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     CoMutex lock;
  94     int desc_offset;
  95     bool cid_updated;
  96     uint32_t parent_cid;
  97     int num_extents;
  98     /* Extent array with num_extents entries, ascend ordered by address */
  99     VmdkExtent *extents;
 100 } BDRVVmdkState;
 101
 102 typedef struct VmdkMetaData {
 103     uint32_t offset;
 104     unsigned int l1_index;
 105     unsigned int l2_index;
 106     unsigned int l2_offset;
 107     int valid;
 108 } VmdkMetaData;
 109
 110 typedef struct VmdkGrainMarker {
 111     uint64_t lba;
 112     uint32_t size;
 113     uint8_t  data[0];
 114 } VmdkGrainMarker;
 115
 116 static int vmdk_probe(const uint8_t *buf, int buf_size, const char *filename)
 117 {
 118     uint32_t magic;
 119
 120     if (buf_size < 4) {
 121         return 0;
 122     }
 123     magic = be32_to_cpu(*(uint32_t *)buf);
 124     if (magic == VMDK3_MAGIC ||
 125         magic == VMDK4_MAGIC) {
 126         return 100;
 127     } else {
 128         const char *p = (const char *)buf;
 129         const char *end = p + buf_size;
 130         while (p < end) {
 131             if (*p == '#') {
 132                 /* skip comment line */
 133                 while (p < end && *p != '\n') {
 134                     p++;
 135                 }
 136                 p++;
 137                 continue;
 138             }
 139             if (*p == ' ') {
 140                 while (p < end && *p == ' ') {
 141                     p++;
 142                 }
 143                 /* skip '\r' if windows line endings used. */
 144                 if (p < end && *p == '\r') {
 145                     p++;
 146                 }
 147                 /* only accept blank lines before 'version=' line */
 148                 if (p == end || *p != '\n') {
 149                     return 0;
 150                 }
 151                 p++;
 152                 continue;
 153             }
 154             if (end - p >= strlen("version=X\n")) {
 155                 if (strncmp("version=1\n", p, strlen("version=1\n")) == 0 ||
 156                     strncmp("version=2\n", p, strlen("version=2\n")) == 0) {
 157                     return 100;
 158                 }
 159             }
 160             if (end - p >= strlen("version=X\r\n")) {
 161                 if (strncmp("version=1\r\n", p, strlen("version=1\r\n")) == 0 ||
 162                     strncmp("version=2\r\n", p, strlen("version=2\r\n")) == 0) {
 163                     return 100;
 164                 }
 165             }
 166             return 0;
 167         }
 168         return 0;
 169     }
 170 }
 171
 172 #define CHECK_CID 1
 173
 174 #define SECTOR_SIZE 512
 175 #define DESC_SIZE (20 * SECTOR_SIZE)    /* 20 sectors of 512 bytes each */
 176 #define BUF_SIZE 4096
 177 #define HEADER_SIZE 512                 /* first sector of 512 bytes */
 178
 179 static void vmdk_free_extents(BlockDriverState *bs)
 180 {
 181     int i;
 182     BDRVVmdkState *s = bs->opaque;
 183     VmdkExtent *e;
 184
 185     for (i = 0; i < s->num_extents; i++) {
 186         e = &s->extents[i];
 187         g_free(e->l1_table);
 188         g_free(e->l2_cache);
 189         g_free(e->l1_backup_table);
 190         if (e->file != bs->file) {
 191             bdrv_delete(e->file);
 192         }
 193     }
 194     g_free(s->extents);
 195 }
 196
 197 static void vmdk_free_last_extent(BlockDriverState *bs)
 198 {
 199     BDRVVmdkState *s = bs->opaque;
 200
 201     if (s->num_extents == 0) {
 202         return;
 203     }
 204     s->num_extents--;
 205     s->extents = g_realloc(s->extents, s->num_extents * sizeof(VmdkExtent));
 206 }
 207
 208 static uint32_t vmdk_read_cid(BlockDriverState *bs, int parent)
 209 {
 210     char desc[DESC_SIZE];
 211     uint32_t cid = 0xffffffff;
 212     const char *p_name, *cid_str;
 213     size_t cid_str_size;
 214     BDRVVmdkState *s = bs->opaque;
 215     int ret;
 216
 217     ret = bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE);
 218     if (ret < 0) {
 219         return 0;
 220     }
 221
 222     if (parent) {
 223         cid_str = "parentCID";
 224         cid_str_size = sizeof("parentCID");
 225     } else {
 226         cid_str = "CID";
 227         cid_str_size = sizeof("CID");
 228     }
 229
 230     desc[DESC_SIZE - 1] = '\0';
 231     p_name = strstr(desc, cid_str);
 232     if (p_name != NULL) {
 233         p_name += cid_str_size;
 234         sscanf(p_name, "%x", &cid);
 235     }
 236
 237     return cid;
 238 }
 239
 240 static int vmdk_write_cid(BlockDriverState *bs, uint32_t cid)
 241 {
 242     char desc[DESC_SIZE], tmp_desc[DESC_SIZE];
 243     char *p_name, *tmp_str;
 244     BDRVVmdkState *s = bs->opaque;
 245     int ret;
 246
 247     ret = bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE);
 248     if (ret < 0) {
 249         return ret;
 250     }
 251
 252     desc[DESC_SIZE - 1] = '\0';
 253     tmp_str = strstr(desc, "parentCID");
 254     if (tmp_str == NULL) {
 255         return -EINVAL;
 256     }
 257
 258     pstrcpy(tmp_desc, sizeof(tmp_desc), tmp_str);
 259     p_name = strstr(desc, "CID");
 260     if (p_name != NULL) {
 261         p_name += sizeof("CID");
 262         snprintf(p_name, sizeof(desc) - (p_name - desc), "%x\n", cid);
 263         pstrcat(desc, sizeof(desc), tmp_desc);
 264     }
 265
 266     ret = bdrv_pwrite_sync(bs->file, s->desc_offset, desc, DESC_SIZE);
 267     if (ret < 0) {
 268         return ret;
 269     }
 270
 271     return 0;
 272 }
 273
 274 static int vmdk_is_cid_valid(BlockDriverState *bs)
 275 {
 276 #ifdef CHECK_CID
 277     BDRVVmdkState *s = bs->opaque;
 278     BlockDriverState *p_bs = bs->backing_hd;
 279     uint32_t cur_pcid;
 280
 281     if (p_bs) {
 282         cur_pcid = vmdk_read_cid(p_bs, 0);
 283         if (s->parent_cid != cur_pcid) {
 284             /* CID not valid */
 285             return 0;
 286         }
 287     }
 288 #endif
 289     /* CID valid */
 290     return 1;
 291 }
 292
 293 static int vmdk_parent_open(BlockDriverState *bs)
 294 {
 295     char *p_name;
 296     char desc[DESC_SIZE + 1];
 297     BDRVVmdkState *s = bs->opaque;
 298     int ret;
 299
 300     desc[DESC_SIZE] = '\0';
 301     ret = bdrv_pread(bs->file, s->desc_offset, desc, DESC_SIZE);
 302     if (ret < 0) {
 303         return ret;
 304     }
 305
 306     p_name = strstr(desc, "parentFileNameHint");
