git.maemo.org Git - qemu/blob - block.c

   1 /*
   2  * QEMU System Emulator block driver
   3  *
   4  * Copyright (c) 2003 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include "config-host.h"
  25 #ifdef HOST_BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "monitor.h"
  32 #include "block_int.h"
  33
  34 #ifdef HOST_BSD
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/ioctl.h>
  38 #ifndef __DragonFly__
  39 #include <sys/disk.h>
  40 #endif
  41 #endif
  42
  43 #ifdef _WIN32
  44 #include <windows.h>
  45 #endif
  46
  47 #define SECTOR_BITS 9
  48 #define SECTOR_SIZE (1 << SECTOR_BITS)
  49
  50 typedef struct BlockDriverAIOCBSync {
  51     BlockDriverAIOCB common;
  52     QEMUBH *bh;
  53     int ret;
  54     /* vector translation state */
  55     QEMUIOVector *qiov;
  56     uint8_t *bounce;
  57     int is_write;
  58 } BlockDriverAIOCBSync;
  59
  60 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
  61         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  62         BlockDriverCompletionFunc *cb, void *opaque);
  63 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
  64         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  65         BlockDriverCompletionFunc *cb, void *opaque);
  66 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  67 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  68                         uint8_t *buf, int nb_sectors);
  69 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  70                          const uint8_t *buf, int nb_sectors);
  71
  72 BlockDriverState *bdrv_first;
  73
  74 static BlockDriver *first_drv;
  75
  76 int path_is_absolute(const char *path)
  77 {
  78     const char *p;
  79 #ifdef _WIN32
  80     /* specific case for names like: "\\.\d:" */
  81     if (*path == '/' || *path == '\\')
  82         return 1;
  83 #endif
  84     p = strchr(path, ':');
  85     if (p)
  86         p++;
  87     else
  88         p = path;
  89 #ifdef _WIN32
  90     return (*p == '/' || *p == '\\');
  91 #else
  92     return (*p == '/');
  93 #endif
  94 }
  95
  96 /* if filename is absolute, just copy it to dest. Otherwise, build a
  97    path to it by considering it is relative to base_path. URL are
  98    supported. */
  99 void path_combine(char *dest, int dest_size,
 100                   const char *base_path,
 101                   const char *filename)
 102 {
 103     const char *p, *p1;
 104     int len;
 105
 106     if (dest_size <= 0)
 107         return;
 108     if (path_is_absolute(filename)) {
 109         pstrcpy(dest, dest_size, filename);
 110     } else {
 111         p = strchr(base_path, ':');
 112         if (p)
 113             p++;
 114         else
 115             p = base_path;
 116         p1 = strrchr(base_path, '/');
 117 #ifdef _WIN32
 118         {
 119             const char *p2;
 120             p2 = strrchr(base_path, '\\');
 121             if (!p1 || p2 > p1)
 122                 p1 = p2;
 123         }
 124 #endif
 125         if (p1)
 126             p1++;
 127         else
 128             p1 = base_path;
 129         if (p1 > p)
 130             p = p1;
 131         len = p - base_path;
 132         if (len > dest_size - 1)
 133             len = dest_size - 1;
 134         memcpy(dest, base_path, len);
 135         dest[len] = '\0';
 136         pstrcat(dest, dest_size, filename);
 137     }
 138 }
 139
 140
 141 static void bdrv_register(BlockDriver *bdrv)
 142 {
 143     if (!bdrv->bdrv_aio_readv) {
 144         /* add AIO emulation layer */
 145         bdrv->bdrv_aio_readv = bdrv_aio_readv_em;
 146         bdrv->bdrv_aio_writev = bdrv_aio_writev_em;
 147         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 148         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 149     } else if (!bdrv->bdrv_read) {
 150         /* add synchronous IO emulation layer */
 151         bdrv->bdrv_read = bdrv_read_em;
 152         bdrv->bdrv_write = bdrv_write_em;
 153     }
 154     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 155     bdrv->next = first_drv;
 156     first_drv = bdrv;
 157 }
 158
 159 /* create a new block device (by default it is empty) */
 160 BlockDriverState *bdrv_new(const char *device_name)
 161 {
 162     BlockDriverState **pbs, *bs;
 163
 164     bs = qemu_mallocz(sizeof(BlockDriverState));
 165     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 166     if (device_name[0] != '\0') {
 167         /* insert at the end */
 168         pbs = &bdrv_first;
 169         while (*pbs != NULL)
 170             pbs = &(*pbs)->next;
 171         *pbs = bs;
 172     }
 173     return bs;
 174 }
 175
 176 BlockDriver *bdrv_find_format(const char *format_name)
 177 {
 178     BlockDriver *drv1;
 179     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 180         if (!strcmp(drv1->format_name, format_name))
 181             return drv1;
 182     }
 183     return NULL;
 184 }
 185
 186 int bdrv_create2(BlockDriver *drv,
 187                 const char *filename, int64_t size_in_sectors,
 188                 const char *backing_file, const char *backing_format,
 189                 int flags)
 190 {
 191     if (drv->bdrv_create2)
 192         return drv->bdrv_create2(filename, size_in_sectors, backing_file,
 193                                  backing_format, flags);
 194     if (drv->bdrv_create)
 195         return drv->bdrv_create(filename, size_in_sectors, backing_file,
 196                                 flags);
 197     return -ENOTSUP;
 198 }
 199
 200 int bdrv_create(BlockDriver *drv,
 201                 const char *filename, int64_t size_in_sectors,
 202                 const char *backing_file, int flags)
 203 {
 204     if (!drv->bdrv_create)
 205         return -ENOTSUP;
 206     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 207 }
 208
 209 #ifdef _WIN32
 210 void get_tmp_filename(char *filename, int size)
 211 {
 212     char temp_dir[MAX_PATH];
