  *
1353  * Returns 0 on success, -errno in error cases.
1354  */
1355 int qcow2_update_header(BlockDriverState *bs)
1356 {
1357     BDRVQcowState *s = bs->opaque;
1358     QCowHeader *header;
1359     char *buf;
1360     size_t buflen = s->cluster_size;
1361     int ret;
1362     uint64_t total_size;
1363     uint32_t refcount_table_clusters;
1364     size_t header_length;
1365     Qcow2UnknownHeaderExtension *uext;
1366
1367     buf = qemu_blockalign(bs, buflen);
1368
1369     /* Header structure */
1370     header = (QCowHeader*) buf;
1371
1372     if (buflen < sizeof(*header)) {
1373         ret = -ENOSPC;
1374         goto fail;
1375     }
1376
1377     header_length = sizeof(*header) + s->unknown_header_fields_size;
1378     total_size = bs->total_sectors * BDRV_SECTOR_SIZE;
1379     refcount_table_clusters = s->refcount_table_size >> (s->cluster_bits - 3);
1380
1381     *header = (QCowHeader) {
1382         /* Version 2 fields */
1383         .magic                  = cpu_to_be32(QCOW_MAGIC),
1384         .version                = cpu_to_be32(s->qcow_version),
1385         .backing_file_offset    = 0,
1386         .backing_file_size      = 0,
1387         .cluster_bits           = cpu_to_be32(s->cluster_bits),
1388         .size                   = cpu_to_be64(total_size),
1389         .crypt_method           = cpu_to_be32(s->crypt_method_header),
1390         .l1_size                = cpu_to_be32(s->l1_size),
1391         .l1_table_offset        = cpu_to_be64(s->l1_table_offset),
1392         .refcount_table_offset  = cpu_to_be64(s->refcount_table_offset),
1393         .refcount_table_clusters = cpu_to_be32(refcount_table_clusters),
1394         .nb_snapshots           = cpu_to_be32(s->nb_snapshots),
1395         .snapshots_offset       = cpu_to_be64(s->snapshots_offset),
1396
1397         /* Version 3 fields */
1398         .incompatible_features  = cpu_to_be64(s->incompatible_features),
1399         .compatible_features    = cpu_to_be64(s->compatible_features),
1400         .autoclear_features     = cpu_to_be64(s->autoclear_features),
1401         .refcount_order         = cpu_to_be32(s->refcount_order),
1402         .header_length          = cpu_to_be32(header_length),
1403     };
1404
1405     /* For older versions, write a shorter header */
1406     switch (s->qcow_version) {
1407     case 2:
1408         ret = offsetof(QCowHeader, incompatible_features);
1409         break;
1410     case 3:
1411         ret = sizeof(*header);
1412         break;
1413     default:
1414         ret = -EINVAL;
1415         goto fail;
1416     }
1417
1418     buf += ret;
1419     buflen -= ret;
1420     memset(buf, 0, buflen);
1421
1422     /* Preserve any unknown field in the header */
1423     if (s->unknown_header_fields_size) {
1424         if (buflen < s->unknown_header_fields_size) {
1425             ret = -ENOSPC;
1426             goto fail;
1427         }
1428
1429         memcpy(buf, s->unknown_header_fields, s->unknown_header_fields_size);
1430         buf += s->unknown_header_fields_size;
1431         buflen -= s->unknown_header_fields_size;
1432     }
1433
1434     /* Backing file format header extension */
1435     if (*bs->backing_format) {
1436         ret = header_ext_add(buf, QCOW2_EXT_MAGIC_BACKING_FORMAT,
1437                              bs->backing_format, strlen(bs->backing_format),
1438                              buflen);
1439         if (ret < 0) {
1440             goto fail;
1441         }
1442
1443         buf += ret;
1444         buflen -= ret;
1445     }
1446
1447     /* Feature table */
1448     Qcow2Feature features[] = {
1449         {
1450             .type = QCOW2_FEAT_TYPE_INCOMPATIBLE,
1451             .bit  = QCOW2_INCOMPAT_DIRTY_BITNR,
1452             .name = "dirty bit",
1453         },
1454         {
1455             .type = QCOW2_FEAT_TYPE_INCOMPATIBLE,
1456             .bit  = QCOW2_INCOMPAT_CORRUPT_BITNR,
1457             .name = "corrupt bit",
1458         },
1459         {
1460             .type = QCOW2_FEAT_TYPE_COMPATIBLE,
1461             .bit  = QCOW2_COMPAT_LAZY_REFCOUNTS_BITNR,
1462             .name = "lazy refcounts",
1463         },
1464     };
1465
1466     ret = header_ext_add(buf, QCOW2_EXT_MAGIC_FEATURE_TABLE,
1467                          features, sizeof(features), buflen);
1468     if (ret < 0) {
1469         goto fail;
1470     }
1471     buf += ret;
1472     buflen -= ret;
1473
1474     /* Keep unknown header extensions */
1475     QLIST_FOREACH(uext, &s->unknown_header_ext, next) {
1476         ret = header_ext_add(buf, uext->magic, uext->data, uext->len, buflen);
1477         if (ret < 0) {
1478             goto fail;
1479         }
1480
1481         buf += ret;
1482         buflen -= ret;
1483     }
1484
1485     /* End of header extensions */
