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 static AIOPool vectored_aio_pool;
  51
  52 typedef struct BlockDriverAIOCBSync {
  53     BlockDriverAIOCB common;
  54     QEMUBH *bh;
  55     int ret;
  56 } BlockDriverAIOCBSync;
  57
  58 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  59         int64_t sector_num, uint8_t *buf, int nb_sectors,
  60         BlockDriverCompletionFunc *cb, void *opaque);
  61 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  62         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  63         BlockDriverCompletionFunc *cb, void *opaque);
  64 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  65 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  66                         uint8_t *buf, int nb_sectors);
  67 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  68                          const uint8_t *buf, int nb_sectors);
  69
  70 BlockDriverState *bdrv_first;
  71
  72 static BlockDriver *first_drv;
  73
  74 int path_is_absolute(const char *path)
  75 {
  76     const char *p;
  77 #ifdef _WIN32
  78     /* specific case for names like: "\\.\d:" */
  79     if (*path == '/' || *path == '\\')
  80         return 1;
  81 #endif
  82     p = strchr(path, ':');
  83     if (p)
  84         p++;
  85     else
  86         p = path;
  87 #ifdef _WIN32
  88     return (*p == '/' || *p == '\\');
  89 #else
  90     return (*p == '/');
  91 #endif
  92 }
  93
  94 /* if filename is absolute, just copy it to dest. Otherwise, build a
  95    path to it by considering it is relative to base_path. URL are
  96    supported. */
  97 void path_combine(char *dest, int dest_size,
  98                   const char *base_path,
  99                   const char *filename)
 100 {
 101     const char *p, *p1;
 102     int len;
 103
 104     if (dest_size <= 0)
 105         return;
 106     if (path_is_absolute(filename)) {
 107         pstrcpy(dest, dest_size, filename);
 108     } else {
 109         p = strchr(base_path, ':');
 110         if (p)
 111             p++;
 112         else
 113             p = base_path;
 114         p1 = strrchr(base_path, '/');
 115 #ifdef _WIN32
 116         {
 117             const char *p2;
 118             p2 = strrchr(base_path, '\\');
 119             if (!p1 || p2 > p1)
 120                 p1 = p2;
 121         }
 122 #endif
 123         if (p1)
 124             p1++;
 125         else
 126             p1 = base_path;
 127         if (p1 > p)
 128             p = p1;
 129         len = p - base_path;
 130         if (len > dest_size - 1)
 131             len = dest_size - 1;
 132         memcpy(dest, base_path, len);
 133         dest[len] = '\0';
 134         pstrcat(dest, dest_size, filename);
 135     }
 136 }
 137
 138
 139 static void bdrv_register(BlockDriver *bdrv)
 140 {
 141     if (!bdrv->bdrv_aio_read) {
 142         /* add AIO emulation layer */
 143         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 144         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 145         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 146         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 147     } else if (!bdrv->bdrv_read) {
 148         /* add synchronous IO emulation layer */
 149         bdrv->bdrv_read = bdrv_read_em;
 150         bdrv->bdrv_write = bdrv_write_em;
 151     }
 152     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 153     bdrv->next = first_drv;
 154     first_drv = bdrv;
 155 }
 156
 157 /* create a new block device (by default it is empty) */
 158 BlockDriverState *bdrv_new(const char *device_name)
 159 {
 160     BlockDriverState **pbs, *bs;
 161
 162     bs = qemu_mallocz(sizeof(BlockDriverState));
 163     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 164     if (device_name[0] != '\0') {
 165         /* insert at the end */
 166         pbs = &bdrv_first;
 167         while (*pbs != NULL)
 168             pbs = &(*pbs)->next;
 169         *pbs = bs;
 170     }
 171     return bs;
 172 }
 173
 174 BlockDriver *bdrv_find_format(const char *format_name)
 175 {
 176     BlockDriver *drv1;
 177     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 178         if (!strcmp(drv1->format_name, format_name))
 179             return drv1;
 180     }
 181     return NULL;
 182 }
 183
 184 int bdrv_create2(BlockDriver *drv,
 185                 const char *filename, int64_t size_in_sectors,
 186                 const char *backing_file, const char *backing_format,
 187                 int flags)
 188 {
 189     if (drv->bdrv_create2)
 190         return drv->bdrv_create2(filename, size_in_sectors, backing_file,
 191                                  backing_format, flags);
 192     if (drv->bdrv_create)
 193         return drv->bdrv_create(filename, size_in_sectors, backing_file,
 194                                 flags);
 195     return -ENOTSUP;
 196 }
 197
 198 int bdrv_create(BlockDriver *drv,
 199                 const char *filename, int64_t size_in_sectors,
 200                 const char *backing_file, int flags)
 201 {
 202     if (!drv->bdrv_create)
 203         return -ENOTSUP;
 204     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 205 }
 206
 207 #ifdef _WIN32
 208 void get_tmp_filename(char *filename, int size)
 209 {
 210     char temp_dir[MAX_PATH];
 211
 212     GetTempPath(MAX_PATH, temp_dir);
 213     GetTempFileName(temp_dir, "qem", 0, filename);
 214 }
 215 #else
 216 void get_tmp_filename(char *filename, int size)
 217 {
 218     int fd;
 219     const char *tmpdir;
 220     /* XXX: race condition possible */
 221     tmpdir = getenv("TMPDIR");
 222     if (!tmpdir)
 223         tmpdir = "/tmp";
 224     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 225     fd = mkstemp(filename);
 226     close(fd);
 227 }
 228 #endif
 229
 230 #ifdef _WIN32
 231 static int is_windows_drive_prefix(const char *filename)
 232 {
 233     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 234              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 235             filename[1] == ':');
 236 }
 237
 238 static int is_windows_drive(const char *filename)
 239 {
 240     if (is_windows_drive_prefix(filename) &&
 241         filename[2] == '\0')
 242         return 1;
 243     if (strstart(filename, "\\\\.\\", NULL) ||
 244         strstart(filename, "//./", NULL))
 245         return 1;
 246     return 0;
 247 }
 248 #endif
 249
 250 static BlockDriver *find_protocol(const char *filename)
 251 {
 252     BlockDriver *drv1;
 253     char protocol[128];
 254     int len;
 255     const char *p;
 256
 257 #ifdef _WIN32
 258     if (is_windows_drive(filename) ||
 259         is_windows_drive_prefix(filename))
 260         return &bdrv_raw;
 261 #endif
 262     p = strchr(filename, ':');
 263     if (!p)
 264         return &bdrv_raw;
 265     len = p - filename;
 266     if (len > sizeof(protocol) - 1)
 267         len = sizeof(protocol) - 1;
