vcs.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 _BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "console.h"
  32 #include "block_int.h"
  33
  34 #ifdef _BSD
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/ioctl.h>
  38 #include <sys/disk.h>
  39 #endif
  40
  41 #define SECTOR_BITS 9
  42 #define SECTOR_SIZE (1 << SECTOR_BITS)
  43
  44 typedef struct BlockDriverAIOCBSync {
  45     BlockDriverAIOCB common;
  46     QEMUBH *bh;
  47     int ret;
  48 } BlockDriverAIOCBSync;
  49
  50 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  51         int64_t sector_num, uint8_t *buf, int nb_sectors,
  52         BlockDriverCompletionFunc *cb, void *opaque);
  53 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  54         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  55         BlockDriverCompletionFunc *cb, void *opaque);
  56 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  57 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  58                         uint8_t *buf, int nb_sectors);
  59 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  60                          const uint8_t *buf, int nb_sectors);
  61
  62 BlockDriverState *bdrv_first;
  63
  64 static BlockDriver *first_drv;
  65
  66 int path_is_absolute(const char *path)
  67 {
  68     const char *p;
  69 #ifdef _WIN32
  70     /* specific case for names like: "\\.\d:" */
  71     if (*path == '/' || *path == '\\')
  72         return 1;
  73 #endif
  74     p = strchr(path, ':');
  75     if (p)
  76         p++;
  77     else
  78         p = path;
  79 #ifdef _WIN32
  80     return (*p == '/' || *p == '\\');
  81 #else
  82     return (*p == '/');
  83 #endif
  84 }
  85
  86 /* if filename is absolute, just copy it to dest. Otherwise, build a
  87    path to it by considering it is relative to base_path. URL are
  88    supported. */
  89 void path_combine(char *dest, int dest_size,
  90                   const char *base_path,
  91                   const char *filename)
  92 {
  93     const char *p, *p1;
  94     int len;
  95
  96     if (dest_size <= 0)
  97         return;
  98     if (path_is_absolute(filename)) {
  99         pstrcpy(dest, dest_size, filename);
 100     } else {
 101         p = strchr(base_path, ':');
 102         if (p)
 103             p++;
 104         else
 105             p = base_path;
 106         p1 = strrchr(base_path, '/');
 107 #ifdef _WIN32
 108         {
 109             const char *p2;
 110             p2 = strrchr(base_path, '\\');
 111             if (!p1 || p2 > p1)
 112                 p1 = p2;
 113         }
 114 #endif
 115         if (p1)
 116             p1++;
 117         else
 118             p1 = base_path;
 119         if (p1 > p)
 120             p = p1;
 121         len = p - base_path;
 122         if (len > dest_size - 1)
 123             len = dest_size - 1;
 124         memcpy(dest, base_path, len);
 125         dest[len] = '\0';
 126         pstrcat(dest, dest_size, filename);
 127     }
 128 }
 129
 130
 131 static void bdrv_register(BlockDriver *bdrv)
 132 {
 133     if (!bdrv->bdrv_aio_read) {
 134         /* add AIO emulation layer */
 135         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 136         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 137         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 138         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 139     } else if (!bdrv->bdrv_read && !bdrv->bdrv_pread) {
 140         /* add synchronous IO emulation layer */
 141         bdrv->bdrv_read = bdrv_read_em;
 142         bdrv->bdrv_write = bdrv_write_em;
 143     }
 144     bdrv->next = first_drv;
 145     first_drv = bdrv;
 146 }
 147
 148 /* create a new block device (by default it is empty) */
 149 BlockDriverState *bdrv_new(const char *device_name)
 150 {
 151     BlockDriverState **pbs, *bs;
 152
 153     bs = qemu_mallocz(sizeof(BlockDriverState));
 154     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 155     if (device_name[0] != '\0') {
 156         /* insert at the end */
 157         pbs = &bdrv_first;
 158         while (*pbs != NULL)
 159             pbs = &(*pbs)->next;
 160         *pbs = bs;
 161     }
 162     return bs;
 163 }
 164
 165 BlockDriver *bdrv_find_format(const char *format_name)
 166 {
 167     BlockDriver *drv1;
 168     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 169         if (!strcmp(drv1->format_name, format_name))
 170             return drv1;
 171     }
 172     return NULL;
 173 }
 174
 175 int bdrv_create(BlockDriver *drv,
 176                 const char *filename, int64_t size_in_sectors,
 177                 const char *backing_file, int flags)
 178 {
 179     if (!drv->bdrv_create)
 180         return -ENOTSUP;
 181     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 182 }
 183
 184 #ifdef _WIN32
 185 void get_tmp_filename(char *filename, int size)
 186 {
 187     char temp_dir[MAX_PATH];
 188
 189     GetTempPath(MAX_PATH, temp_dir);
 190     GetTempFileName(temp_dir, "qem", 0, filename);
 191 }
 192 #else
 193 void get_tmp_filename(char *filename, int size)
 194 {
 195     int fd;
 196     const char *tmpdir;
 197     /* XXX: race condition possible */
 198     tmpdir = getenv("TMPDIR");
 199     if (!tmpdir)
 200         tmpdir = "/tmp";
 201     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 202     fd = mkstemp(filename);
 203     close(fd);
 204 }
 205 #endif
 206
 207 #ifdef _WIN32
 208 static int is_windows_drive_prefix(const char *filename)
 209 {
 210     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 211              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 212             filename[1] == ':');
 213 }
 214
 215 static int is_windows_drive(const char *filename)
 216 {
 217     if (is_windows_drive_prefix(filename) &&
 218         filename[2] == '\0')
 219         return 1;
 220     if (strstart(filename, "\\\\.\\", NULL) ||
 221         strstart(filename, "//./", NULL))
 222         return 1;
 223     return 0;
 224 }
 225 #endif
 226
 227 static BlockDriver *find_protocol(const char *filename)
 228 {
 229     BlockDriver *drv1;
 230     char protocol[128];
 231     int len;
 232     const char *p;
 233
 234 #ifdef _WIN32
 235     if (is_windows_drive(filename) ||
 236         is_windows_drive_prefix(filename))
 237         return &bdrv_raw;
 238 #endif
 239     p = strchr(filename, ':');
 240     if (!p)
 241         return &bdrv_raw;
