block.c

   1 /*
   2  * QEMU System Emulator block driver
   3  *
   4  * Copyright (c) 2003 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include "config-host.h"
  25 #ifdef HOST_BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "monitor.h"
  32 #include "block_int.h"
  33
  34 #ifdef HOST_BSD
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/ioctl.h>
  38 #ifndef __DragonFly__
  39 #include <sys/disk.h>
  40 #endif
  41 #endif
  42
  43 #ifdef _WIN32
  44 #include <windows.h>
  45 #endif
  46
  47 #define SECTOR_BITS 9
  48 #define SECTOR_SIZE (1 << SECTOR_BITS)
  49
  50 static AIOPool vectored_aio_pool;
  51
  52 typedef struct BlockDriverAIOCBSync {
  53     BlockDriverAIOCB common;
  54     QEMUBH *bh;
  55     int ret;
  56 } BlockDriverAIOCBSync;
  57
  58 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  59         int64_t sector_num, uint8_t *buf, int nb_sectors,
  60         BlockDriverCompletionFunc *cb, void *opaque);
  61 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  62         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  63         BlockDriverCompletionFunc *cb, void *opaque);
  64 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  65 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  66                         uint8_t *buf, int nb_sectors);
  67 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  68                          const uint8_t *buf, int nb_sectors);
  69
  70 BlockDriverState *bdrv_first;
  71
  72 static BlockDriver *first_drv;
  73
  74 int path_is_absolute(const char *path)
  75 {
  76     const char *p;
  77 #ifdef _WIN32
  78     /* specific case for names like: "\\.\d:" */
  79     if (*path == '/' || *path == '\\')
  80         return 1;
  81 #endif
  82     p = strchr(path, ':');
  83     if (p)
  84         p++;
  85     else
  86         p = path;
  87 #ifdef _WIN32
  88     return (*p == '/' || *p == '\\');
  89 #else
  90     return (*p == '/');
  91 #endif
  92 }
  93
  94 /* if filename is absolute, just copy it to dest. Otherwise, build a
  95    path to it by considering it is relative to base_path. URL are
  96    supported. */
  97 void path_combine(char *dest, int dest_size,
  98                   const char *base_path,
  99                   const char *filename)
 100 {
 101     const char *p, *p1;
 102     int len;
 103
 104     if (dest_size <= 0)
 105         return;
 106     if (path_is_absolute(filename)) {
 107         pstrcpy(dest, dest_size, filename);
 108     } else {
 109         p = strchr(base_path, ':');
 110         if (p)
 111             p++;
 112         else
 113             p = base_path;
 114         p1 = strrchr(base_path, '/');
 115 #ifdef _WIN32
 116         {
 117             const char *p2;
 118             p2 = strrchr(base_path, '\\');
 119             if (!p1 || p2 > p1)
 120                 p1 = p2;
 121         }
 122 #endif
 123         if (p1)
 124             p1++;
 125         else
 126             p1 = base_path;
 127         if (p1 > p)
 128             p = p1;
 129         len = p - base_path;
 130         if (len > dest_size - 1)
 131             len = dest_size - 1;
 132         memcpy(dest, base_path, len);
 133         dest[len] = '\0';
 134         pstrcat(dest, dest_size, filename);
 135     }
 136 }
 137
 138
 139 static void bdrv_register(BlockDriver *bdrv)
 140 {
 141     if (!bdrv->bdrv_aio_read) {
 142         /* add AIO emulation layer */
 143         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 144         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 145         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 146         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 147     } else if (!bdrv->bdrv_read) {
 148         /* add synchronous IO emulation layer */
 149         bdrv->bdrv_read = bdrv_read_em;
 150         bdrv->bdrv_write = bdrv_write_em;
 151     }
 152     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 153     bdrv->next = first_drv;
 154     first_drv = bdrv;
 155 }
 156
 157 /* create a new block device (by default it is empty) */
 158 BlockDriverState *bdrv_new(const char *device_name)
 159 {
 160     BlockDriverState **pbs, *bs;
 161
 162     bs = qemu_mallocz(sizeof(BlockDriverState));
 163     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 164     if (device_name[0] != '\0') {
 165         /* insert at the end */
 166         pbs = &bdrv_first;
 167         while (*pbs != NULL)
 168             pbs = &(*pbs)->next;
 169         *pbs = bs;
 170     }
 171     return bs;
 172 }
 173
 174 BlockDriver *bdrv_find_format(const char *format_name)
 175 {
 176     BlockDriver *drv1;
 177     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 178         if (!strcmp(drv1->format_name, format_name))
 179             return drv1;
 180     }
 181     return NULL;
 182 }
 183
 184 int bdrv_create2(BlockDriver *drv,
 185                 const char *filename, int64_t size_in_sectors,
 186                 const char *backing_file, const char *backing_format,
 187                 int flags)
 188 {
 189     if (drv->bdrv_create2)
 190         return drv->bdrv_create2(filename, size_in_sectors, backing_file,
 191                                  backing_format, flags);
 192     if (drv->bdrv_create)
 193         return drv->bdrv_create(filename, size_in_sectors, backing_file,
 194                                 flags);
 195     return -ENOTSUP;
 196 }
 197
 198 int bdrv_create(BlockDriver *drv,
 199                 const char *filename, int64_t size_in_sectors,
 200                 const char *backing_file, int flags)
 201 {
 202     if (!drv->bdrv_create)
 203         return -ENOTSUP;
 204     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 205 }
 206
 207 #ifdef _WIN32
 208 void get_tmp_filename(char *filename, int size)
 209 {
