block/vmdk.c

   1 /*
   2  * Block driver for the VMDK format
   3  *
   4  * Copyright (c) 2004 Fabrice Bellard
   5  * Copyright (c) 2005 Filip Navara
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "qemu-common.h"
  27 #include "block_int.h"
  28 #include "module.h"
  29 #include "zlib.h"
  30
  31 #define VMDK3_MAGIC (('C' << 24) | ('O' << 16) | ('W' << 8) | 'D')
  32 #define VMDK4_MAGIC (('K' << 24) | ('D' << 16) | ('M' << 8) | 'V')
  33 #define VMDK4_COMPRESSION_DEFLATE 1
  34 #define VMDK4_FLAG_RGD (1 << 1)
  35 #define VMDK4_FLAG_COMPRESS (1 << 16)
  36 #define VMDK4_FLAG_MARKER (1 << 17)
  37
  38 typedef struct {
  39     uint32_t version;
  40     uint32_t flags;
  41     uint32_t disk_sectors;
  42     uint32_t granularity;
  43     uint32_t l1dir_offset;
  44     uint32_t l1dir_size;
  45     uint32_t file_sectors;
  46     uint32_t cylinders;
  47     uint32_t heads;
  48     uint32_t sectors_per_track;
  49 } VMDK3Header;
  50
  51 typedef struct {
  52     uint32_t version;
  53     uint32_t flags;
  54     int64_t capacity;
  55     int64_t granularity;
  56     int64_t desc_offset;
  57     int64_t desc_size;
  58     int32_t num_gtes_per_gte;
  59     int64_t gd_offset;
  60     int64_t rgd_offset;
  61     int64_t grain_offset;
  62     char filler[1];
  63     char check_bytes[4];
  64     uint16_t compressAlgorithm;
  65 } QEMU_PACKED VMDK4Header;
  66
  67 #define L2_CACHE_SIZE 16
  68
  69 typedef struct VmdkExtent {
  70     BlockDriverState *file;
  71     bool flat;
  72     bool compressed;
  73     bool has_marker;
  74     int64_t sectors;
  75     int64_t end_sector;
  76     int64_t flat_start_offset;
  77     int64_t l1_table_offset;
  78     int64_t l1_backup_table_offset;
  79     uint32_t *l1_table;
  80     uint32_t *l1_backup_table;
  81     unsigned int l1_size;
  82     uint32_t l1_entry_sectors;
  83
  84     unsigned int l2_size;
  85     uint32_t *l2_cache;
  86     uint32_t l2_cache_offsets[L2_CACHE_SIZE];
  87     uint32_t l2_cache_counts[L2_CACHE_SIZE];
  88
  89     unsigned int cluster_sectors;
  90 } VmdkExtent;
  91
  92 typedef struct BDRVVmdkState {
  93     int desc_offset;
  94     bool cid_updated;
  95     uint32_t parent_cid;
  96     int num_extents;
  97     /* Extent array with num_extents entries, ascend ordered by address */
  98     VmdkExtent *extents;
  99 } BDRVVmdkState;
 100
 101 typedef struct VmdkMetaData {
 102     uint32_t offset;
 103     unsigned int l1_index;
 104     unsigned int l2_index;
 105     unsigned int l2_offset;
 106     int valid;
 107 } VmdkMetaData;
 108
 109 typedef struct VmdkGrainMarker {
 110     uint64_t lba;
 111     uint32_t size;
 112     uint8_t  data[0];
 113 } VmdkGrainMarker;
 114
 115 static int vmdk_probe(const uint8_t *buf, int buf_size, const char *filename)
 116 {
 117     uint32_t magic;
 118
 119     if (buf_size < 4) {
 120         return 0;
 121     }
 122     magic = be32_to_cpu(*(uint32_t *)buf);
 123     if (magic == VMDK3_MAGIC ||
 124         magic == VMDK4_MAGIC) {
 125         return 100;
 126     } else {
 127         const char *p = (const char *)buf;
 128         const char *end = p + buf_size;
 129         while (p < end) {
 130             if (*p == '#') {
 131                 /* skip comment line */
 132                 while (p < end && *p != '\n') {
 133                     p++;
 134                 }
 135                 p++;
 136                 continue;
 137             }
 138             if (*p == ' ') {
 139                 while (p < end && *p == ' ') {
 140                     p++;
 141                 }
 142                 /* skip '\r' if windows line endings used. */
 143                 if (p < end && *p == '\r') {
 144                     p++;
 145                 }
 146                 /* only accept blank lines before 'version=' line */
 147                 if (p == end || *p != '\n') {
 148                     return 0;
 149                 }
 150                 p++;
 151                 continue;
 152             }
 153             if (end - p >= strlen("version=X\n")) {
 154                 if (strncmp("version=1\n", p, strlen("version=1\n")) == 0 ||
 155                     strncmp("version=2\n", p, strlen("version=2\n")) == 0) {
 156                     return 100;
 157                 }
 158             }
 159             if (end - p >= strlen("version=X\r\n")) {
 160                 if (strncmp("version=1\r\n", p, strlen("version=1\r\n")) == 0 ||
 161                     strncmp("version=2\r\n", p, strlen("version=2\r\n")) == 0) {
 162                     return 100;
 163                 }
 164             }
 165             return 0;
 166         }
 167         return 0;
 168     }
 169 }
 170
 171 #define CHECK_CID 1
 172
 173 #define SECTOR_SIZE 512
 174 #define DESC_SIZE (20 * SECTOR_SIZE)    /* 20 sectors of 512 bytes each */
 175 #define BUF_SIZE 4096
 176 #define HEADER_SIZE 512                 /* first sector of 512 bytes */
 177
 178 static void vmdk_free_extents(BlockDriverState *bs)
 179 {
 180     int i;
 181     BDRVVmdkState *s = bs->opaque;
 182     VmdkExtent *e;
 183
 184     for (i = 0; i < s->num_extents; i++) {
 185         e = &s->extents[i];
 186         g_free(e->l1_table);
 187         g_free(e->l2_cache);
 188         g_free(e->l1_backup_table);
 189         if (e->file != bs->file) {
 190             bdrv_delete(e->file);
 191         }
 192     }
 193     g_free(s->extents);
 194 }
 195
 196 static void vmdk_free_last_extent(BlockDriverState *bs)
 197 {
 198     BDRVVmdkState *s = bs->opaque;
 199
 200     if (s->num_extents == 0) {
 201         return;
 202     }
 203     s->num_extents--;
 204     s->extents = g_realloc(s->extents, s->num_extents * sizeof(VmdkExtent));
 205 }
 206
 207 static uint32_t vmdk_read_cid(BlockDriverState *bs, int parent)
 208 {
 209     char desc[DESC_SIZE];
 210     uint32_t cid;
 211     const char *p_name, *cid_str;
 212     size_t cid_str_size;
 213     BDRVVmdkState *s = bs->opaque;
 214
 215     if (bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE) != DESC_SIZE) {
 216         return 0;
 217     }
 218
 219     if (parent) {
 220         cid_str = "parentCID";
 221         cid_str_size = sizeof("parentCID");
 222     } else {
 223         cid_str = "CID";
 224         cid_str_size = sizeof("CID");
 225     }
 226
 227     p_name = strstr(desc, cid_str);
 228     if (p_name != NULL) {
 229         p_name += cid_str_size;
 230         sscanf(p_name, "%x", &cid);
 231     }
 232
 233     return cid;
 234 }
 235
 236 static int vmdk_write_cid(BlockDriverState *bs, uint32_t cid)