 213
 214     GetTempPath(MAX_PATH, temp_dir);
 215     GetTempFileName(temp_dir, "qem", 0, filename);
 216 }
 217 #else
 218 void get_tmp_filename(char *filename, int size)
 219 {
 220     int fd;
 221     const char *tmpdir;
 222     /* XXX: race condition possible */
 223     tmpdir = getenv("TMPDIR");
 224     if (!tmpdir)
 225         tmpdir = "/tmp";
 226     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 227     fd = mkstemp(filename);
 228     close(fd);
 229 }
 230 #endif
 231
 232 #ifdef _WIN32
 233 static int is_windows_drive_prefix(const char *filename)
 234 {
 235     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 236              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 237             filename[1] == ':');
 238 }
 239
 240 static int is_windows_drive(const char *filename)
 241 {
 242     if (is_windows_drive_prefix(filename) &&
 243         filename[2] == '\0')
 244         return 1;
 245     if (strstart(filename, "\\\\.\\", NULL) ||
 246         strstart(filename, "//./", NULL))
 247         return 1;
 248     return 0;
 249 }
 250 #endif
 251
 252 static BlockDriver *find_protocol(const char *filename)
 253 {
 254     BlockDriver *drv1;
 255     char protocol[128];
 256     int len;
 257     const char *p;
 258
 259 #ifdef _WIN32
 260     if (is_windows_drive(filename) ||
 261         is_windows_drive_prefix(filename))
 262         return &bdrv_raw;
 263 #endif
 264     p = strchr(filename, ':');
 265     if (!p)
 266         return &bdrv_raw;
 267     len = p - filename;
 268     if (len > sizeof(protocol) - 1)
 269         len = sizeof(protocol) - 1;
 270     memcpy(protocol, filename, len);
 271     protocol[len] = '\0';
 272     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 273         if (drv1->protocol_name &&
 274             !strcmp(drv1->protocol_name, protocol))
 275             return drv1;
 276     }
 277     return NULL;
 278 }
 279
 280 /* XXX: force raw format if block or character device ? It would
 281    simplify the BSD case */
 282 static BlockDriver *find_image_format(const char *filename)
 283 {
 284     int ret, score, score_max;
 285     BlockDriver *drv1, *drv;
 286     uint8_t buf[2048];
 287     BlockDriverState *bs;
 288
 289     /* detect host devices. By convention, /dev/cdrom[N] is always
 290        recognized as a host CDROM */
 291     if (strstart(filename, "/dev/cdrom", NULL))
 292         return &bdrv_host_device;
 293 #ifdef _WIN32
 294     if (is_windows_drive(filename))
 295         return &bdrv_host_device;
 296 #else
 297     {
 298         struct stat st;
 299         if (stat(filename, &st) >= 0 &&
 300             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 301             return &bdrv_host_device;
 302         }
 303     }
 304 #endif
 305
 306     drv = find_protocol(filename);
 307     /* no need to test disk image formats for vvfat */
 308     if (drv == &bdrv_vvfat)
 309         return drv;
 310
 311     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 312     if (ret < 0)
 313         return NULL;
 314     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 315     bdrv_delete(bs);
 316     if (ret < 0) {
 317         return NULL;
 318     }
 319
 320     score_max = 0;
 321     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 322         if (drv1->bdrv_probe) {
 323             score = drv1->bdrv_probe(buf, ret, filename);
 324             if (score > score_max) {
 325                 score_max = score;
 326                 drv = drv1;
 327             }
 328         }
 329     }
 330     return drv;
 331 }
 332
 333 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 334 {
 335     BlockDriverState *bs;
 336     int ret;
 337
 338     bs = bdrv_new("");
 339     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 340     if (ret < 0) {
 341         bdrv_delete(bs);
 342         return ret;
 343     }
 344     bs->growable = 1;
 345     *pbs = bs;
 346     return 0;
 347 }
 348
 349 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 350 {
 351     return bdrv_open2(bs, filename, flags, NULL);
 352 }
 353
 354 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 355                BlockDriver *drv)
 356 {
 357     int ret, open_flags;
 358     char tmp_filename[PATH_MAX];
 359     char backing_filename[PATH_MAX];
 360
 361     bs->read_only = 0;
 362     bs->is_temporary = 0;
 363     bs->encrypted = 0;
 364     bs->valid_key = 0;
 365     /* buffer_alignment defaulted to 512, drivers can change this value */
 366     bs->buffer_alignment = 512;
 367
 368     if (flags & BDRV_O_SNAPSHOT) {
 369         BlockDriverState *bs1;
 370         int64_t total_size;
 371         int is_protocol = 0;
 372
 373         /* if snapshot, we create a temporary backing file and open it
 374            instead of opening 'filename' directly */
 375
 376         /* if there is a backing file, use it */
 377         bs1 = bdrv_new("");
 378         ret = bdrv_open2(bs1, filename, 0, drv);
 379         if (ret < 0) {
 380             bdrv_delete(bs1);
 381             return ret;
 382         }
 383         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 384
 385         if (bs1->drv && bs1->drv->protocol_name)
 386             is_protocol = 1;
 387
 388         bdrv_delete(bs1);
 389
 390         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 391
 392         /* Real path is meaningless for protocols */
 393         if (is_protocol)
 394             snprintf(backing_filename, sizeof(backing_filename),
 395                      "%s", filename);
 396         else
 397             realpath(filename, backing_filename);
 398
 399         ret = bdrv_create2(&bdrv_qcow2, tmp_filename,
 400                            total_size, backing_filename,
 401                            (drv ? drv->format_name : NULL), 0);
 402         if (ret < 0) {
 403             return ret;
 404         }
 405         filename = tmp_filename;