 242     len = p - filename;
 243     if (len > sizeof(protocol) - 1)
 244         len = sizeof(protocol) - 1;
 245     memcpy(protocol, filename, len);
 246     protocol[len] = '\0';
 247     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 248         if (drv1->protocol_name &&
 249             !strcmp(drv1->protocol_name, protocol))
 250             return drv1;
 251     }
 252     return NULL;
 253 }
 254
 255 /* XXX: force raw format if block or character device ? It would
 256    simplify the BSD case */
 257 static BlockDriver *find_image_format(const char *filename)
 258 {
 259     int ret, score, score_max;
 260     BlockDriver *drv1, *drv;
 261     uint8_t buf[2048];
 262     BlockDriverState *bs;
 263
 264     /* detect host devices. By convention, /dev/cdrom[N] is always
 265        recognized as a host CDROM */
 266     if (strstart(filename, "/dev/cdrom", NULL))
 267         return &bdrv_host_device;
 268 #ifdef _WIN32
 269     if (is_windows_drive(filename))
 270         return &bdrv_host_device;
 271 #else
 272     {
 273         struct stat st;
 274         if (stat(filename, &st) >= 0 &&
 275             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 276             return &bdrv_host_device;
 277         }
 278     }
 279 #endif
 280
 281     drv = find_protocol(filename);
 282     /* no need to test disk image formats for vvfat */
 283     if (drv == &bdrv_vvfat)
 284         return drv;
 285
 286     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 287     if (ret < 0)
 288         return NULL;
 289     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 290     bdrv_delete(bs);
 291     if (ret < 0) {
 292         return NULL;
 293     }
 294
 295     score_max = 0;
 296     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 297         if (drv1->bdrv_probe) {
 298             score = drv1->bdrv_probe(buf, ret, filename);
 299             if (score > score_max) {
 300                 score_max = score;
 301                 drv = drv1;
 302             }
 303         }
 304     }
 305     return drv;
 306 }
 307
 308 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 309 {
 310     BlockDriverState *bs;
 311     int ret;
 312
 313     bs = bdrv_new("");
 314     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 315     if (ret < 0) {
 316         bdrv_delete(bs);
 317         return ret;
 318     }
 319     bs->growable = 1;
 320     *pbs = bs;
 321     return 0;
 322 }
 323
 324 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 325 {
 326     return bdrv_open2(bs, filename, flags, NULL);
 327 }
 328
 329 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 330                BlockDriver *drv)
 331 {
 332     int ret, open_flags;
 333     char tmp_filename[PATH_MAX];
 334     char backing_filename[PATH_MAX];
 335
 336     bs->read_only = 0;
 337     bs->is_temporary = 0;
 338     bs->encrypted = 0;
 339     bs->valid_key = 0;
 340
 341     if (flags & BDRV_O_SNAPSHOT) {
 342         BlockDriverState *bs1;
 343         int64_t total_size;
 344         int is_protocol = 0;
 345
 346         /* if snapshot, we create a temporary backing file and open it
 347            instead of opening 'filename' directly */
 348
 349         /* if there is a backing file, use it */
 350         bs1 = bdrv_new("");
 351         ret = bdrv_open(bs1, filename, 0);
 352         if (ret < 0) {
 353             bdrv_delete(bs1);
 354             return ret;
 355         }
 356         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 357
 358         if (bs1->drv && bs1->drv->protocol_name)
 359             is_protocol = 1;
 360
 361         bdrv_delete(bs1);
 362
 363         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 364
 365         /* Real path is meaningless for protocols */
 366         if (is_protocol)
 367             snprintf(backing_filename, sizeof(backing_filename),
 368                      "%s", filename);
 369         else
 370             realpath(filename, backing_filename);
 371
 372         ret = bdrv_create(&bdrv_qcow2, tmp_filename,
 373                           total_size, backing_filename, 0);
 374         if (ret < 0) {
 375             return ret;
 376         }
 377         filename = tmp_filename;
 378         bs->is_temporary = 1;
 379     }
 380
 381     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 382     if (flags & BDRV_O_FILE) {
 383         drv = find_protocol(filename);
 384     } else if (!drv) {
 385         drv = find_image_format(filename);
 386     }
 387     if (!drv) {
 388         ret = -ENOENT;
 389         goto unlink_and_fail;
 390     }
 391     bs->drv = drv;
 392     bs->opaque = qemu_mallocz(drv->instance_size);
 393     /* Note: for compatibility, we open disk image files as RDWR, and
 394        RDONLY as fallback */
 395     if (!(flags & BDRV_O_FILE))
 396         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 397     else
 398         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 399     ret = drv->bdrv_open(bs, filename, open_flags);
 400     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 401         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 402         bs->read_only = 1;
 403     }
 404     if (ret < 0) {
 405         qemu_free(bs->opaque);
 406         bs->opaque = NULL;
 407         bs->drv = NULL;
 408     unlink_and_fail:
 409         if (bs->is_temporary)
 410             unlink(filename);
 411         return ret;
 412     }
 413     if (drv->bdrv_getlength) {
 414         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 415     }
 416 #ifndef _WIN32
 417     if (bs->is_temporary) {
 418         unlink(filename);
 419     }
 420 #endif
 421     if (bs->backing_file[0] != '\0') {
 422         /* if there is a backing file, use it */
 423         bs->backing_hd = bdrv_new("");
 424         path_combine(backing_filename, sizeof(backing_filename),
 425                      filename, bs->backing_file);
 426         ret = bdrv_open(bs->backing_hd, backing_filename, open_flags);
 427         if (ret < 0) {
 428             bdrv_close(bs);
 429             return ret;
 430         }
 431     }
 432
 433     /* call the change callback */
 434     bs->media_changed = 1;
 435     if (bs->change_cb)
 436         bs->change_cb(bs->change_opaque);
 437
 438     return 0;
 439 }
 440
 441 void bdrv_close(BlockDriverState *bs)
 442 {
 443     if (bs->drv) {
 444         if (bs->backing_hd)