1486     ret = header_ext_add(buf, QCOW2_EXT_MAGIC_END, NULL, 0, buflen);
1487     if (ret < 0) {
1488         goto fail;
1489     }
1490
1491     buf += ret;
1492     buflen -= ret;
1493
1494     /* Backing file name */
1495     if (*bs->backing_file) {
1496         size_t backing_file_len = strlen(bs->backing_file);
1497
1498         if (buflen < backing_file_len) {
1499             ret = -ENOSPC;
1500             goto fail;
1501         }
1502
1503         /* Using strncpy is ok here, since buf is not NUL-terminated. */
1504         strncpy(buf, bs->backing_file, buflen);
1505
1506         header->backing_file_offset = cpu_to_be64(buf - ((char*) header));
1507         header->backing_file_size   = cpu_to_be32(backing_file_len);
1508     }
1509
1510     /* Write the new header */
1511     ret = bdrv_pwrite(bs->file, 0, header, s->cluster_size);
1512     if (ret < 0) {
1513         goto fail;
1514     }
1515
1516     ret = 0;
1517 fail:
1518     qemu_vfree(header);
1519     return ret;
1520 }
1521
1522 static int qcow2_change_backing_file(BlockDriverState *bs,
1523     const char *backing_file, const char *backing_fmt)
1524 {
1525     pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: "");
1526     pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: "");
1527
1528     return qcow2_update_header(bs);
1529 }
1530
1531 static int preallocate(BlockDriverState *bs)
1532 {
1533     uint64_t nb_sectors;
1534     uint64_t offset;
1535     uint64_t host_offset = 0;
1536     int num;
1537     int ret;
1538     QCowL2Meta *meta;
1539
1540     nb_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
1541     offset = 0;
1542
1543     while (nb_sectors) {
1544         num = MIN(nb_sectors, INT_MAX >> BDRV_SECTOR_BITS);
1545         ret = qcow2_alloc_cluster_offset(bs, offset, &num,
1546                                          &host_offset, &meta);
1547         if (ret < 0) {
1548             return ret;
1549         }
1550
1551         while (meta) {
1552             QCowL2Meta *next = meta->next;
1553
1554             ret = qcow2_alloc_cluster_link_l2(bs, meta);
1555             if (ret < 0) {
1556                 qcow2_free_any_clusters(bs, meta->alloc_offset,
1557                                         meta->nb_clusters, QCOW2_DISCARD_NEVER);
1558                 return ret;
1559             }
1560
1561             /* There are no dependent requests, but we need to remove our
1562              * request from the list of in-flight requests */
1563             QLIST_REMOVE(meta, next_in_flight);
1564
1565             g_free(meta);
1566             meta = next;
1567         }
1568
1569         /* TODO Preallocate data if requested */
1570
1571         nb_sectors -= num;
1572         offset += num << BDRV_SECTOR_BITS;
1573     }
1574
1575     /*
1576      * It is expected that the image file is large enough to actually contain
1577      * all of the allocated clusters (otherwise we get failing reads after
1578      * EOF). Extend the image to the last allocated sector.
1579      */
1580     if (host_offset != 0) {
1581         uint8_t buf[BDRV_SECTOR_SIZE];
1582         memset(buf, 0, BDRV_SECTOR_SIZE);
1583         ret = bdrv_write(bs->file, (host_offset >> BDRV_SECTOR_BITS) + num - 1,
1584                          buf, 1);
1585         if (ret < 0) {
1586             return ret;
1587         }
1588     }
1589
1590     return 0;
1591 }
1592
1593 static int qcow2_create2(const char *filename, int64_t total_size,
1594                          const char *backing_file, const char *backing_format,
1595                          int flags, size_t cluster_size, int prealloc,
1596                          QEMUOptionParameter *options, int version,
1597                          Error **errp)
1598 {
1599     /* Calculate cluster_bits */
1600     int cluster_bits;
1601     cluster_bits = ffs(cluster_size) - 1;
1602     if (cluster_bits < MIN_CLUSTER_BITS || cluster_bits > MAX_CLUSTER_BITS ||
1603         (1 << cluster_bits) != cluster_size)
1604     {
1605         error_setg(errp, "Cluster size must be a power of two between %d and "
1606                    "%dk", 1 << MIN_CLUSTER_BITS, 1 << (MAX_CLUSTER_BITS - 10));
1607         return -EINVAL;
1608     }
1609
1610     /*
1611      * Open the image file and write a minimal qcow2 header.
1612      *
1613      * We keep things simple and start with a zero-sized image. We also
1614      * do without refcount blocks or a L1 table for now. We'll fix the
1615      * inconsistency later.
1616      *
1617      * We do need a refcount table because growing the refcount table means
1618      * allocating two new refcount blocks - the seconds of which would be at
1619      * 2 GB for 64k clusters, and we don't want to have a 2 GB initial file
1620      * size for any qcow2 image.