 406         drv = &bdrv_qcow2;
 407         bs->is_temporary = 1;
 408     }
 409
 410     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 411     if (flags & BDRV_O_FILE) {
 412         drv = find_protocol(filename);
 413     } else if (!drv) {
 414         drv = find_image_format(filename);
 415     }
 416     if (!drv) {
 417         ret = -ENOENT;
 418         goto unlink_and_fail;
 419     }
 420     bs->drv = drv;
 421     bs->opaque = qemu_mallocz(drv->instance_size);
 422     /* Note: for compatibility, we open disk image files as RDWR, and
 423        RDONLY as fallback */
 424     if (!(flags & BDRV_O_FILE))
 425         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 426     else
 427         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 428     ret = drv->bdrv_open(bs, filename, open_flags);
 429     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 430         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 431         bs->read_only = 1;
 432     }
 433     if (ret < 0) {
 434         qemu_free(bs->opaque);
 435         bs->opaque = NULL;
 436         bs->drv = NULL;
 437     unlink_and_fail:
 438         if (bs->is_temporary)
 439             unlink(filename);
 440         return ret;
 441     }
 442     if (drv->bdrv_getlength) {
 443         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 444     }
 445 #ifndef _WIN32
 446     if (bs->is_temporary) {
 447         unlink(filename);
 448     }
 449 #endif
 450     if (bs->backing_file[0] != '\0') {
 451         /* if there is a backing file, use it */
 452         BlockDriver *back_drv = NULL;
 453         bs->backing_hd = bdrv_new("");
 454         path_combine(backing_filename, sizeof(backing_filename),
 455                      filename, bs->backing_file);
 456         if (bs->backing_format[0] != '\0')
 457             back_drv = bdrv_find_format(bs->backing_format);
 458         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 459                          back_drv);
 460         if (ret < 0) {
 461             bdrv_close(bs);
 462             return ret;
 463         }
 464     }
 465
 466     if (!bdrv_key_required(bs)) {
 467         /* call the change callback */
 468         bs->media_changed = 1;
 469         if (bs->change_cb)
 470             bs->change_cb(bs->change_opaque);
 471     }
 472     return 0;
 473 }
 474
 475 void bdrv_close(BlockDriverState *bs)
 476 {
 477     if (bs->drv) {
 478         if (bs->backing_hd)
 479             bdrv_delete(bs->backing_hd);
 480         bs->drv->bdrv_close(bs);
 481         qemu_free(bs->opaque);
 482 #ifdef _WIN32
 483         if (bs->is_temporary) {
 484             unlink(bs->filename);
 485         }
 486 #endif
 487         bs->opaque = NULL;
 488         bs->drv = NULL;
 489
 490         /* call the change callback */
 491         bs->media_changed = 1;
 492         if (bs->change_cb)
 493             bs->change_cb(bs->change_opaque);
 494     }
 495 }
 496
 497 void bdrv_delete(BlockDriverState *bs)
 498 {
 499     BlockDriverState **pbs;
 500
 501     pbs = &bdrv_first;
 502     while (*pbs != bs && *pbs != NULL)
 503         pbs = &(*pbs)->next;
 504     if (*pbs == bs)
 505         *pbs = bs->next;
 506
 507     bdrv_close(bs);
 508     qemu_free(bs);
 509 }
 510
 511 /*
 512  * Run consistency checks on an image
 513  *
 514  * Returns the number of errors or -errno when an internal error occurs
 515  */
 516 int bdrv_check(BlockDriverState *bs)
 517 {
 518     if (bs->drv->bdrv_check == NULL) {
 519         return -ENOTSUP;
 520     }
 521
 522     return bs->drv->bdrv_check(bs);
 523 }
 524
 525 /* commit COW file into the raw image */
 526 int bdrv_commit(BlockDriverState *bs)
 527 {
 528     BlockDriver *drv = bs->drv;
 529     int64_t i, total_sectors;
 530     int n, j;
 531     unsigned char sector[512];
 532
 533     if (!drv)
 534         return -ENOMEDIUM;
 535
 536     if (bs->read_only) {
 537         return -EACCES;
 538     }
 539
 540     if (!bs->backing_hd) {
 541         return -ENOTSUP;
 542     }
 543
 544     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 545     for (i = 0; i < total_sectors;) {
 546         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 547             for(j = 0; j < n; j++) {
 548                 if (bdrv_read(bs, i, sector, 1) != 0) {
 549                     return -EIO;
 550                 }
 551
 552                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 553                     return -EIO;
 554                 }
 555                 i++;
 556             }
 557         } else {
 558             i += n;
 559         }
 560     }
 561
 562     if (drv->bdrv_make_empty)
 563         return drv->bdrv_make_empty(bs);
 564
 565     return 0;
 566 }
 567
 568 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 569                                    size_t size)
 570 {
 571     int64_t len;
 572
 573     if (!bdrv_is_inserted(bs))
 574         return -ENOMEDIUM;
 575
 576     if (bs->growable)
 577         return 0;
 578
 579     len = bdrv_getlength(bs);
 580
 581     if ((offset + size) > len)
 582         return -EIO;
 583
 584     return 0;
 585 }
 586
 587 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 588                               int nb_sectors)
 589 {
 590     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 591 }
 592
 593 /* return < 0 if error. See bdrv_write() for the return codes */
 594 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 595               uint8_t *buf, int nb_sectors)
 596 {
 597     BlockDriver *drv = bs->drv;
 598
 599     if (!drv)
 600         return -ENOMEDIUM;
 601     if (bdrv_check_request(bs, sector_num, nb_sectors))
 602         return -EIO;
 603
 604     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 605 }
 606
 607 /* Return < 0 if error. Important errors are:
 608   -EIO         generic I/O error (may happen for all errors)
 609   -ENOMEDIUM   No media inserted.