 268     memcpy(protocol, filename, len);
 269     protocol[len] = '\0';
 270     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 271         if (drv1->protocol_name &&
 272             !strcmp(drv1->protocol_name, protocol))
 273             return drv1;
 274     }
 275     return NULL;
 276 }
 277
 278 /* XXX: force raw format if block or character device ? It would
 279    simplify the BSD case */
 280 static BlockDriver *find_image_format(const char *filename)
 281 {
 282     int ret, score, score_max;
 283     BlockDriver *drv1, *drv;
 284     uint8_t buf[2048];
 285     BlockDriverState *bs;
 286
 287     /* detect host devices. By convention, /dev/cdrom[N] is always
 288        recognized as a host CDROM */
 289     if (strstart(filename, "/dev/cdrom", NULL))
 290         return &bdrv_host_device;
 291 #ifdef _WIN32
 292     if (is_windows_drive(filename))
 293         return &bdrv_host_device;
 294 #else
 295     {
 296         struct stat st;
 297         if (stat(filename, &st) >= 0 &&
 298             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 299             return &bdrv_host_device;
 300         }
 301     }
 302 #endif
 303
 304     drv = find_protocol(filename);
 305     /* no need to test disk image formats for vvfat */
 306     if (drv == &bdrv_vvfat)
 307         return drv;
 308
 309     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 310     if (ret < 0)
 311         return NULL;
 312     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 313     bdrv_delete(bs);
 314     if (ret < 0) {
 315         return NULL;
 316     }
 317
 318     score_max = 0;
 319     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 320         if (drv1->bdrv_probe) {
 321             score = drv1->bdrv_probe(buf, ret, filename);
 322             if (score > score_max) {
 323                 score_max = score;
 324                 drv = drv1;
 325             }
 326         }
 327     }
 328     return drv;
 329 }
 330
 331 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 332 {
 333     BlockDriverState *bs;
 334     int ret;
 335
 336     bs = bdrv_new("");
 337     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 338     if (ret < 0) {
 339         bdrv_delete(bs);
 340         return ret;
 341     }
 342     bs->growable = 1;
 343     *pbs = bs;
 344     return 0;
 345 }
 346
 347 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 348 {
 349     return bdrv_open2(bs, filename, flags, NULL);
 350 }
 351
 352 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 353                BlockDriver *drv)
 354 {
 355     int ret, open_flags;
 356     char tmp_filename[PATH_MAX];
 357     char backing_filename[PATH_MAX];
 358
 359     bs->read_only = 0;
 360     bs->is_temporary = 0;
 361     bs->encrypted = 0;
 362     bs->valid_key = 0;
 363
 364     if (flags & BDRV_O_SNAPSHOT) {
 365         BlockDriverState *bs1;
 366         int64_t total_size;
 367         int is_protocol = 0;
 368
 369         /* if snapshot, we create a temporary backing file and open it
 370            instead of opening 'filename' directly */
 371
 372         /* if there is a backing file, use it */
 373         bs1 = bdrv_new("");
 374         ret = bdrv_open2(bs1, filename, 0, drv);
 375         if (ret < 0) {
 376             bdrv_delete(bs1);
 377             return ret;
 378         }
 379         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 380
 381         if (bs1->drv && bs1->drv->protocol_name)
 382             is_protocol = 1;
 383
 384         bdrv_delete(bs1);
 385
 386         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 387
 388         /* Real path is meaningless for protocols */
 389         if (is_protocol)
 390             snprintf(backing_filename, sizeof(backing_filename),
 391                      "%s", filename);
 392         else
 393             realpath(filename, backing_filename);
 394
 395         ret = bdrv_create2(&bdrv_qcow2, tmp_filename,
 396                            total_size, backing_filename,
 397                            (drv ? drv->format_name : NULL), 0);
 398         if (ret < 0) {
 399             return ret;
 400         }
 401         filename = tmp_filename;
 402         drv = &bdrv_qcow2;
 403         bs->is_temporary = 1;
 404     }
 405
 406     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 407     if (flags & BDRV_O_FILE) {
 408         drv = find_protocol(filename);
 409     } else if (!drv) {
 410         drv = find_image_format(filename);
 411     }
 412     if (!drv) {
 413         ret = -ENOENT;
 414         goto unlink_and_fail;
 415     }
 416     bs->drv = drv;
 417     bs->opaque = qemu_mallocz(drv->instance_size);
 418     /* Note: for compatibility, we open disk image files as RDWR, and
 419        RDONLY as fallback */
 420     if (!(flags & BDRV_O_FILE))
 421         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 422     else
 423         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 424     ret = drv->bdrv_open(bs, filename, open_flags);
 425     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 426         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 427         bs->read_only = 1;
 428     }
 429     if (ret < 0) {
 430         qemu_free(bs->opaque);
 431         bs->opaque = NULL;
 432         bs->drv = NULL;
 433     unlink_and_fail:
 434         if (bs->is_temporary)
 435             unlink(filename);
 436         return ret;
 437     }
 438     if (drv->bdrv_getlength) {
 439         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 440     }
 441 #ifndef _WIN32
 442     if (bs->is_temporary) {
 443         unlink(filename);
 444     }
 445 #endif
 446     if (bs->backing_file[0] != '\0') {
 447         /* if there is a backing file, use it */
 448         BlockDriver *back_drv = NULL;
 449         bs->backing_hd = bdrv_new("");
 450         path_combine(backing_filename, sizeof(backing_filename),
 451                      filename, bs->backing_file);
 452         if (bs->backing_format[0] != '\0')
 453             back_drv = bdrv_find_format(bs->backing_format);
 454         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 455                          back_drv);
 456         if (ret < 0) {
 457             bdrv_close(bs);
 458             return ret;
 459         }
 460     }
 461
 462     if (!bdrv_key_required(bs)) {
 463         /* call the change callback */
 464         bs->media_changed = 1;
 465         if (bs->change_cb)
 466             bs->change_cb(bs->change_opaque);
 467     }
 468     return 0;
 469 }
 470
 471 void bdrv_close(BlockDriverState *bs)
 472 {
 473     if (bs->drv) {
 474         if (bs->backing_hd)
 475             bdrv_delete(bs->backing_hd);
 476         bs->drv->bdrv_close(bs);
 477         qemu_free(bs->opaque);
 478 #ifdef _WIN32
 479         if (bs->is_temporary) {
 480             unlink(bs->filename);
 481         }
 482 #endif
