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