block.c

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