 483         bs->opaque = NULL;
 484         bs->drv = NULL;
 485
 486         /* call the change callback */
 487         bs->media_changed = 1;
 488         if (bs->change_cb)
 489             bs->change_cb(bs->change_opaque);
 490     }
 491 }
 492
 493 void bdrv_delete(BlockDriverState *bs)
 494 {
 495     BlockDriverState **pbs;
 496
 497     pbs = &bdrv_first;
 498     while (*pbs != bs && *pbs != NULL)
 499         pbs = &(*pbs)->next;
 500     if (*pbs == bs)
 501         *pbs = bs->next;
 502
 503     bdrv_close(bs);
 504     qemu_free(bs);
 505 }
 506
 507 /* commit COW file into the raw image */
 508 int bdrv_commit(BlockDriverState *bs)
 509 {
 510     BlockDriver *drv = bs->drv;
 511     int64_t i, total_sectors;
 512     int n, j;
 513     unsigned char sector[512];
 514
 515     if (!drv)
 516         return -ENOMEDIUM;
 517
 518     if (bs->read_only) {
 519         return -EACCES;
 520     }
 521
 522     if (!bs->backing_hd) {
 523         return -ENOTSUP;
 524     }
 525
 526     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 527     for (i = 0; i < total_sectors;) {
 528         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 529             for(j = 0; j < n; j++) {
 530                 if (bdrv_read(bs, i, sector, 1) != 0) {
 531                     return -EIO;
 532                 }
 533
 534                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 535                     return -EIO;
 536                 }
 537                 i++;
 538             }
 539         } else {
 540             i += n;
 541         }
 542     }
 543
 544     if (drv->bdrv_make_empty)
 545         return drv->bdrv_make_empty(bs);
 546
 547     return 0;
 548 }
 549
 550 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 551                                    size_t size)
 552 {
 553     int64_t len;
 554
 555     if (!bdrv_is_inserted(bs))
 556         return -ENOMEDIUM;
 557
 558     if (bs->growable)
 559         return 0;
 560
 561     len = bdrv_getlength(bs);
 562
 563     if ((offset + size) > len)
 564         return -EIO;
 565
 566     return 0;
 567 }
 568
 569 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 570                               int nb_sectors)
 571 {
 572     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 573 }
 574
 575 /* return < 0 if error. See bdrv_write() for the return codes */
 576 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 577               uint8_t *buf, int nb_sectors)
 578 {
 579     BlockDriver *drv = bs->drv;
 580
 581     if (!drv)
 582         return -ENOMEDIUM;
 583     if (bdrv_check_request(bs, sector_num, nb_sectors))
 584         return -EIO;
 585
 586     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 587 }
 588
 589 /* Return < 0 if error. Important errors are:
 590   -EIO         generic I/O error (may happen for all errors)
 591   -ENOMEDIUM   No media inserted.
 592   -EINVAL      Invalid sector number or nb_sectors
 593   -EACCES      Trying to write a read-only device
 594 */
 595 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 596                const uint8_t *buf, int nb_sectors)
 597 {
 598     BlockDriver *drv = bs->drv;
 599     if (!bs->drv)
 600         return -ENOMEDIUM;
 601     if (bs->read_only)
 602         return -EACCES;
 603     if (bdrv_check_request(bs, sector_num, nb_sectors))
 604         return -EIO;
 605
 606     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 607 }
 608
 609 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 610                void *buf, int count1)
 611 {
 612     uint8_t tmp_buf[SECTOR_SIZE];
 613     int len, nb_sectors, count;
 614     int64_t sector_num;
 615
 616     count = count1;
 617     /* first read to align to sector start */
 618     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 619     if (len > count)
 620         len = count;
 621     sector_num = offset >> SECTOR_BITS;
 622     if (len > 0) {
 623         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 624             return -EIO;
 625         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 626         count -= len;
 627         if (count == 0)
 628             return count1;
 629         sector_num++;
 630         buf += len;
 631     }
 632
 633     /* read the sectors "in place" */
 634     nb_sectors = count >> SECTOR_BITS;
 635     if (nb_sectors > 0) {
 636         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 637             return -EIO;
 638         sector_num += nb_sectors;
 639         len = nb_sectors << SECTOR_BITS;
 640         buf += len;
 641         count -= len;
 642     }
 643
 644     /* add data from the last sector */
 645     if (count > 0) {
 646         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 647             return -EIO;
 648         memcpy(buf, tmp_buf, count);
 649     }
 650     return count1;
 651 }
 652
 653 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 654                 const void *buf, int count1)
 655 {
 656     uint8_t tmp_buf[SECTOR_SIZE];
 657     int len, nb_sectors, count;
 658     int64_t sector_num;
 659
 660     count = count1;
 661     /* first write to align to sector start */
 662     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 663     if (len > count)
 664         len = count;
 665     sector_num = offset >> SECTOR_BITS;
 666     if (len > 0) {
 667         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 668             return -EIO;
 669         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 670         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 671             return -EIO;
 672         count -= len;
 673         if (count == 0)
 674             return count1;
 675         sector_num++;
 676         buf += len;
 677     }
 678
 679     /* write the sectors "in place" */
 680     nb_sectors = count >> SECTOR_BITS;
 681     if (nb_sectors > 0) {
 682         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 683             return -EIO;
 684         sector_num += nb_sectors;
 685         len = nb_sectors << SECTOR_BITS;
 686         buf += len;
 687         count -= len;
 688     }
 689
 690     /* add data from the last sector */
 691     if (count > 0) {
 692         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 693             return -EIO;
 694         memcpy(tmp_buf, buf, count);
 695         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 696             return -EIO;
 697     }
 698     return count1;
 699 }
 700
 701 /**
 702  * Truncate file to 'offset' bytes (needed only for file protocols)
 703  */
 704 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 705 {
 706     BlockDriver *drv = bs->drv;
 707     if (!drv)
 708         return -ENOMEDIUM;
 709     if (!drv->bdrv_truncate)
 710         return -ENOTSUP;
 711     return drv->bdrv_truncate(bs, offset);
 712 }
 713
 714 /**
 715  * Length of a file in bytes. Return < 0 if error or unknown.