 210     char temp_dir[MAX_PATH];
 211
 212     GetTempPath(MAX_PATH, temp_dir);
 213     GetTempFileName(temp_dir, "qem", 0, filename);
 214 }
 215 #else
 216 void get_tmp_filename(char *filename, int size)
 217 {
 218     int fd;
 219     const char *tmpdir;
 220     /* XXX: race condition possible */
 221     tmpdir = getenv("TMPDIR");
 222     if (!tmpdir)
 223         tmpdir = "/tmp";
 224     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 225     fd = mkstemp(filename);
 226     close(fd);
 227 }
 228 #endif
 229
 230 #ifdef _WIN32
 231 static int is_windows_drive_prefix(const char *filename)
 232 {
 233     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 234              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 235             filename[1] == ':');
 236 }
 237
 238 static int is_windows_drive(const char *filename)
 239 {
 240     if (is_windows_drive_prefix(filename) &&
 241         filename[2] == '\0')
 242         return 1;
 243     if (strstart(filename, "\\\\.\\", NULL) ||
 244         strstart(filename, "//./", NULL))
 245         return 1;
 246     return 0;
 247 }
 248 #endif
 249
 250 static BlockDriver *find_protocol(const char *filename)
 251 {
 252     BlockDriver *drv1;
 253     char protocol[128];
 254     int len;
 255     const char *p;
 256
 257 #ifdef _WIN32
 258     if (is_windows_drive(filename) ||
 259         is_windows_drive_prefix(filename))
 260         return &bdrv_raw;
 261 #endif
 262     p = strchr(filename, ':');
 263     if (!p)
 264         return &bdrv_raw;
 265     len = p - filename;
 266     if (len > sizeof(protocol) - 1)
 267         len = sizeof(protocol) - 1;
 268     memcpy(protocol, filename, len);
 269     protocol[len] = '\0';
 270     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 271         if (drv1->protocol_name &&
 272             !strcmp(drv1->protocol_name, protocol))
 273             return drv1;
 274     }
 275     return NULL;
 276 }
 277
 278 /* XXX: force raw format if block or character device ? It would
 279    simplify the BSD case */
 280 static BlockDriver *find_image_format(const char *filename)
 281 {
 282     int ret, score, score_max;
 283     BlockDriver *drv1, *drv;
 284     uint8_t buf[2048];
 285     BlockDriverState *bs;
 286
 287     /* detect host devices. By convention, /dev/cdrom[N] is always
 288        recognized as a host CDROM */
 289     if (strstart(filename, "/dev/cdrom", NULL))
 290         return &bdrv_host_device;
 291 #ifdef _WIN32
 292     if (is_windows_drive(filename))
 293         return &bdrv_host_device;
 294 #else
 295     {
 296         struct stat st;
 297         if (stat(filename, &st) >= 0 &&
 298             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 299             return &bdrv_host_device;
 300         }
 301     }
 302 #endif
 303
 304     drv = find_protocol(filename);
 305     /* no need to test disk image formats for vvfat */
 306     if (drv == &bdrv_vvfat)
 307         return drv;
 308
 309     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 310     if (ret < 0)
 311         return NULL;
 312     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 313     bdrv_delete(bs);
 314     if (ret < 0) {
 315         return NULL;
 316     }
 317
 318     score_max = 0;
 319     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 320         if (drv1->bdrv_probe) {
 321             score = drv1->bdrv_probe(buf, ret, filename);
 322             if (score > score_max) {
 323                 score_max = score;
 324                 drv = drv1;
 325             }
 326         }
 327     }
 328     return drv;
 329 }
 330
 331 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 332 {
 333     BlockDriverState *bs;
 334     int ret;
 335
 336     bs = bdrv_new("");
 337     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 338     if (ret < 0) {
 339         bdrv_delete(bs);
 340         return ret;
 341     }
 342     bs->growable = 1;
 343     *pbs = bs;
 344     return 0;
 345 }
 346
 347 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 348 {
 349     return bdrv_open2(bs, filename, flags, NULL);
 350 }
 351
 352 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 353                BlockDriver *drv)
 354 {
 355     int ret, open_flags;
 356     char tmp_filename[PATH_MAX];
 357     char backing_filename[PATH_MAX];
 358
 359     bs->read_only = 0;
 360     bs->is_temporary = 0;
 361     bs->encrypted = 0;
 362     bs->valid_key = 0;
 363
 364     if (flags & BDRV_O_SNAPSHOT) {
 365         BlockDriverState *bs1;
 366         int64_t total_size;
 367         int is_protocol = 0;
 368
 369         /* if snapshot, we create a temporary backing file and open it
 370            instead of opening 'filename' directly */
 371
 372         /* if there is a backing file, use it */
 373         bs1 = bdrv_new("");
 374         ret = bdrv_open2(bs1, filename, 0, drv);
 375         if (ret < 0) {
 376             bdrv_delete(bs1);
 377             return ret;
 378         }
 379         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 380
 381         if (bs1->drv && bs1->drv->protocol_name)
 382             is_protocol = 1;
 383
 384         bdrv_delete(bs1);
 385
 386         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 387
 388         /* Real path is meaningless for protocols */
 389         if (is_protocol)
 390             snprintf(backing_filename, sizeof(backing_filename),
 391                      "%s", filename);
 392         else
 393             realpath(filename, backing_filename);
 394
 395         ret = bdrv_create2(&bdrv_qcow2, tmp_filename,
 396                            total_size, backing_filename,
 397                            (drv ? drv->format_name : NULL), 0);
 398         if (ret < 0) {
 399             return ret;
 400         }
 401         filename = tmp_filename;