 610   -EINVAL      Invalid sector number or nb_sectors
 611   -EACCES      Trying to write a read-only device
 612 */
 613 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 614                const uint8_t *buf, int nb_sectors)
 615 {
 616     BlockDriver *drv = bs->drv;
 617     if (!bs->drv)
 618         return -ENOMEDIUM;
 619     if (bs->read_only)
 620         return -EACCES;
 621     if (bdrv_check_request(bs, sector_num, nb_sectors))
 622         return -EIO;
 623
 624     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 625 }
 626
 627 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 628                void *buf, int count1)
 629 {
 630     uint8_t tmp_buf[SECTOR_SIZE];
 631     int len, nb_sectors, count;
 632     int64_t sector_num;
 633
 634     count = count1;
 635     /* first read to align to sector start */
 636     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 637     if (len > count)
 638         len = count;
 639     sector_num = offset >> SECTOR_BITS;
 640     if (len > 0) {
 641         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 642             return -EIO;
 643         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 644         count -= len;
 645         if (count == 0)
 646             return count1;
 647         sector_num++;
 648         buf += len;
 649     }
 650
 651     /* read the sectors "in place" */
 652     nb_sectors = count >> SECTOR_BITS;
 653     if (nb_sectors > 0) {
 654         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 655             return -EIO;
 656         sector_num += nb_sectors;
 657         len = nb_sectors << SECTOR_BITS;
 658         buf += len;
 659         count -= len;
 660     }
 661
 662     /* add data from the last sector */
 663     if (count > 0) {
 664         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 665             return -EIO;
 666         memcpy(buf, tmp_buf, count);
 667     }
 668     return count1;
 669 }
 670
 671 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 672                 const void *buf, int count1)
 673 {
 674     uint8_t tmp_buf[SECTOR_SIZE];
 675     int len, nb_sectors, count;
 676     int64_t sector_num;
 677
 678     count = count1;
 679     /* first write to align to sector start */
 680     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 681     if (len > count)
 682         len = count;
 683     sector_num = offset >> SECTOR_BITS;
 684     if (len > 0) {
 685         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 686             return -EIO;
 687         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 688         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 689             return -EIO;
 690         count -= len;
 691         if (count == 0)
 692             return count1;
 693         sector_num++;
 694         buf += len;
 695     }
 696
 697     /* write the sectors "in place" */
 698     nb_sectors = count >> SECTOR_BITS;
 699     if (nb_sectors > 0) {
 700         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 701             return -EIO;
 702         sector_num += nb_sectors;
 703         len = nb_sectors << SECTOR_BITS;
 704         buf += len;
 705         count -= len;
 706     }
 707
 708     /* add data from the last sector */
 709     if (count > 0) {
 710         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 711             return -EIO;
 712         memcpy(tmp_buf, buf, count);
 713         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 714             return -EIO;
 715     }
 716     return count1;
 717 }
 718
 719 /**
 720  * Truncate file to 'offset' bytes (needed only for file protocols)
 721  */
 722 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 723 {
 724     BlockDriver *drv = bs->drv;
 725     if (!drv)
 726         return -ENOMEDIUM;
 727     if (!drv->bdrv_truncate)
 728         return -ENOTSUP;
 729     return drv->bdrv_truncate(bs, offset);
 730 }
 731
 732 /**
 733  * Length of a file in bytes. Return < 0 if error or unknown.
 734  */
 735 int64_t bdrv_getlength(BlockDriverState *bs)
 736 {
 737     BlockDriver *drv = bs->drv;
 738     if (!drv)
 739         return -ENOMEDIUM;
 740     if (!drv->bdrv_getlength) {
 741         /* legacy mode */
 742         return bs->total_sectors * SECTOR_SIZE;
 743     }
 744     return drv->bdrv_getlength(bs);
 745 }
 746
 747 /* return 0 as number of sectors if no device present or error */
 748 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 749 {
 750     int64_t length;
 751     length = bdrv_getlength(bs);
 752     if (length < 0)
 753         length = 0;
 754     else
 755         length = length >> SECTOR_BITS;
 756     *nb_sectors_ptr = length;
 757 }
 758
 759 struct partition {
 760         uint8_t boot_ind;           /* 0x80 - active */
 761         uint8_t head;               /* starting head */
 762         uint8_t sector;             /* starting sector */
 763         uint8_t cyl;                /* starting cylinder */
 764         uint8_t sys_ind;            /* What partition type */
 765         uint8_t end_head;           /* end head */
 766         uint8_t end_sector;         /* end sector */
 767         uint8_t end_cyl;            /* end cylinder */
 768         uint32_t start_sect;        /* starting sector counting from 0 */
 769         uint32_t nr_sects;          /* nr of sectors in partition */
 770 } __attribute__((packed));
 771
 772 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 773 static int guess_disk_lchs(BlockDriverState *bs,
 774                            int *pcylinders, int *pheads, int *psectors)
 775 {
 776     uint8_t buf[512];
 777     int ret, i, heads, sectors, cylinders;
 778     struct partition *p;
 779     uint32_t nr_sects;
 780     uint64_t nb_sectors;
 781
 782     bdrv_get_geometry(bs, &nb_sectors);
 783
 784     ret = bdrv_read(bs, 0, buf, 1);
 785     if (ret < 0)
 786         return -1;
 787     /* test msdos magic */
 788     if (buf[510] != 0x55 || buf[511] != 0xaa)
 789         return -1;
 790     for(i = 0; i < 4; i++) {
 791         p = ((struct partition *)(buf + 0x1be)) + i;
 792         nr_sects = le32_to_cpu(p->nr_sects);
 793         if (nr_sects && p->end_head) {
 794             /* We make the assumption that the partition terminates on
 795                a cylinder boundary */
 796             heads = p->end_head + 1;
 797             sectors = p->end_sector & 63;
 798             if (sectors == 0)
 799                 continue;
 800             cylinders = nb_sectors / (heads * sectors);
 801             if (cylinders < 1 || cylinders > 16383)
 802                 continue;
 803             *pheads = heads;
 804             *psectors = sectors;
 805             *pcylinders = cylinders;
 806 #if 0
 807             printf("guessed geometry: LCHS=%d %d %d\n",
 808                    cylinders, heads, sectors);
 809 #endif
 810             return 0;
 811         }
 812     }
 813     return -1;
 814 }
 815
 816 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 817 {