 716  */
 717 int64_t bdrv_getlength(BlockDriverState *bs)
 718 {
 719     BlockDriver *drv = bs->drv;
 720     if (!drv)
 721         return -ENOMEDIUM;
 722     if (!drv->bdrv_getlength) {
 723         /* legacy mode */
 724         return bs->total_sectors * SECTOR_SIZE;
 725     }
 726     return drv->bdrv_getlength(bs);
 727 }
 728
 729 /* return 0 as number of sectors if no device present or error */
 730 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 731 {
 732     int64_t length;
 733     length = bdrv_getlength(bs);
 734     if (length < 0)
 735         length = 0;
 736     else
 737         length = length >> SECTOR_BITS;
 738     *nb_sectors_ptr = length;
 739 }
 740
 741 struct partition {
 742         uint8_t boot_ind;           /* 0x80 - active */
 743         uint8_t head;               /* starting head */
 744         uint8_t sector;             /* starting sector */
 745         uint8_t cyl;                /* starting cylinder */
 746         uint8_t sys_ind;            /* What partition type */
 747         uint8_t end_head;           /* end head */
 748         uint8_t end_sector;         /* end sector */
 749         uint8_t end_cyl;            /* end cylinder */
 750         uint32_t start_sect;        /* starting sector counting from 0 */
 751         uint32_t nr_sects;          /* nr of sectors in partition */
 752 } __attribute__((packed));
 753
 754 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 755 static int guess_disk_lchs(BlockDriverState *bs,
 756                            int *pcylinders, int *pheads, int *psectors)
 757 {
 758     uint8_t buf[512];
 759     int ret, i, heads, sectors, cylinders;
 760     struct partition *p;
 761     uint32_t nr_sects;
 762     uint64_t nb_sectors;
 763
 764     bdrv_get_geometry(bs, &nb_sectors);
 765
 766     ret = bdrv_read(bs, 0, buf, 1);
 767     if (ret < 0)
 768         return -1;
 769     /* test msdos magic */
 770     if (buf[510] != 0x55 || buf[511] != 0xaa)
 771         return -1;
 772     for(i = 0; i < 4; i++) {
 773         p = ((struct partition *)(buf + 0x1be)) + i;
 774         nr_sects = le32_to_cpu(p->nr_sects);
 775         if (nr_sects && p->end_head) {
 776             /* We make the assumption that the partition terminates on
 777                a cylinder boundary */
 778             heads = p->end_head + 1;
 779             sectors = p->end_sector & 63;
 780             if (sectors == 0)
 781                 continue;
 782             cylinders = nb_sectors / (heads * sectors);
 783             if (cylinders < 1 || cylinders > 16383)
 784                 continue;
 785             *pheads = heads;
 786             *psectors = sectors;
 787             *pcylinders = cylinders;
 788 #if 0
 789             printf("guessed geometry: LCHS=%d %d %d\n",
 790                    cylinders, heads, sectors);
 791 #endif
 792             return 0;
 793         }
 794     }
 795     return -1;
 796 }
 797
 798 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 799 {
 800     int translation, lba_detected = 0;
 801     int cylinders, heads, secs;
 802     uint64_t nb_sectors;
 803
 804     /* if a geometry hint is available, use it */
 805     bdrv_get_geometry(bs, &nb_sectors);
 806     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 807     translation = bdrv_get_translation_hint(bs);
 808     if (cylinders != 0) {
 809         *pcyls = cylinders;
 810         *pheads = heads;
 811         *psecs = secs;
 812     } else {
 813         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 814             if (heads > 16) {
 815                 /* if heads > 16, it means that a BIOS LBA
 816                    translation was active, so the default
 817                    hardware geometry is OK */
 818                 lba_detected = 1;
 819                 goto default_geometry;
 820             } else {
 821                 *pcyls = cylinders;
 822                 *pheads = heads;
 823                 *psecs = secs;
 824                 /* disable any translation to be in sync with
 825                    the logical geometry */
 826                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 827                     bdrv_set_translation_hint(bs,
 828                                               BIOS_ATA_TRANSLATION_NONE);
 829                 }
 830             }
 831         } else {
 832         default_geometry:
 833             /* if no geometry, use a standard physical disk geometry */
 834             cylinders = nb_sectors / (16 * 63);
 835
 836             if (cylinders > 16383)
 837                 cylinders = 16383;
 838             else if (cylinders < 2)
 839                 cylinders = 2;
 840             *pcyls = cylinders;
 841             *pheads = 16;
 842             *psecs = 63;
 843             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 844                 if ((*pcyls * *pheads) <= 131072) {
 845                     bdrv_set_translation_hint(bs,
 846                                               BIOS_ATA_TRANSLATION_LARGE);
 847                 } else {
 848                     bdrv_set_translation_hint(bs,