1621      */
1622     BlockDriverState* bs;
1623     QCowHeader *header;
1624     uint64_t* refcount_table;
1625     Error *local_err = NULL;
1626     int ret;
1627
1628     ret = bdrv_create_file(filename, options, &local_err);
1629     if (ret < 0) {
1630         error_propagate(errp, local_err);
1631         return ret;
1632     }
1633
1634     bs = NULL;
1635     ret = bdrv_open(&bs, filename, NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1636                     NULL, &local_err);
1637     if (ret < 0) {
1638         error_propagate(errp, local_err);
1639         return ret;
1640     }
1641
1642     /* Write the header */
1643     QEMU_BUILD_BUG_ON((1 << MIN_CLUSTER_BITS) < sizeof(*header));
1644     header = g_malloc0(cluster_size);
1645     *header = (QCowHeader) {
1646         .magic                      = cpu_to_be32(QCOW_MAGIC),
1647         .version                    = cpu_to_be32(version),
1648         .cluster_bits               = cpu_to_be32(cluster_bits),
1649         .size                       = cpu_to_be64(0),
1650         .l1_table_offset            = cpu_to_be64(0),
1651         .l1_size                    = cpu_to_be32(0),
1652         .refcount_table_offset      = cpu_to_be64(cluster_size),
1653         .refcount_table_clusters    = cpu_to_be32(1),
1654         .refcount_order             = cpu_to_be32(3 + REFCOUNT_SHIFT),
1655         .header_length              = cpu_to_be32(sizeof(*header)),
1656     };
1657
1658     if (flags & BLOCK_FLAG_ENCRYPT) {
1659         header->crypt_method = cpu_to_be32(QCOW_CRYPT_AES);
1660     } else {
1661         header->crypt_method = cpu_to_be32(QCOW_CRYPT_NONE);
1662     }
1663
1664     if (flags & BLOCK_FLAG_LAZY_REFCOUNTS) {
1665         header->compatible_features |=
1666             cpu_to_be64(QCOW2_COMPAT_LAZY_REFCOUNTS);
1667     }
1668
1669     ret = bdrv_pwrite(bs, 0, header, cluster_size);
1670     g_free(header);
1671     if (ret < 0) {
1672         error_setg_errno(errp, -ret, "Could not write qcow2 header");
1673         goto out;
1674     }
1675
1676     /* Write a refcount table with one refcount block */
1677     refcount_table = g_malloc0(2 * cluster_size);
1678     refcount_table[0] = cpu_to_be64(2 * cluster_size);
1679     ret = bdrv_pwrite(bs, cluster_size, refcount_table, 2 * cluster_size);
1680     g_free(refcount_table);
1681
1682     if (ret < 0) {
1683         error_setg_errno(errp, -ret, "Could not write refcount table");
1684         goto out;
1685     }
1686
1687     bdrv_unref(bs);
1688     bs = NULL;
1689
1690     /*
1691      * And now open the image and make it consistent first (i.e. increase the
1692      * refcount of the cluster that is occupied by the header and the refcount
1693      * table)
1694      */
1695     BlockDriver* drv = bdrv_find_format("qcow2");
1696     assert(drv != NULL);
1697     ret = bdrv_open(&bs, filename, NULL, NULL,
1698         BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_NO_FLUSH, drv, &local_err);
1699     if (ret < 0) {
1700         error_propagate(errp, local_err);
1701         goto out;
1702     }
1703
1704     ret = qcow2_alloc_clusters(bs, 3 * cluster_size);
1705     if (ret < 0) {
1706         error_setg_errno(errp, -ret, "Could not allocate clusters for qcow2 "
1707                          "header and refcount table");
1708         goto out;
1709
1710     } else if (ret != 0) {
1711         error_report("Huh, first cluster in empty image is already in use?");
1712         abort();
1713     }
1714
1715     /* Okay, now that we have a valid image, let's give it the right size */
1716     ret = bdrv_truncate(bs, total_size * BDRV_SECTOR_SIZE);
1717     if (ret < 0) {
1718         error_setg_errno(errp, -ret, "Could not resize image");
1719         goto out;
1720     }
1721
1722     /* Want a backing file? There you go.*/
1723     if (backing_file) {
1724         ret = bdrv_change_backing_file(bs, backing_file, backing_format);
1725         if (ret < 0) {
1726             error_setg_errno(errp, -ret, "Could not assign backing file '%s' "
1727                              "with format '%s'", backing_file, backing_format);
1728             goto out;
1729         }
1730     }
1731
1732     /* And if we're supposed to preallocate metadata, do that now */
1733     if (prealloc) {
1734         BDRVQcowState *s = bs->opaque;
1735         qemu_co_mutex_lock(&s->lock);
1736         ret = preallocate(bs);
1737         qemu_co_mutex_unlock(&s->lock);
1738         if (ret < 0) {
1739             error_setg_errno(errp, -ret, "Could not preallocate metadata");
1740             goto out;
1741         }