 818     int translation, lba_detected = 0;
 819     int cylinders, heads, secs;
 820     uint64_t nb_sectors;
 821
 822     /* if a geometry hint is available, use it */
 823     bdrv_get_geometry(bs, &nb_sectors);
 824     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 825     translation = bdrv_get_translation_hint(bs);
 826     if (cylinders != 0) {
 827         *pcyls = cylinders;
 828         *pheads = heads;
 829         *psecs = secs;
 830     } else {
 831         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 832             if (heads > 16) {
 833                 /* if heads > 16, it means that a BIOS LBA
 834                    translation was active, so the default
 835                    hardware geometry is OK */
 836                 lba_detected = 1;
 837                 goto default_geometry;
 838             } else {
 839                 *pcyls = cylinders;
 840                 *pheads = heads;
 841                 *psecs = secs;
 842                 /* disable any translation to be in sync with
 843                    the logical geometry */
 844                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 845                     bdrv_set_translation_hint(bs,
 846                                               BIOS_ATA_TRANSLATION_NONE);
 847                 }
 848             }
 849         } else {
 850         default_geometry:
 851             /* if no geometry, use a standard physical disk geometry */
 852             cylinders = nb_sectors / (16 * 63);
 853
 854             if (cylinders > 16383)
 855                 cylinders = 16383;
 856             else if (cylinders < 2)
 857                 cylinders = 2;
 858             *pcyls = cylinders;
 859             *pheads = 16;
 860             *psecs = 63;
 861             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 862                 if ((*pcyls * *pheads) <= 131072) {
 863                     bdrv_set_translation_hint(bs,
 864                                               BIOS_ATA_TRANSLATION_LARGE);
 865                 } else {
 866                     bdrv_set_translation_hint(bs,
 867                                               BIOS_ATA_TRANSLATION_LBA);
 868                 }
 869             }
 870         }
 871         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 872     }
 873 }
 874
 875 void bdrv_set_geometry_hint(BlockDriverState *bs,
 876                             int cyls, int heads, int secs)
 877 {
 878     bs->cyls = cyls;
 879     bs->heads = heads;
 880     bs->secs = secs;
 881 }
 882
 883 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 884 {
 885     bs->type = type;
 886     bs->removable = ((type == BDRV_TYPE_CDROM ||
 887                       type == BDRV_TYPE_FLOPPY));
 888 }
 889
 890 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 891 {
 892     bs->translation = translation;
 893 }
 894
 895 void bdrv_get_geometry_hint(BlockDriverState *bs,
 896                             int *pcyls, int *pheads, int *psecs)
 897 {
 898     *pcyls = bs->cyls;
 899     *pheads = bs->heads;
 900     *psecs = bs->secs;
 901 }
 902
 903 int bdrv_get_type_hint(BlockDriverState *bs)
 904 {
 905     return bs->type;
 906 }
 907
 908 int bdrv_get_translation_hint(BlockDriverState *bs)
 909 {
 910     return bs->translation;
 911 }
 912
 913 int bdrv_is_removable(BlockDriverState *bs)
 914 {
 915     return bs->removable;
 916 }
 917
 918 int bdrv_is_read_only(BlockDriverState *bs)
 919 {
 920     return bs->read_only;
 921 }
 922
 923 int bdrv_is_sg(BlockDriverState *bs)
 924 {
 925     return bs->sg;
 926 }
 927
 928 /* XXX: no longer used */
 929 void bdrv_set_change_cb(BlockDriverState *bs,
 930                         void (*change_cb)(void *opaque), void *opaque)
 931 {
 932     bs->change_cb = change_cb;
 933     bs->change_opaque = opaque;
 934 }
 935
 936 int bdrv_is_encrypted(BlockDriverState *bs)
 937 {
 938     if (bs->backing_hd && bs->backing_hd->encrypted)
 939         return 1;
 940     return bs->encrypted;
 941 }
 942
 943 int bdrv_key_required(BlockDriverState *bs)
 944 {
 945     BlockDriverState *backing_hd = bs->backing_hd;
 946
 947     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 948         return 1;
 949     return (bs->encrypted && !bs->valid_key);
 950 }
 951
 952 int bdrv_set_key(BlockDriverState *bs, const char *key)
 953 {
 954     int ret;
 955     if (bs->backing_hd && bs->backing_hd->encrypted) {
 956         ret = bdrv_set_key(bs->backing_hd, key);
 957         if (ret < 0)
 958             return ret;
 959         if (!bs->encrypted)
 960             return 0;
 961     }
 962     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 963         return -1;
 964     ret = bs->drv->bdrv_set_key(bs, key);
 965     if (ret < 0) {
 966         bs->valid_key = 0;
 967     } else if (!bs->valid_key) {
 968         bs->valid_key = 1;
 969         /* call the change callback now, we skipped it on open */
 970         bs->media_changed = 1;
 971         if (bs->change_cb)
 972             bs->change_cb(bs->change_opaque);
 973     }
 974     return ret;
 975 }
 976
 977 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 978 {
 979     if (!bs->drv) {
 980         buf[0] = '\0';
 981     } else {
 982         pstrcpy(buf, buf_size, bs->drv->format_name);
 983     }
 984 }
 985
 986 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 987                          void *opaque)
 988 {
 989     BlockDriver *drv;
 990
 991     for (drv = first_drv; drv != NULL; drv = drv->next) {
 992         it(opaque, drv->format_name);
 993     }
 994 }
 995
 996 BlockDriverState *bdrv_find(const char *name)
 997 {
 998     BlockDriverState *bs;
 999
1000     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1001         if (!strcmp(name, bs->device_name))
1002             return bs;
1003     }
1004     return NULL;
1005 }
1006
1007 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1008 {
1009     BlockDriverState *bs;
1010
1011     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1012         it(opaque, bs);
1013     }
1014 }
1015
1016 const char *bdrv_get_device_name(BlockDriverState *bs)
1017 {
1018     return bs->device_name;
1019 }
1020
1021 void bdrv_flush(BlockDriverState *bs)
1022 {
1023     if (!bs->drv)
1024         return;
1025     if (bs->drv->bdrv_flush)
1026         bs->drv->bdrv_flush(bs);
1027     if (bs->backing_hd)
1028         bdrv_flush(bs->backing_hd);
1029 }
1030
1031 void bdrv_flush_all(void)
1032 {
1033     BlockDriverState *bs;
1034
1035     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1036         if (bs->drv && !bdrv_is_read_only(bs) &&
1037             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1038             bdrv_flush(bs);
1039 }
1040
1041 /*
1042  * Returns true iff the specified sector is present in the disk image. Drivers
1043  * not implementing the functionality are assumed to not support backing files,
1044  * hence all their sectors are reported as allocated.