 445             bdrv_delete(bs->backing_hd);
 446         bs->drv->bdrv_close(bs);
 447         qemu_free(bs->opaque);
 448 #ifdef _WIN32
 449         if (bs->is_temporary) {
 450             unlink(bs->filename);
 451         }
 452 #endif
 453         bs->opaque = NULL;
 454         bs->drv = NULL;
 455
 456         /* call the change callback */
 457         bs->media_changed = 1;
 458         if (bs->change_cb)
 459             bs->change_cb(bs->change_opaque);
 460     }
 461 }
 462
 463 void bdrv_delete(BlockDriverState *bs)
 464 {
 465     BlockDriverState **pbs;
 466
 467     pbs = &bdrv_first;
 468     while (*pbs != bs && *pbs != NULL)
 469         pbs = &(*pbs)->next;
 470     if (*pbs == bs)
 471         *pbs = bs->next;
 472
 473     bdrv_close(bs);
 474     qemu_free(bs);
 475 }
 476
 477 /* commit COW file into the raw image */
 478 int bdrv_commit(BlockDriverState *bs)
 479 {
 480     BlockDriver *drv = bs->drv;
 481     int64_t i, total_sectors;
 482     int n, j;
 483     unsigned char sector[512];
 484
 485     if (!drv)
 486         return -ENOMEDIUM;
 487
 488     if (bs->read_only) {
 489         return -EACCES;
 490     }
 491
 492     if (!bs->backing_hd) {
 493         return -ENOTSUP;
 494     }
 495
 496     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 497     for (i = 0; i < total_sectors;) {
 498         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 499             for(j = 0; j < n; j++) {
 500                 if (bdrv_read(bs, i, sector, 1) != 0) {
 501                     return -EIO;
 502                 }
 503
 504                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 505                     return -EIO;
 506                 }
 507                 i++;
 508             }
 509         } else {
 510             i += n;
 511         }
 512     }
 513
 514     if (drv->bdrv_make_empty)
 515         return drv->bdrv_make_empty(bs);
 516
 517     return 0;
 518 }
 519
 520 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 521                                    size_t size)
 522 {
 523     int64_t len;
 524
 525     if (!bdrv_is_inserted(bs))
 526         return -ENOMEDIUM;
 527
 528     if (bs->growable)
 529         return 0;
 530
 531     len = bdrv_getlength(bs);
 532
 533     if ((offset + size) > len)
 534         return -EIO;
 535
 536     return 0;
 537 }
 538
 539 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 540                               int nb_sectors)
 541 {
 542     int64_t offset;
 543
 544     /* Deal with byte accesses */
 545     if (sector_num < 0)
 546         offset = -sector_num;
 547     else
 548         offset = sector_num * 512;
 549
 550     return bdrv_check_byte_request(bs, offset, nb_sectors * 512);
 551 }
 552
 553 /* return < 0 if error. See bdrv_write() for the return codes */
 554 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 555               uint8_t *buf, int nb_sectors)
 556 {
 557     BlockDriver *drv = bs->drv;
 558
 559     if (!drv)
 560         return -ENOMEDIUM;
 561     if (bdrv_check_request(bs, sector_num, nb_sectors))
 562         return -EIO;
 563
 564     if (drv->bdrv_pread) {
 565         int ret, len;
 566         len = nb_sectors * 512;
 567         ret = drv->bdrv_pread(bs, sector_num * 512, buf, len);
 568         if (ret < 0)
 569             return ret;
 570         else if (ret != len)
 571             return -EINVAL;
 572         else {
 573             bs->rd_bytes += (unsigned) len;
 574             bs->rd_ops ++;
 575             return 0;
 576         }
 577     } else {
 578         return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 579     }
 580 }
 581
 582 /* Return < 0 if error. Important errors are:
 583   -EIO         generic I/O error (may happen for all errors)
 584   -ENOMEDIUM   No media inserted.
 585   -EINVAL      Invalid sector number or nb_sectors
 586   -EACCES      Trying to write a read-only device
 587 */
 588 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 589                const uint8_t *buf, int nb_sectors)
 590 {
 591     BlockDriver *drv = bs->drv;
 592     if (!bs->drv)
 593         return -ENOMEDIUM;
 594     if (bs->read_only)
 595         return -EACCES;
 596     if (bdrv_check_request(bs, sector_num, nb_sectors))
 597         return -EIO;
 598
 599     if (drv->bdrv_pwrite) {
 600         int ret, len, count = 0;
 601         len = nb_sectors * 512;
 602         do {
 603             ret = drv->bdrv_pwrite(bs, sector_num * 512, buf, len - count);
 604             if (ret < 0) {
 605                 printf("bdrv_write ret=%d\n", ret);
 606                 return ret;
 607             }
 608             count += ret;
 609             buf += ret;
 610         } while (count != len);
 611         bs->wr_bytes += (unsigned) len;
 612         bs->wr_ops ++;
 613         return 0;
 614     }
 615     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 616 }
 617
 618 static int bdrv_pread_em(BlockDriverState *bs, int64_t offset,
 619                          uint8_t *buf, int count1)
 620 {
 621     uint8_t tmp_buf[SECTOR_SIZE];
 622     int len, nb_sectors, count;
 623     int64_t sector_num;
 624
 625     count = count1;
 626     /* first read to align to sector start */
 627     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 628     if (len > count)
 629         len = count;
 630     sector_num = offset >> SECTOR_BITS;
 631     if (len > 0) {
 632         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 633             return -EIO;
 634         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 635         count -= len;
 636         if (count == 0)
 637             return count1;
 638         sector_num++;
 639         buf += len;
 640     }
 641
 642     /* read the sectors "in place" */
 643     nb_sectors = count >> SECTOR_BITS;
 644     if (nb_sectors > 0) {
 645         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 646             return -EIO;
 647         sector_num += nb_sectors;
 648         len = nb_sectors << SECTOR_BITS;
 649         buf += len;
 650         count -= len;
 651     }
 652
 653     /* add data from the last sector */
 654     if (count > 0) {
 655         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 656             return -EIO;
 657         memcpy(buf, tmp_buf, count);
 658     }
 659     return count1;
 660 }
 661
 662 static int bdrv_pwrite_em(BlockDriverState *bs, int64_t offset,
 663                           const uint8_t *buf, int count1)