 237 {
 238     char desc[DESC_SIZE], tmp_desc[DESC_SIZE];
 239     char *p_name, *tmp_str;
 240     BDRVVmdkState *s = bs->opaque;
 241
 242     memset(desc, 0, sizeof(desc));
 243     if (bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE) != DESC_SIZE) {
 244         return -EIO;
 245     }
 246
 247     tmp_str = strstr(desc, "parentCID");
 248     pstrcpy(tmp_desc, sizeof(tmp_desc), tmp_str);
 249     p_name = strstr(desc, "CID");
 250     if (p_name != NULL) {
 251         p_name += sizeof("CID");
 252         snprintf(p_name, sizeof(desc) - (p_name - desc), "%x\n", cid);
 253         pstrcat(desc, sizeof(desc), tmp_desc);
 254     }
 255
 256     if (bdrv_pwrite_sync(bs->file, s->desc_offset, desc, DESC_SIZE) < 0) {
 257         return -EIO;
 258     }
 259     return 0;
 260 }
 261
 262 static int vmdk_is_cid_valid(BlockDriverState *bs)
 263 {
 264 #ifdef CHECK_CID
 265     BDRVVmdkState *s = bs->opaque;
 266     BlockDriverState *p_bs = bs->backing_hd;
 267     uint32_t cur_pcid;
 268
 269     if (p_bs) {
 270         cur_pcid = vmdk_read_cid(p_bs, 0);
 271         if (s->parent_cid != cur_pcid) {
 272             /* CID not valid */
 273             return 0;
 274         }
 275     }
 276 #endif
 277     /* CID valid */
 278     return 1;
 279 }
 280
 281 static int vmdk_parent_open(BlockDriverState *bs)
 282 {
 283     char *p_name;
 284     char desc[DESC_SIZE + 1];
 285     BDRVVmdkState *s = bs->opaque;
 286
 287     desc[DESC_SIZE] = '\0';
 288     if (bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE) != DESC_SIZE) {
 289         return -1;
 290     }
 291
 292     p_name = strstr(desc, "parentFileNameHint");
 293     if (p_name != NULL) {
 294         char *end_name;
 295
 296         p_name += sizeof("parentFileNameHint") + 1;
 297         end_name = strchr(p_name, '\"');
 298         if (end_name == NULL) {
 299             return -1;
 300         }
 301         if ((end_name - p_name) > sizeof(bs->backing_file) - 1) {
 302             return -1;
 303         }
 304
 305         pstrcpy(bs->backing_file, end_name - p_name + 1, p_name);
 306     }
 307
 308     return 0;
 309 }
 310
 311 /* Create and append extent to the extent array. Return the added VmdkExtent
 312  * address. return NULL if allocation failed. */
 313 static VmdkExtent *vmdk_add_extent(BlockDriverState *bs,
 314                            BlockDriverState *file, bool flat, int64_t sectors,
 315                            int64_t l1_offset, int64_t l1_backup_offset,
 316                            uint32_t l1_size,
 317                            int l2_size, unsigned int cluster_sectors)
 318 {
 319     VmdkExtent *extent;
 320     BDRVVmdkState *s = bs->opaque;
 321
 322     s->extents = g_realloc(s->extents,
 323                               (s->num_extents + 1) * sizeof(VmdkExtent));
 324     extent = &s->extents[s->num_extents];
 325     s->num_extents++;
 326
 327     memset(extent, 0, sizeof(VmdkExtent));
 328     extent->file = file;
 329     extent->flat = flat;
 330     extent->sectors = sectors;
 331     extent->l1_table_offset = l1_offset;
 332     extent->l1_backup_table_offset = l1_backup_offset;
 333     extent->l1_size = l1_size;
 334     extent->l1_entry_sectors = l2_size * cluster_sectors;
 335     extent->l2_size = l2_size;
 336     extent->cluster_sectors = cluster_sectors;
 337
 338     if (s->num_extents > 1) {
 339         extent->end_sector = (*(extent - 1)).end_sector + extent->sectors;
 340     } else {
 341         extent->end_sector = extent->sectors;
 342     }
 343     bs->total_sectors = extent->end_sector;
 344     return extent;
 345 }
 346
 347 static int vmdk_init_tables(BlockDriverState *bs, VmdkExtent *extent)
 348 {
 349     int ret;
 350     int l1_size, i;
 351
 352     /* read the L1 table */
 353     l1_size = extent->l1_size * sizeof(uint32_t);
 354     extent->l1_table = g_malloc(l1_size);
 355     ret = bdrv_pread(extent->file,
 356                     extent->l1_table_offset,
 357                     extent->l1_table,
 358                     l1_size);
 359     if (ret < 0) {
 360         goto fail_l1;
 361     }
 362     for (i = 0; i < extent->l1_size; i++) {
 363         le32_to_cpus(&extent->l1_table[i]);
 364     }
 365
 366     if (extent->l1_backup_table_offset) {
 367         extent->l1_backup_table = g_malloc(l1_size);
 368         ret = bdrv_pread(extent->file,
 369                         extent->l1_backup_table_offset,
 370                         extent->l1_backup_table,
 371                         l1_size);
 372         if (ret < 0) {
 373             goto fail_l1b;
 374         }
 375         for (i = 0; i < extent->l1_size; i++) {
 376             le32_to_cpus(&extent->l1_backup_table[i]);
 377         }
 378     }
 379
 380     extent->l2_cache =
 381         g_malloc(extent->l2_size * L2_CACHE_SIZE * sizeof(uint32_t));
 382     return 0;
 383  fail_l1b:
 384     g_free(extent->l1_backup_table);
 385  fail_l1:
 386     g_free(extent->l1_table);
 387     return ret;
 388 }
 389
 390 static int vmdk_open_vmdk3(BlockDriverState *bs,
 391                            BlockDriverState *file,
 392                            int flags)
 393 {
 394     int ret;
 395     uint32_t magic;
 396     VMDK3Header header;
 397     VmdkExtent *extent;
 398
 399     ret = bdrv_pread(file, sizeof(magic), &header, sizeof(header));
 400     if (ret < 0) {
 401         return ret;
 402     }
 403     extent = vmdk_add_extent(bs,
 404                              bs->file, false,
 405                              le32_to_cpu(header.disk_sectors),
 406                              le32_to_cpu(header.l1dir_offset) << 9,
 407                              0, 1 << 6, 1 << 9,
 408                              le32_to_cpu(header.granularity));
 409     ret = vmdk_init_tables(bs, extent);
 410     if (ret) {
 411         /* free extent allocated by vmdk_add_extent */
 412         vmdk_free_last_extent(bs);
 413     }
 414     return ret;
 415 }
 416
 417 static int vmdk_open_desc_file(BlockDriverState *bs, int flags,
 418                                int64_t desc_offset);
 419
 420 static int vmdk_open_vmdk4(BlockDriverState *bs,
 421                            BlockDriverState *file,
 422                            int flags)
 423 {
 424     int ret;
 425     uint32_t magic;
 426     uint32_t l1_size, l1_entry_sectors;
 427     VMDK4Header header;
 428     VmdkExtent *extent;
 429     int64_t l1_backup_offset = 0;
 430
 431     ret = bdrv_pread(file, sizeof(magic), &header, sizeof(header));
 432     if (ret < 0) {
 433         return ret;