1045  *
1046  * 'pnum' is set to the number of sectors (including and immediately following
1047  * the specified sector) that are known to be in the same
1048  * allocated/unallocated state.
1049  *
1050  * 'nb_sectors' is the max value 'pnum' should be set to.
1051  */
1052 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1053         int *pnum)
1054 {
1055     int64_t n;
1056     if (!bs->drv->bdrv_is_allocated) {
1057         if (sector_num >= bs->total_sectors) {
1058             *pnum = 0;
1059             return 0;
1060         }
1061         n = bs->total_sectors - sector_num;
1062         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1063         return 1;
1064     }
1065     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1066 }
1067
1068 void bdrv_info(Monitor *mon)
1069 {
1070     BlockDriverState *bs;
1071
1072     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1073         monitor_printf(mon, "%s:", bs->device_name);
1074         monitor_printf(mon, " type=");
1075         switch(bs->type) {
1076         case BDRV_TYPE_HD:
1077             monitor_printf(mon, "hd");
1078             break;
1079         case BDRV_TYPE_CDROM:
1080             monitor_printf(mon, "cdrom");
1081             break;
1082         case BDRV_TYPE_FLOPPY:
1083             monitor_printf(mon, "floppy");
1084             break;
1085         }
1086         monitor_printf(mon, " removable=%d", bs->removable);
1087         if (bs->removable) {
1088             monitor_printf(mon, " locked=%d", bs->locked);
1089         }
1090         if (bs->drv) {
1091             monitor_printf(mon, " file=");
1092             monitor_print_filename(mon, bs->filename);
1093             if (bs->backing_file[0] != '\0') {
1094                 monitor_printf(mon, " backing_file=");
1095                 monitor_print_filename(mon, bs->backing_file);
1096             }
1097             monitor_printf(mon, " ro=%d", bs->read_only);
1098             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1099             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1100         } else {
1101             monitor_printf(mon, " [not inserted]");
1102         }
1103         monitor_printf(mon, "\n");
1104     }
1105 }
1106
1107 /* The "info blockstats" command. */
1108 void bdrv_info_stats(Monitor *mon)
1109 {
1110     BlockDriverState *bs;
1111
1112     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1113         monitor_printf(mon, "%s:"
1114                        " rd_bytes=%" PRIu64
1115                        " wr_bytes=%" PRIu64
1116                        " rd_operations=%" PRIu64
1117                        " wr_operations=%" PRIu64
1118                        "\n",
1119                        bs->device_name,
1120                        bs->rd_bytes, bs->wr_bytes,
1121                        bs->rd_ops, bs->wr_ops);
1122     }
1123 }
1124
1125 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1126 {
1127     if (bs->backing_hd && bs->backing_hd->encrypted)
1128         return bs->backing_file;
1129     else if (bs->encrypted)
1130         return bs->filename;
1131     else
1132         return NULL;
1133 }
1134
1135 void bdrv_get_backing_filename(BlockDriverState *bs,
1136                                char *filename, int filename_size)
1137 {
1138     if (!bs->backing_hd) {
1139         pstrcpy(filename, filename_size, "");
1140     } else {
1141         pstrcpy(filename, filename_size, bs->backing_file);
1142     }
1143 }
1144
1145 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1146                           const uint8_t *buf, int nb_sectors)
1147 {
1148     BlockDriver *drv = bs->drv;
1149     if (!drv)
1150         return -ENOMEDIUM;
1151     if (!drv->bdrv_write_compressed)
1152         return -ENOTSUP;
1153     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1154 }
1155
1156 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1157 {
1158     BlockDriver *drv = bs->drv;
1159     if (!drv)
1160         return -ENOMEDIUM;
1161     if (!drv->bdrv_get_info)
1162         return -ENOTSUP;
1163     memset(bdi, 0, sizeof(*bdi));
1164     return drv->bdrv_get_info(bs, bdi);
1165 }
1166
1167 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1168 {
1169     BlockDriver *drv = bs->drv;
1170     if (!drv)
1171         return -ENOMEDIUM;
1172     if (!drv->bdrv_put_buffer)
1173         return -ENOTSUP;
1174     return drv->bdrv_put_buffer(bs, buf, pos, size);
1175 }
1176
1177 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1178 {
1179     BlockDriver *drv = bs->drv;
1180     if (!drv)
1181         return -ENOMEDIUM;
1182     if (!drv->bdrv_get_buffer)
1183         return -ENOTSUP;
1184     return drv->bdrv_get_buffer(bs, buf, pos, size);
1185 }
1186
1187 /**************************************************************/
1188 /* handling of snapshots */
1189
1190 int bdrv_snapshot_create(BlockDriverState *bs,
1191                          QEMUSnapshotInfo *sn_info)
1192 {
1193     BlockDriver *drv = bs->drv;
1194     if (!drv)
1195         return -ENOMEDIUM;
1196     if (!drv->bdrv_snapshot_create)
1197         return -ENOTSUP;
1198     return drv->bdrv_snapshot_create(bs, sn_info);
1199 }
1200
1201 int bdrv_snapshot_goto(BlockDriverState *bs,
1202                        const char *snapshot_id)
1203 {
1204     BlockDriver *drv = bs->drv;
1205     if (!drv)
1206         return -ENOMEDIUM;
1207     if (!drv->bdrv_snapshot_goto)
1208         return -ENOTSUP;
1209     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1210 }
1211
1212 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1213 {
1214     BlockDriver *drv = bs->drv;
1215     if (!drv)
1216         return -ENOMEDIUM;
1217     if (!drv->bdrv_snapshot_delete)
1218         return -ENOTSUP;
1219     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1220 }
1221
1222 int bdrv_snapshot_list(BlockDriverState *bs,
1223                        QEMUSnapshotInfo **psn_info)
1224 {
1225     BlockDriver *drv = bs->drv;
1226     if (!drv)
1227         return -ENOMEDIUM;
1228     if (!drv->bdrv_snapshot_list)
1229         return -ENOTSUP;
1230     return drv->bdrv_snapshot_list(bs, psn_info);
1231 }
1232
1233 #define NB_SUFFIXES 4
1234
1235 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1236 {