 307     if (p_name != NULL) {
 308         char *end_name;
 309
 310         p_name += sizeof("parentFileNameHint") + 1;
 311         end_name = strchr(p_name, '\"');
 312         if (end_name == NULL) {
 313             return -EINVAL;
 314         }
 315         if ((end_name - p_name) > sizeof(bs->backing_file) - 1) {
 316             return -EINVAL;
 317         }
 318
 319         pstrcpy(bs->backing_file, end_name - p_name + 1, p_name);
 320     }
 321
 322     return 0;
 323 }
 324
 325 /* Create and append extent to the extent array. Return the added VmdkExtent
 326  * address. return NULL if allocation failed. */
 327 static VmdkExtent *vmdk_add_extent(BlockDriverState *bs,
 328                            BlockDriverState *file, bool flat, int64_t sectors,
 329                            int64_t l1_offset, int64_t l1_backup_offset,
 330                            uint32_t l1_size,
 331                            int l2_size, unsigned int cluster_sectors)
 332 {
 333     VmdkExtent *extent;
 334     BDRVVmdkState *s = bs->opaque;
 335
 336     s->extents = g_realloc(s->extents,
 337                               (s->num_extents + 1) * sizeof(VmdkExtent));
 338     extent = &s->extents[s->num_extents];
 339     s->num_extents++;
 340
 341     memset(extent, 0, sizeof(VmdkExtent));
 342     extent->file = file;
 343     extent->flat = flat;
 344     extent->sectors = sectors;
 345     extent->l1_table_offset = l1_offset;
 346     extent->l1_backup_table_offset = l1_backup_offset;
 347     extent->l1_size = l1_size;
 348     extent->l1_entry_sectors = l2_size * cluster_sectors;
 349     extent->l2_size = l2_size;
 350     extent->cluster_sectors = cluster_sectors;
 351
 352     if (s->num_extents > 1) {
 353         extent->end_sector = (*(extent - 1)).end_sector + extent->sectors;
 354     } else {
 355         extent->end_sector = extent->sectors;
 356     }
 357     bs->total_sectors = extent->end_sector;
 358     return extent;
 359 }
 360
 361 static int vmdk_init_tables(BlockDriverState *bs, VmdkExtent *extent)
 362 {
 363     int ret;
 364     int l1_size, i;
 365
 366     /* read the L1 table */
 367     l1_size = extent->l1_size * sizeof(uint32_t);
 368     extent->l1_table = g_malloc(l1_size);
 369     ret = bdrv_pread(extent->file,
 370                     extent->l1_table_offset,
 371                     extent->l1_table,
 372                     l1_size);
 373     if (ret < 0) {
 374         goto fail_l1;
 375     }
 376     for (i = 0; i < extent->l1_size; i++) {
 377         le32_to_cpus(&extent->l1_table[i]);
 378     }
 379
 380     if (extent->l1_backup_table_offset) {
 381         extent->l1_backup_table = g_malloc(l1_size);
 382         ret = bdrv_pread(extent->file,
 383                         extent->l1_backup_table_offset,
 384                         extent->l1_backup_table,
 385                         l1_size);
 386         if (ret < 0) {
 387             goto fail_l1b;
 388         }
 389         for (i = 0; i < extent->l1_size; i++) {
 390             le32_to_cpus(&extent->l1_backup_table[i]);
 391         }
 392     }
 393
 394     extent->l2_cache =
 395         g_malloc(extent->l2_size * L2_CACHE_SIZE * sizeof(uint32_t));
 396     return 0;
 397  fail_l1b:
 398     g_free(extent->l1_backup_table);
 399  fail_l1:
 400     g_free(extent->l1_table);
 401     return ret;
 402 }
 403
 404 static int vmdk_open_vmdk3(BlockDriverState *bs,
 405                            BlockDriverState *file,
 406                            int flags)
 407 {
 408     int ret;
 409     uint32_t magic;
 410     VMDK3Header header;
 411     VmdkExtent *extent;
 412
 413     ret = bdrv_pread(file, sizeof(magic), &header, sizeof(header));
 414     if (ret < 0) {
 415         return ret;
 416     }
 417     extent = vmdk_add_extent(bs,
 418                              bs->file, false,
 419                              le32_to_cpu(header.disk_sectors),
 420                              le32_to_cpu(header.l1dir_offset) << 9,
 421                              0, 1 << 6, 1 << 9,
 422                              le32_to_cpu(header.granularity));
 423     ret = vmdk_init_tables(bs, extent);
 424     if (ret) {
 425         /* free extent allocated by vmdk_add_extent */
 426         vmdk_free_last_extent(bs);
 427     }
 428     return ret;
 429 }
 430
 431 static int vmdk_open_desc_file(BlockDriverState *bs, int flags,
 432                                int64_t desc_offset);
 433
 434 static int vmdk_open_vmdk4(BlockDriverState *bs,
 435                            BlockDriverState *file,
 436                            int flags)
 437 {
 438     int ret;
 439     uint32_t magic;
 440     uint32_t l1_size, l1_entry_sectors;
 441     VMDK4Header header;
 442     VmdkExtent *extent;
 443     int64_t l1_backup_offset = 0;
 444
 445     ret = bdrv_pread(file, sizeof(magic), &header, sizeof(header));
 446     if (ret < 0) {
 447         return ret;
 448     }
 449     if (header.capacity == 0 && header.desc_offset) {
 450         return vmdk_open_desc_file(bs, flags, header.desc_offset << 9);
 451     }
 452     l1_entry_sectors = le32_to_cpu(header.num_gtes_per_gte)
 453                         * le64_to_cpu(header.granularity);
 454     if (l1_entry_sectors <= 0) {
 455         return -EINVAL;
 456     }
 457     l1_size = (le64_to_cpu(header.capacity) + l1_entry_sectors - 1)
 458                 / l1_entry_sectors;
 459     if (le32_to_cpu(header.flags) & VMDK4_FLAG_RGD) {
 460         l1_backup_offset = le64_to_cpu(header.rgd_offset) << 9;
 461     }
 462     extent = vmdk_add_extent(bs, file, false,
 463                           le64_to_cpu(header.capacity),
 464                           le64_to_cpu(header.gd_offset) << 9,
 465                           l1_backup_offset,
 466                           l1_size,
 467                           le32_to_cpu(header.num_gtes_per_gte),
 468                           le64_to_cpu(header.granularity));
 469     extent->compressed =
 470         le16_to_cpu(header.compressAlgorithm) == VMDK4_COMPRESSION_DEFLATE;
 471     extent->has_marker = le32_to_cpu(header.flags) & VMDK4_FLAG_MARKER;
 472     ret = vmdk_init_tables(bs, extent);
 473     if (ret) {
 474         /* free extent allocated by vmdk_add_extent */
 475         vmdk_free_last_extent(bs);
 476     }
 477     return ret;
 478 }
 479
 480 /* find an option value out of descriptor file */