 402         drv = &bdrv_qcow2;
 403         bs->is_temporary = 1;
 404     }
 405
 406     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 407     if (flags & BDRV_O_FILE) {
 408         drv = find_protocol(filename);
 409     } else if (!drv) {
 410         drv = find_image_format(filename);
 411     }
 412     if (!drv) {
 413         ret = -ENOENT;
 414         goto unlink_and_fail;
 415     }
 416     bs->drv = drv;
 417     bs->opaque = qemu_mallocz(drv->instance_size);
 418     /* Note: for compatibility, we open disk image files as RDWR, and
 419        RDONLY as fallback */
 420     if (!(flags & BDRV_O_FILE))
 421         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 422     else
 423         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 424     ret = drv->bdrv_open(bs, filename, open_flags);
 425     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 426         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 427         bs->read_only = 1;
 428     }
 429     if (ret < 0) {
 430         qemu_free(bs->opaque);
 431         bs->opaque = NULL;
 432         bs->drv = NULL;
 433     unlink_and_fail:
 434         if (bs->is_temporary)
 435             unlink(filename);
 436         return ret;
 437     }
 438     if (drv->bdrv_getlength) {
 439         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 440     }
 441 #ifndef _WIN32
 442     if (bs->is_temporary) {
 443         unlink(filename);
 444     }
 445 #endif
 446     if (bs->backing_file[0] != '\0') {
 447         /* if there is a backing file, use it */
 448         BlockDriver *back_drv = NULL;
 449         bs->backing_hd = bdrv_new("");
 450         path_combine(backing_filename, sizeof(backing_filename),
 451                      filename, bs->backing_file);
 452         if (bs->backing_format[0] != '\0')
 453             back_drv = bdrv_find_format(bs->backing_format);
 454         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 455                          back_drv);
 456         if (ret < 0) {
 457             bdrv_close(bs);
 458             return ret;
 459         }
 460     }
 461
 462     if (!bdrv_key_required(bs)) {
 463         /* call the change callback */
 464         bs->media_changed = 1;
 465         if (bs->change_cb)
 466             bs->change_cb(bs->change_opaque);
 467     }
 468     return 0;
 469 }
 470
 471 void bdrv_close(BlockDriverState *bs)
 472 {
 473     if (bs->drv) {
 474         if (bs->backing_hd)
 475             bdrv_delete(bs->backing_hd);
 476         bs->drv->bdrv_close(bs);
 477         qemu_free(bs->opaque);
 478 #ifdef _WIN32
 479         if (bs->is_temporary) {
 480             unlink(bs->filename);
 481         }
 482 #endif
 483         bs->opaque = NULL;
 484         bs->drv = NULL;
 485
 486         /* call the change callback */
 487         bs->media_changed = 1;
 488         if (bs->change_cb)
 489             bs->change_cb(bs->change_opaque);
 490     }
 491 }
 492
 493 void bdrv_delete(BlockDriverState *bs)
 494 {
 495     BlockDriverState **pbs;
 496
 497     pbs = &bdrv_first;
 498     while (*pbs != bs && *pbs != NULL)
 499         pbs = &(*pbs)->next;
 500     if (*pbs == bs)
 501         *pbs = bs->next;
 502
 503     bdrv_close(bs);
 504     qemu_free(bs);
 505 }
 506
 507 /* commit COW file into the raw image */
 508 int bdrv_commit(BlockDriverState *bs)
 509 {
 510     BlockDriver *drv = bs->drv;
 511     int64_t i, total_sectors;
 512     int n, j;
 513     unsigned char sector[512];
 514
 515     if (!drv)
 516         return -ENOMEDIUM;
 517
 518     if (bs->read_only) {
 519         return -EACCES;
 520     }
 521
 522     if (!bs->backing_hd) {
 523         return -ENOTSUP;
 524     }
 525
 526     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 527     for (i = 0; i < total_sectors;) {
 528         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 529             for(j = 0; j < n; j++) {
 530                 if (bdrv_read(bs, i, sector, 1) != 0) {
 531                     return -EIO;
 532                 }
 533
 534                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 535                     return -EIO;
 536                 }
 537                 i++;
 538             }
 539         } else {
 540             i += n;
 541         }
 542     }
 543
 544     if (drv->bdrv_make_empty)
 545         return drv->bdrv_make_empty(bs);
 546
 547     return 0;
 548 }
 549
 550 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 551                                    size_t size)
 552 {
 553     int64_t len;
 554
 555     if (!bdrv_is_inserted(bs))
 556         return -ENOMEDIUM;
 557
 558     if (bs->growable)
 559         return 0;
 560
 561     len = bdrv_getlength(bs);
 562
 563     if ((offset + size) > len)
 564         return -EIO;
 565
 566     return 0;
 567 }
 568
 569 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 570                               int nb_sectors)
 571 {
 572     int64_t offset;
 573
 574     /* Deal with byte accesses */
 575     if (sector_num < 0)
 576         offset = -sector_num;
 577     else
 578         offset = sector_num * 512;
 579
 580     return bdrv_check_byte_request(bs, offset, nb_sectors * 512);
 581 }
 582
 583 /* return < 0 if error. See bdrv_write() for the return codes */
 584 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 585               uint8_t *buf, int nb_sectors)
 586 {
 587     BlockDriver *drv = bs->drv;
 588
 589     if (!drv)
 590         return -ENOMEDIUM;
 591     if (bdrv_check_request(bs, sector_num, nb_sectors))
 592         return -EIO;
 593
 594     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 595 }
 596
 597 /* Return < 0 if error. Important errors are:
 598   -EIO         generic I/O error (may happen for all errors)
 599   -ENOMEDIUM   No media inserted.