1742     }
1743
1744     bdrv_unref(bs);
1745     bs = NULL;
1746
1747     /* Reopen the image without BDRV_O_NO_FLUSH to flush it before returning */
1748     ret = bdrv_open(&bs, filename, NULL, NULL,
1749                     BDRV_O_RDWR | BDRV_O_CACHE_WB | BDRV_O_NO_BACKING,
1750                     drv, &local_err);
1751     if (local_err) {
1752         error_propagate(errp, local_err);
1753         goto out;
1754     }
1755
1756     ret = 0;
1757 out:
1758     if (bs) {
1759         bdrv_unref(bs);
1760     }
1761     return ret;
1762 }
1763
1764 static int qcow2_create(const char *filename, QEMUOptionParameter *options,
1765                         Error **errp)
1766 {
1767     const char *backing_file = NULL;
1768     const char *backing_fmt = NULL;
1769     uint64_t sectors = 0;
1770     int flags = 0;
1771     size_t cluster_size = DEFAULT_CLUSTER_SIZE;
1772     int prealloc = 0;
1773     int version = 3;
1774     Error *local_err = NULL;
1775     int ret;
1776
1777     /* Read out options */
1778     while (options && options->name) {
1779         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
1780             sectors = options->value.n / 512;
1781         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
1782             backing_file = options->value.s;
1783         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FMT)) {
1784             backing_fmt = options->value.s;
1785         } else if (!strcmp(options->name, BLOCK_OPT_ENCRYPT)) {
1786             flags |= options->value.n ? BLOCK_FLAG_ENCRYPT : 0;
1787         } else if (!strcmp(options->name, BLOCK_OPT_CLUSTER_SIZE)) {
1788             if (options->value.n) {
1789                 cluster_size = options->value.n;
1790             }
1791         } else if (!strcmp(options->name, BLOCK_OPT_PREALLOC)) {
1792             if (!options->value.s || !strcmp(options->value.s, "off")) {
1793                 prealloc = 0;
1794             } else if (!strcmp(options->value.s, "metadata")) {
1795                 prealloc = 1;
1796             } else {
1797                 error_setg(errp, "Invalid preallocation mode: '%s'",
1798                            options->value.s);
1799                 return -EINVAL;
1800             }
1801         } else if (!strcmp(options->name, BLOCK_OPT_COMPAT_LEVEL)) {
1802             if (!options->value.s) {
1803                 /* keep the default */
1804             } else if (!strcmp(options->value.s, "0.10")) {
1805                 version = 2;
1806             } else if (!strcmp(options->value.s, "1.1")) {
1807                 version = 3;
1808             } else {
1809                 error_setg(errp, "Invalid compatibility level: '%s'",
1810                            options->value.s);
1811                 return -EINVAL;
1812             }
1813         } else if (!strcmp(options->name, BLOCK_OPT_LAZY_REFCOUNTS)) {
1814             flags |= options->value.n ? BLOCK_FLAG_LAZY_REFCOUNTS : 0;
1815         }
1816         options++;
1817     }
1818
1819     if (backing_file && prealloc) {
1820         error_setg(errp, "Backing file and preallocation cannot be used at "
1821                    "the same time");
1822         return -EINVAL;
1823     }
1824
1825     if (version < 3 && (flags & BLOCK_FLAG_LAZY_REFCOUNTS)) {
1826         error_setg(errp, "Lazy refcounts only supported with compatibility "
1827                    "level 1.1 and above (use compat=1.1 or greater)");
1828         return -EINVAL;
1829     }
1830
1831     ret = qcow2_create2(filename, sectors, backing_file, backing_fmt, flags,
1832                         cluster_size, prealloc, options, version, &local_err);
1833     if (local_err) {
1834         error_propagate(errp, local_err);
1835     }
1836     return ret;
1837 }
1838
1839 static coroutine_fn int qcow2_co_write_zeroes(BlockDriverState *bs,
1840     int64_t sector_num, int nb_sectors, BdrvRequestFlags flags)
1841 {
1842     int ret;
1843     BDRVQcowState *s = bs->opaque;
1844
1845     /* Emulate misaligned zero writes */
1846     if (sector_num % s->cluster_sectors || nb_sectors % s->cluster_sectors) {
1847         return -ENOTSUP;
1848     }
1849
1850     /* Whatever is left can use real zero clusters */
1851     qemu_co_mutex_lock(&s->lock);
1852     ret = qcow2_zero_clusters(bs, sector_num << BDRV_SECTOR_BITS,
1853         nb_sectors);
1854     qemu_co_mutex_unlock(&s->lock);
1855
1856     return ret;
1857 }
1858
1859 static coroutine_fn int qcow2_co_discard(BlockDriverState *bs,
1860     int64_t sector_num, int nb_sectors)
1861 {
1862     int ret;
1863     BDRVQcowState *s = bs->opaque;
1864