 434     }
 435     if (header.capacity == 0 && header.desc_offset) {
 436         return vmdk_open_desc_file(bs, flags, header.desc_offset << 9);
 437     }
 438     l1_entry_sectors = le32_to_cpu(header.num_gtes_per_gte)
 439                         * le64_to_cpu(header.granularity);
 440     if (l1_entry_sectors <= 0) {
 441         return -EINVAL;
 442     }
 443     l1_size = (le64_to_cpu(header.capacity) + l1_entry_sectors - 1)
 444                 / l1_entry_sectors;
 445     if (le32_to_cpu(header.flags) & VMDK4_FLAG_RGD) {
 446         l1_backup_offset = le64_to_cpu(header.rgd_offset) << 9;
 447     }
 448     extent = vmdk_add_extent(bs, file, false,
 449                           le64_to_cpu(header.capacity),
 450                           le64_to_cpu(header.gd_offset) << 9,
 451                           l1_backup_offset,
 452                           l1_size,
 453                           le32_to_cpu(header.num_gtes_per_gte),
 454                           le64_to_cpu(header.granularity));
 455     extent->compressed =
 456         le16_to_cpu(header.compressAlgorithm) == VMDK4_COMPRESSION_DEFLATE;
 457     extent->has_marker = le32_to_cpu(header.flags) & VMDK4_FLAG_MARKER;
 458     ret = vmdk_init_tables(bs, extent);
 459     if (ret) {
 460         /* free extent allocated by vmdk_add_extent */
 461         vmdk_free_last_extent(bs);
 462     }
 463     return ret;
 464 }
 465
 466 /* find an option value out of descriptor file */
 467 static int vmdk_parse_description(const char *desc, const char *opt_name,
 468         char *buf, int buf_size)
 469 {
 470     char *opt_pos, *opt_end;
 471     const char *end = desc + strlen(desc);
 472
 473     opt_pos = strstr(desc, opt_name);
 474     if (!opt_pos) {
 475         return -1;
 476     }
 477     /* Skip "=\"" following opt_name */
 478     opt_pos += strlen(opt_name) + 2;
 479     if (opt_pos >= end) {
 480         return -1;
 481     }
 482     opt_end = opt_pos;
 483     while (opt_end < end && *opt_end != '"') {
 484         opt_end++;
 485     }
 486     if (opt_end == end || buf_size < opt_end - opt_pos + 1) {
 487         return -1;
 488     }
 489     pstrcpy(buf, opt_end - opt_pos + 1, opt_pos);
 490     return 0;
 491 }
 492
 493 /* Open an extent file and append to bs array */
 494 static int vmdk_open_sparse(BlockDriverState *bs,
 495                             BlockDriverState *file,
 496                             int flags)
 497 {
 498     uint32_t magic;
 499
 500     if (bdrv_pread(file, 0, &magic, sizeof(magic)) != sizeof(magic)) {
 501         return -EIO;
 502     }
 503
 504     magic = be32_to_cpu(magic);
 505     switch (magic) {
 506         case VMDK3_MAGIC:
 507             return vmdk_open_vmdk3(bs, file, flags);
 508             break;
 509         case VMDK4_MAGIC:
 510             return vmdk_open_vmdk4(bs, file, flags);
 511             break;
 512         default:
 513             return -EINVAL;
 514             break;
 515     }
 516 }
 517
 518 static int vmdk_parse_extents(const char *desc, BlockDriverState *bs,
 519         const char *desc_file_path)
 520 {
 521     int ret;
 522     char access[11];
 523     char type[11];
 524     char fname[512];
 525     const char *p = desc;
 526     int64_t sectors = 0;
 527     int64_t flat_offset;
 528     char extent_path[PATH_MAX];
 529     BlockDriverState *extent_file;
 530
 531     while (*p) {
 532         /* parse extent line:
 533          * RW [size in sectors] FLAT "file-name.vmdk" OFFSET
 534          * or
 535          * RW [size in sectors] SPARSE "file-name.vmdk"
 536          */
 537         flat_offset = -1;
 538         ret = sscanf(p, "%10s %" SCNd64 " %10s %511s %" SCNd64,
 539                 access, &sectors, type, fname, &flat_offset);
 540         if (ret < 4 || strcmp(access, "RW")) {
 541             goto next_line;
 542         } else if (!strcmp(type, "FLAT")) {
 543             if (ret != 5 || flat_offset < 0) {
 544                 return -EINVAL;
 545             }
 546         } else if (ret != 4) {
 547             return -EINVAL;
 548         }
 549
 550         /* trim the quotation marks around */
 551         if (fname[0] == '"') {
 552             memmove(fname, fname + 1, strlen(fname));
 553             if (strlen(fname) <= 1 || fname[strlen(fname) - 1] != '"') {
 554                 return -EINVAL;
 555             }
 556             fname[strlen(fname) - 1] = '\0';
 557         }
 558         if (sectors <= 0 ||
 559             (strcmp(type, "FLAT") && strcmp(type, "SPARSE")) ||
 560             (strcmp(access, "RW"))) {
 561             goto next_line;
 562         }
 563
 564         path_combine(extent_path, sizeof(extent_path),
 565                 desc_file_path, fname);
 566         ret = bdrv_file_open(&extent_file, extent_path, bs->open_flags);
 567         if (ret) {
 568             return ret;
 569         }
 570
 571         /* save to extents array */
 572         if (!strcmp(type, "FLAT")) {
 573             /* FLAT extent */
 574             VmdkExtent *extent;
 575
 576             extent = vmdk_add_extent(bs, extent_file, true, sectors,
 577                             0, 0, 0, 0, sectors);
 578             extent->flat_start_offset = flat_offset << 9;
 579         } else if (!strcmp(type, "SPARSE")) {
 580             /* SPARSE extent */
 581             ret = vmdk_open_sparse(bs, extent_file, bs->open_flags);
 582             if (ret) {
 583                 bdrv_delete(extent_file);
 584                 return ret;
 585             }
 586         } else {
 587             fprintf(stderr,
 588                 "VMDK: Not supported extent type \"%s\""".\n", type);
 589             return -ENOTSUP;
 590         }
 591 next_line:
 592         /* move to next line */
 593         while (*p && *p != '\n') {
 594             p++;
 595         }
 596         p++;
 597     }
 598     return 0;
 599 }
 600
 601 static int vmdk_open_desc_file(BlockDriverState *bs, int flags,
 602                                int64_t desc_offset)
 603 {
 604     int ret;
 605     char buf[2048];
 606     char ct[128];
 607     BDRVVmdkState *s = bs->opaque;
 608
 609     ret = bdrv_pread(bs->file, desc_offset, buf, sizeof(buf));
 610     if (ret < 0) {
 611         return ret;
 612     }
 613     buf[2047] = '\0';
 614     if (vmdk_parse_description(buf, "createType", ct, sizeof(ct))) {
 615         return -EINVAL;
 616     }
 617     if (strcmp(ct, "monolithicFlat") &&
 618         strcmp(ct, "twoGbMaxExtentSparse") &&
 619         strcmp(ct, "twoGbMaxExtentFlat")) {
 620         fprintf(stderr,