 600   -EINVAL      Invalid sector number or nb_sectors
 601   -EACCES      Trying to write a read-only device
 602 */
 603 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 604                const uint8_t *buf, int nb_sectors)
 605 {
 606     BlockDriver *drv = bs->drv;
 607     if (!bs->drv)
 608         return -ENOMEDIUM;
 609     if (bs->read_only)
 610         return -EACCES;
 611     if (bdrv_check_request(bs, sector_num, nb_sectors))
 612         return -EIO;
 613
 614     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 615 }
 616
 617 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 618                void *buf, int count1)
 619 {
 620     uint8_t tmp_buf[SECTOR_SIZE];
 621     int len, nb_sectors, count;
 622     int64_t sector_num;
 623
 624     count = count1;
 625     /* first read to align to sector start */
 626     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 627     if (len > count)
 628         len = count;
 629     sector_num = offset >> SECTOR_BITS;
 630     if (len > 0) {
 631         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 632             return -EIO;
 633         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 634         count -= len;
 635         if (count == 0)
 636             return count1;
 637         sector_num++;
 638         buf += len;
 639     }
 640
 641     /* read the sectors "in place" */
 642     nb_sectors = count >> SECTOR_BITS;
 643     if (nb_sectors > 0) {
 644         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 645             return -EIO;
 646         sector_num += nb_sectors;
 647         len = nb_sectors << SECTOR_BITS;
 648         buf += len;
 649         count -= len;
 650     }
 651
 652     /* add data from the last sector */
 653     if (count > 0) {
 654         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 655             return -EIO;
 656         memcpy(buf, tmp_buf, count);
 657     }
 658     return count1;
 659 }
 660
 661 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 662                 const void *buf, int count1)
 663 {
 664     uint8_t tmp_buf[SECTOR_SIZE];
 665     int len, nb_sectors, count;
 666     int64_t sector_num;
 667
 668     count = count1;
 669     /* first write to align to sector start */
 670     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 671     if (len > count)
 672         len = count;
 673     sector_num = offset >> SECTOR_BITS;
 674     if (len > 0) {
 675         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 676             return -EIO;
 677         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 678         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 679             return -EIO;
 680         count -= len;
 681         if (count == 0)
 682             return count1;
 683         sector_num++;
 684         buf += len;
 685     }
 686
 687     /* write the sectors "in place" */
 688     nb_sectors = count >> SECTOR_BITS;
 689     if (nb_sectors > 0) {
 690         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 691             return -EIO;
 692         sector_num += nb_sectors;
 693         len = nb_sectors << SECTOR_BITS;
 694         buf += len;
 695         count -= len;
 696     }
 697
 698     /* add data from the last sector */
 699     if (count > 0) {
 700         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 701             return -EIO;
 702         memcpy(tmp_buf, buf, count);
 703         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 704             return -EIO;
 705     }
 706     return count1;
 707 }
 708
 709 /**
 710  * Truncate file to 'offset' bytes (needed only for file protocols)
 711  */
 712 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 713 {
 714     BlockDriver *drv = bs->drv;
 715     if (!drv)
 716         return -ENOMEDIUM;
 717     if (!drv->bdrv_truncate)
 718         return -ENOTSUP;
 719     return drv->bdrv_truncate(bs, offset);
 720 }
 721
 722 /**
 723  * Length of a file in bytes. Return < 0 if error or unknown.
 724  */
 725 int64_t bdrv_getlength(BlockDriverState *bs)
 726 {
 727     BlockDriver *drv = bs->drv;
 728     if (!drv)
 729         return -ENOMEDIUM;
 730     if (!drv->bdrv_getlength) {
 731         /* legacy mode */
 732         return bs->total_sectors * SECTOR_SIZE;
 733     }
 734     return drv->bdrv_getlength(bs);
 735 }
 736
 737 /* return 0 as number of sectors if no device present or error */
 738 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 739 {
 740     int64_t length;
 741     length = bdrv_getlength(bs);
 742     if (length < 0)
 743         length = 0;
 744     else
 745         length = length >> SECTOR_BITS;
 746     *nb_sectors_ptr = length;
 747 }
 748
 749 struct partition {
 750         uint8_t boot_ind;           /* 0x80 - active */
 751         uint8_t head;               /* starting head */
 752         uint8_t sector;             /* starting sector */
 753         uint8_t cyl;                /* starting cylinder */
 754         uint8_t sys_ind;            /* What partition type */
 755         uint8_t end_head;           /* end head */
 756         uint8_t end_sector;         /* end sector */
 757         uint8_t end_cyl;            /* end cylinder */
 758         uint32_t start_sect;        /* starting sector counting from 0 */
 759         uint32_t nr_sects;          /* nr of sectors in partition */
 760 } __attribute__((packed));
 761
 762 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 763 static int guess_disk_lchs(BlockDriverState *bs,
 764                            int *pcylinders, int *pheads, int *psectors)
 765 {
 766     uint8_t buf[512];
 767     int ret, i, heads, sectors, cylinders;
 768     struct partition *p;
 769     uint32_t nr_sects;
 770     uint64_t nb_sectors;
 771
 772     bdrv_get_geometry(bs, &nb_sectors);
 773
 774     ret = bdrv_read(bs, 0, buf, 1);
 775     if (ret < 0)
 776         return -1;
 777     /* test msdos magic */
 778     if (buf[510] != 0x55 || buf[511] != 0xaa)
 779         return -1;
 780     for(i = 0; i < 4; i++) {
 781         p = ((struct partition *)(buf + 0x1be)) + i;
 782         nr_sects = le32_to_cpu(p->nr_sects);
 783         if (nr_sects && p->end_head) {
 784             /* We make the assumption that the partition terminates on
 785                a cylinder boundary */
 786             heads = p->end_head + 1;
 787             sectors = p->end_sector & 63;
 788             if (sectors == 0)
 789                 continue;
 790             cylinders = nb_sectors / (heads * sectors);
 791             if (cylinders < 1 || cylinders > 16383)
 792                 continue;
 793             *pheads = heads;
 794             *psectors = sectors;
 795             *pcylinders = cylinders;
 796 #if 0
 797             printf("guessed geometry: LCHS=%d %d %d\n",
 798                    cylinders, heads, sectors);
 799 #endif
 800             return 0;
 801         }
 802     }
 803     return -1;
 804 }
 805