 849                                               BIOS_ATA_TRANSLATION_LBA);
 850                 }
 851             }
 852         }
 853         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 854     }
 855 }
 856
 857 void bdrv_set_geometry_hint(BlockDriverState *bs,
 858                             int cyls, int heads, int secs)
 859 {
 860     bs->cyls = cyls;
 861     bs->heads = heads;
 862     bs->secs = secs;
 863 }
 864
 865 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 866 {
 867     bs->type = type;
 868     bs->removable = ((type == BDRV_TYPE_CDROM ||
 869                       type == BDRV_TYPE_FLOPPY));
 870 }
 871
 872 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 873 {
 874     bs->translation = translation;
 875 }
 876
 877 void bdrv_get_geometry_hint(BlockDriverState *bs,
 878                             int *pcyls, int *pheads, int *psecs)
 879 {
 880     *pcyls = bs->cyls;
 881     *pheads = bs->heads;
 882     *psecs = bs->secs;
 883 }
 884
 885 int bdrv_get_type_hint(BlockDriverState *bs)
 886 {
 887     return bs->type;
 888 }
 889
 890 int bdrv_get_translation_hint(BlockDriverState *bs)
 891 {
 892     return bs->translation;
 893 }
 894
 895 int bdrv_is_removable(BlockDriverState *bs)
 896 {
 897     return bs->removable;
 898 }
 899
 900 int bdrv_is_read_only(BlockDriverState *bs)
 901 {
 902     return bs->read_only;
 903 }
 904
 905 int bdrv_is_sg(BlockDriverState *bs)
 906 {
 907     return bs->sg;
 908 }
 909
 910 /* XXX: no longer used */
 911 void bdrv_set_change_cb(BlockDriverState *bs,
 912                         void (*change_cb)(void *opaque), void *opaque)
 913 {
 914     bs->change_cb = change_cb;
 915     bs->change_opaque = opaque;
 916 }
 917
 918 int bdrv_is_encrypted(BlockDriverState *bs)
 919 {
 920     if (bs->backing_hd && bs->backing_hd->encrypted)
 921         return 1;
 922     return bs->encrypted;
 923 }
 924
 925 int bdrv_key_required(BlockDriverState *bs)
 926 {
 927     BlockDriverState *backing_hd = bs->backing_hd;
 928
 929     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 930         return 1;
 931     return (bs->encrypted && !bs->valid_key);
 932 }
 933
 934 int bdrv_set_key(BlockDriverState *bs, const char *key)
 935 {
 936     int ret;
 937     if (bs->backing_hd && bs->backing_hd->encrypted) {
 938         ret = bdrv_set_key(bs->backing_hd, key);
 939         if (ret < 0)
 940             return ret;
 941         if (!bs->encrypted)
 942             return 0;
 943     }
 944     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 945         return -1;
 946     ret = bs->drv->bdrv_set_key(bs, key);
 947     if (ret < 0) {
 948         bs->valid_key = 0;
 949     } else if (!bs->valid_key) {
 950         bs->valid_key = 1;
 951         /* call the change callback now, we skipped it on open */
 952         bs->media_changed = 1;
 953         if (bs->change_cb)
 954             bs->change_cb(bs->change_opaque);
 955     }
 956     return ret;
 957 }
 958
 959 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 960 {
 961     if (!bs->drv) {
 962         buf[0] = '\0';
 963     } else {
 964         pstrcpy(buf, buf_size, bs->drv->format_name);
 965     }
 966 }
 967
 968 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 969                          void *opaque)
 970 {
 971     BlockDriver *drv;
 972
 973     for (drv = first_drv; drv != NULL; drv = drv->next) {
 974         it(opaque, drv->format_name);
 975     }
 976 }
 977
 978 BlockDriverState *bdrv_find(const char *name)
 979 {
 980     BlockDriverState *bs;
 981
 982     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 983         if (!strcmp(name, bs->device_name))
 984             return bs;
 985     }
 986     return NULL;
 987 }
 988
 989 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
 990 {
 991     BlockDriverState *bs;
 992
 993     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 994         it(opaque, bs);
 995     }
 996 }
 997
 998 const char *bdrv_get_device_name(BlockDriverState *bs)
 999 {
1000     return bs->device_name;
1001 }
1002
1003 void bdrv_flush(BlockDriverState *bs)
1004 {
1005     if (bs->drv->bdrv_flush)
1006         bs->drv->bdrv_flush(bs);
1007     if (bs->backing_hd)
1008         bdrv_flush(bs->backing_hd);
1009 }
1010
1011 void bdrv_flush_all(void)
1012 {
1013     BlockDriverState *bs;
1014
1015     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1016         if (bs->drv && !bdrv_is_read_only(bs) &&
1017             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1018             bdrv_flush(bs);
1019 }
1020
1021 /*
1022  * Returns true iff the specified sector is present in the disk image. Drivers
1023  * not implementing the functionality are assumed to not support backing files,
1024  * hence all their sectors are reported as allocated.
1025  *
1026  * 'pnum' is set to the number of sectors (including and immediately following
1027  * the specified sector) that are known to be in the same
1028  * allocated/unallocated state.
1029  *
1030  * 'nb_sectors' is the max value 'pnum' should be set to.