 481 static int vmdk_parse_description(const char *desc, const char *opt_name,
 482         char *buf, int buf_size)
 483 {
 484     char *opt_pos, *opt_end;
 485     const char *end = desc + strlen(desc);
 486
 487     opt_pos = strstr(desc, opt_name);
 488     if (!opt_pos) {
 489         return -1;
 490     }
 491     /* Skip "=\"" following opt_name */
 492     opt_pos += strlen(opt_name) + 2;
 493     if (opt_pos >= end) {
 494         return -1;
 495     }
 496     opt_end = opt_pos;
 497     while (opt_end < end && *opt_end != '"') {
 498         opt_end++;
 499     }
 500     if (opt_end == end || buf_size < opt_end - opt_pos + 1) {
 501         return -1;
 502     }
 503     pstrcpy(buf, opt_end - opt_pos + 1, opt_pos);
 504     return 0;
 505 }
 506
 507 /* Open an extent file and append to bs array */
 508 static int vmdk_open_sparse(BlockDriverState *bs,
 509                             BlockDriverState *file,
 510                             int flags)
 511 {
 512     uint32_t magic;
 513
 514     if (bdrv_pread(file, 0, &magic, sizeof(magic)) != sizeof(magic)) {
 515         return -EIO;
 516     }
 517
 518     magic = be32_to_cpu(magic);
 519     switch (magic) {
 520         case VMDK3_MAGIC:
 521             return vmdk_open_vmdk3(bs, file, flags);
 522             break;
 523         case VMDK4_MAGIC:
 524             return vmdk_open_vmdk4(bs, file, flags);
 525             break;
 526         default:
 527             return -EINVAL;
 528             break;
 529     }
 530 }
 531
 532 static int vmdk_parse_extents(const char *desc, BlockDriverState *bs,
 533         const char *desc_file_path)
 534 {
 535     int ret;
 536     char access[11];
 537     char type[11];
 538     char fname[512];
 539     const char *p = desc;
 540     int64_t sectors = 0;
 541     int64_t flat_offset;
 542     char extent_path[PATH_MAX];
 543     BlockDriverState *extent_file;
 544
 545     while (*p) {
 546         /* parse extent line:
 547          * RW [size in sectors] FLAT "file-name.vmdk" OFFSET
 548          * or
 549          * RW [size in sectors] SPARSE "file-name.vmdk"
 550          */
 551         flat_offset = -1;
 552         ret = sscanf(p, "%10s %" SCNd64 " %10s %511s %" SCNd64,
 553                 access, &sectors, type, fname, &flat_offset);
 554         if (ret < 4 || strcmp(access, "RW")) {
 555             goto next_line;
 556         } else if (!strcmp(type, "FLAT")) {
 557             if (ret != 5 || flat_offset < 0) {
 558                 return -EINVAL;
 559             }
 560         } else if (ret != 4) {
 561             return -EINVAL;
 562         }
 563
 564         /* trim the quotation marks around */
 565         if (fname[0] == '"') {
 566             memmove(fname, fname + 1, strlen(fname));
 567             if (strlen(fname) <= 1 || fname[strlen(fname) - 1] != '"') {
 568                 return -EINVAL;
 569             }
 570             fname[strlen(fname) - 1] = '\0';
 571         }
 572         if (sectors <= 0 ||
 573             (strcmp(type, "FLAT") && strcmp(type, "SPARSE")) ||
 574             (strcmp(access, "RW"))) {
 575             goto next_line;
 576         }
 577
 578         path_combine(extent_path, sizeof(extent_path),
 579                 desc_file_path, fname);
 580         ret = bdrv_file_open(&extent_file, extent_path, bs->open_flags);
 581         if (ret) {
 582             return ret;
 583         }
 584
 585         /* save to extents array */
 586         if (!strcmp(type, "FLAT")) {
 587             /* FLAT extent */
 588             VmdkExtent *extent;
 589
 590             extent = vmdk_add_extent(bs, extent_file, true, sectors,
 591                             0, 0, 0, 0, sectors);
 592             extent->flat_start_offset = flat_offset << 9;
 593         } else if (!strcmp(type, "SPARSE")) {
 594             /* SPARSE extent */
 595             ret = vmdk_open_sparse(bs, extent_file, bs->open_flags);
 596             if (ret) {
 597                 bdrv_delete(extent_file);
 598                 return ret;
 599             }
 600         } else {
 601             fprintf(stderr,
 602                 "VMDK: Not supported extent type \"%s\""".\n", type);
 603             return -ENOTSUP;
 604         }
 605 next_line:
 606         /* move to next line */
 607         while (*p && *p != '\n') {
 608             p++;
 609         }
 610         p++;
 611     }
 612     return 0;
 613 }
 614
 615 static int vmdk_open_desc_file(BlockDriverState *bs, int flags,
 616                                int64_t desc_offset)
 617 {
 618     int ret;
 619     char buf[2048];
 620     char ct[128];
 621     BDRVVmdkState *s = bs->opaque;
 622
 623     ret = bdrv_pread(bs->file, desc_offset, buf, sizeof(buf));
 624     if (ret < 0) {
 625         return ret;
 626     }
 627     buf[2047] = '\0';
 628     if (vmdk_parse_description(buf, "createType", ct, sizeof(ct))) {
 629         return -EINVAL;
 630     }
 631     if (strcmp(ct, "monolithicFlat") &&
 632         strcmp(ct, "twoGbMaxExtentSparse") &&
 633         strcmp(ct, "twoGbMaxExtentFlat")) {
 634         fprintf(stderr,
 635                 "VMDK: Not supported image type \"%s\""".\n", ct);
 636         return -ENOTSUP;
 637     }
 638     s->desc_offset = 0;
 639     return vmdk_parse_extents(buf, bs, bs->file->filename);
 640 }
 641
 642 static int vmdk_open(BlockDriverState *bs, int flags)
 643 {
 644     int ret;
 645     BDRVVmdkState *s = bs->opaque;
 646
 647     if (vmdk_open_sparse(bs, bs->file, flags) == 0) {
 648         s->desc_offset = 0x200;
 649     } else {
 650         ret = vmdk_open_desc_file(bs, flags, 0);
 651         if (ret) {
 652             goto fail;
 653         }
 654     }
 655     /* try to open parent images, if exist */
 656     ret = vmdk_parent_open(bs);
 657     if (ret) {
 658         goto fail;
 659     }
 660     s->parent_cid = vmdk_read_cid(bs, 1);
 661     qemu_co_mutex_init(&s->lock);
 662     return ret;
 663
 664 fail:
 665     vmdk_free_extents(bs);
 666     return ret;
 667 }
 668
 669 static int get_whole_cluster(BlockDriverState *bs,
 670                 VmdkExtent *extent,
 671                 uint64_t cluster_offset,
 672                 uint64_t offset,
 673                 bool allocate)
 674 {
 675     /* 128 sectors * 512 bytes each = grain size 64KB */
 676     uint8_t  whole_grain[extent->cluster_sectors * 512];
 677
 678     /* we will be here if it's first write on non-exist grain(cluster).