 621                 "VMDK: Not supported image type \"%s\""".\n", ct);
 622         return -ENOTSUP;
 623     }
 624     s->desc_offset = 0;
 625     ret = vmdk_parse_extents(buf, bs, bs->file->filename);
 626     if (ret) {
 627         vmdk_free_extents(bs);
 628         return ret;
 629     }
 630
 631     /* try to open parent images, if exist */
 632     if (vmdk_parent_open(bs)) {
 633         vmdk_free_extents(bs);
 634         return -EINVAL;
 635     }
 636     s->parent_cid = vmdk_read_cid(bs, 1);
 637     return 0;
 638 }
 639
 640 static int vmdk_open(BlockDriverState *bs, int flags)
 641 {
 642     int ret;
 643     BDRVVmdkState *s = bs->opaque;
 644
 645     if (vmdk_open_sparse(bs, bs->file, flags) == 0) {
 646         s->desc_offset = 0x200;
 647         /* try to open parent images, if exist */
 648         ret = vmdk_parent_open(bs);
 649         if (ret) {
 650             vmdk_free_extents(bs);
 651             return ret;
 652         }
 653         s->parent_cid = vmdk_read_cid(bs, 1);
 654         return 0;
 655     } else {
 656         return vmdk_open_desc_file(bs, flags, 0);
 657     }
 658 }
 659
 660 static int get_whole_cluster(BlockDriverState *bs,
 661                 VmdkExtent *extent,
 662                 uint64_t cluster_offset,
 663                 uint64_t offset,
 664                 bool allocate)
 665 {
 666     /* 128 sectors * 512 bytes each = grain size 64KB */
 667     uint8_t  whole_grain[extent->cluster_sectors * 512];
 668
 669     /* we will be here if it's first write on non-exist grain(cluster).
 670      * try to read from parent image, if exist */
 671     if (bs->backing_hd) {
 672         int ret;
 673
 674         if (!vmdk_is_cid_valid(bs)) {
 675             return -1;
 676         }
 677
 678         /* floor offset to cluster */
 679         offset -= offset % (extent->cluster_sectors * 512);
 680         ret = bdrv_read(bs->backing_hd, offset >> 9, whole_grain,
 681                 extent->cluster_sectors);
 682         if (ret < 0) {
 683             return -1;
 684         }
 685
 686         /* Write grain only into the active image */
 687         ret = bdrv_write(extent->file, cluster_offset, whole_grain,
 688                 extent->cluster_sectors);
 689         if (ret < 0) {
 690             return -1;
 691         }
 692     }
 693     return 0;
 694 }
 695
 696 static int vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data)
 697 {
 698     /* update L2 table */
 699     if (bdrv_pwrite_sync(
 700                 extent->file,
 701                 ((int64_t)m_data->l2_offset * 512)
 702                     + (m_data->l2_index * sizeof(m_data->offset)),
 703                 &(m_data->offset),
 704                 sizeof(m_data->offset)
 705             ) < 0) {
 706         return -1;
 707     }
 708     /* update backup L2 table */
 709     if (extent->l1_backup_table_offset != 0) {
 710         m_data->l2_offset = extent->l1_backup_table[m_data->l1_index];
 711         if (bdrv_pwrite_sync(
 712                     extent->file,
 713                     ((int64_t)m_data->l2_offset * 512)
 714                         + (m_data->l2_index * sizeof(m_data->offset)),
 715                     &(m_data->offset), sizeof(m_data->offset)
 716                 ) < 0) {
 717             return -1;
 718         }
 719     }
 720
 721     return 0;
 722 }
 723
 724 static int get_cluster_offset(BlockDriverState *bs,
 725                                     VmdkExtent *extent,
 726                                     VmdkMetaData *m_data,
 727                                     uint64_t offset,
 728                                     int allocate,
 729                                     uint64_t *cluster_offset)
 730 {
 731     unsigned int l1_index, l2_offset, l2_index;
 732     int min_index, i, j;
 733     uint32_t min_count, *l2_table, tmp = 0;
 734
 735     if (m_data) {
 736         m_data->valid = 0;
 737     }
 738     if (extent->flat) {
 739         *cluster_offset = extent->flat_start_offset;
 740         return 0;
 741     }
 742
 743     offset -= (extent->end_sector - extent->sectors) * SECTOR_SIZE;
 744     l1_index = (offset >> 9) / extent->l1_entry_sectors;
 745     if (l1_index >= extent->l1_size) {
 746         return -1;
 747     }
 748     l2_offset = extent->l1_table[l1_index];
 749     if (!l2_offset) {
 750         return -1;
 751     }
 752     for (i = 0; i < L2_CACHE_SIZE; i++) {
 753         if (l2_offset == extent->l2_cache_offsets[i]) {
 754             /* increment the hit count */
 755             if (++extent->l2_cache_counts[i] == 0xffffffff) {
 756                 for (j = 0; j < L2_CACHE_SIZE; j++) {
 757                     extent->l2_cache_counts[j] >>= 1;
 758                 }
 759             }
 760             l2_table = extent->l2_cache + (i * extent->l2_size);
 761             goto found;
 762         }
 763     }
 764     /* not found: load a new entry in the least used one */
 765     min_index = 0;
 766     min_count = 0xffffffff;
 767     for (i = 0; i < L2_CACHE_SIZE; i++) {
 768         if (extent->l2_cache_counts[i] < min_count) {
 769             min_count = extent->l2_cache_counts[i];
 770             min_index = i;
 771         }
 772     }
 773     l2_table = extent->l2_cache + (min_index * extent->l2_size);
 774     if (bdrv_pread(
 775                 extent->file,
 776                 (int64_t)l2_offset * 512,
 777                 l2_table,
 778                 extent->l2_size * sizeof(uint32_t)
 779             ) != extent->l2_size * sizeof(uint32_t)) {
 780         return -1;
 781     }
 782
 783     extent->l2_cache_offsets[min_index] = l2_offset;
 784     extent->l2_cache_counts[min_index] = 1;
 785  found:
 786     l2_index = ((offset >> 9) / extent->cluster_sectors) % extent->l2_size;
 787     *cluster_offset = le32_to_cpu(l2_table[l2_index]);
 788
 789     if (!*cluster_offset) {
 790         if (!allocate) {
 791             return -1;
 792         }
 793
 794         /* Avoid the L2 tables update for the images that have snapshots. */
 795         *cluster_offset = bdrv_getlength(extent->file);
 796         if (!extent->compressed) {
 797             bdrv_truncate(
 798                 extent->file,
 799                 *cluster_offset + (extent->cluster_sectors << 9)
 800             );
 801         }
 802
 803         *cluster_offset >>= 9;
 804         tmp = cpu_to_le32(*cluster_offset);
 805         l2_table[l2_index] = tmp;
 806
 807         /* First of all we write grain itself, to avoid race condition
 808          * that may to corrupt the image.
