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