 664 {
 665     uint8_t tmp_buf[SECTOR_SIZE];
 666     int len, nb_sectors, count;
 667     int64_t sector_num;
 668
 669     count = count1;
 670     /* first write to align to sector start */
 671     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 672     if (len > count)
 673         len = count;
 674     sector_num = offset >> SECTOR_BITS;
 675     if (len > 0) {
 676         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 677             return -EIO;
 678         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 679         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 680             return -EIO;
 681         count -= len;
 682         if (count == 0)
 683             return count1;
 684         sector_num++;
 685         buf += len;
 686     }
 687
 688     /* write the sectors "in place" */
 689     nb_sectors = count >> SECTOR_BITS;
 690     if (nb_sectors > 0) {
 691         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 692             return -EIO;
 693         sector_num += nb_sectors;
 694         len = nb_sectors << SECTOR_BITS;
 695         buf += len;
 696         count -= len;
 697     }
 698
 699     /* add data from the last sector */
 700     if (count > 0) {
 701         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 702             return -EIO;
 703         memcpy(tmp_buf, buf, count);
 704         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 705             return -EIO;
 706     }
 707     return count1;
 708 }
 709
 710 /**
 711  * Read with byte offsets (needed only for file protocols)
 712  */
 713 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 714                void *buf1, int count1)
 715 {
 716     BlockDriver *drv = bs->drv;
 717
 718     if (!drv)
 719         return -ENOMEDIUM;
 720     if (bdrv_check_byte_request(bs, offset, count1))
 721         return -EIO;
 722
 723     if (!drv->bdrv_pread)
 724         return bdrv_pread_em(bs, offset, buf1, count1);
 725     return drv->bdrv_pread(bs, offset, buf1, count1);
 726 }
 727
 728 /**
 729  * Write with byte offsets (needed only for file protocols)
 730  */
 731 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 732                 const void *buf1, int count1)
 733 {
 734     BlockDriver *drv = bs->drv;
 735
 736     if (!drv)
 737         return -ENOMEDIUM;
 738     if (bdrv_check_byte_request(bs, offset, count1))
 739         return -EIO;
 740
 741     if (!drv->bdrv_pwrite)
 742         return bdrv_pwrite_em(bs, offset, buf1, count1);
 743     return drv->bdrv_pwrite(bs, offset, buf1, count1);
 744 }
 745
 746 /**
 747  * Truncate file to 'offset' bytes (needed only for file protocols)
 748  */
 749 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 750 {
 751     BlockDriver *drv = bs->drv;
 752     if (!drv)
 753         return -ENOMEDIUM;
 754     if (!drv->bdrv_truncate)
 755         return -ENOTSUP;
 756     return drv->bdrv_truncate(bs, offset);
 757 }
 758
 759 /**
 760  * Length of a file in bytes. Return < 0 if error or unknown.
 761  */
 762 int64_t bdrv_getlength(BlockDriverState *bs)
 763 {
 764     BlockDriver *drv = bs->drv;
 765     if (!drv)
 766         return -ENOMEDIUM;
 767     if (!drv->bdrv_getlength) {
 768         /* legacy mode */
 769         return bs->total_sectors * SECTOR_SIZE;
 770     }
 771     return drv->bdrv_getlength(bs);
 772 }
 773
 774 /* return 0 as number of sectors if no device present or error */
 775 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 776 {
 777     int64_t length;
 778     length = bdrv_getlength(bs);
 779     if (length < 0)
 780         length = 0;
 781     else
 782         length = length >> SECTOR_BITS;
 783     *nb_sectors_ptr = length;
 784 }
 785
 786 struct partition {
 787         uint8_t boot_ind;           /* 0x80 - active */
 788         uint8_t head;               /* starting head */
 789         uint8_t sector;             /* starting sector */
 790         uint8_t cyl;                /* starting cylinder */
 791         uint8_t sys_ind;            /* What partition type */
 792         uint8_t end_head;           /* end head */
 793         uint8_t end_sector;         /* end sector */
 794         uint8_t end_cyl;            /* end cylinder */
 795         uint32_t start_sect;        /* starting sector counting from 0 */
 796         uint32_t nr_sects;          /* nr of sectors in partition */
 797 } __attribute__((packed));
 798
 799 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 800 static int guess_disk_lchs(BlockDriverState *bs,
 801                            int *pcylinders, int *pheads, int *psectors)
 802 {
 803     uint8_t buf[512];
 804     int ret, i, heads, sectors, cylinders;
 805     struct partition *p;
 806     uint32_t nr_sects;
 807     uint64_t nb_sectors;
 808
 809     bdrv_get_geometry(bs, &nb_sectors);
 810
 811     ret = bdrv_read(bs, 0, buf, 1);
 812     if (ret < 0)
 813         return -1;
 814     /* test msdos magic */
 815     if (buf[510] != 0x55 || buf[511] != 0xaa)
 816         return -1;
 817     for(i = 0; i < 4; i++) {
 818         p = ((struct partition *)(buf + 0x1be)) + i;
 819         nr_sects = le32_to_cpu(p->nr_sects);
 820         if (nr_sects && p->end_head) {
 821             /* We make the assumption that the partition terminates on
 822                a cylinder boundary */
 823             heads = p->end_head + 1;
 824             sectors = p->end_sector & 63;
 825             if (sectors == 0)
 826                 continue;
 827             cylinders = nb_sectors / (heads * sectors);
 828             if (cylinders < 1 || cylinders > 16383)
 829                 continue;
 830             *pheads = heads;
 831             *psectors = sectors;
 832             *pcylinders = cylinders;
 833 #if 0
 834             printf("guessed geometry: LCHS=%d %d %d\n",
 835                    cylinders, heads, sectors);
 836 #endif
 837             return 0;
 838         }
 839     }
 840     return -1;
 841 }
 842
 843 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 844 {
 845     int translation, lba_detected = 0;
 846     int cylinders, heads, secs;
 847     uint64_t nb_sectors;
 848
 849     /* if a geometry hint is available, use it */
 850     bdrv_get_geometry(bs, &nb_sectors);
 851     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 852     translation = bdrv_get_translation_hint(bs);
 853     if (cylinders != 0) {
 854         *pcyls = cylinders;