 809          * This problem may occur because of insufficient space on host disk
 810          * or inappropriate VM shutdown.
 811          */
 812         if (get_whole_cluster(
 813                 bs, extent, *cluster_offset, offset, allocate) == -1) {
 814             return -1;
 815         }
 816
 817         if (m_data) {
 818             m_data->offset = tmp;
 819             m_data->l1_index = l1_index;
 820             m_data->l2_index = l2_index;
 821             m_data->l2_offset = l2_offset;
 822             m_data->valid = 1;
 823         }
 824     }
 825     *cluster_offset <<= 9;
 826     return 0;
 827 }
 828
 829 static VmdkExtent *find_extent(BDRVVmdkState *s,
 830                                 int64_t sector_num, VmdkExtent *start_hint)
 831 {
 832     VmdkExtent *extent = start_hint;
 833
 834     if (!extent) {
 835         extent = &s->extents[0];
 836     }
 837     while (extent < &s->extents[s->num_extents]) {
 838         if (sector_num < extent->end_sector) {
 839             return extent;
 840         }
 841         extent++;
 842     }
 843     return NULL;
 844 }
 845
 846 static int vmdk_is_allocated(BlockDriverState *bs, int64_t sector_num,
 847                              int nb_sectors, int *pnum)
 848 {
 849     BDRVVmdkState *s = bs->opaque;
 850     int64_t index_in_cluster, n, ret;
 851     uint64_t offset;
 852     VmdkExtent *extent;
 853
 854     extent = find_extent(s, sector_num, NULL);
 855     if (!extent) {
 856         return 0;
 857     }
 858     ret = get_cluster_offset(bs, extent, NULL,
 859                             sector_num * 512, 0, &offset);
 860     /* get_cluster_offset returning 0 means success */
 861     ret = !ret;
 862
 863     index_in_cluster = sector_num % extent->cluster_sectors;
 864     n = extent->cluster_sectors - index_in_cluster;
 865     if (n > nb_sectors) {
 866         n = nb_sectors;
 867     }
 868     *pnum = n;
 869     return ret;
 870 }
 871
 872 static int vmdk_write_extent(VmdkExtent *extent, int64_t cluster_offset,
 873                             int64_t offset_in_cluster, const uint8_t *buf,
 874                             int nb_sectors, int64_t sector_num)
 875 {
 876     int ret;
 877     VmdkGrainMarker *data = NULL;
 878     uLongf buf_len;
 879     const uint8_t *write_buf = buf;
 880     int write_len = nb_sectors * 512;
 881
 882     if (extent->compressed) {
 883         if (!extent->has_marker) {
 884             ret = -EINVAL;
 885             goto out;
 886         }
 887         buf_len = (extent->cluster_sectors << 9) * 2;
 888         data = g_malloc(buf_len + sizeof(VmdkGrainMarker));
 889         if (compress(data->data, &buf_len, buf, nb_sectors << 9) != Z_OK ||
 890                 buf_len == 0) {
 891             ret = -EINVAL;
 892             goto out;
 893         }
 894         data->lba = sector_num;
 895         data->size = buf_len;
 896         write_buf = (uint8_t *)data;
 897         write_len = buf_len + sizeof(VmdkGrainMarker);
 898     }
 899     ret = bdrv_pwrite(extent->file,
 900                         cluster_offset + offset_in_cluster,
 901                         write_buf,
 902                         write_len);
 903     if (ret != write_len) {
 904         ret = ret < 0 ? ret : -EIO;
 905         goto out;
 906     }
 907     ret = 0;
 908  out:
 909     g_free(data);
 910     return ret;
 911 }
 912
 913 static int vmdk_read_extent(VmdkExtent *extent, int64_t cluster_offset,
 914                             int64_t offset_in_cluster, uint8_t *buf,
 915                             int nb_sectors)
 916 {
 917     int ret;
 918     int cluster_bytes, buf_bytes;
 919     uint8_t *cluster_buf, *compressed_data;
 920     uint8_t *uncomp_buf;
 921     uint32_t data_len;
 922     VmdkGrainMarker *marker;
 923     uLongf buf_len;
 924
 925
 926     if (!extent->compressed) {
 927         ret = bdrv_pread(extent->file,
 928                           cluster_offset + offset_in_cluster,
 929                           buf, nb_sectors * 512);
 930         if (ret == nb_sectors * 512) {
 931             return 0;
 932         } else {
 933             return -EIO;
 934         }
 935     }
 936     cluster_bytes = extent->cluster_sectors * 512;
 937     /* Read two clusters in case GrainMarker + compressed data > one cluster */
 938     buf_bytes = cluster_bytes * 2;
 939     cluster_buf = g_malloc(buf_bytes);
 940     uncomp_buf = g_malloc(cluster_bytes);
 941     ret = bdrv_pread(extent->file,
 942                 cluster_offset,
 943                 cluster_buf, buf_bytes);
 944     if (ret < 0) {
 945         goto out;
 946     }
 947     compressed_data = cluster_buf;
 948     buf_len = cluster_bytes;
 949     data_len = cluster_bytes;
 950     if (extent->has_marker) {
 951         marker = (VmdkGrainMarker *)cluster_buf;
 952         compressed_data = marker->data;
 953         data_len = le32_to_cpu(marker->size);
 954     }
 955     if (!data_len || data_len > buf_bytes) {
 956         ret = -EINVAL;
 957         goto out;
 958     }
 959     ret = uncompress(uncomp_buf, &buf_len, compressed_data, data_len);
 960     if (ret != Z_OK) {
 961         ret = -EINVAL;
 962         goto out;
 963
 964     }
 965     if (offset_in_cluster < 0 ||
 966             offset_in_cluster + nb_sectors * 512 > buf_len) {
 967         ret = -EINVAL;
 968         goto out;
 969     }
 970     memcpy(buf, uncomp_buf + offset_in_cluster, nb_sectors * 512);
 971     ret = 0;
 972
 973  out:
 974     g_free(uncomp_buf);
 975     g_free(cluster_buf);
 976     return ret;
 977 }
 978
 979 static int vmdk_read(BlockDriverState *bs, int64_t sector_num,
 980                     uint8_t *buf, int nb_sectors)
 981 {
 982     BDRVVmdkState *s = bs->opaque;
 983     int ret;
 984     uint64_t n, index_in_cluster;
 985     VmdkExtent *extent = NULL;
 986     uint64_t cluster_offset;
 987
 988     while (nb_sectors > 0) {
 989         extent = find_extent(s, sector_num, extent);
 990         if (!extent) {
 991             return -EIO;
 992         }
 993         ret = get_cluster_offset(
 994                             bs, extent, NULL,
 995                             sector_num << 9, 0, &cluster_offset);
 996         index_in_cluster = sector_num % extent->cluster_sectors;
 997         n = extent->cluster_sectors - index_in_cluster;
 998         if (n > nb_sectors) {
 999             n = nb_sectors;