1237     static const char suffixes[NB_SUFFIXES] = "KMGT";
1238     int64_t base;
1239     int i;
1240
1241     if (size <= 999) {
1242         snprintf(buf, buf_size, "%" PRId64, size);
1243     } else {
1244         base = 1024;
1245         for(i = 0; i < NB_SUFFIXES; i++) {
1246             if (size < (10 * base)) {
1247                 snprintf(buf, buf_size, "%0.1f%c",
1248                          (double)size / base,
1249                          suffixes[i]);
1250                 break;
1251             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1252                 snprintf(buf, buf_size, "%" PRId64 "%c",
1253                          ((size + (base >> 1)) / base),
1254                          suffixes[i]);
1255                 break;
1256             }
1257             base = base * 1024;
1258         }
1259     }
1260     return buf;
1261 }
1262
1263 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1264 {
1265     char buf1[128], date_buf[128], clock_buf[128];
1266 #ifdef _WIN32
1267     struct tm *ptm;
1268 #else
1269     struct tm tm;
1270 #endif
1271     time_t ti;
1272     int64_t secs;
1273
1274     if (!sn) {
1275         snprintf(buf, buf_size,
1276                  "%-10s%-20s%7s%20s%15s",
1277                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1278     } else {
1279         ti = sn->date_sec;
1280 #ifdef _WIN32
1281         ptm = localtime(&ti);
1282         strftime(date_buf, sizeof(date_buf),
1283                  "%Y-%m-%d %H:%M:%S", ptm);
1284 #else
1285         localtime_r(&ti, &tm);
1286         strftime(date_buf, sizeof(date_buf),
1287                  "%Y-%m-%d %H:%M:%S", &tm);
1288 #endif
1289         secs = sn->vm_clock_nsec / 1000000000;
1290         snprintf(clock_buf, sizeof(clock_buf),
1291                  "%02d:%02d:%02d.%03d",
1292                  (int)(secs / 3600),
1293                  (int)((secs / 60) % 60),
1294                  (int)(secs % 60),
1295                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1296         snprintf(buf, buf_size,
1297                  "%-10s%-20s%7s%20s%15s",
1298                  sn->id_str, sn->name,
1299                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1300                  date_buf,
1301                  clock_buf);
1302     }
1303     return buf;
1304 }
1305
1306
1307 /**************************************************************/
1308 /* async I/Os */
1309
1310 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1311                                  QEMUIOVector *qiov, int nb_sectors,
1312                                  BlockDriverCompletionFunc *cb, void *opaque)
1313 {
1314     BlockDriver *drv = bs->drv;
1315     BlockDriverAIOCB *ret;
1316
1317     if (!drv)
1318         return NULL;
1319     if (bdrv_check_request(bs, sector_num, nb_sectors))
1320         return NULL;
1321
1322     ret = drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
1323                               cb, opaque);
1324
1325     if (ret) {
1326         /* Update stats even though technically transfer has not happened. */
1327         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1328         bs->rd_ops ++;
1329     }
1330
1331     return ret;
1332 }
1333
1334 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1335                                   QEMUIOVector *qiov, int nb_sectors,
1336                                   BlockDriverCompletionFunc *cb, void *opaque)
1337 {
1338     BlockDriver *drv = bs->drv;
1339     BlockDriverAIOCB *ret;
1340
1341     if (!drv)
1342         return NULL;
1343     if (bs->read_only)
1344         return NULL;
1345     if (bdrv_check_request(bs, sector_num, nb_sectors))
1346         return NULL;
1347
1348     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1349                                cb, opaque);
1350
1351     if (ret) {
1352         /* Update stats even though technically transfer has not happened. */
1353         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1354         bs->wr_ops ++;
1355     }
1356
1357     return ret;
1358 }
1359
1360 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1361 {
1362     acb->pool->cancel(acb);
1363 }
1364
1365
1366 /**************************************************************/
1367 /* async block device emulation */
1368
1369 static void bdrv_aio_bh_cb(void *opaque)
1370 {
1371     BlockDriverAIOCBSync *acb = opaque;
1372
1373     if (!acb->is_write)
1374         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1375     qemu_vfree(acb->bounce);
1376     acb->common.cb(acb->common.opaque, acb->ret);
1377
1378     qemu_aio_release(acb);
1379 }
1380
1381 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1382                                             int64_t sector_num,
1383                                             QEMUIOVector *qiov,
1384                                             int nb_sectors,
1385                                             BlockDriverCompletionFunc *cb,
1386                                             void *opaque,
1387                                             int is_write)
1388
1389 {
1390     BlockDriverAIOCBSync *acb;
1391
1392     acb = qemu_aio_get(bs, cb, opaque);
1393     acb->is_write = is_write;
1394     acb->qiov = qiov;
1395     acb->bounce = qemu_blockalign(bs, qiov->size);
1396
1397     if (!acb->bh)
1398         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1399
1400     if (is_write) {
1401         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1402         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1403     } else {
1404         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1405     }
1406
1407     qemu_bh_schedule(acb->bh);
1408
1409     return &acb->common;
1410 }
1411
1412 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1413         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1414         BlockDriverCompletionFunc *cb, void *opaque)
1415 {
1416     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1417 }
1418
1419 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1420         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1421         BlockDriverCompletionFunc *cb, void *opaque)
1422 {
1423     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1424 }