 806 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 807 {
 808     int translation, lba_detected = 0;
 809     int cylinders, heads, secs;
 810     uint64_t nb_sectors;
 811
 812     /* if a geometry hint is available, use it */
 813     bdrv_get_geometry(bs, &nb_sectors);
 814     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 815     translation = bdrv_get_translation_hint(bs);
 816     if (cylinders != 0) {
 817         *pcyls = cylinders;
 818         *pheads = heads;
 819         *psecs = secs;
 820     } else {
 821         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 822             if (heads > 16) {
 823                 /* if heads > 16, it means that a BIOS LBA
 824                    translation was active, so the default
 825                    hardware geometry is OK */
 826                 lba_detected = 1;
 827                 goto default_geometry;
 828             } else {
 829                 *pcyls = cylinders;
 830                 *pheads = heads;
 831                 *psecs = secs;
 832                 /* disable any translation to be in sync with
 833                    the logical geometry */
 834                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 835                     bdrv_set_translation_hint(bs,
 836                                               BIOS_ATA_TRANSLATION_NONE);
 837                 }
 838             }
 839         } else {
 840         default_geometry:
 841             /* if no geometry, use a standard physical disk geometry */
 842             cylinders = nb_sectors / (16 * 63);
 843
 844             if (cylinders > 16383)
 845                 cylinders = 16383;
 846             else if (cylinders < 2)
 847                 cylinders = 2;
 848             *pcyls = cylinders;
 849             *pheads = 16;
 850             *psecs = 63;
 851             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 852                 if ((*pcyls * *pheads) <= 131072) {
 853                     bdrv_set_translation_hint(bs,
 854                                               BIOS_ATA_TRANSLATION_LARGE);
 855                 } else {
 856                     bdrv_set_translation_hint(bs,
 857                                               BIOS_ATA_TRANSLATION_LBA);
 858                 }
 859             }
 860         }
 861         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 862     }
 863 }
 864
 865 void bdrv_set_geometry_hint(BlockDriverState *bs,
 866                             int cyls, int heads, int secs)
 867 {
 868     bs->cyls = cyls;
 869     bs->heads = heads;
 870     bs->secs = secs;
 871 }
 872
 873 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 874 {
 875     bs->type = type;
 876     bs->removable = ((type == BDRV_TYPE_CDROM ||
 877                       type == BDRV_TYPE_FLOPPY));
 878 }
 879
 880 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 881 {
 882     bs->translation = translation;
 883 }
 884
 885 void bdrv_get_geometry_hint(BlockDriverState *bs,
 886                             int *pcyls, int *pheads, int *psecs)
 887 {
 888     *pcyls = bs->cyls;
 889     *pheads = bs->heads;
 890     *psecs = bs->secs;
 891 }
 892
 893 int bdrv_get_type_hint(BlockDriverState *bs)
 894 {
 895     return bs->type;
 896 }
 897
 898 int bdrv_get_translation_hint(BlockDriverState *bs)
 899 {
 900     return bs->translation;
 901 }
 902
 903 int bdrv_is_removable(BlockDriverState *bs)
 904 {
 905     return bs->removable;
 906 }
 907
 908 int bdrv_is_read_only(BlockDriverState *bs)
 909 {
 910     return bs->read_only;
 911 }
 912
 913 int bdrv_is_sg(BlockDriverState *bs)
 914 {
 915     return bs->sg;
 916 }
 917
 918 /* XXX: no longer used */
 919 void bdrv_set_change_cb(BlockDriverState *bs,
 920                         void (*change_cb)(void *opaque), void *opaque)
 921 {
 922     bs->change_cb = change_cb;
 923     bs->change_opaque = opaque;
 924 }
 925
 926 int bdrv_is_encrypted(BlockDriverState *bs)
 927 {
 928     if (bs->backing_hd && bs->backing_hd->encrypted)
 929         return 1;
 930     return bs->encrypted;
 931 }
 932
 933 int bdrv_key_required(BlockDriverState *bs)
 934 {
 935     BlockDriverState *backing_hd = bs->backing_hd;
 936
 937     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 938         return 1;
 939     return (bs->encrypted && !bs->valid_key);
 940 }
 941
 942 int bdrv_set_key(BlockDriverState *bs, const char *key)
 943 {
 944     int ret;
 945     if (bs->backing_hd && bs->backing_hd->encrypted) {
 946         ret = bdrv_set_key(bs->backing_hd, key);
 947         if (ret < 0)
 948             return ret;
 949         if (!bs->encrypted)
 950             return 0;
 951     }
 952     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 953         return -1;
 954     ret = bs->drv->bdrv_set_key(bs, key);
 955     if (ret < 0) {
 956         bs->valid_key = 0;
 957     } else if (!bs->valid_key) {
 958         bs->valid_key = 1;
 959         /* call the change callback now, we skipped it on open */
 960         bs->media_changed = 1;
 961         if (bs->change_cb)
 962             bs->change_cb(bs->change_opaque);
 963     }
 964     return ret;
 965 }
 966
 967 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 968 {
 969     if (!bs->drv) {
 970         buf[0] = '\0';
 971     } else {
 972         pstrcpy(buf, buf_size, bs->drv->format_name);
 973     }
 974 }
 975
 976 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 977                          void *opaque)
 978 {
 979     BlockDriver *drv;
 980
 981     for (drv = first_drv; drv != NULL; drv = drv->next) {
 982         it(opaque, drv->format_name);
 983     }
 984 }
 985
 986 BlockDriverState *bdrv_find(const char *name)
 987 {
 988     BlockDriverState *bs;
 989
 990     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 991         if (!strcmp(name, bs->device_name))
 992             return bs;
 993     }
 994     return NULL;
 995 }
 996
 997 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
 998 {
 999     BlockDriverState *bs;
1000
1001     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1002         it(opaque, bs);
1003     }
1004 }
1005
1006 const char *bdrv_get_device_name(BlockDriverState *bs)
1007 {
1008     return bs->device_name;
1009 }
1010
1011 void bdrv_flush(BlockDriverState *bs)
1012 {
1013     if (bs->drv->bdrv_flush)
1014         bs->drv->bdrv_flush(bs);
1015     if (bs->backing_hd)
1016         bdrv_flush(bs->backing_hd);
1017 }
1018
1019 void bdrv_flush_all(void)
1020 {
1021     BlockDriverState *bs;
1022
1023     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1024         if (bs->drv && !bdrv_is_read_only(bs) &&
1025             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1026             bdrv_flush(bs);
1027 }
1028
1029 /*
1030  * Returns true iff the specified sector is present in the disk image. Drivers
1031  * not implementing the functionality are assumed to not support backing files,
1032  * hence all their sectors are reported as allocated.