1865     qemu_co_mutex_lock(&s->lock);
1866     ret = qcow2_discard_clusters(bs, sector_num << BDRV_SECTOR_BITS,
1867         nb_sectors, QCOW2_DISCARD_REQUEST);
1868     qemu_co_mutex_unlock(&s->lock);
1869     return ret;
1870 }
1871
1872 static int qcow2_truncate(BlockDriverState *bs, int64_t offset)
1873 {
1874     BDRVQcowState *s = bs->opaque;
1875     int64_t new_l1_size;
1876     int ret;
1877
1878     if (offset & 511) {
1879         error_report("The new size must be a multiple of 512");
1880         return -EINVAL;
1881     }
1882
1883     /* cannot proceed if image has snapshots */
1884     if (s->nb_snapshots) {
1885         error_report("Can't resize an image which has snapshots");
1886         return -ENOTSUP;
1887     }
1888
1889     /* shrinking is currently not supported */
1890     if (offset < bs->total_sectors * 512) {
1891         error_report("qcow2 doesn't support shrinking images yet");
1892         return -ENOTSUP;
1893     }
1894
1895     new_l1_size = size_to_l1(s, offset);
1896     ret = qcow2_grow_l1_table(bs, new_l1_size, true);
1897     if (ret < 0) {
1898         return ret;
1899     }
1900
1901     /* write updated header.size */
1902     offset = cpu_to_be64(offset);
1903     ret = bdrv_pwrite_sync(bs->file, offsetof(QCowHeader, size),
1904                            &offset, sizeof(uint64_t));
1905     if (ret < 0) {
1906         return ret;
1907     }
1908
1909     s->l1_vm_state_index = new_l1_size;
1910     return 0;
1911 }
1912
1913 /* XXX: put compressed sectors first, then all the cluster aligned
1914    tables to avoid losing bytes in alignment */
1915 static int qcow2_write_compressed(BlockDriverState *bs, int64_t sector_num,
1916                                   const uint8_t *buf, int nb_sectors)
1917 {
1918     BDRVQcowState *s = bs->opaque;
1919     z_stream strm;
1920     int ret, out_len;
1921     uint8_t *out_buf;
1922     uint64_t cluster_offset;
1923
1924     if (nb_sectors == 0) {
1925         /* align end of file to a sector boundary to ease reading with
1926            sector based I/Os */
1927         cluster_offset = bdrv_getlength(bs->file);
1928         cluster_offset = (cluster_offset + 511) & ~511;
1929         bdrv_truncate(bs->file, cluster_offset);
1930         return 0;
1931     }
1932
1933     if (nb_sectors != s->cluster_sectors) {
1934         ret = -EINVAL;
1935
1936         /* Zero-pad last write if image size is not cluster aligned */
1937         if (sector_num + nb_sectors == bs->total_sectors &&
1938             nb_sectors < s->cluster_sectors) {
1939             uint8_t *pad_buf = qemu_blockalign(bs, s->cluster_size);
1940             memset(pad_buf, 0, s->cluster_size);
1941             memcpy(pad_buf, buf, nb_sectors * BDRV_SECTOR_SIZE);
1942             ret = qcow2_write_compressed(bs, sector_num,
1943                                          pad_buf, s->cluster_sectors);
1944             qemu_vfree(pad_buf);
1945         }
1946         return ret;
1947     }
1948
1949     out_buf = g_malloc(s->cluster_size + (s->cluster_size / 1000) + 128);
1950
1951     /* best compression, small window, no zlib header */
1952     memset(&strm, 0, sizeof(strm));
1953     ret = deflateInit2(&strm, Z_DEFAULT_COMPRESSION,
1954                        Z_DEFLATED, -12,
1955                        9, Z_DEFAULT_STRATEGY);
1956     if (ret != 0) {
1957         ret = -EINVAL;
1958         goto fail;
1959     }
1960
1961     strm.avail_in = s->cluster_size;
1962     strm.next_in = (uint8_t *)buf;
1963     strm.avail_out = s->cluster_size;
1964     strm.next_out = out_buf;
1965
1966     ret = deflate(&strm, Z_FINISH);
1967     if (ret != Z_STREAM_END && ret != Z_OK) {
1968         deflateEnd(&strm);
1969         ret = -EINVAL;
1970         goto fail;
1971     }
1972     out_len = strm.next_out - out_buf;
1973
1974     deflateEnd(&strm);
1975
1976     if (ret != Z_STREAM_END || out_len >= s->cluster_size) {
1977         /* could not compress: write normal cluster */
1978         ret = bdrv_write(bs, sector_num, buf, s->cluster_sectors);
1979         if (ret < 0) {
1980             goto fail;
1981         }
1982     } else {
1983         cluster_offset = qcow2_alloc_compressed_cluster_offset(bs,
1984             sector_num << 9, out_len);
1985         if (!cluster_offset) {
1986             ret = -EIO;
1987             goto fail;
1988         }
1989         cluster_offset &= s->cluster_offset_mask;
1990
1991         ret = qcow2_pre_write_overlap_check(bs, 0, cluster_offset, out_len);
1992         if (ret < 0) {
1993             goto fail;
1994         }
1995