1425
1426
1427 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1428 {
1429     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1430     qemu_bh_cancel(acb->bh);
1431     qemu_aio_release(acb);
1432 }
1433
1434 /**************************************************************/
1435 /* sync block device emulation */
1436
1437 static void bdrv_rw_em_cb(void *opaque, int ret)
1438 {
1439     *(int *)opaque = ret;
1440 }
1441
1442 #define NOT_DONE 0x7fffffff
1443
1444 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1445                         uint8_t *buf, int nb_sectors)
1446 {
1447     int async_ret;
1448     BlockDriverAIOCB *acb;
1449     struct iovec iov;
1450     QEMUIOVector qiov;
1451
1452     async_ret = NOT_DONE;
1453     iov.iov_base = (void *)buf;
1454     iov.iov_len = nb_sectors * 512;
1455     qemu_iovec_init_external(&qiov, &iov, 1);
1456     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1457         bdrv_rw_em_cb, &async_ret);
1458     if (acb == NULL)
1459         return -1;
1460
1461     while (async_ret == NOT_DONE) {
1462         qemu_aio_wait();
1463     }
1464
1465     return async_ret;
1466 }
1467
1468 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1469                          const uint8_t *buf, int nb_sectors)
1470 {
1471     int async_ret;
1472     BlockDriverAIOCB *acb;
1473     struct iovec iov;
1474     QEMUIOVector qiov;
1475
1476     async_ret = NOT_DONE;
1477     iov.iov_base = (void *)buf;
1478     iov.iov_len = nb_sectors * 512;
1479     qemu_iovec_init_external(&qiov, &iov, 1);
1480     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1481         bdrv_rw_em_cb, &async_ret);
1482     if (acb == NULL)
1483         return -1;
1484     while (async_ret == NOT_DONE) {
1485         qemu_aio_wait();
1486     }
1487     return async_ret;
1488 }
1489
1490 void bdrv_init(void)
1491 {
1492     bdrv_register(&bdrv_raw);
1493     bdrv_register(&bdrv_host_device);
1494 #ifndef _WIN32
1495     bdrv_register(&bdrv_cow);
1496 #endif
1497     bdrv_register(&bdrv_qcow);
1498     bdrv_register(&bdrv_vmdk);
1499     bdrv_register(&bdrv_cloop);
1500     bdrv_register(&bdrv_dmg);
1501     bdrv_register(&bdrv_bochs);
1502     bdrv_register(&bdrv_vpc);
1503     bdrv_register(&bdrv_vvfat);
1504     bdrv_register(&bdrv_qcow2);
1505     bdrv_register(&bdrv_parallels);
1506     bdrv_register(&bdrv_nbd);
1507 }
1508
1509 void aio_pool_init(AIOPool *pool, int aiocb_size,
1510                    void (*cancel)(BlockDriverAIOCB *acb))
1511 {
1512     pool->aiocb_size = aiocb_size;
1513     pool->cancel = cancel;
1514     pool->free_aiocb = NULL;
1515 }
1516
1517 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1518                         BlockDriverCompletionFunc *cb, void *opaque)
1519 {
1520     BlockDriverAIOCB *acb;
1521
1522     if (pool->free_aiocb) {
1523         acb = pool->free_aiocb;
1524         pool->free_aiocb = acb->next;
1525     } else {
1526         acb = qemu_mallocz(pool->aiocb_size);
1527         acb->pool = pool;
1528     }
1529     acb->bs = bs;
1530     acb->cb = cb;
1531     acb->opaque = opaque;
1532     return acb;
1533 }
1534
1535 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1536                    void *opaque)
1537 {
1538     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1539 }
1540
1541 void qemu_aio_release(void *p)
1542 {
1543     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1544     AIOPool *pool = acb->pool;
1545     acb->next = pool->free_aiocb;
1546     pool->free_aiocb = acb;
1547 }
1548
1549 /**************************************************************/
1550 /* removable device support */
1551
1552 /**
1553  * Return TRUE if the media is present
1554  */
1555 int bdrv_is_inserted(BlockDriverState *bs)
1556 {
1557     BlockDriver *drv = bs->drv;
1558     int ret;
1559     if (!drv)
1560         return 0;
1561     if (!drv->bdrv_is_inserted)
1562         return 1;
1563     ret = drv->bdrv_is_inserted(bs);
1564     return ret;
1565 }
1566
1567 /**
1568  * Return TRUE if the media changed since the last call to this
1569  * function. It is currently only used for floppy disks
1570  */
1571 int bdrv_media_changed(BlockDriverState *bs)
1572 {
1573     BlockDriver *drv = bs->drv;
1574     int ret;
1575
1576     if (!drv || !drv->bdrv_media_changed)
1577         ret = -ENOTSUP;
1578     else
1579         ret = drv->bdrv_media_changed(bs);
1580     if (ret == -ENOTSUP)
1581         ret = bs->media_changed;
1582     bs->media_changed = 0;
1583     return ret;
1584 }
1585
1586 /**
1587  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1588  */
1589 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1590 {
1591     BlockDriver *drv = bs->drv;
1592     int ret;
1593
1594     if (!drv || !drv->bdrv_eject) {
1595         ret = -ENOTSUP;
1596     } else {
1597         ret = drv->bdrv_eject(bs, eject_flag);
1598     }
1599     if (ret == -ENOTSUP) {
1600         if (eject_flag)
1601             bdrv_close(bs);
1602     }
1603 }
1604
1605 int bdrv_is_locked(BlockDriverState *bs)
1606 {
1607     return bs->locked;
1608 }
1609
1610 /**
1611  * Lock or unlock the media (if it is locked, the user won't be able
1612  * to eject it manually).
1613  */
1614 void bdrv_set_locked(BlockDriverState *bs, int locked)
1615 {
1616     BlockDriver *drv = bs->drv;
1617
1618     bs->locked = locked;
1619     if (drv && drv->bdrv_set_locked) {
1620         drv->bdrv_set_locked(bs, locked);
1621     }
1622 }
1623
1624 /* needed for generic scsi interface */
1625
1626 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1627 {
1628     BlockDriver *drv = bs->drv;
1629
1630     if (drv && drv->bdrv_ioctl)
1631         return drv->bdrv_ioctl(bs, req, buf);
1632     return -ENOTSUP;
1633 }
1634
1635 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1636         unsigned long int req, void *buf,
1637         BlockDriverCompletionFunc *cb, void *opaque)
1638 {
1639     BlockDriver *drv = bs->drv;
1640
1641     if (drv && drv->bdrv_aio_ioctl)
1642         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1643     return NULL;
1644 }
1645
1646 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1647 {
1648     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1649 }