 855         *pheads = heads;
 856         *psecs = secs;
 857     } else {
 858         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 859             if (heads > 16) {
 860                 /* if heads > 16, it means that a BIOS LBA
 861                    translation was active, so the default
 862                    hardware geometry is OK */
 863                 lba_detected = 1;
 864                 goto default_geometry;
 865             } else {
 866                 *pcyls = cylinders;
 867                 *pheads = heads;
 868                 *psecs = secs;
 869                 /* disable any translation to be in sync with
 870                    the logical geometry */
 871                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 872                     bdrv_set_translation_hint(bs,
 873                                               BIOS_ATA_TRANSLATION_NONE);
 874                 }
 875             }
 876         } else {
 877         default_geometry:
 878             /* if no geometry, use a standard physical disk geometry */
 879             cylinders = nb_sectors / (16 * 63);
 880
 881             if (cylinders > 16383)
 882                 cylinders = 16383;
 883             else if (cylinders < 2)
 884                 cylinders = 2;
 885             *pcyls = cylinders;
 886             *pheads = 16;
 887             *psecs = 63;
 888             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 889                 if ((*pcyls * *pheads) <= 131072) {
 890                     bdrv_set_translation_hint(bs,
 891                                               BIOS_ATA_TRANSLATION_LARGE);
 892                 } else {
 893                     bdrv_set_translation_hint(bs,
 894                                               BIOS_ATA_TRANSLATION_LBA);
 895                 }
 896             }
 897         }
 898         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 899     }
 900 }
 901
 902 void bdrv_set_geometry_hint(BlockDriverState *bs,
 903                             int cyls, int heads, int secs)
 904 {
 905     bs->cyls = cyls;
 906     bs->heads = heads;
 907     bs->secs = secs;
 908 }
 909
 910 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 911 {
 912     bs->type = type;
 913     bs->removable = ((type == BDRV_TYPE_CDROM ||
 914                       type == BDRV_TYPE_FLOPPY));
 915 }
 916
 917 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 918 {
 919     bs->translation = translation;
 920 }
 921
 922 void bdrv_get_geometry_hint(BlockDriverState *bs,
 923                             int *pcyls, int *pheads, int *psecs)
 924 {
 925     *pcyls = bs->cyls;
 926     *pheads = bs->heads;
 927     *psecs = bs->secs;
 928 }
 929
 930 int bdrv_get_type_hint(BlockDriverState *bs)
 931 {
 932     return bs->type;
 933 }
 934
 935 int bdrv_get_translation_hint(BlockDriverState *bs)
 936 {
 937     return bs->translation;
 938 }
 939
 940 int bdrv_is_removable(BlockDriverState *bs)
 941 {
 942     return bs->removable;
 943 }
 944
 945 int bdrv_is_read_only(BlockDriverState *bs)
 946 {
 947     return bs->read_only;
 948 }
 949
 950 int bdrv_is_sg(BlockDriverState *bs)
 951 {
 952     return bs->sg;
 953 }
 954
 955 /* XXX: no longer used */
 956 void bdrv_set_change_cb(BlockDriverState *bs,
 957                         void (*change_cb)(void *opaque), void *opaque)
 958 {
 959     bs->change_cb = change_cb;
 960     bs->change_opaque = opaque;
 961 }
 962
 963 int bdrv_is_encrypted(BlockDriverState *bs)
 964 {
 965     if (bs->backing_hd && bs->backing_hd->encrypted)
 966         return 1;
 967     return bs->encrypted;
 968 }
 969
 970 int bdrv_key_required(BlockDriverState *bs)
 971 {
 972     BlockDriverState *backing_hd = bs->backing_hd;
 973
 974     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 975         return 1;
 976     return (bs->encrypted && !bs->valid_key);
 977 }
 978
 979 int bdrv_set_key(BlockDriverState *bs, const char *key)
 980 {
 981     int ret;
 982     if (bs->backing_hd && bs->backing_hd->encrypted) {
 983         ret = bdrv_set_key(bs->backing_hd, key);
 984         if (ret < 0)
 985             return ret;
 986         if (!bs->encrypted)
 987             return 0;
 988     }
 989     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 990         return -1;
 991     ret = bs->drv->bdrv_set_key(bs, key);
 992     bs->valid_key = (ret == 0);
 993     return ret;
 994 }
 995
 996 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 997 {
 998     if (!bs->drv) {
 999         buf[0] = '\0';
1000     } else {
1001         pstrcpy(buf, buf_size, bs->drv->format_name);
1002     }
1003 }
1004
1005 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
1006                          void *opaque)
1007 {
1008     BlockDriver *drv;
1009
1010     for (drv = first_drv; drv != NULL; drv = drv->next) {
1011         it(opaque, drv->format_name);
1012     }
1013 }
1014
1015 BlockDriverState *bdrv_find(const char *name)
1016 {
1017     BlockDriverState *bs;
1018
1019     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1020         if (!strcmp(name, bs->device_name))
1021             return bs;
1022     }
1023     return NULL;
1024 }
1025
1026 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1027 {
1028     BlockDriverState *bs;
1029
1030     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1031         it(opaque, bs);
1032     }
1033 }
1034
1035 const char *bdrv_get_device_name(BlockDriverState *bs)
1036 {
1037     return bs->device_name;
1038 }
1039
1040 void bdrv_flush(BlockDriverState *bs)
1041 {
1042     if (bs->drv->bdrv_flush)
1043         bs->drv->bdrv_flush(bs);
1044     if (bs->backing_hd)
1045         bdrv_flush(bs->backing_hd);
1046 }
1047
1048 void bdrv_flush_all(void)
1049 {
1050     BlockDriverState *bs;
1051
1052     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1053         if (bs->drv && !bdrv_is_read_only(bs) &&
1054             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1055             bdrv_flush(bs);
1056 }
1057
1058 /*
1059  * Returns true iff the specified sector is present in the disk image. Drivers
1060  * not implementing the functionality are assumed to not support backing files,
1061  * hence all their sectors are reported as allocated.
1062  *
1063  * 'pnum' is set to the number of sectors (including and immediately following
1064  * the specified sector) that are known to be in the same
1065  * allocated/unallocated state.