1000         }
1001         if (ret) {
1002             /* if not allocated, try to read from parent image, if exist */
1003             if (bs->backing_hd) {
1004                 if (!vmdk_is_cid_valid(bs)) {
1005                     return -EINVAL;
1006                 }
1007                 ret = bdrv_read(bs->backing_hd, sector_num, buf, n);
1008                 if (ret < 0) {
1009                     return ret;
1010                 }
1011             } else {
1012                 memset(buf, 0, 512 * n);
1013             }
1014         } else {
1015             ret = vmdk_read_extent(extent,
1016                             cluster_offset, index_in_cluster * 512,
1017                             buf, n);
1018             if (ret) {
1019                 return ret;
1020             }
1021         }
1022         nb_sectors -= n;
1023         sector_num += n;
1024         buf += n * 512;
1025     }
1026     return 0;
1027 }
1028
1029 static int vmdk_write(BlockDriverState *bs, int64_t sector_num,
1030                      const uint8_t *buf, int nb_sectors)
1031 {
1032     BDRVVmdkState *s = bs->opaque;
1033     VmdkExtent *extent = NULL;
1034     int n, ret;
1035     int64_t index_in_cluster;
1036     uint64_t cluster_offset;
1037     VmdkMetaData m_data;
1038
1039     if (sector_num > bs->total_sectors) {
1040         fprintf(stderr,
1041                 "(VMDK) Wrong offset: sector_num=0x%" PRIx64
1042                 " total_sectors=0x%" PRIx64 "\n",
1043                 sector_num, bs->total_sectors);
1044         return -EIO;
1045     }
1046
1047     while (nb_sectors > 0) {
1048         extent = find_extent(s, sector_num, extent);
1049         if (!extent) {
1050             return -EIO;
1051         }
1052         ret = get_cluster_offset(
1053                                 bs,
1054                                 extent,
1055                                 &m_data,
1056                                 sector_num << 9, !extent->compressed,
1057                                 &cluster_offset);
1058         if (extent->compressed) {
1059             if (ret == 0) {
1060                 /* Refuse write to allocated cluster for streamOptimized */
1061                 fprintf(stderr,
1062                         "VMDK: can't write to allocated cluster"
1063                         " for streamOptimized\n");
1064                 return -EIO;
1065             } else {
1066                 /* allocate */
1067                 ret = get_cluster_offset(
1068                                         bs,
1069                                         extent,
1070                                         &m_data,
1071                                         sector_num << 9, 1,
1072                                         &cluster_offset);
1073             }
1074         }
1075         if (ret) {
1076             return -EINVAL;
1077         }
1078         index_in_cluster = sector_num % extent->cluster_sectors;
1079         n = extent->cluster_sectors - index_in_cluster;
1080         if (n > nb_sectors) {
1081             n = nb_sectors;
1082         }
1083
1084         ret = vmdk_write_extent(extent,
1085                         cluster_offset, index_in_cluster * 512,
1086                         buf, n, sector_num);
1087         if (ret) {
1088             return ret;
1089         }
1090         if (m_data.valid) {
1091             /* update L2 tables */
1092             if (vmdk_L2update(extent, &m_data) == -1) {
1093                 return -EIO;
1094             }
1095         }
1096         nb_sectors -= n;
1097         sector_num += n;
1098         buf += n * 512;
1099
1100         /* update CID on the first write every time the virtual disk is
1101          * opened */
1102         if (!s->cid_updated) {
1103             vmdk_write_cid(bs, time(NULL));
1104             s->cid_updated = true;
1105         }
1106     }
1107     return 0;
1108 }
1109
1110
1111 static int vmdk_create_extent(const char *filename, int64_t filesize,
1112                               bool flat, bool compress)
1113 {
1114     int ret, i;
1115     int fd = 0;
1116     VMDK4Header header;
1117     uint32_t tmp, magic, grains, gd_size, gt_size, gt_count;
1118
1119     fd = open(
1120         filename,
1121         O_WRONLY | O_CREAT | O_TRUNC | O_BINARY | O_LARGEFILE,
1122         0644);
1123     if (fd < 0) {
1124         return -errno;
1125     }
1126     if (flat) {
1127         ret = ftruncate(fd, filesize);
1128         if (ret < 0) {
1129             ret = -errno;
1130         }
1131         goto exit;
1132     }
1133     magic = cpu_to_be32(VMDK4_MAGIC);
1134     memset(&header, 0, sizeof(header));
1135     header.version = 1;
1136     header.flags =
1137         3 | (compress ? VMDK4_FLAG_COMPRESS | VMDK4_FLAG_MARKER : 0);
1138     header.compressAlgorithm = compress ? VMDK4_COMPRESSION_DEFLATE : 0;
1139     header.capacity = filesize / 512;
1140     header.granularity = 128;
1141     header.num_gtes_per_gte = 512;
1142
1143     grains = (filesize / 512 + header.granularity - 1) / header.granularity;
1144     gt_size = ((header.num_gtes_per_gte * sizeof(uint32_t)) + 511) >> 9;
1145     gt_count =
1146         (grains + header.num_gtes_per_gte - 1) / header.num_gtes_per_gte;
1147     gd_size = (gt_count * sizeof(uint32_t) + 511) >> 9;
1148
1149     header.desc_offset = 1;
1150     header.desc_size = 20;
1151     header.rgd_offset = header.desc_offset + header.desc_size;
1152     header.gd_offset = header.rgd_offset + gd_size + (gt_size * gt_count);
1153     header.grain_offset =
1154        ((header.gd_offset + gd_size + (gt_size * gt_count) +
1155          header.granularity - 1) / header.granularity) *
1156         header.granularity;
1157     /* swap endianness for all header fields */
1158     header.version = cpu_to_le32(header.version);
1159     header.flags = cpu_to_le32(header.flags);
1160     header.capacity = cpu_to_le64(header.capacity);
1161     header.granularity = cpu_to_le64(header.granularity);
1162     header.num_gtes_per_gte = cpu_to_le32(header.num_gtes_per_gte);
1163     header.desc_offset = cpu_to_le64(header.desc_offset);
1164     header.desc_size = cpu_to_le64(header.desc_size);
1165     header.rgd_offset = cpu_to_le64(header.rgd_offset);
1166     header.gd_offset = cpu_to_le64(header.gd_offset);
1167     header.grain_offset = cpu_to_le64(header.grain_offset);
1168     header.compressAlgorithm = cpu_to_le16(header.compressAlgorithm);