1033  *
1034  * 'pnum' is set to the number of sectors (including and immediately following
1035  * the specified sector) that are known to be in the same
1036  * allocated/unallocated state.
1037  *
1038  * 'nb_sectors' is the max value 'pnum' should be set to.
1039  */
1040 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1041         int *pnum)
1042 {
1043     int64_t n;
1044     if (!bs->drv->bdrv_is_allocated) {
1045         if (sector_num >= bs->total_sectors) {
1046             *pnum = 0;
1047             return 0;
1048         }
1049         n = bs->total_sectors - sector_num;
1050         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1051         return 1;
1052     }
1053     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1054 }
1055
1056 void bdrv_info(Monitor *mon)
1057 {
1058     BlockDriverState *bs;
1059
1060     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1061         monitor_printf(mon, "%s:", bs->device_name);
1062         monitor_printf(mon, " type=");
1063         switch(bs->type) {
1064         case BDRV_TYPE_HD:
1065             monitor_printf(mon, "hd");
1066             break;
1067         case BDRV_TYPE_CDROM:
1068             monitor_printf(mon, "cdrom");
1069             break;
1070         case BDRV_TYPE_FLOPPY:
1071             monitor_printf(mon, "floppy");
1072             break;
1073         }
1074         monitor_printf(mon, " removable=%d", bs->removable);
1075         if (bs->removable) {
1076             monitor_printf(mon, " locked=%d", bs->locked);
1077         }
1078         if (bs->drv) {
1079             monitor_printf(mon, " file=");
1080             monitor_print_filename(mon, bs->filename);
1081             if (bs->backing_file[0] != '\0') {
1082                 monitor_printf(mon, " backing_file=");
1083                 monitor_print_filename(mon, bs->backing_file);
1084             }
1085             monitor_printf(mon, " ro=%d", bs->read_only);
1086             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1087             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1088         } else {
1089             monitor_printf(mon, " [not inserted]");
1090         }
1091         monitor_printf(mon, "\n");
1092     }
1093 }
1094
1095 /* The "info blockstats" command. */
1096 void bdrv_info_stats(Monitor *mon)
1097 {
1098     BlockDriverState *bs;
1099
1100     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1101         monitor_printf(mon, "%s:"
1102                        " rd_bytes=%" PRIu64
1103                        " wr_bytes=%" PRIu64
1104                        " rd_operations=%" PRIu64
1105                        " wr_operations=%" PRIu64
1106                        "\n",
1107                        bs->device_name,
1108                        bs->rd_bytes, bs->wr_bytes,
1109                        bs->rd_ops, bs->wr_ops);
1110     }
1111 }
1112
1113 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1114 {
1115     if (bs->backing_hd && bs->backing_hd->encrypted)
1116         return bs->backing_file;
1117     else if (bs->encrypted)
1118         return bs->filename;
1119     else
1120         return NULL;
1121 }
1122
1123 void bdrv_get_backing_filename(BlockDriverState *bs,
1124                                char *filename, int filename_size)
1125 {
1126     if (!bs->backing_hd) {
1127         pstrcpy(filename, filename_size, "");
1128     } else {
1129         pstrcpy(filename, filename_size, bs->backing_file);
1130     }
1131 }
1132
1133 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1134                           const uint8_t *buf, int nb_sectors)
1135 {
1136     BlockDriver *drv = bs->drv;
1137     if (!drv)
1138         return -ENOMEDIUM;
1139     if (!drv->bdrv_write_compressed)
1140         return -ENOTSUP;
1141     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1142 }
1143
1144 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1145 {
1146     BlockDriver *drv = bs->drv;
1147     if (!drv)
1148         return -ENOMEDIUM;
1149     if (!drv->bdrv_get_info)
1150         return -ENOTSUP;
1151     memset(bdi, 0, sizeof(*bdi));
1152     return drv->bdrv_get_info(bs, bdi);
1153 }
1154
1155 /**************************************************************/
1156 /* handling of snapshots */
1157
1158 int bdrv_snapshot_create(BlockDriverState *bs,
1159                          QEMUSnapshotInfo *sn_info)
1160 {
1161     BlockDriver *drv = bs->drv;
1162     if (!drv)
1163         return -ENOMEDIUM;
1164     if (!drv->bdrv_snapshot_create)
1165         return -ENOTSUP;
1166     return drv->bdrv_snapshot_create(bs, sn_info);
1167 }
1168
1169 int bdrv_snapshot_goto(BlockDriverState *bs,
1170                        const char *snapshot_id)
1171 {
1172     BlockDriver *drv = bs->drv;
1173     if (!drv)
1174         return -ENOMEDIUM;
1175     if (!drv->bdrv_snapshot_goto)
1176         return -ENOTSUP;
1177     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1178 }
1179
1180 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1181 {
1182     BlockDriver *drv = bs->drv;
1183     if (!drv)
1184         return -ENOMEDIUM;
1185     if (!drv->bdrv_snapshot_delete)
1186         return -ENOTSUP;
1187     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1188 }
1189
1190 int bdrv_snapshot_list(BlockDriverState *bs,
1191                        QEMUSnapshotInfo **psn_info)
1192 {
1193     BlockDriver *drv = bs->drv;
1194     if (!drv)
1195         return -ENOMEDIUM;
1196     if (!drv->bdrv_snapshot_list)
1197         return -ENOTSUP;
1198     return drv->bdrv_snapshot_list(bs, psn_info);
1199 }
1200
1201 #define NB_SUFFIXES 4
1202
1203 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1204 {
1205     static const char suffixes[NB_SUFFIXES] = "KMGT";
1206     int64_t base;
1207     int i;
1208
1209     if (size <= 999) {
1210         snprintf(buf, buf_size, "%" PRId64, size);
1211     } else {
1212         base = 1024;
1213         for(i = 0; i < NB_SUFFIXES; i++) {
1214             if (size < (10 * base)) {
1215                 snprintf(buf, buf_size, "%0.1f%c",
1216                          (double)size / base,