 679      * try to read from parent image, if exist */
 680     if (bs->backing_hd) {
 681         int ret;
 682
 683         if (!vmdk_is_cid_valid(bs)) {
 684             return -1;
 685         }
 686
 687         /* floor offset to cluster */
 688         offset -= offset % (extent->cluster_sectors * 512);
 689         ret = bdrv_read(bs->backing_hd, offset >> 9, whole_grain,
 690                 extent->cluster_sectors);
 691         if (ret < 0) {
 692             return -1;
 693         }
 694
 695         /* Write grain only into the active image */
 696         ret = bdrv_write(extent->file, cluster_offset, whole_grain,
 697                 extent->cluster_sectors);
 698         if (ret < 0) {
 699             return -1;
 700         }
 701     }
 702     return 0;
 703 }
 704
 705 static int vmdk_L2update(VmdkExtent *extent, VmdkMetaData *m_data)
 706 {
 707     /* update L2 table */
 708     if (bdrv_pwrite_sync(
 709                 extent->file,
 710                 ((int64_t)m_data->l2_offset * 512)
 711                     + (m_data->l2_index * sizeof(m_data->offset)),
 712                 &(m_data->offset),
 713                 sizeof(m_data->offset)
 714             ) < 0) {
 715         return -1;
 716     }
 717     /* update backup L2 table */
 718     if (extent->l1_backup_table_offset != 0) {
 719         m_data->l2_offset = extent->l1_backup_table[m_data->l1_index];
 720         if (bdrv_pwrite_sync(
 721                     extent->file,
 722                     ((int64_t)m_data->l2_offset * 512)
 723                         + (m_data->l2_index * sizeof(m_data->offset)),
 724                     &(m_data->offset), sizeof(m_data->offset)
 725                 ) < 0) {
 726             return -1;
 727         }
 728     }
 729
 730     return 0;
 731 }
 732
 733 static int get_cluster_offset(BlockDriverState *bs,
 734                                     VmdkExtent *extent,
 735                                     VmdkMetaData *m_data,
 736                                     uint64_t offset,
 737                                     int allocate,
 738                                     uint64_t *cluster_offset)
 739 {
 740     unsigned int l1_index, l2_offset, l2_index;
 741     int min_index, i, j;
 742     uint32_t min_count, *l2_table, tmp = 0;
 743
 744     if (m_data) {
 745         m_data->valid = 0;
 746     }
 747     if (extent->flat) {
 748         *cluster_offset = extent->flat_start_offset;
 749         return 0;
 750     }
 751
 752     offset -= (extent->end_sector - extent->sectors) * SECTOR_SIZE;
 753     l1_index = (offset >> 9) / extent->l1_entry_sectors;
 754     if (l1_index >= extent->l1_size) {
 755         return -1;
 756     }
 757     l2_offset = extent->l1_table[l1_index];
 758     if (!l2_offset) {
 759         return -1;
 760     }
 761     for (i = 0; i < L2_CACHE_SIZE; i++) {
 762         if (l2_offset == extent->l2_cache_offsets[i]) {
 763             /* increment the hit count */
 764             if (++extent->l2_cache_counts[i] == 0xffffffff) {
 765                 for (j = 0; j < L2_CACHE_SIZE; j++) {
 766                     extent->l2_cache_counts[j] >>= 1;
 767                 }
 768             }
 769             l2_table = extent->l2_cache + (i * extent->l2_size);
 770             goto found;
 771         }
 772     }
 773     /* not found: load a new entry in the least used one */
 774     min_index = 0;
 775     min_count = 0xffffffff;
 776     for (i = 0; i < L2_CACHE_SIZE; i++) {
 777         if (extent->l2_cache_counts[i] < min_count) {
 778             min_count = extent->l2_cache_counts[i];
 779             min_index = i;
 780         }
 781     }
 782     l2_table = extent->l2_cache + (min_index * extent->l2_size);
 783     if (bdrv_pread(
 784                 extent->file,
 785                 (int64_t)l2_offset * 512,
 786                 l2_table,
 787                 extent->l2_size * sizeof(uint32_t)
 788             ) != extent->l2_size * sizeof(uint32_t)) {
 789         return -1;
 790     }
 791
 792     extent->l2_cache_offsets[min_index] = l2_offset;
 793     extent->l2_cache_counts[min_index] = 1;
 794  found:
 795     l2_index = ((offset >> 9) / extent->cluster_sectors) % extent->l2_size;
 796     *cluster_offset = le32_to_cpu(l2_table[l2_index]);
 797
 798     if (!*cluster_offset) {
 799         if (!allocate) {
 800             return -1;
 801         }
 802
 803         /* Avoid the L2 tables update for the images that have snapshots. */
 804         *cluster_offset = bdrv_getlength(extent->file);
 805         if (!extent->compressed) {
 806             bdrv_truncate(
 807                 extent->file,
 808                 *cluster_offset + (extent->cluster_sectors << 9)
 809             );
 810         }
 811
 812         *cluster_offset >>= 9;
 813         tmp = cpu_to_le32(*cluster_offset);
 814         l2_table[l2_index] = tmp;
 815
 816         /* First of all we write grain itself, to avoid race condition
 817          * that may to corrupt the image.
 818          * This problem may occur because of insufficient space on host disk
 819          * or inappropriate VM shutdown.