1066  *
1067  * 'nb_sectors' is the max value 'pnum' should be set to.
1068  */
1069 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1070         int *pnum)
1071 {
1072     int64_t n;
1073     if (!bs->drv->bdrv_is_allocated) {
1074         if (sector_num >= bs->total_sectors) {
1075             *pnum = 0;
1076             return 0;
1077         }
1078         n = bs->total_sectors - sector_num;
1079         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1080         return 1;
1081     }
1082     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1083 }
1084
1085 void bdrv_info(void)
1086 {
1087     BlockDriverState *bs;
1088
1089     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1090         term_printf("%s:", bs->device_name);
1091         term_printf(" type=");
1092         switch(bs->type) {
1093         case BDRV_TYPE_HD:
1094             term_printf("hd");
1095             break;
1096         case BDRV_TYPE_CDROM:
1097             term_printf("cdrom");
1098             break;
1099         case BDRV_TYPE_FLOPPY:
1100             term_printf("floppy");
1101             break;
1102         }
1103         term_printf(" removable=%d", bs->removable);
1104         if (bs->removable) {
1105             term_printf(" locked=%d", bs->locked);
1106         }
1107         if (bs->drv) {
1108             term_printf(" file=");
1109             term_print_filename(bs->filename);
1110             if (bs->backing_file[0] != '\0') {
1111                 term_printf(" backing_file=");
1112                 term_print_filename(bs->backing_file);
1113             }
1114             term_printf(" ro=%d", bs->read_only);
1115             term_printf(" drv=%s", bs->drv->format_name);
1116             term_printf(" encrypted=%d", bdrv_is_encrypted(bs));
1117         } else {
1118             term_printf(" [not inserted]");
1119         }
1120         term_printf("\n");
1121     }
1122 }
1123
1124 /* The "info blockstats" command. */
1125 void bdrv_info_stats (void)
1126 {
1127     BlockDriverState *bs;
1128     BlockDriverInfo bdi;
1129
1130     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1131         term_printf ("%s:"
1132                      " rd_bytes=%" PRIu64
1133                      " wr_bytes=%" PRIu64
1134                      " rd_operations=%" PRIu64
1135                      " wr_operations=%" PRIu64
1136                      ,
1137                      bs->device_name,
1138                      bs->rd_bytes, bs->wr_bytes,
1139                      bs->rd_ops, bs->wr_ops);
1140         if (bdrv_get_info(bs, &bdi) == 0)
1141             term_printf(" high=%" PRId64
1142                         " bytes_free=%" PRId64,
1143                         bdi.highest_alloc, bdi.num_free_bytes);
1144         term_printf("\n");
1145     }
1146 }
1147
1148 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1149 {
1150     if (bs->backing_hd && bs->backing_hd->encrypted)
1151         return bs->backing_file;
1152     else if (bs->encrypted)
1153         return bs->filename;
1154     else
1155         return NULL;
1156 }
1157
1158 void bdrv_get_backing_filename(BlockDriverState *bs,
1159                                char *filename, int filename_size)
1160 {
1161     if (!bs->backing_hd) {
1162         pstrcpy(filename, filename_size, "");
1163     } else {
1164         pstrcpy(filename, filename_size, bs->backing_file);
1165     }
1166 }
1167
1168 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1169                           const uint8_t *buf, int nb_sectors)
1170 {
1171     BlockDriver *drv = bs->drv;
1172     if (!drv)
1173         return -ENOMEDIUM;
1174     if (!drv->bdrv_write_compressed)
1175         return -ENOTSUP;
1176     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1177 }
1178
1179 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1180 {
1181     BlockDriver *drv = bs->drv;
1182     if (!drv)
1183         return -ENOMEDIUM;
1184     if (!drv->bdrv_get_info)
1185         return -ENOTSUP;
1186     memset(bdi, 0, sizeof(*bdi));
1187     return drv->bdrv_get_info(bs, bdi);
1188 }
1189
1190 /**************************************************************/
1191 /* handling of snapshots */
1192
1193 int bdrv_snapshot_create(BlockDriverState *bs,
1194                          QEMUSnapshotInfo *sn_info)
1195 {
1196     BlockDriver *drv = bs->drv;
1197     if (!drv)
1198         return -ENOMEDIUM;
1199     if (!drv->bdrv_snapshot_create)
1200         return -ENOTSUP;
1201     return drv->bdrv_snapshot_create(bs, sn_info);
1202 }
1203
1204 int bdrv_snapshot_goto(BlockDriverState *bs,
1205                        const char *snapshot_id)
1206 {
1207     BlockDriver *drv = bs->drv;
1208     if (!drv)
1209         return -ENOMEDIUM;
1210     if (!drv->bdrv_snapshot_goto)
1211         return -ENOTSUP;
1212     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1213 }
1214
1215 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1216 {
1217     BlockDriver *drv = bs->drv;
1218     if (!drv)
1219         return -ENOMEDIUM;
1220     if (!drv->bdrv_snapshot_delete)
1221         return -ENOTSUP;
1222     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1223 }
1224
1225 int bdrv_snapshot_list(BlockDriverState *bs,
1226                        QEMUSnapshotInfo **psn_info)
1227 {
1228     BlockDriver *drv = bs->drv;
1229     if (!drv)
1230         return -ENOMEDIUM;
1231     if (!drv->bdrv_snapshot_list)
1232         return -ENOTSUP;
1233     return drv->bdrv_snapshot_list(bs, psn_info);
1234 }
1235
1236 #define NB_SUFFIXES 4
1237
1238 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1239 {
1240     static const char suffixes[NB_SUFFIXES] = "KMGT";
1241     int64_t base;
1242     int i;
1243
1244     if (size <= 999) {
1245         snprintf(buf, buf_size, "%" PRId64, size);
1246     } else {
1247         base = 1024;
1248         for(i = 0; i < NB_SUFFIXES; i++) {
1249             if (size < (10 * base)) {
1250                 snprintf(buf, buf_size, "%0.1f%c",
1251                          (double)size / base,
1252                          suffixes[i]);
1253                 break;
1254             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1255                 snprintf(buf, buf_size, "%" PRId64 "%c",
1256                          ((size + (base >> 1)) / base),