1217                          suffixes[i]);
1218                 break;
1219             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1220                 snprintf(buf, buf_size, "%" PRId64 "%c",
1221                          ((size + (base >> 1)) / base),
1222                          suffixes[i]);
1223                 break;
1224             }
1225             base = base * 1024;
1226         }
1227     }
1228     return buf;
1229 }
1230
1231 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1232 {
1233     char buf1[128], date_buf[128], clock_buf[128];
1234 #ifdef _WIN32
1235     struct tm *ptm;
1236 #else
1237     struct tm tm;
1238 #endif
1239     time_t ti;
1240     int64_t secs;
1241
1242     if (!sn) {
1243         snprintf(buf, buf_size,
1244                  "%-10s%-20s%7s%20s%15s",
1245                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1246     } else {
1247         ti = sn->date_sec;
1248 #ifdef _WIN32
1249         ptm = localtime(&ti);
1250         strftime(date_buf, sizeof(date_buf),
1251                  "%Y-%m-%d %H:%M:%S", ptm);
1252 #else
1253         localtime_r(&ti, &tm);
1254         strftime(date_buf, sizeof(date_buf),
1255                  "%Y-%m-%d %H:%M:%S", &tm);
1256 #endif
1257         secs = sn->vm_clock_nsec / 1000000000;
1258         snprintf(clock_buf, sizeof(clock_buf),
1259                  "%02d:%02d:%02d.%03d",
1260                  (int)(secs / 3600),
1261                  (int)((secs / 60) % 60),
1262                  (int)(secs % 60),
1263                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1264         snprintf(buf, buf_size,
1265                  "%-10s%-20s%7s%20s%15s",
1266                  sn->id_str, sn->name,
1267                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1268                  date_buf,
1269                  clock_buf);
1270     }
1271     return buf;
1272 }
1273
1274
1275 /**************************************************************/
1276 /* async I/Os */
1277
1278 typedef struct VectorTranslationAIOCB {
1279     BlockDriverAIOCB common;
1280     QEMUIOVector *iov;
1281     uint8_t *bounce;
1282     int is_write;
1283     BlockDriverAIOCB *aiocb;
1284 } VectorTranslationAIOCB;
1285
1286 static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
1287 {
1288     VectorTranslationAIOCB *acb
1289         = container_of(_acb, VectorTranslationAIOCB, common);
1290
1291     bdrv_aio_cancel(acb->aiocb);
1292 }
1293
1294 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1295 {
1296     VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
1297
1298     if (!s->is_write) {
1299         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1300     }
1301     qemu_vfree(s->bounce);
1302     s->common.cb(s->common.opaque, ret);
1303     qemu_aio_release(s);
1304 }
1305
1306 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1307                                             int64_t sector_num,
1308                                             QEMUIOVector *iov,
1309                                             int nb_sectors,
1310                                             BlockDriverCompletionFunc *cb,
1311                                             void *opaque,
1312                                             int is_write)
1313
1314 {
1315     VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
1316                                                   cb, opaque);
1317
1318     s->iov = iov;
1319     s->bounce = qemu_memalign(512, nb_sectors * 512);
1320     s->is_write = is_write;
1321     if (is_write) {
1322         qemu_iovec_to_buffer(s->iov, s->bounce);
1323         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1324                                   bdrv_aio_rw_vector_cb, s);
1325     } else {
1326         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1327                                  bdrv_aio_rw_vector_cb, s);
1328     }
1329     if (!s->aiocb) {
1330         qemu_vfree(s->bounce);
1331         qemu_aio_release(s);
1332         return NULL;
1333     }
1334     return &s->common;
1335 }
1336
1337 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1338                                  QEMUIOVector *iov, int nb_sectors,
1339                                  BlockDriverCompletionFunc *cb, void *opaque)
1340 {
1341     if (bdrv_check_request(bs, sector_num, nb_sectors))
1342         return NULL;
1343
1344     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1345                               cb, opaque, 0);
1346 }
1347
1348 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1349                                   QEMUIOVector *iov, int nb_sectors,
1350                                   BlockDriverCompletionFunc *cb, void *opaque)
1351 {
1352     if (bdrv_check_request(bs, sector_num, nb_sectors))
1353         return NULL;
1354
1355     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1356                               cb, opaque, 1);
1357 }
1358
1359 BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1360                                 uint8_t *buf, int nb_sectors,
1361                                 BlockDriverCompletionFunc *cb, void *opaque)
1362 {
1363     BlockDriver *drv = bs->drv;
1364     BlockDriverAIOCB *ret;
1365
1366     if (!drv)
1367         return NULL;
1368     if (bdrv_check_request(bs, sector_num, nb_sectors))
1369         return NULL;
1370
1371     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1372
1373     if (ret) {
1374         /* Update stats even though technically transfer has not happened. */
1375         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1376         bs->rd_ops ++;
1377     }
1378
1379     return ret;
1380 }
1381
1382 BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1383                                  const uint8_t *buf, int nb_sectors,
1384                                  BlockDriverCompletionFunc *cb, void *opaque)
1385 {
1386     BlockDriver *drv = bs->drv;
1387     BlockDriverAIOCB *ret;
1388
1389     if (!drv)
1390         return NULL;
1391     if (bs->read_only)
1392         return NULL;
1393     if (bdrv_check_request(bs, sector_num, nb_sectors))
