block/vmdk.c

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