 820          */
 821         if (get_whole_cluster(
 822                 bs, extent, *cluster_offset, offset, allocate) == -1) {
 823             return -1;
 824         }
 825
 826         if (m_data) {
 827             m_data->offset = tmp;
 828             m_data->l1_index = l1_index;
 829             m_data->l2_index = l2_index;
 830             m_data->l2_offset = l2_offset;
 831             m_data->valid = 1;
 832         }
 833     }
 834     *cluster_offset <<= 9;
 835     return 0;
 836 }
 837
 838 static VmdkExtent *find_extent(BDRVVmdkState *s,
 839                                 int64_t sector_num, VmdkExtent *start_hint)
 840 {
 841     VmdkExtent *extent = start_hint;
 842
 843     if (!extent) {
 844         extent = &s->extents[0];
 845     }
 846     while (extent < &s->extents[s->num_extents]) {
 847         if (sector_num < extent->end_sector) {
 848             return extent;
 849         }
 850         extent++;
 851     }
 852     return NULL;
 853 }
 854
 855 static int vmdk_is_allocated(BlockDriverState *bs, int64_t sector_num,
 856                              int nb_sectors, int *pnum)
 857 {
 858     BDRVVmdkState *s = bs->opaque;
 859     int64_t index_in_cluster, n, ret;
 860     uint64_t offset;
 861     VmdkExtent *extent;
 862
 863     extent = find_extent(s, sector_num, NULL);
 864     if (!extent) {
 865         return 0;
 866     }
 867     ret = get_cluster_offset(bs, extent, NULL,
 868                             sector_num * 512, 0, &offset);
 869     /* get_cluster_offset returning 0 means success */
 870     ret = !ret;
 871
 872     index_in_cluster = sector_num % extent->cluster_sectors;
 873     n = extent->cluster_sectors - index_in_cluster;
 874     if (n > nb_sectors) {
 875         n = nb_sectors;
 876     }
 877     *pnum = n;
 878     return ret;
 879 }
 880
 881 static int vmdk_write_extent(VmdkExtent *extent, int64_t cluster_offset,
 882                             int64_t offset_in_cluster, const uint8_t *buf,
 883                             int nb_sectors, int64_t sector_num)
 884 {
 885     int ret;
 886     VmdkGrainMarker *data = NULL;
 887     uLongf buf_len;
 888     const uint8_t *write_buf = buf;
 889     int write_len = nb_sectors * 512;
 890
 891     if (extent->compressed) {
 892         if (!extent->has_marker) {
 893             ret = -EINVAL;
 894             goto out;
 895         }
 896         buf_len = (extent->cluster_sectors << 9) * 2;
 897         data = g_malloc(buf_len + sizeof(VmdkGrainMarker));
 898         if (compress(data->data, &buf_len, buf, nb_sectors << 9) != Z_OK ||
 899                 buf_len == 0) {
 900             ret = -EINVAL;
 901             goto out;
 902         }
 903         data->lba = sector_num;
 904         data->size = buf_len;
 905         write_buf = (uint8_t *)data;
 906         write_len = buf_len + sizeof(VmdkGrainMarker);
 907     }
 908     ret = bdrv_pwrite(extent->file,
 909                         cluster_offset + offset_in_cluster,
 910                         write_buf,
 911                         write_len);
 912     if (ret != write_len) {
 913         ret = ret < 0 ? ret : -EIO;
 914         goto out;
 915     }
 916     ret = 0;
 917  out:
 918     g_free(data);
 919     return ret;
 920 }
 921
 922 static int vmdk_read_extent(VmdkExtent *extent, int64_t cluster_offset,
 923                             int64_t offset_in_cluster, uint8_t *buf,
 924                             int nb_sectors)
 925 {
 926     int ret;
 927     int cluster_bytes, buf_bytes;
 928     uint8_t *cluster_buf, *compressed_data;
 929     uint8_t *uncomp_buf;
 930     uint32_t data_len;
 931     VmdkGrainMarker *marker;
 932     uLongf buf_len;
 933
 934
 935     if (!extent->compressed) {
 936         ret = bdrv_pread(extent->file,
 937                           cluster_offset + offset_in_cluster,
 938                           buf, nb_sectors * 512);
 939         if (ret == nb_sectors * 512) {
 940             return 0;
 941         } else {
 942             return -EIO;
 943         }
 944     }
 945     cluster_bytes = extent->cluster_sectors * 512;
 946     /* Read two clusters in case GrainMarker + compressed data > one cluster */
 947     buf_bytes = cluster_bytes * 2;
 948     cluster_buf = g_malloc(buf_bytes);
 949     uncomp_buf = g_malloc(cluster_bytes);
 950     ret = bdrv_pread(extent->file,
 951                 cluster_offset,
 952                 cluster_buf, buf_bytes);
 953     if (ret < 0) {
 954         goto out;
 955     }
 956     compressed_data = cluster_buf;
 957     buf_len = cluster_bytes;
 958     data_len = cluster_bytes;
 959     if (extent->has_marker) {
 960         marker = (VmdkGrainMarker *)cluster_buf;
 961         compressed_data = marker->data;
 962         data_len = le32_to_cpu(marker->size);
 963     }
 964     if (!data_len || data_len > buf_bytes) {
 965         ret = -EINVAL;
 966         goto out;
 967     }
 968     ret = uncompress(uncomp_buf, &buf_len, compressed_data, data_len);
 969     if (ret != Z_OK) {
 970         ret = -EINVAL;
 971         goto out;
 972
 973     }
 974     if (offset_in_cluster < 0 ||
 975             offset_in_cluster + nb_sectors * 512 > buf_len) {
 976         ret = -EINVAL;
 977         goto out;
 978     }
 979     memcpy(buf, uncomp_buf + offset_in_cluster, nb_sectors * 512);
 980     ret = 0;
 981
 982  out:
 983     g_free(uncomp_buf);
 984     g_free(cluster_buf);
 985     return ret;
 986 }
 987
 988 static int vmdk_read(BlockDriverState *bs, int64_t sector_num,
 989                     uint8_t *buf, int nb_sectors)
 990 {
 991     BDRVVmdkState *s = bs->opaque;
 992     int ret;
 993     uint64_t n, index_in_cluster;
 994     VmdkExtent *extent = NULL;
 995     uint64_t cluster_offset;
 996
 997     while (nb_sectors > 0) {
 998         extent = find_extent(s, sector_num, extent);
 999         if (!extent) {
1000             return -EIO;
1001         }
1002         ret = get_cluster_offset(
1003                             bs, extent, NULL,
1004                             sector_num << 9, 0, &cluster_offset);
1005         index_in_cluster = sector_num % extent->cluster_sectors;
1006         n = extent->cluster_sectors - index_in_cluster;
1007         if (n > nb_sectors) {
1008             n = nb_sectors;
1009         }
1010         if (ret) {
1011             /* if not allocated, try to read from parent image, if exist */
1012             if (bs->backing_hd) {
1013                 if (!vmdk_is_cid_valid(bs)) {
1014                     return -EINVAL;
1015                 }
1016                 ret = bdrv_read(bs->backing_hd, sector_num, buf, n);