1169
1170     header.check_bytes[0] = 0xa;
1171     header.check_bytes[1] = 0x20;
1172     header.check_bytes[2] = 0xd;
1173     header.check_bytes[3] = 0xa;
1174
1175     /* write all the data */
1176     ret = qemu_write_full(fd, &magic, sizeof(magic));
1177     if (ret != sizeof(magic)) {
1178         ret = -errno;
1179         goto exit;
1180     }
1181     ret = qemu_write_full(fd, &header, sizeof(header));
1182     if (ret != sizeof(header)) {
1183         ret = -errno;
1184         goto exit;
1185     }
1186
1187     ret = ftruncate(fd, le64_to_cpu(header.grain_offset) << 9);
1188     if (ret < 0) {
1189         ret = -errno;
1190         goto exit;
1191     }
1192
1193     /* write grain directory */
1194     lseek(fd, le64_to_cpu(header.rgd_offset) << 9, SEEK_SET);
1195     for (i = 0, tmp = le64_to_cpu(header.rgd_offset) + gd_size;
1196          i < gt_count; i++, tmp += gt_size) {
1197         ret = qemu_write_full(fd, &tmp, sizeof(tmp));
1198         if (ret != sizeof(tmp)) {
1199             ret = -errno;
1200             goto exit;
1201         }
1202     }
1203
1204     /* write backup grain directory */
1205     lseek(fd, le64_to_cpu(header.gd_offset) << 9, SEEK_SET);
1206     for (i = 0, tmp = le64_to_cpu(header.gd_offset) + gd_size;
1207          i < gt_count; i++, tmp += gt_size) {
1208         ret = qemu_write_full(fd, &tmp, sizeof(tmp));
1209         if (ret != sizeof(tmp)) {
1210             ret = -errno;
1211             goto exit;
1212         }
1213     }
1214
1215     ret = 0;
1216  exit:
1217     close(fd);
1218     return ret;
1219 }
1220
1221 static int filename_decompose(const char *filename, char *path, char *prefix,
1222         char *postfix, size_t buf_len)
1223 {
1224     const char *p, *q;
1225
1226     if (filename == NULL || !strlen(filename)) {
1227         fprintf(stderr, "Vmdk: no filename provided.\n");
1228         return -1;
1229     }
1230     p = strrchr(filename, '/');
1231     if (p == NULL) {
1232         p = strrchr(filename, '\\');
1233     }
1234     if (p == NULL) {
1235         p = strrchr(filename, ':');
1236     }
1237     if (p != NULL) {
1238         p++;
1239         if (p - filename >= buf_len) {
1240             return -1;
1241         }
1242         pstrcpy(path, p - filename + 1, filename);
1243     } else {
1244         p = filename;
1245         path[0] = '\0';
1246     }
1247     q = strrchr(p, '.');
1248     if (q == NULL) {
1249         pstrcpy(prefix, buf_len, p);
1250         postfix[0] = '\0';
1251     } else {
1252         if (q - p >= buf_len) {
1253             return -1;
1254         }
1255         pstrcpy(prefix, q - p + 1, p);
1256         pstrcpy(postfix, buf_len, q);
1257     }
1258     return 0;
1259 }
1260
1261 static int relative_path(char *dest, int dest_size,
1262         const char *base, const char *target)
1263 {
1264     int i = 0;
1265     int n = 0;
1266     const char *p, *q;
1267 #ifdef _WIN32
1268     const char *sep = "\\";
1269 #else
1270     const char *sep = "/";
1271 #endif
1272
1273     if (!(dest && base && target)) {
1274         return -1;
1275     }
1276     if (path_is_absolute(target)) {
1277         dest[dest_size - 1] = '\0';
1278         strncpy(dest, target, dest_size - 1);
1279         return 0;
1280     }
1281     while (base[i] == target[i]) {
1282         i++;
1283     }
1284     p = &base[i];
1285     q = &target[i];
1286     while (*p) {
1287         if (*p == *sep) {
1288             n++;
1289         }
1290         p++;
1291     }
1292     dest[0] = '\0';
1293     for (; n; n--) {
1294         pstrcat(dest, dest_size, "..");
1295         pstrcat(dest, dest_size, sep);
1296     }
1297     pstrcat(dest, dest_size, q);
1298     return 0;
1299 }
1300
1301 static int vmdk_create(const char *filename, QEMUOptionParameter *options)
1302 {
1303     int fd, idx = 0;
1304     char desc[BUF_SIZE];
1305     int64_t total_size = 0, filesize;
1306     const char *backing_file = NULL;
1307     const char *fmt = NULL;
1308     int flags = 0;
1309     int ret = 0;
1310     bool flat, split, compress;
1311     char ext_desc_lines[BUF_SIZE] = "";
1312     char path[PATH_MAX], prefix[PATH_MAX], postfix[PATH_MAX];
1313     const int64_t split_size = 0x80000000;  /* VMDK has constant split size */
1314     const char *desc_extent_line;
1315     char parent_desc_line[BUF_SIZE] = "";
1316     uint32_t parent_cid = 0xffffffff;
1317     const char desc_template[] =
1318         "# Disk DescriptorFile\n"
1319         "version=1\n"
1320         "CID=%x\n"
1321         "parentCID=%x\n"
1322         "createType=\"%s\"\n"
1323         "%s"
1324         "\n"
1325         "# Extent description\n"
1326         "%s"
1327         "\n"
1328         "# The Disk Data Base\n"
1329         "#DDB\n"
1330         "\n"
1331         "ddb.virtualHWVersion = \"%d\"\n"
1332         "ddb.geometry.cylinders = \"%" PRId64 "\"\n"
1333         "ddb.geometry.heads = \"16\"\n"
1334         "ddb.geometry.sectors = \"63\"\n"
1335         "ddb.adapterType = \"ide\"\n";
1336
1337     if (filename_decompose(filename, path, prefix, postfix, PATH_MAX)) {
1338         return -EINVAL;
1339     }
1340     /* Read out options */
1341     while (options && options->name) {
1342         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
1343             total_size = options->value.n;
1344         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
1345             backing_file = options->value.s;
1346         } else if (!strcmp(options->name, BLOCK_OPT_COMPAT6)) {
1347             flags |= options->value.n ? BLOCK_FLAG_COMPAT6 : 0;
1348         } else if (!strcmp(options->name, BLOCK_OPT_SUBFMT)) {
1349             fmt = options->value.s;
1350         }
1351         options++;
1352     }
1353     if (!fmt) {
1354         /* Default format to monolithicSparse */
1355         fmt = "monolithicSparse";
1356     } else if (strcmp(fmt, "monolithicFlat") &&
1357                strcmp(fmt, "monolithicSparse") &&
1358                strcmp(fmt, "twoGbMaxExtentSparse") &&
1359                strcmp(fmt, "twoGbMaxExtentFlat") &&
1360                strcmp(fmt, "streamOptimized")) {
1361         fprintf(stderr, "VMDK: Unknown subformat: %s\n", fmt);
1362         return -EINVAL;
1363     }
1364     split = !(strcmp(fmt, "twoGbMaxExtentFlat") &&
1365               strcmp(fmt, "twoGbMaxExtentSparse"));
1366     flat = !(strcmp(fmt, "monolithicFlat") &&