1257                          suffixes[i]);
1258                 break;
1259             }
1260             base = base * 1024;
1261         }
1262     }
1263     return buf;
1264 }
1265
1266 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1267 {
1268     char buf1[128], date_buf[128], clock_buf[128];
1269 #ifdef _WIN32
1270     struct tm *ptm;
1271 #else
1272     struct tm tm;
1273 #endif
1274     time_t ti;
1275     int64_t secs;
1276
1277     if (!sn) {
1278         snprintf(buf, buf_size,
1279                  "%-10s%-20s%7s%20s%15s",
1280                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1281     } else {
1282         ti = sn->date_sec;
1283 #ifdef _WIN32
1284         ptm = localtime(&ti);
1285         strftime(date_buf, sizeof(date_buf),
1286                  "%Y-%m-%d %H:%M:%S", ptm);
1287 #else
1288         localtime_r(&ti, &tm);
1289         strftime(date_buf, sizeof(date_buf),
1290                  "%Y-%m-%d %H:%M:%S", &tm);
1291 #endif
1292         secs = sn->vm_clock_nsec / 1000000000;
1293         snprintf(clock_buf, sizeof(clock_buf),
1294                  "%02d:%02d:%02d.%03d",
1295                  (int)(secs / 3600),
1296                  (int)((secs / 60) % 60),
1297                  (int)(secs % 60),
1298                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1299         snprintf(buf, buf_size,
1300                  "%-10s%-20s%7s%20s%15s",
1301                  sn->id_str, sn->name,
1302                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1303                  date_buf,
1304                  clock_buf);
1305     }
1306     return buf;
1307 }
1308
1309
1310 /**************************************************************/
1311 /* async I/Os */
1312
1313 typedef struct VectorTranslationState {
1314     QEMUIOVector *iov;
1315     uint8_t *bounce;
1316     int is_write;
1317     BlockDriverAIOCB *aiocb;
1318     BlockDriverAIOCB *this_aiocb;
1319 } VectorTranslationState;
1320
1321 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1322 {
1323     VectorTranslationState *s = opaque;
1324
1325     if (!s->is_write) {
1326         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1327     }
1328     qemu_vfree(s->bounce);
1329     s->this_aiocb->cb(s->this_aiocb->opaque, ret);
1330     qemu_aio_release(s->this_aiocb);
1331 }
1332
1333 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1334                                             int64_t sector_num,
1335                                             QEMUIOVector *iov,
1336                                             int nb_sectors,
1337                                             BlockDriverCompletionFunc *cb,
1338                                             void *opaque,
1339                                             int is_write)
1340
1341 {
1342     VectorTranslationState *s = qemu_mallocz(sizeof(*s));
1343     BlockDriverAIOCB *aiocb = qemu_aio_get(bs, cb, opaque);
1344
1345     s->this_aiocb = aiocb;
1346     s->iov = iov;
1347     s->bounce = qemu_memalign(512, nb_sectors * 512);
1348     s->is_write = is_write;
1349     if (is_write) {
1350         qemu_iovec_to_buffer(s->iov, s->bounce);
1351         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1352                                   bdrv_aio_rw_vector_cb, s);
1353     } else {
1354         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1355                                  bdrv_aio_rw_vector_cb, s);
1356     }
1357     return aiocb;
1358 }
1359
1360 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1361                                  QEMUIOVector *iov, int nb_sectors,
1362                                  BlockDriverCompletionFunc *cb, void *opaque)
1363 {
1364     if (bdrv_check_request(bs, sector_num, nb_sectors))
1365         return NULL;
1366
1367     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1368                               cb, opaque, 0);
1369 }
1370
1371 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1372                                   QEMUIOVector *iov, int nb_sectors,
1373                                   BlockDriverCompletionFunc *cb, void *opaque)
1374 {
1375     if (bdrv_check_request(bs, sector_num, nb_sectors))
1376         return NULL;
1377
1378     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1379                               cb, opaque, 1);
1380 }
1381
1382 BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1383                                 uint8_t *buf, int nb_sectors,
1384                                 BlockDriverCompletionFunc *cb, void *opaque)
1385 {
1386     BlockDriver *drv = bs->drv;
1387     BlockDriverAIOCB *ret;
1388
1389     if (!drv)
1390         return NULL;
1391     if (bdrv_check_request(bs, sector_num, nb_sectors))
1392         return NULL;
1393
1394     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1395
1396     if (ret) {
1397         /* Update stats even though technically transfer has not happened. */
1398         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1399         bs->rd_ops ++;
1400     }
1401
1402     return ret;
1403 }
1404
1405 BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1406                                  const uint8_t *buf, int nb_sectors,
1407                                  BlockDriverCompletionFunc *cb, void *opaque)
1408 {
1409     BlockDriver *drv = bs->drv;
1410     BlockDriverAIOCB *ret;
1411
1412     if (!drv)
1413         return NULL;
1414     if (bs->read_only)
1415         return NULL;
1416     if (bdrv_check_request(bs, sector_num, nb_sectors))
1417         return NULL;
1418
1419     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1420
1421     if (ret) {
1422         /* Update stats even though technically transfer has not happened. */
1423         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1424         bs->wr_ops ++;
1425     }
1426
1427     return ret;
1428 }
1429
1430 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1431 {
1432     BlockDriver *drv = acb->bs->drv;
1433
1434     if (acb->cb == bdrv_aio_rw_vector_cb) {
1435         VectorTranslationState *s = acb->opaque;
1436         acb = s->aiocb;
1437     }
1438
1439     drv->bdrv_aio_cancel(acb);
1440 }
1441
1442
1443 /**************************************************************/