1031  */
1032 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1033         int *pnum)
1034 {
1035     int64_t n;
1036     if (!bs->drv->bdrv_is_allocated) {
1037         if (sector_num >= bs->total_sectors) {
1038             *pnum = 0;
1039             return 0;
1040         }
1041         n = bs->total_sectors - sector_num;
1042         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1043         return 1;
1044     }
1045     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1046 }
1047
1048 void bdrv_info(Monitor *mon)
1049 {
1050     BlockDriverState *bs;
1051
1052     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1053         monitor_printf(mon, "%s:", bs->device_name);
1054         monitor_printf(mon, " type=");
1055         switch(bs->type) {
1056         case BDRV_TYPE_HD:
1057             monitor_printf(mon, "hd");
1058             break;
1059         case BDRV_TYPE_CDROM:
1060             monitor_printf(mon, "cdrom");
1061             break;
1062         case BDRV_TYPE_FLOPPY:
1063             monitor_printf(mon, "floppy");
1064             break;
1065         }
1066         monitor_printf(mon, " removable=%d", bs->removable);
1067         if (bs->removable) {
1068             monitor_printf(mon, " locked=%d", bs->locked);
1069         }
1070         if (bs->drv) {
1071             monitor_printf(mon, " file=");
1072             monitor_print_filename(mon, bs->filename);
1073             if (bs->backing_file[0] != '\0') {
1074                 monitor_printf(mon, " backing_file=");
1075                 monitor_print_filename(mon, bs->backing_file);
1076             }
1077             monitor_printf(mon, " ro=%d", bs->read_only);
1078             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1079             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1080         } else {
1081             monitor_printf(mon, " [not inserted]");
1082         }
1083         monitor_printf(mon, "\n");
1084     }
1085 }
1086
1087 /* The "info blockstats" command. */
1088 void bdrv_info_stats(Monitor *mon)
1089 {
1090     BlockDriverState *bs;
1091
1092     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1093         monitor_printf(mon, "%s:"
1094                        " rd_bytes=%" PRIu64
1095                        " wr_bytes=%" PRIu64
1096                        " rd_operations=%" PRIu64
1097                        " wr_operations=%" PRIu64
1098                        "\n",
1099                        bs->device_name,
1100                        bs->rd_bytes, bs->wr_bytes,
1101                        bs->rd_ops, bs->wr_ops);
1102     }
1103 }
1104
1105 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1106 {
1107     if (bs->backing_hd && bs->backing_hd->encrypted)
1108         return bs->backing_file;
1109     else if (bs->encrypted)
1110         return bs->filename;
1111     else
1112         return NULL;
1113 }
1114
1115 void bdrv_get_backing_filename(BlockDriverState *bs,
1116                                char *filename, int filename_size)
1117 {
1118     if (!bs->backing_hd) {
1119         pstrcpy(filename, filename_size, "");
1120     } else {
1121         pstrcpy(filename, filename_size, bs->backing_file);
1122     }
1123 }
1124
1125 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1126                           const uint8_t *buf, int nb_sectors)
1127 {
1128     BlockDriver *drv = bs->drv;
1129     if (!drv)
1130         return -ENOMEDIUM;
1131     if (!drv->bdrv_write_compressed)
1132         return -ENOTSUP;
1133     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1134 }
1135
1136 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1137 {
1138     BlockDriver *drv = bs->drv;
1139     if (!drv)
1140         return -ENOMEDIUM;
1141     if (!drv->bdrv_get_info)
1142         return -ENOTSUP;
1143     memset(bdi, 0, sizeof(*bdi));
1144     return drv->bdrv_get_info(bs, bdi);
1145 }
1146
1147 /**************************************************************/
1148 /* handling of snapshots */
1149
1150 int bdrv_snapshot_create(BlockDriverState *bs,
1151                          QEMUSnapshotInfo *sn_info)
1152 {
1153     BlockDriver *drv = bs->drv;
1154     if (!drv)
1155         return -ENOMEDIUM;
1156     if (!drv->bdrv_snapshot_create)
1157         return -ENOTSUP;
1158     return drv->bdrv_snapshot_create(bs, sn_info);
1159 }
1160
1161 int bdrv_snapshot_goto(BlockDriverState *bs,
1162                        const char *snapshot_id)
1163 {
1164     BlockDriver *drv = bs->drv;
1165     if (!drv)
1166         return -ENOMEDIUM;
1167     if (!drv->bdrv_snapshot_goto)
1168         return -ENOTSUP;
1169     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1170 }
1171
1172 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1173 {
1174     BlockDriver *drv = bs->drv;
1175     if (!drv)
1176         return -ENOMEDIUM;
1177     if (!drv->bdrv_snapshot_delete)
1178         return -ENOTSUP;
1179     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1180 }
1181
1182 int bdrv_snapshot_list(BlockDriverState *bs,
1183                        QEMUSnapshotInfo **psn_info)
1184 {
1185     BlockDriver *drv = bs->drv;
1186     if (!drv)
1187         return -ENOMEDIUM;
1188     if (!drv->bdrv_snapshot_list)
1189         return -ENOTSUP;
1190     return drv->bdrv_snapshot_list(bs, psn_info);
1191 }
1192
1193 #define NB_SUFFIXES 4
1194
1195 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1196 {
1197     static const char suffixes[NB_SUFFIXES] = "KMGT";
1198     int64_t base;
1199     int i;
1200
1201     if (size <= 999) {
1202         snprintf(buf, buf_size, "%" PRId64, size);
1203     } else {
1204         base = 1024;
1205         for(i = 0; i < NB_SUFFIXES; i++) {
1206             if (size < (10 * base)) {
1207                 snprintf(buf, buf_size, "%0.1f%c",
1208                          (double)size / base,
1209                          suffixes[i]);
1210                 break;
1211             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1212                 snprintf(buf, buf_size, "%" PRId64 "%c",
1213                          ((size + (base >> 1)) / base),
1214                          suffixes[i]);
1215                 break;
1216             }
1217             base = base * 1024;
1218         }
1219     }
1220     return buf;
1221 }
1222
1223 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1224 {
1225     char buf1[128], date_buf[128], clock_buf[128];
1226 #ifdef _WIN32
1227     struct tm *ptm;
1228 #else
1229     struct tm tm;
1230 #endif
1231     time_t ti;
1232     int64_t secs;
1233
1234     if (!sn) {