1017                 if (ret < 0) {
1018                     return ret;
1019                 }
1020             } else {
1021                 memset(buf, 0, 512 * n);
1022             }
1023         } else {
1024             ret = vmdk_read_extent(extent,
1025                             cluster_offset, index_in_cluster * 512,
1026                             buf, n);
1027             if (ret) {
1028                 return ret;
1029             }
1030         }
1031         nb_sectors -= n;
1032         sector_num += n;
1033         buf += n * 512;
1034     }
1035     return 0;
1036 }
1037
1038 static coroutine_fn int vmdk_co_read(BlockDriverState *bs, int64_t sector_num,
1039                                      uint8_t *buf, int nb_sectors)
1040 {
1041     int ret;
1042     BDRVVmdkState *s = bs->opaque;
1043     qemu_co_mutex_lock(&s->lock);
1044     ret = vmdk_read(bs, sector_num, buf, nb_sectors);
1045     qemu_co_mutex_unlock(&s->lock);
1046     return ret;
1047 }
1048
1049 static int vmdk_write(BlockDriverState *bs, int64_t sector_num,
1050                      const uint8_t *buf, int nb_sectors)
1051 {
1052     BDRVVmdkState *s = bs->opaque;
1053     VmdkExtent *extent = NULL;
1054     int n, ret;
1055     int64_t index_in_cluster;
1056     uint64_t cluster_offset;
1057     VmdkMetaData m_data;
1058
1059     if (sector_num > bs->total_sectors) {
1060         fprintf(stderr,
1061                 "(VMDK) Wrong offset: sector_num=0x%" PRIx64
1062                 " total_sectors=0x%" PRIx64 "\n",
1063                 sector_num, bs->total_sectors);
1064         return -EIO;
1065     }
1066
1067     while (nb_sectors > 0) {
1068         extent = find_extent(s, sector_num, extent);
1069         if (!extent) {
1070             return -EIO;
1071         }
1072         ret = get_cluster_offset(
1073                                 bs,
1074                                 extent,
1075                                 &m_data,
1076                                 sector_num << 9, !extent->compressed,
1077                                 &cluster_offset);
1078         if (extent->compressed) {
1079             if (ret == 0) {
1080                 /* Refuse write to allocated cluster for streamOptimized */
1081                 fprintf(stderr,
1082                         "VMDK: can't write to allocated cluster"
1083                         " for streamOptimized\n");
1084                 return -EIO;
1085             } else {
1086                 /* allocate */
1087                 ret = get_cluster_offset(
1088                                         bs,
1089                                         extent,
1090                                         &m_data,
1091                                         sector_num << 9, 1,
1092                                         &cluster_offset);
1093             }
1094         }
1095         if (ret) {
1096             return -EINVAL;
1097         }
1098         index_in_cluster = sector_num % extent->cluster_sectors;
1099         n = extent->cluster_sectors - index_in_cluster;
1100         if (n > nb_sectors) {
1101             n = nb_sectors;
1102         }
1103
1104         ret = vmdk_write_extent(extent,
1105                         cluster_offset, index_in_cluster * 512,
1106                         buf, n, sector_num);
1107         if (ret) {
1108             return ret;
1109         }
1110         if (m_data.valid) {
1111             /* update L2 tables */
1112             if (vmdk_L2update(extent, &m_data) == -1) {
1113                 return -EIO;
1114             }
1115         }
1116         nb_sectors -= n;
1117         sector_num += n;
1118         buf += n * 512;
1119
1120         /* update CID on the first write every time the virtual disk is
1121          * opened */
1122         if (!s->cid_updated) {
1123             ret = vmdk_write_cid(bs, time(NULL));
1124             if (ret < 0) {
1125                 return ret;
1126             }
1127             s->cid_updated = true;
1128         }
1129     }
1130     return 0;
1131 }
1132
1133 static coroutine_fn int vmdk_co_write(BlockDriverState *bs, int64_t sector_num,
1134                                       const uint8_t *buf, int nb_sectors)
1135 {
1136     int ret;
1137     BDRVVmdkState *s = bs->opaque;
1138     qemu_co_mutex_lock(&s->lock);
1139     ret = vmdk_write(bs, sector_num, buf, nb_sectors);
1140     qemu_co_mutex_unlock(&s->lock);
1141     return ret;
1142 }
1143
1144
1145 static int vmdk_create_extent(const char *filename, int64_t filesize,
1146                               bool flat, bool compress)
1147 {
1148     int ret, i;
1149     int fd = 0;
1150     VMDK4Header header;
1151     uint32_t tmp, magic, grains, gd_size, gt_size, gt_count;
1152
1153     fd = open(
1154         filename,
1155         O_WRONLY | O_CREAT | O_TRUNC | O_BINARY | O_LARGEFILE,
1156         0644);
1157     if (fd < 0) {
1158         return -errno;
1159     }
1160     if (flat) {
1161         ret = ftruncate(fd, filesize);
1162         if (ret < 0) {
1163             ret = -errno;
1164         }
1165         goto exit;
1166     }
1167     magic = cpu_to_be32(VMDK4_MAGIC);
1168     memset(&header, 0, sizeof(header));
1169     header.version = 1;
1170     header.flags =
1171         3 | (compress ? VMDK4_FLAG_COMPRESS | VMDK4_FLAG_MARKER : 0);
1172     header.compressAlgorithm = compress ? VMDK4_COMPRESSION_DEFLATE : 0;
1173     header.capacity = filesize / 512;
1174     header.granularity = 128;
1175     header.num_gtes_per_gte = 512;
1176
1177     grains = (filesize / 512 + header.granularity - 1) / header.granularity;
1178     gt_size = ((header.num_gtes_per_gte * sizeof(uint32_t)) + 511) >> 9;
1179     gt_count =
1180         (grains + header.num_gtes_per_gte - 1) / header.num_gtes_per_gte;
1181     gd_size = (gt_count * sizeof(uint32_t) + 511) >> 9;
1182
1183     header.desc_offset = 1;
1184     header.desc_size = 20;
1185     header.rgd_offset = header.desc_offset + header.desc_size;
1186     header.gd_offset = header.rgd_offset + gd_size + (gt_size * gt_count);
1187     header.grain_offset =
1188        ((header.gd_offset + gd_size + (gt_size * gt_count) +
1189          header.granularity - 1) / header.granularity) *
1190         header.granularity;
1191     /* swap endianness for all header fields */
1192     header.version = cpu_to_le32(header.version);
1193     header.flags = cpu_to_le32(header.flags);
1194     header.capacity = cpu_to_le64(header.capacity);
1195     header.granularity = cpu_to_le64(header.granularity);
1196     header.num_gtes_per_gte = cpu_to_le32(header.num_gtes_per_gte);