1444 /* async block device emulation */
1445
1446 static void bdrv_aio_bh_cb(void *opaque)
1447 {
1448     BlockDriverAIOCBSync *acb = opaque;
1449     acb->common.cb(acb->common.opaque, acb->ret);
1450     qemu_aio_release(acb);
1451 }
1452
1453 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1454         int64_t sector_num, uint8_t *buf, int nb_sectors,
1455         BlockDriverCompletionFunc *cb, void *opaque)
1456 {
1457     BlockDriverAIOCBSync *acb;
1458     int ret;
1459
1460     acb = qemu_aio_get(bs, cb, opaque);
1461     if (!acb->bh)
1462         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1463     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1464     acb->ret = ret;
1465     qemu_bh_schedule(acb->bh);
1466     return &acb->common;
1467 }
1468
1469 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1470         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1471         BlockDriverCompletionFunc *cb, void *opaque)
1472 {
1473     BlockDriverAIOCBSync *acb;
1474     int ret;
1475
1476     acb = qemu_aio_get(bs, cb, opaque);
1477     if (!acb->bh)
1478         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1479     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1480     acb->ret = ret;
1481     qemu_bh_schedule(acb->bh);
1482     return &acb->common;
1483 }
1484
1485 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1486 {
1487     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1488     qemu_bh_cancel(acb->bh);
1489     qemu_aio_release(acb);
1490 }
1491
1492 /**************************************************************/
1493 /* sync block device emulation */
1494
1495 static void bdrv_rw_em_cb(void *opaque, int ret)
1496 {
1497     *(int *)opaque = ret;
1498 }
1499
1500 #define NOT_DONE 0x7fffffff
1501
1502 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1503                         uint8_t *buf, int nb_sectors)
1504 {
1505     int async_ret;
1506     BlockDriverAIOCB *acb;
1507
1508     async_ret = NOT_DONE;
1509     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1510                         bdrv_rw_em_cb, &async_ret);
1511     if (acb == NULL)
1512         return -1;
1513
1514     while (async_ret == NOT_DONE) {
1515         qemu_aio_wait();
1516     }
1517
1518     return async_ret;
1519 }
1520
1521 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1522                          const uint8_t *buf, int nb_sectors)
1523 {
1524     int async_ret;
1525     BlockDriverAIOCB *acb;
1526
1527     async_ret = NOT_DONE;
1528     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1529                          bdrv_rw_em_cb, &async_ret);
1530     if (acb == NULL)
1531         return -1;
1532     while (async_ret == NOT_DONE) {
1533         qemu_aio_wait();
1534     }
1535     return async_ret;
1536 }
1537
1538 void bdrv_init(void)
1539 {
1540     bdrv_register(&bdrv_raw);
1541     bdrv_register(&bdrv_host_device);
1542 #ifndef _WIN32
1543     bdrv_register(&bdrv_cow);
1544 #endif
1545     bdrv_register(&bdrv_qcow);
1546     bdrv_register(&bdrv_vmdk);
1547     bdrv_register(&bdrv_cloop);
1548     bdrv_register(&bdrv_dmg);
1549     bdrv_register(&bdrv_bochs);
1550     bdrv_register(&bdrv_vpc);
1551     bdrv_register(&bdrv_vvfat);
1552     bdrv_register(&bdrv_qcow2);
1553     bdrv_register(&bdrv_parallels);
1554     bdrv_register(&bdrv_nbd);
1555 }
1556
1557 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1558                    void *opaque)
1559 {
1560     BlockDriver *drv;
1561     BlockDriverAIOCB *acb;
1562
1563     drv = bs->drv;
1564     if (drv->free_aiocb) {
1565         acb = drv->free_aiocb;
1566         drv->free_aiocb = acb->next;
1567     } else {
1568         acb = qemu_mallocz(drv->aiocb_size);
1569     }
1570     acb->bs = bs;
1571     acb->cb = cb;
1572     acb->opaque = opaque;
1573     return acb;
1574 }
1575
1576 void qemu_aio_release(void *p)
1577 {
1578     BlockDriverAIOCB *acb = p;
1579     BlockDriver *drv = acb->bs->drv;
1580     acb->next = drv->free_aiocb;
1581     drv->free_aiocb = acb;
1582 }
1583
1584 /**************************************************************/
1585 /* removable device support */
1586
1587 /**
1588  * Return TRUE if the media is present
1589  */
1590 int bdrv_is_inserted(BlockDriverState *bs)
1591 {
1592     BlockDriver *drv = bs->drv;
1593     int ret;
1594     if (!drv)
1595         return 0;
1596     if (!drv->bdrv_is_inserted)
1597         return 1;
1598     ret = drv->bdrv_is_inserted(bs);
1599     return ret;
1600 }
1601
1602 /**
1603  * Return TRUE if the media changed since the last call to this
1604  * function. It is currently only used for floppy disks
1605  */
1606 int bdrv_media_changed(BlockDriverState *bs)
1607 {
1608     BlockDriver *drv = bs->drv;
1609     int ret;
1610
1611     if (!drv || !drv->bdrv_media_changed)
1612         ret = -ENOTSUP;
1613     else
1614         ret = drv->bdrv_media_changed(bs);
1615     if (ret == -ENOTSUP)
1616         ret = bs->media_changed;
1617     bs->media_changed = 0;
1618     return ret;
1619 }
1620
1621 /**
1622  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1623  */
1624 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1625 {
1626     BlockDriver *drv = bs->drv;
1627     int ret;
1628
1629     if (!drv || !drv->bdrv_eject) {
1630         ret = -ENOTSUP;
1631     } else {
1632         ret = drv->bdrv_eject(bs, eject_flag);
1633     }
1634     if (ret == -ENOTSUP) {
1635         if (eject_flag)
1636             bdrv_close(bs);
1637     }
1638 }
1639
1640 int bdrv_is_locked(BlockDriverState *bs)
1641 {
1642     return bs->locked;
1643 }
1644
1645 /**
1646  * Lock or unlock the media (if it is locked, the user won't be able
1647  * to eject it manually).
1648  */
1649 void bdrv_set_locked(BlockDriverState *bs, int locked)
1650 {
1651     BlockDriver *drv = bs->drv;
1652
1653     bs->locked = locked;
1654     if (drv && drv->bdrv_set_locked) {
1655         drv->bdrv_set_locked(bs, locked);
1656     }
1657 }
1658
1659 /* needed for generic scsi interface */
1660
1661 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1662 {
1663     BlockDriver *drv = bs->drv;
1664
1665     if (drv && drv->bdrv_ioctl)
1666         return drv->bdrv_ioctl(bs, req, buf);
1667     return -ENOTSUP;
1668 }