1367              strcmp(fmt, "twoGbMaxExtentFlat"));
1368     compress = !strcmp(fmt, "streamOptimized");
1369     if (flat) {
1370         desc_extent_line = "RW %lld FLAT \"%s\" 0\n";
1371     } else {
1372         desc_extent_line = "RW %lld SPARSE \"%s\"\n";
1373     }
1374     if (flat && backing_file) {
1375         /* not supporting backing file for flat image */
1376         return -ENOTSUP;
1377     }
1378     if (backing_file) {
1379         char parent_filename[PATH_MAX];
1380         BlockDriverState *bs = bdrv_new("");
1381         ret = bdrv_open(bs, backing_file, 0, NULL);
1382         if (ret != 0) {
1383             bdrv_delete(bs);
1384             return ret;
1385         }
1386         if (strcmp(bs->drv->format_name, "vmdk")) {
1387             bdrv_delete(bs);
1388             return -EINVAL;
1389         }
1390         filesize = bdrv_getlength(bs);
1391         parent_cid = vmdk_read_cid(bs, 0);
1392         bdrv_delete(bs);
1393         relative_path(parent_filename, sizeof(parent_filename),
1394                       filename, backing_file);
1395         snprintf(parent_desc_line, sizeof(parent_desc_line),
1396                 "parentFileNameHint=\"%s\"", parent_filename);
1397     }
1398
1399     /* Create extents */
1400     filesize = total_size;
1401     while (filesize > 0) {
1402         char desc_line[BUF_SIZE];
1403         char ext_filename[PATH_MAX];
1404         char desc_filename[PATH_MAX];
1405         int64_t size = filesize;
1406
1407         if (split && size > split_size) {
1408             size = split_size;
1409         }
1410         if (split) {
1411             snprintf(desc_filename, sizeof(desc_filename), "%s-%c%03d%s",
1412                     prefix, flat ? 'f' : 's', ++idx, postfix);
1413         } else if (flat) {
1414             snprintf(desc_filename, sizeof(desc_filename), "%s-flat%s",
1415                     prefix, postfix);
1416         } else {
1417             snprintf(desc_filename, sizeof(desc_filename), "%s%s",
1418                     prefix, postfix);
1419         }
1420         snprintf(ext_filename, sizeof(ext_filename), "%s%s",
1421                 path, desc_filename);
1422
1423         if (vmdk_create_extent(ext_filename, size, flat, compress)) {
1424             return -EINVAL;
1425         }
1426         filesize -= size;
1427
1428         /* Format description line */
1429         snprintf(desc_line, sizeof(desc_line),
1430                     desc_extent_line, size / 512, desc_filename);
1431         pstrcat(ext_desc_lines, sizeof(ext_desc_lines), desc_line);
1432     }
1433     /* generate descriptor file */
1434     snprintf(desc, sizeof(desc), desc_template,
1435             (unsigned int)time(NULL),
1436             parent_cid,
1437             fmt,
1438             parent_desc_line,
1439             ext_desc_lines,
1440             (flags & BLOCK_FLAG_COMPAT6 ? 6 : 4),
1441             total_size / (int64_t)(63 * 16 * 512));
1442     if (split || flat) {
1443         fd = open(
1444                 filename,
1445                 O_WRONLY | O_CREAT | O_TRUNC | O_BINARY | O_LARGEFILE,
1446                 0644);
1447     } else {
1448         fd = open(
1449                 filename,
1450                 O_WRONLY | O_BINARY | O_LARGEFILE,
1451                 0644);
1452     }
1453     if (fd < 0) {
1454         return -errno;
1455     }
1456     /* the descriptor offset = 0x200 */
1457     if (!split && !flat && 0x200 != lseek(fd, 0x200, SEEK_SET)) {
1458         ret = -errno;
1459         goto exit;
1460     }
1461     ret = qemu_write_full(fd, desc, strlen(desc));
1462     if (ret != strlen(desc)) {
1463         ret = -errno;
1464         goto exit;
1465     }
1466     ret = 0;
1467 exit:
1468     close(fd);
1469     return ret;
1470 }
1471
1472 static void vmdk_close(BlockDriverState *bs)
1473 {
1474     vmdk_free_extents(bs);
1475 }
1476
1477 static int vmdk_flush(BlockDriverState *bs)
1478 {
1479     int i, ret, err;
1480     BDRVVmdkState *s = bs->opaque;
1481
1482     ret = bdrv_flush(bs->file);
1483     for (i = 0; i < s->num_extents; i++) {
1484         err = bdrv_flush(s->extents[i].file);
1485         if (err < 0) {
1486             ret = err;
1487         }
1488     }
1489     return ret;
1490 }
1491
1492 static int64_t vmdk_get_allocated_file_size(BlockDriverState *bs)
1493 {
1494     int i;
1495     int64_t ret = 0;
1496     int64_t r;
1497     BDRVVmdkState *s = bs->opaque;
1498
1499     ret = bdrv_get_allocated_file_size(bs->file);
1500     if (ret < 0) {
1501         return ret;
1502     }
1503     for (i = 0; i < s->num_extents; i++) {
1504         if (s->extents[i].file == bs->file) {
1505             continue;
1506         }
1507         r = bdrv_get_allocated_file_size(s->extents[i].file);
1508         if (r < 0) {
1509             return r;
1510         }
1511         ret += r;
1512     }
1513     return ret;
1514 }
1515
1516 static QEMUOptionParameter vmdk_create_options[] = {
1517     {
1518         .name = BLOCK_OPT_SIZE,
1519         .type = OPT_SIZE,
1520         .help = "Virtual disk size"
1521     },
1522     {
1523         .name = BLOCK_OPT_BACKING_FILE,
1524         .type = OPT_STRING,
1525         .help = "File name of a base image"
1526     },
1527     {
1528         .name = BLOCK_OPT_COMPAT6,
1529         .type = OPT_FLAG,
1530         .help = "VMDK version 6 image"
1531     },
1532     {
1533         .name = BLOCK_OPT_SUBFMT,
1534         .type = OPT_STRING,
1535         .help =
1536             "VMDK flat extent format, can be one of "
1537             "{monolithicSparse (default) | monolithicFlat | twoGbMaxExtentSparse | twoGbMaxExtentFlat | streamOptimized} "
1538     },
1539     { NULL }
1540 };
1541
1542 static BlockDriver bdrv_vmdk = {
1543     .format_name    = "vmdk",
1544     .instance_size  = sizeof(BDRVVmdkState),
1545     .bdrv_probe     = vmdk_probe,
1546     .bdrv_open      = vmdk_open,
1547     .bdrv_read      = vmdk_read,
1548     .bdrv_write     = vmdk_write,
1549     .bdrv_close     = vmdk_close,
1550     .bdrv_create    = vmdk_create,
1551     .bdrv_flush     = vmdk_flush,
1552     .bdrv_is_allocated  = vmdk_is_allocated,
1553     .bdrv_get_allocated_file_size  = vmdk_get_allocated_file_size,
1554
1555     .create_options = vmdk_create_options,
1556 };
1557
1558 static void bdrv_vmdk_init(void)
1559 {
1560     bdrv_register(&bdrv_vmdk);
1561 }
1562
1563 block_init(bdrv_vmdk_init);