1197     header.desc_offset = cpu_to_le64(header.desc_offset);
1198     header.desc_size = cpu_to_le64(header.desc_size);
1199     header.rgd_offset = cpu_to_le64(header.rgd_offset);
1200     header.gd_offset = cpu_to_le64(header.gd_offset);
1201     header.grain_offset = cpu_to_le64(header.grain_offset);
1202     header.compressAlgorithm = cpu_to_le16(header.compressAlgorithm);
1203
1204     header.check_bytes[0] = 0xa;
1205     header.check_bytes[1] = 0x20;
1206     header.check_bytes[2] = 0xd;
1207     header.check_bytes[3] = 0xa;
1208
1209     /* write all the data */
1210     ret = qemu_write_full(fd, &magic, sizeof(magic));
1211     if (ret != sizeof(magic)) {
1212         ret = -errno;
1213         goto exit;
1214     }
1215     ret = qemu_write_full(fd, &header, sizeof(header));
1216     if (ret != sizeof(header)) {
1217         ret = -errno;
1218         goto exit;
1219     }
1220
1221     ret = ftruncate(fd, le64_to_cpu(header.grain_offset) << 9);
1222     if (ret < 0) {
1223         ret = -errno;
1224         goto exit;
1225     }
1226
1227     /* write grain directory */
1228     lseek(fd, le64_to_cpu(header.rgd_offset) << 9, SEEK_SET);
1229     for (i = 0, tmp = le64_to_cpu(header.rgd_offset) + gd_size;
1230          i < gt_count; i++, tmp += gt_size) {
1231         ret = qemu_write_full(fd, &tmp, sizeof(tmp));
1232         if (ret != sizeof(tmp)) {
1233             ret = -errno;
1234             goto exit;
1235         }
1236     }
1237
1238     /* write backup grain directory */
1239     lseek(fd, le64_to_cpu(header.gd_offset) << 9, SEEK_SET);
1240     for (i = 0, tmp = le64_to_cpu(header.gd_offset) + gd_size;
1241          i < gt_count; i++, tmp += gt_size) {
1242         ret = qemu_write_full(fd, &tmp, sizeof(tmp));
1243         if (ret != sizeof(tmp)) {
1244             ret = -errno;
1245             goto exit;
1246         }
1247     }
1248
1249     ret = 0;
1250  exit:
1251     close(fd);
1252     return ret;
1253 }
1254
1255 static int filename_decompose(const char *filename, char *path, char *prefix,
1256         char *postfix, size_t buf_len)
1257 {
1258     const char *p, *q;
1259
1260     if (filename == NULL || !strlen(filename)) {
1261         fprintf(stderr, "Vmdk: no filename provided.\n");
1262         return -1;
1263     }
1264     p = strrchr(filename, '/');
1265     if (p == NULL) {
1266         p = strrchr(filename, '\\');
1267     }
1268     if (p == NULL) {
1269         p = strrchr(filename, ':');
1270     }
1271     if (p != NULL) {
1272         p++;
1273         if (p - filename >= buf_len) {
1274             return -1;
1275         }
1276         pstrcpy(path, p - filename + 1, filename);
1277     } else {
1278         p = filename;
1279         path[0] = '\0';
1280     }
1281     q = strrchr(p, '.');
1282     if (q == NULL) {
1283         pstrcpy(prefix, buf_len, p);
1284         postfix[0] = '\0';
1285     } else {
1286         if (q - p >= buf_len) {
1287             return -1;
1288         }
1289         pstrcpy(prefix, q - p + 1, p);
1290         pstrcpy(postfix, buf_len, q);
1291     }
1292     return 0;
1293 }
1294
1295 static int relative_path(char *dest, int dest_size,
1296         const char *base, const char *target)
1297 {
1298     int i = 0;
1299     int n = 0;
1300     const char *p, *q;
1301 #ifdef _WIN32
1302     const char *sep = "\\";
1303 #else
1304     const char *sep = "/";
1305 #endif
1306
1307     if (!(dest && base && target)) {
1308         return -1;
1309     }
1310     if (path_is_absolute(target)) {
1311         dest[dest_size - 1] = '\0';
1312         strncpy(dest, target, dest_size - 1);
1313         return 0;
1314     }
1315     while (base[i] == target[i]) {
1316         i++;
1317     }
1318     p = &base[i];
1319     q = &target[i];
1320     while (*p) {
1321         if (*p == *sep) {
1322             n++;
1323         }
1324         p++;
1325     }
1326     dest[0] = '\0';
1327     for (; n; n--) {
1328         pstrcat(dest, dest_size, "..");
1329         pstrcat(dest, dest_size, sep);
1330     }
1331     pstrcat(dest, dest_size, q);
1332     return 0;
1333 }
1334
1335 static int vmdk_create(const char *filename, QEMUOptionParameter *options)
1336 {
1337     int fd, idx = 0;
1338     char desc[BUF_SIZE];
1339     int64_t total_size = 0, filesize;
1340     const char *backing_file = NULL;
1341     const char *fmt = NULL;
1342     int flags = 0;
1343     int ret = 0;
1344     bool flat, split, compress;
1345     char ext_desc_lines[BUF_SIZE] = "";
1346     char path[PATH_MAX], prefix[PATH_MAX], postfix[PATH_MAX];
1347     const int64_t split_size = 0x80000000;  /* VMDK has constant split size */
1348     const char *desc_extent_line;
1349     char parent_desc_line[BUF_SIZE] = "";
1350     uint32_t parent_cid = 0xffffffff;
1351     const char desc_template[] =
1352         "# Disk DescriptorFile\n"
1353         "version=1\n"
1354         "CID=%x\n"
1355         "parentCID=%x\n"
1356         "createType=\"%s\"\n"
1357         "%s"
1358         "\n"
1359         "# Extent description\n"
1360         "%s"
1361         "\n"
1362         "# The Disk Data Base\n"
1363         "#DDB\n"
1364         "\n"
1365         "ddb.virtualHWVersion = \"%d\"\n"
1366         "ddb.geometry.cylinders = \"%" PRId64 "\"\n"
1367         "ddb.geometry.heads = \"16\"\n"
1368         "ddb.geometry.sectors = \"63\"\n"
1369         "ddb.adapterType = \"ide\"\n";
1370
1371     if (filename_decompose(filename, path, prefix, postfix, PATH_MAX)) {
1372         return -EINVAL;
1373     }
1374     /* Read out options */
1375     while (options && options->name) {
1376         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
1377             total_size = options->value.n;
1378         } else if (!strcmp(options->name, BLOCK_OPT_BACKING_FILE)) {
1379             backing_file = options->value.s;
1380         } else if (!strcmp(options->name, BLOCK_OPT_COMPAT6)) {
1381             flags |= options->value.n ? BLOCK_FLAG_COMPAT6 : 0;
1382         } else if (!strcmp(options->name, BLOCK_OPT_SUBFMT)) {
1383             fmt = options->value.s;
1384         }
1385         options++;
1386     }
1387     if (!fmt) {
1388         /* Default format to monolithicSparse */
1389         fmt = "monolithicSparse";
1390     } else if (strcmp(fmt, "monolithicFlat") &&
1391                strcmp(fmt, "monolithicSparse") &&
1392                strcmp(fmt, "twoGbMaxExtentSparse") &&
1393                strcmp(fmt, "twoGbMaxExtentFlat") &&
1394                strcmp(fmt, "streamOptimized")) {
1395         fprintf(stderr, "VMDK: Unknown subformat: %s\n", fmt);