1235         snprintf(buf, buf_size,
1236                  "%-10s%-20s%7s%20s%15s",
1237                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1238     } else {
1239         ti = sn->date_sec;
1240 #ifdef _WIN32
1241         ptm = localtime(&ti);
1242         strftime(date_buf, sizeof(date_buf),
1243                  "%Y-%m-%d %H:%M:%S", ptm);
1244 #else
1245         localtime_r(&ti, &tm);
1246         strftime(date_buf, sizeof(date_buf),
1247                  "%Y-%m-%d %H:%M:%S", &tm);
1248 #endif
1249         secs = sn->vm_clock_nsec / 1000000000;
1250         snprintf(clock_buf, sizeof(clock_buf),
1251                  "%02d:%02d:%02d.%03d",
1252                  (int)(secs / 3600),
1253                  (int)((secs / 60) % 60),
1254                  (int)(secs % 60),
1255                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1256         snprintf(buf, buf_size,
1257                  "%-10s%-20s%7s%20s%15s",
1258                  sn->id_str, sn->name,
1259                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1260                  date_buf,
1261                  clock_buf);
1262     }
1263     return buf;
1264 }
1265
1266
1267 /**************************************************************/
1268 /* async I/Os */
1269
1270 typedef struct VectorTranslationAIOCB {
1271     BlockDriverAIOCB common;
1272     QEMUIOVector *iov;
1273     uint8_t *bounce;
1274     int is_write;
1275     BlockDriverAIOCB *aiocb;
1276 } VectorTranslationAIOCB;
1277
1278 static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
1279 {
1280     VectorTranslationAIOCB *acb
1281         = container_of(_acb, VectorTranslationAIOCB, common);
1282
1283     bdrv_aio_cancel(acb->aiocb);
1284 }
1285
1286 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1287 {
1288     VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
1289
1290     if (!s->is_write) {
1291         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1292     }
1293     qemu_vfree(s->bounce);
1294     s->common.cb(s->common.opaque, ret);
1295     qemu_aio_release(s);
1296 }
1297
1298 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1299                                             int64_t sector_num,
1300                                             QEMUIOVector *iov,
1301                                             int nb_sectors,
1302                                             BlockDriverCompletionFunc *cb,
1303                                             void *opaque,
1304                                             int is_write)
1305
1306 {
1307     VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
1308                                                   cb, opaque);
1309
1310     s->iov = iov;
1311     s->bounce = qemu_memalign(512, nb_sectors * 512);
1312     s->is_write = is_write;
1313     if (is_write) {
1314         qemu_iovec_to_buffer(s->iov, s->bounce);
1315         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1316                                   bdrv_aio_rw_vector_cb, s);
1317     } else {
1318         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1319                                  bdrv_aio_rw_vector_cb, s);
1320     }
1321     if (!s->aiocb) {
1322         qemu_vfree(s->bounce);
1323         qemu_aio_release(s);
1324         return NULL;
1325     }
1326     return &s->common;
1327 }
1328
1329 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1330                                  QEMUIOVector *iov, int nb_sectors,
1331                                  BlockDriverCompletionFunc *cb, void *opaque)
1332 {
1333     if (bdrv_check_request(bs, sector_num, nb_sectors))
1334         return NULL;
1335
1336     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1337                               cb, opaque, 0);
1338 }
1339
1340 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1341                                   QEMUIOVector *iov, int nb_sectors,
1342                                   BlockDriverCompletionFunc *cb, void *opaque)
1343 {
1344     if (bdrv_check_request(bs, sector_num, nb_sectors))
1345         return NULL;
1346
1347     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1348                               cb, opaque, 1);
1349 }
1350
1351 BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1352                                 uint8_t *buf, int nb_sectors,
1353                                 BlockDriverCompletionFunc *cb, void *opaque)
1354 {
1355     BlockDriver *drv = bs->drv;
1356     BlockDriverAIOCB *ret;
1357
1358     if (!drv)
1359         return NULL;
1360     if (bdrv_check_request(bs, sector_num, nb_sectors))
1361         return NULL;
1362
1363     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1364
1365     if (ret) {
1366         /* Update stats even though technically transfer has not happened. */
1367         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1368         bs->rd_ops ++;
1369     }
1370
1371     return ret;
1372 }
1373
1374 BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1375                                  const uint8_t *buf, int nb_sectors,
1376                                  BlockDriverCompletionFunc *cb, void *opaque)
1377 {
1378     BlockDriver *drv = bs->drv;
1379     BlockDriverAIOCB *ret;
1380
1381     if (!drv)
1382         return NULL;
1383     if (bs->read_only)
1384         return NULL;
1385     if (bdrv_check_request(bs, sector_num, nb_sectors))
1386         return NULL;
1387
1388     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1389
1390     if (ret) {
1391         /* Update stats even though technically transfer has not happened. */
1392         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1393         bs->wr_ops ++;
1394     }
1395
1396     return ret;
1397 }
1398
1399 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1400 {
1401     acb->pool->cancel(acb);
1402 }
1403
1404
1405 /**************************************************************/
1406 /* async block device emulation */
1407
1408 static void bdrv_aio_bh_cb(void *opaque)
1409 {
1410     BlockDriverAIOCBSync *acb = opaque;
1411     acb->common.cb(acb->common.opaque, acb->ret);
1412     qemu_aio_release(acb);
1413 }
1414
1415 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1416         int64_t sector_num, uint8_t *buf, int nb_sectors,
1417         BlockDriverCompletionFunc *cb, void *opaque)
1418 {
1419     BlockDriverAIOCBSync *acb;
1420     int ret;
1421
1422     acb = qemu_aio_get(bs, cb, opaque);
1423     if (!acb->bh)