1996         BLKDBG_EVENT(bs->file, BLKDBG_WRITE_COMPRESSED);
1997         ret = bdrv_pwrite(bs->file, cluster_offset, out_buf, out_len);
1998         if (ret < 0) {
1999             goto fail;
2000         }
2001     }
2002
2003     ret = 0;
2004 fail:
2005     g_free(out_buf);
2006     return ret;
2007 }
2008
2009 static coroutine_fn int qcow2_co_flush_to_os(BlockDriverState *bs)
2010 {
2011     BDRVQcowState *s = bs->opaque;
2012     int ret;
2013
2014     qemu_co_mutex_lock(&s->lock);
2015     ret = qcow2_cache_flush(bs, s->l2_table_cache);
2016     if (ret < 0) {
2017         qemu_co_mutex_unlock(&s->lock);
2018         return ret;
2019     }
2020
2021     if (qcow2_need_accurate_refcounts(s)) {
2022         ret = qcow2_cache_flush(bs, s->refcount_block_cache);
2023         if (ret < 0) {
2024             qemu_co_mutex_unlock(&s->lock);
2025             return ret;
2026         }
2027     }
2028     qemu_co_mutex_unlock(&s->lock);
2029
2030     return 0;
2031 }
2032
2033 static int qcow2_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2034 {
2035     BDRVQcowState *s = bs->opaque;
2036     bdi->unallocated_blocks_are_zero = true;
2037     bdi->can_write_zeroes_with_unmap = (s->qcow_version >= 3);
2038     bdi->cluster_size = s->cluster_size;
2039     bdi->vm_state_offset = qcow2_vm_state_offset(s);
2040     return 0;
2041 }
2042
2043 static ImageInfoSpecific *qcow2_get_specific_info(BlockDriverState *bs)
2044 {
2045     BDRVQcowState *s = bs->opaque;
2046     ImageInfoSpecific *spec_info = g_new(ImageInfoSpecific, 1);
2047
2048     *spec_info = (ImageInfoSpecific){
2049         .kind  = IMAGE_INFO_SPECIFIC_KIND_QCOW2,
2050         {
2051             .qcow2 = g_new(ImageInfoSpecificQCow2, 1),
2052         },
2053     };
2054     if (s->qcow_version == 2) {
2055         *spec_info->qcow2 = (ImageInfoSpecificQCow2){
2056             .compat = g_strdup("0.10"),
2057         };
2058     } else if (s->qcow_version == 3) {
2059         *spec_info->qcow2 = (ImageInfoSpecificQCow2){
2060             .compat             = g_strdup("1.1"),
2061             .lazy_refcounts     = s->compatible_features &
2062                                   QCOW2_COMPAT_LAZY_REFCOUNTS,
2063             .has_lazy_refcounts = true,
2064         };
2065     }
2066
2067     return spec_info;
2068 }
2069
2070 #if 0
2071 static void dump_refcounts(BlockDriverState *bs)
2072 {
2073     BDRVQcowState *s = bs->opaque;
2074     int64_t nb_clusters, k, k1, size;
2075     int refcount;
2076
2077     size = bdrv_getlength(bs->file);
2078     nb_clusters = size_to_clusters(s, size);
2079     for(k = 0; k < nb_clusters;) {
2080         k1 = k;
2081         refcount = get_refcount(bs, k);
2082         k++;
2083         while (k < nb_clusters && get_refcount(bs, k) == refcount)
2084             k++;
2085         printf("%" PRId64 ": refcount=%d nb=%" PRId64 "\n", k, refcount,
2086                k - k1);
2087     }
2088 }
2089 #endif
2090
2091 static int qcow2_save_vmstate(BlockDriverState *bs, QEMUIOVector *qiov,
2092                               int64_t pos)
2093 {
2094     BDRVQcowState *s = bs->opaque;
2095     int64_t total_sectors = bs->total_sectors;
2096     int growable = bs->growable;
2097     bool zero_beyond_eof = bs->zero_beyond_eof;
2098     int ret;
2099
2100     BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_SAVE);
2101     bs->growable = 1;
2102     bs->zero_beyond_eof = false;
2103     ret = bdrv_pwritev(bs, qcow2_vm_state_offset(s) + pos, qiov);
2104     bs->growable = growable;
2105     bs->zero_beyond_eof = zero_beyond_eof;
2106
2107     /* bdrv_co_do_writev will have increased the total_sectors value to include
2108      * the VM state - the VM state is however not an actual part of the block
2109      * device, therefore, we need to restore the old value. */
2110     bs->total_sectors = total_sectors;
2111
2112     return ret;
2113 }
2114
2115 static int qcow2_load_vmstate(BlockDriverState *bs, uint8_t *buf,
2116                               int64_t pos, int size)
2117 {
2118     BDRVQcowState *s = bs->opaque;
2119     int growable = bs->growable;
2120     bool zero_beyond_eof = bs->zero_beyond_eof;
2121     int ret;
2122
2123     BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_LOAD);
2124     bs->growable = 1;
2125     bs->zero_beyond_eof = false;
2126     ret = bdrv_pread(bs, qcow2_vm_state_offset(s) + pos, buf, size);
2127     bs->growable = growable;
2128     bs->zero_beyond_eof = zero_beyond_eof;
2129
2130     return ret;
2131 }
2132
2133 /*
2134  * Downgrades an image's version. To achieve this, any incompatible features
2135  * have to be removed.