1396         return -EINVAL;
1397     }
1398     split = !(strcmp(fmt, "twoGbMaxExtentFlat") &&
1399               strcmp(fmt, "twoGbMaxExtentSparse"));
1400     flat = !(strcmp(fmt, "monolithicFlat") &&
1401              strcmp(fmt, "twoGbMaxExtentFlat"));
1402     compress = !strcmp(fmt, "streamOptimized");
1403     if (flat) {
1404         desc_extent_line = "RW %lld FLAT \"%s\" 0\n";
1405     } else {
1406         desc_extent_line = "RW %lld SPARSE \"%s\"\n";
1407     }
1408     if (flat && backing_file) {
1409         /* not supporting backing file for flat image */
1410         return -ENOTSUP;
1411     }
1412     if (backing_file) {
1413         char parent_filename[PATH_MAX];
1414         BlockDriverState *bs = bdrv_new("");
1415         ret = bdrv_open(bs, backing_file, 0, NULL);
1416         if (ret != 0) {
1417             bdrv_delete(bs);
1418             return ret;
1419         }
1420         if (strcmp(bs->drv->format_name, "vmdk")) {
1421             bdrv_delete(bs);
1422             return -EINVAL;
1423         }
1424         parent_cid = vmdk_read_cid(bs, 0);
1425         bdrv_delete(bs);
1426         relative_path(parent_filename, sizeof(parent_filename),
1427                       filename, backing_file);
1428         snprintf(parent_desc_line, sizeof(parent_desc_line),
1429                 "parentFileNameHint=\"%s\"", parent_filename);
1430     }
1431
1432     /* Create extents */
1433     filesize = total_size;
1434     while (filesize > 0) {
1435         char desc_line[BUF_SIZE];
1436         char ext_filename[PATH_MAX];
1437         char desc_filename[PATH_MAX];
1438         int64_t size = filesize;
1439
1440         if (split && size > split_size) {
1441             size = split_size;
1442         }
1443         if (split) {
1444             snprintf(desc_filename, sizeof(desc_filename), "%s-%c%03d%s",
1445                     prefix, flat ? 'f' : 's', ++idx, postfix);
1446         } else if (flat) {
1447             snprintf(desc_filename, sizeof(desc_filename), "%s-flat%s",
1448                     prefix, postfix);
1449         } else {
1450             snprintf(desc_filename, sizeof(desc_filename), "%s%s",
1451                     prefix, postfix);
1452         }
1453         snprintf(ext_filename, sizeof(ext_filename), "%s%s",
1454                 path, desc_filename);
1455
1456         if (vmdk_create_extent(ext_filename, size, flat, compress)) {
1457             return -EINVAL;
1458         }
1459         filesize -= size;
1460
1461         /* Format description line */
1462         snprintf(desc_line, sizeof(desc_line),
1463                     desc_extent_line, size / 512, desc_filename);
1464         pstrcat(ext_desc_lines, sizeof(ext_desc_lines), desc_line);
1465     }
1466     /* generate descriptor file */
1467     snprintf(desc, sizeof(desc), desc_template,
1468             (unsigned int)time(NULL),
1469             parent_cid,
1470             fmt,
1471             parent_desc_line,
1472             ext_desc_lines,
1473             (flags & BLOCK_FLAG_COMPAT6 ? 6 : 4),
1474             total_size / (int64_t)(63 * 16 * 512));
1475     if (split || flat) {
1476         fd = open(
1477                 filename,
1478                 O_WRONLY | O_CREAT | O_TRUNC | O_BINARY | O_LARGEFILE,
1479                 0644);
1480     } else {
1481         fd = open(
1482                 filename,
1483                 O_WRONLY | O_BINARY | O_LARGEFILE,
1484                 0644);
1485     }
1486     if (fd < 0) {
1487         return -errno;
1488     }
1489     /* the descriptor offset = 0x200 */
1490     if (!split && !flat && 0x200 != lseek(fd, 0x200, SEEK_SET)) {
1491         ret = -errno;
1492         goto exit;
1493     }
1494     ret = qemu_write_full(fd, desc, strlen(desc));
1495     if (ret != strlen(desc)) {
1496         ret = -errno;
1497         goto exit;
1498     }
1499     ret = 0;
1500 exit:
1501     close(fd);
1502     return ret;
1503 }
1504
1505 static void vmdk_close(BlockDriverState *bs)
1506 {
1507     vmdk_free_extents(bs);
1508 }
1509
1510 static coroutine_fn int vmdk_co_flush(BlockDriverState *bs)
1511 {
1512     int i, ret, err;
1513     BDRVVmdkState *s = bs->opaque;
1514
1515     ret = bdrv_co_flush(bs->file);
1516     for (i = 0; i < s->num_extents; i++) {
1517         err = bdrv_co_flush(s->extents[i].file);
1518         if (err < 0) {
1519             ret = err;
1520         }
1521     }
1522     return ret;
1523 }
1524
1525 static int64_t vmdk_get_allocated_file_size(BlockDriverState *bs)
1526 {
1527     int i;
1528     int64_t ret = 0;
1529     int64_t r;
1530     BDRVVmdkState *s = bs->opaque;
1531
1532     ret = bdrv_get_allocated_file_size(bs->file);
1533     if (ret < 0) {
1534         return ret;
1535     }
1536     for (i = 0; i < s->num_extents; i++) {
1537         if (s->extents[i].file == bs->file) {
1538             continue;
1539         }
1540         r = bdrv_get_allocated_file_size(s->extents[i].file);
1541         if (r < 0) {
1542             return r;
1543         }
1544         ret += r;
1545     }
1546     return ret;
1547 }
1548
1549 static QEMUOptionParameter vmdk_create_options[] = {
1550     {
1551         .name = BLOCK_OPT_SIZE,
1552         .type = OPT_SIZE,
1553         .help = "Virtual disk size"
1554     },
1555     {
1556         .name = BLOCK_OPT_BACKING_FILE,
1557         .type = OPT_STRING,
1558         .help = "File name of a base image"
1559     },
1560     {
1561         .name = BLOCK_OPT_COMPAT6,
1562         .type = OPT_FLAG,
1563         .help = "VMDK version 6 image"
1564     },
1565     {
1566         .name = BLOCK_OPT_SUBFMT,
1567         .type = OPT_STRING,
1568         .help =
1569             "VMDK flat extent format, can be one of "
1570             "{monolithicSparse (default) | monolithicFlat | twoGbMaxExtentSparse | twoGbMaxExtentFlat | streamOptimized} "
1571     },
1572     { NULL }
1573 };
1574
1575 static BlockDriver bdrv_vmdk = {
1576     .format_name    = "vmdk",
1577     .instance_size  = sizeof(BDRVVmdkState),
1578     .bdrv_probe     = vmdk_probe,
1579     .bdrv_open      = vmdk_open,
1580     .bdrv_read      = vmdk_co_read,
1581     .bdrv_write     = vmdk_co_write,
1582     .bdrv_close     = vmdk_close,
1583     .bdrv_create    = vmdk_create,
1584     .bdrv_co_flush  = vmdk_co_flush,
1585     .bdrv_is_allocated  = vmdk_is_allocated,
1586     .bdrv_get_allocated_file_size  = vmdk_get_allocated_file_size,
1587
1588     .create_options = vmdk_create_options,
1589 };
1590
1591 static void bdrv_vmdk_init(void)
1592 {
1593     bdrv_register(&bdrv_vmdk);
1594 }
1595
1596 block_init(bdrv_vmdk_init);