1424         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1425     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1426     acb->ret = ret;
1427     qemu_bh_schedule(acb->bh);
1428     return &acb->common;
1429 }
1430
1431 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1432         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1433         BlockDriverCompletionFunc *cb, void *opaque)
1434 {
1435     BlockDriverAIOCBSync *acb;
1436     int ret;
1437
1438     acb = qemu_aio_get(bs, cb, opaque);
1439     if (!acb->bh)
1440         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1441     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1442     acb->ret = ret;
1443     qemu_bh_schedule(acb->bh);
1444     return &acb->common;
1445 }
1446
1447 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1448 {
1449     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1450     qemu_bh_cancel(acb->bh);
1451     qemu_aio_release(acb);
1452 }
1453
1454 /**************************************************************/
1455 /* sync block device emulation */
1456
1457 static void bdrv_rw_em_cb(void *opaque, int ret)
1458 {
1459     *(int *)opaque = ret;
1460 }
1461
1462 #define NOT_DONE 0x7fffffff
1463
1464 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1465                         uint8_t *buf, int nb_sectors)
1466 {
1467     int async_ret;
1468     BlockDriverAIOCB *acb;
1469
1470     async_ret = NOT_DONE;
1471     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1472                         bdrv_rw_em_cb, &async_ret);
1473     if (acb == NULL)
1474         return -1;
1475
1476     while (async_ret == NOT_DONE) {
1477         qemu_aio_wait();
1478     }
1479
1480     return async_ret;
1481 }
1482
1483 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1484                          const uint8_t *buf, int nb_sectors)
1485 {
1486     int async_ret;
1487     BlockDriverAIOCB *acb;
1488
1489     async_ret = NOT_DONE;
1490     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1491                          bdrv_rw_em_cb, &async_ret);
1492     if (acb == NULL)
1493         return -1;
1494     while (async_ret == NOT_DONE) {
1495         qemu_aio_wait();
1496     }
1497     return async_ret;
1498 }
1499
1500 void bdrv_init(void)
1501 {
1502     aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
1503                   bdrv_aio_cancel_vector);
1504
1505     bdrv_register(&bdrv_raw);
1506     bdrv_register(&bdrv_host_device);
1507 #ifndef _WIN32
1508     bdrv_register(&bdrv_cow);
1509 #endif
1510     bdrv_register(&bdrv_qcow);
1511     bdrv_register(&bdrv_vmdk);
1512     bdrv_register(&bdrv_cloop);
1513     bdrv_register(&bdrv_dmg);
1514     bdrv_register(&bdrv_bochs);
1515     bdrv_register(&bdrv_vpc);
1516     bdrv_register(&bdrv_vvfat);
1517     bdrv_register(&bdrv_qcow2);
1518     bdrv_register(&bdrv_parallels);
1519     bdrv_register(&bdrv_nbd);
1520 }
1521
1522 void aio_pool_init(AIOPool *pool, int aiocb_size,
1523                    void (*cancel)(BlockDriverAIOCB *acb))
1524 {
1525     pool->aiocb_size = aiocb_size;
1526     pool->cancel = cancel;
1527     pool->free_aiocb = NULL;
1528 }
1529
1530 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1531                         BlockDriverCompletionFunc *cb, void *opaque)
1532 {
1533     BlockDriverAIOCB *acb;
1534
1535     if (pool->free_aiocb) {
1536         acb = pool->free_aiocb;
1537         pool->free_aiocb = acb->next;
1538     } else {
1539         acb = qemu_mallocz(pool->aiocb_size);
1540         acb->pool = pool;
1541     }
1542     acb->bs = bs;
1543     acb->cb = cb;
1544     acb->opaque = opaque;
1545     return acb;
1546 }
1547
1548 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1549                    void *opaque)
1550 {
1551     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1552 }
1553
1554 void qemu_aio_release(void *p)
1555 {
1556     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1557     AIOPool *pool = acb->pool;
1558     acb->next = pool->free_aiocb;
1559     pool->free_aiocb = acb;
1560 }
1561
1562 /**************************************************************/
1563 /* removable device support */
1564
1565 /**
1566  * Return TRUE if the media is present
1567  */
1568 int bdrv_is_inserted(BlockDriverState *bs)
1569 {
1570     BlockDriver *drv = bs->drv;
1571     int ret;
1572     if (!drv)
1573         return 0;
1574     if (!drv->bdrv_is_inserted)
1575         return 1;
1576     ret = drv->bdrv_is_inserted(bs);
1577     return ret;
1578 }
1579
1580 /**
1581  * Return TRUE if the media changed since the last call to this
1582  * function. It is currently only used for floppy disks
1583  */
1584 int bdrv_media_changed(BlockDriverState *bs)
1585 {
1586     BlockDriver *drv = bs->drv;
1587     int ret;
1588
1589     if (!drv || !drv->bdrv_media_changed)
1590         ret = -ENOTSUP;
1591     else
1592         ret = drv->bdrv_media_changed(bs);
1593     if (ret == -ENOTSUP)
1594         ret = bs->media_changed;
1595     bs->media_changed = 0;
1596     return ret;
1597 }
1598
1599 /**
1600  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1601  */
1602 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1603 {
1604     BlockDriver *drv = bs->drv;
1605     int ret;
1606
1607     if (!drv || !drv->bdrv_eject) {
1608         ret = -ENOTSUP;
1609     } else {
1610         ret = drv->bdrv_eject(bs, eject_flag);
1611     }
1612     if (ret == -ENOTSUP) {
1613         if (eject_flag)
1614             bdrv_close(bs);
1615     }
1616 }
1617
1618 int bdrv_is_locked(BlockDriverState *bs)
1619 {
1620     return bs->locked;
1621 }
1622
1623 /**
1624  * Lock or unlock the media (if it is locked, the user won't be able
1625  * to eject it manually).
1626  */
1627 void bdrv_set_locked(BlockDriverState *bs, int locked)
1628 {
1629     BlockDriver *drv = bs->drv;
1630
1631     bs->locked = locked;
1632     if (drv && drv->bdrv_set_locked) {
1633         drv->bdrv_set_locked(bs, locked);
1634     }
1635 }
1636
1637 /* needed for generic scsi interface */
1638
1639 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1640 {
1641     BlockDriver *drv = bs->drv;
1642
1643     if (drv && drv->bdrv_ioctl)
1644         return drv->bdrv_ioctl(bs, req, buf);
1645     return -ENOTSUP;
1646 }
1647
1648 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1649         unsigned long int req, void *buf,
1650         BlockDriverCompletionFunc *cb, void *opaque)
1651 {
1652     BlockDriver *drv = bs->drv;
1653
1654     if (drv && drv->bdrv_aio_ioctl)
1655         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1656     return NULL;
1657 }