2136  */
2137 static int qcow2_downgrade(BlockDriverState *bs, int target_version)
2138 {
2139     BDRVQcowState *s = bs->opaque;
2140     int current_version = s->qcow_version;
2141     int ret;
2142
2143     if (target_version == current_version) {
2144         return 0;
2145     } else if (target_version > current_version) {
2146         return -EINVAL;
2147     } else if (target_version != 2) {
2148         return -EINVAL;
2149     }
2150
2151     if (s->refcount_order != 4) {
2152         /* we would have to convert the image to a refcount_order == 4 image
2153          * here; however, since qemu (at the time of writing this) does not
2154          * support anything different than 4 anyway, there is no point in doing
2155          * so right now; however, we should error out (if qemu supports this in
2156          * the future and this code has not been adapted) */
2157         error_report("qcow2_downgrade: Image refcount orders other than 4 are "
2158                      "currently not supported.");
2159         return -ENOTSUP;
2160     }
2161
2162     /* clear incompatible features */
2163     if (s->incompatible_features & QCOW2_INCOMPAT_DIRTY) {
2164         ret = qcow2_mark_clean(bs);
2165         if (ret < 0) {
2166             return ret;
2167         }
2168     }
2169
2170     /* with QCOW2_INCOMPAT_CORRUPT, it is pretty much impossible to get here in
2171      * the first place; if that happens nonetheless, returning -ENOTSUP is the
2172      * best thing to do anyway */
2173
2174     if (s->incompatible_features) {
2175         return -ENOTSUP;
2176     }
2177
2178     /* since we can ignore compatible features, we can set them to 0 as well */
2179     s->compatible_features = 0;
2180     /* if lazy refcounts have been used, they have already been fixed through
2181      * clearing the dirty flag */
2182
2183     /* clearing autoclear features is trivial */
2184     s->autoclear_features = 0;
2185
2186     ret = qcow2_expand_zero_clusters(bs);
2187     if (ret < 0) {
2188         return ret;
2189     }
2190
2191     s->qcow_version = target_version;
2192     ret = qcow2_update_header(bs);
2193     if (ret < 0) {
2194         s->qcow_version = current_version;
2195         return ret;
2196     }
2197     return 0;
2198 }
2199
2200 static int qcow2_amend_options(BlockDriverState *bs,
2201                                QEMUOptionParameter *options)
2202 {
2203     BDRVQcowState *s = bs->opaque;
2204     int old_version = s->qcow_version, new_version = old_version;
2205     uint64_t new_size = 0;
2206     const char *backing_file = NULL, *backing_format = NULL;
2207     bool lazy_refcounts = s->use_lazy_refcounts;
2208     int ret;
2209     int i;
2210
2211     for (i = 0; options[i].name; i++)
2212     {
2213         if (!options[i].assigned) {
2214             /* only change explicitly defined options */
2215             continue;
2216         }
2217
2218         if (!strcmp(options[i].name, "compat")) {
2219             if (!options[i].value.s) {
2220                 /* preserve default */
2221             } else if (!strcmp(options[i].value.s, "0.10")) {
2222                 new_version = 2;
2223             } else if (!strcmp(options[i].value.s, "1.1")) {
2224                 new_version = 3;
2225             } else {
2226                 fprintf(stderr, "Unknown compatibility level %s.\n",
2227                         options[i].value.s);
2228                 return -EINVAL;
2229             }
2230         } else if (!strcmp(options[i].name, "preallocation")) {
2231             fprintf(stderr, "Cannot change preallocation mode.\n");
2232             return -ENOTSUP;
2233         } else if (!strcmp(options[i].name, "size")) {
2234             new_size = options[i].value.n;
2235         } else if (!strcmp(options[i].name, "backing_file")) {
2236             backing_file = options[i].value.s;
2237         } else if (!strcmp(options[i].name, "backing_fmt")) {
2238             backing_format = options[i].value.s;
2239         } else if (!strcmp(options[i].name, "encryption")) {
2240             if ((options[i].value.n != !!s->crypt_method)) {
2241                 fprintf(stderr, "Changing the encryption flag is not "
2242                         "supported.\n");
2243                 return -ENOTSUP;
2244             }
2245         } else if (!strcmp(options[i].name, "cluster_size")) {
2246             if (options[i].value.n != s->cluster_size) {
2247                 fprintf(stderr, "Changing the cluster size is not "
2248                         "supported.\n");
2249                 return -ENOTSUP;
2250             }
2251         } else if (!strcmp(options[i].name, "lazy_refcounts")) {
2252             lazy_refcounts = options[i].value.n;
2253         } else {
2254             /* if this assertion fails, this probably means a new option was
2255              * added without having it covered here */
2256             assert(false);
2257         }
2258     }
2259
2260     if (new_version != old_version) {
2261         if (new_version > old_version) {
2262             /* Upgrade */
2263             s->qcow_version = new_version;
2264             ret = qcow2_update_header(bs);
2265             if (ret < 0) {
2266                 s->qcow_version = old_version;
2267                 return ret;
2268             }
2269         } else {
2270             ret = qcow2_downgrade(bs, new_version);
2271             if (ret < 0) {
2272                 return ret;
2273             }
2274         }
2275     }
2276
2277     if (backing_file || backing_format) {
2278         ret = qcow2_change_backing_file(bs, backing_file ?: bs->backing_file,
2279                                         backing_format ?: bs->backing_format);
2280         if (ret < 0) {
2281             return ret;
2282         }
2283     }
2284
2285     if (s->use_lazy_refcounts != lazy_refcounts) {
2286         if (lazy_refcounts) {
2287             if (s->qcow_version < 3) {
2288                 fprintf(stderr, "Lazy refcounts only supported with compatibility "
2289                         "level 1.1 and above (use compat=1.1 or greater)\n");
2290                 return -EINVAL;
2291             }
2292             s->compatible_features |= QCOW2_COMPAT_LAZY_REFCOUNTS;
2293             ret = qcow2_update_header(bs);
2294             if (ret < 0) {
2295                 s->compatible_features &= ~QCOW2_COMPAT_LAZY_REFCOUNTS;
2296                 return ret;
2297             }
2298             s->use_lazy_refcounts = true;
2299         } else {
2300             /* make image clean first */
2301             ret = qcow2_mark_clean(bs);
2302             if (ret < 0) {
2303                 return ret;
2304             }
2305             /* now disallow lazy refcounts */
2306             s->compatible_features &= ~QCOW2_COMPAT_LAZY_REFCOUNTS;
2307             ret = qcow2_update_header(bs);
2308             if (ret < 0) {
2309                 s->compatible_features |= QCOW2_COMPAT_LAZY_REFCOUNTS;
2310                 return ret;
2311             }
2312             s->use_lazy_refcounts = false;
2313         }
2314     }
2315
2316     if (new_size) {
2317         ret = bdrv_truncate(bs, new_size);
2318         if (ret < 0) {
2319             return ret;
2320         }
2321     }
2322
2323     return 0;
2324 }
2325
2326 static QEMUOptionParameter qcow2_create_options[] = {
2327     {
2328         .name = BLOCK_OPT_SIZE,
2329         .type = OPT_SIZE,
2330         .help = "Virtual disk size"
2331     },
2332     {
2333         .name = BLOCK_OPT_COMPAT_LEVEL,
2334         .type = OPT_STRING,
2335         .help = "Compatibility level (0.10 or 1.1)"
2336     },
2337     {
2338         .name = BLOCK_OPT_BACKING_FILE,
2339         .type = OPT_STRING,
2340         .help = "File name of a base image"
2341     },
2342     {
2343         .name = BLOCK_OPT_BACKING_FMT,
2344         .type = OPT_STRING,
2345         .help = "Image format of the base image"
2346     },
2347     {
2348         .name = BLOCK_OPT_ENCRYPT,
2349         .type = OPT_FLAG,
2350         .help = "Encrypt the image"
2351     },
2352     {
2353         .name = BLOCK_OPT_CLUSTER_SIZE,
2354         .type = OPT_SIZE,
2355         .help = "qcow2 cluster size",
2356         .value = { .n = DEFAULT_CLUSTER_SIZE },
2357     },
2358     {
2359         .name = BLOCK_OPT_PREALLOC,
2360         .type = OPT_STRING,
2361         .help = "Preallocation mode (allowed values: off, metadata)"
2362     },
2363     {
2364         .name = BLOCK_OPT_LAZY_REFCOUNTS,
2365         .type = OPT_FLAG,
2366         .help = "Postpone refcount updates",
2367     },
2368     { NULL }
2369 };
2370
2371 static BlockDriver bdrv_qcow2 = {
2372     .format_name        = "qcow2",
2373     .instance_size      = sizeof(BDRVQcowState),
2374     .bdrv_probe         = qcow2_probe,
2375     .bdrv_open          = qcow2_open,
2376     .bdrv_close         = qcow2_close,
2377     .bdrv_reopen_prepare  = qcow2_reopen_prepare,
2378     .bdrv_create        = qcow2_create,
2379     .bdrv_has_zero_init = bdrv_has_zero_init_1,
2380     .bdrv_co_get_block_status = qcow2_co_get_block_status,
2381     .bdrv_set_key       = qcow2_set_key,
2382
2383     .bdrv_co_readv          = qcow2_co_readv,
2384     .bdrv_co_writev         = qcow2_co_writev,
2385     .bdrv_co_flush_to_os    = qcow2_co_flush_to_os,
2386
2387     .bdrv_co_write_zeroes   = qcow2_co_write_zeroes,
2388     .bdrv_co_discard        = qcow2_co_discard,
2389     .bdrv_truncate          = qcow2_truncate,
2390     .bdrv_write_compressed  = qcow2_write_compressed,
2391
2392     .bdrv_snapshot_create   = qcow2_snapshot_create,
2393     .bdrv_snapshot_goto     = qcow2_snapshot_goto,
2394     .bdrv_snapshot_delete   = qcow2_snapshot_delete,
2395     .bdrv_snapshot_list     = qcow2_snapshot_list,
2396     .bdrv_snapshot_load_tmp     = qcow2_snapshot_load_tmp,
2397     .bdrv_get_info      = qcow2_get_info,
2398     .bdrv_get_specific_info = qcow2_get_specific_info,
2399
2400     .bdrv_save_vmstate    = qcow2_save_vmstate,
2401     .bdrv_load_vmstate    = qcow2_load_vmstate,
2402
2403     .bdrv_change_backing_file   = qcow2_change_backing_file,
2404
2405     .bdrv_refresh_limits        = qcow2_refresh_limits,
2406     .bdrv_invalidate_cache      = qcow2_invalidate_cache,
2407
2408     .create_options = qcow2_create_options,
2409     .bdrv_check = qcow2_check,
2410     .bdrv_amend_options = qcow2_amend_options,
2411 };
2412
2413 static void bdrv_qcow2_init(void)
2414 {
2415     bdrv_register(&bdrv_qcow2);
2416 }
2417
2418 block_init(bdrv_qcow2_init);