1394         return NULL;
1395
1396     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1397
1398     if (ret) {
1399         /* Update stats even though technically transfer has not happened. */
1400         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1401         bs->wr_ops ++;
1402     }
1403
1404     return ret;
1405 }
1406
1407 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1408 {
1409     acb->pool->cancel(acb);
1410 }
1411
1412
1413 /**************************************************************/
1414 /* async block device emulation */
1415
1416 static void bdrv_aio_bh_cb(void *opaque)
1417 {
1418     BlockDriverAIOCBSync *acb = opaque;
1419     acb->common.cb(acb->common.opaque, acb->ret);
1420     qemu_aio_release(acb);
1421 }
1422
1423 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1424         int64_t sector_num, uint8_t *buf, int nb_sectors,
1425         BlockDriverCompletionFunc *cb, void *opaque)
1426 {
1427     BlockDriverAIOCBSync *acb;
1428     int ret;
1429
1430     acb = qemu_aio_get(bs, cb, opaque);
1431     if (!acb->bh)
1432         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1433     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1434     acb->ret = ret;
1435     qemu_bh_schedule(acb->bh);
1436     return &acb->common;
1437 }
1438
1439 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1440         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1441         BlockDriverCompletionFunc *cb, void *opaque)
1442 {
1443     BlockDriverAIOCBSync *acb;
1444     int ret;
1445
1446     acb = qemu_aio_get(bs, cb, opaque);
1447     if (!acb->bh)
1448         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1449     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1450     acb->ret = ret;
1451     qemu_bh_schedule(acb->bh);
1452     return &acb->common;
1453 }
1454
1455 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1456 {
1457     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1458     qemu_bh_cancel(acb->bh);
1459     qemu_aio_release(acb);
1460 }
1461
1462 /**************************************************************/
1463 /* sync block device emulation */
1464
1465 static void bdrv_rw_em_cb(void *opaque, int ret)
1466 {
1467     *(int *)opaque = ret;
1468 }
1469
1470 #define NOT_DONE 0x7fffffff
1471
1472 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1473                         uint8_t *buf, int nb_sectors)
1474 {
1475     int async_ret;
1476     BlockDriverAIOCB *acb;
1477
1478     async_ret = NOT_DONE;
1479     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1480                         bdrv_rw_em_cb, &async_ret);
1481     if (acb == NULL)
1482         return -1;
1483
1484     while (async_ret == NOT_DONE) {
1485         qemu_aio_wait();
1486     }
1487
1488     return async_ret;
1489 }
1490
1491 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1492                          const uint8_t *buf, int nb_sectors)
1493 {
1494     int async_ret;
1495     BlockDriverAIOCB *acb;
1496
1497     async_ret = NOT_DONE;
1498     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1499                          bdrv_rw_em_cb, &async_ret);
1500     if (acb == NULL)
1501         return -1;
1502     while (async_ret == NOT_DONE) {
1503         qemu_aio_wait();
1504     }
1505     return async_ret;
1506 }
1507
1508 void bdrv_init(void)
1509 {
1510     aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
1511                   bdrv_aio_cancel_vector);
1512
1513     bdrv_register(&bdrv_raw);
1514     bdrv_register(&bdrv_host_device);
1515 #ifndef _WIN32
1516     bdrv_register(&bdrv_cow);
1517 #endif
1518     bdrv_register(&bdrv_qcow);
1519     bdrv_register(&bdrv_vmdk);
1520     bdrv_register(&bdrv_cloop);
1521     bdrv_register(&bdrv_dmg);
1522     bdrv_register(&bdrv_bochs);
1523     bdrv_register(&bdrv_vpc);
1524     bdrv_register(&bdrv_vvfat);
1525     bdrv_register(&bdrv_qcow2);
1526     bdrv_register(&bdrv_parallels);
1527     bdrv_register(&bdrv_nbd);
1528 }
1529
1530 void aio_pool_init(AIOPool *pool, int aiocb_size,
1531                    void (*cancel)(BlockDriverAIOCB *acb))
1532 {
1533     pool->aiocb_size = aiocb_size;
1534     pool->cancel = cancel;
1535     pool->free_aiocb = NULL;
1536 }
1537
1538 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1539                         BlockDriverCompletionFunc *cb, void *opaque)
1540 {
1541     BlockDriverAIOCB *acb;
1542
1543     if (pool->free_aiocb) {
1544         acb = pool->free_aiocb;
1545         pool->free_aiocb = acb->next;
1546     } else {
1547         acb = qemu_mallocz(pool->aiocb_size);
1548         acb->pool = pool;
1549     }
1550     acb->bs = bs;
1551     acb->cb = cb;
1552     acb->opaque = opaque;
1553     return acb;
1554 }
1555
1556 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1557                    void *opaque)
1558 {
1559     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1560 }
1561
1562 void qemu_aio_release(void *p)
1563 {
1564     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1565     AIOPool *pool = acb->pool;
1566     acb->next = pool->free_aiocb;
1567     pool->free_aiocb = acb;
1568 }
1569
1570 /**************************************************************/
1571 /* removable device support */
1572
1573 /**
1574  * Return TRUE if the media is present
1575  */
1576 int bdrv_is_inserted(BlockDriverState *bs)
1577 {
1578     BlockDriver *drv = bs->drv;
1579     int ret;
1580     if (!drv)
1581         return 0;
1582     if (!drv->bdrv_is_inserted)
1583         return 1;
1584     ret = drv->bdrv_is_inserted(bs);
1585     return ret;
1586 }
1587
1588 /**
1589  * Return TRUE if the media changed since the last call to this
1590  * function. It is currently only used for floppy disks
1591  */
1592 int bdrv_media_changed(BlockDriverState *bs)
1593 {
1594     BlockDriver *drv = bs->drv;
1595     int ret;
1596
1597     if (!drv || !drv->bdrv_media_changed)
1598         ret = -ENOTSUP;
1599     else
1600         ret = drv->bdrv_media_changed(bs);
1601     if (ret == -ENOTSUP)
1602         ret = bs->media_changed;
1603     bs->media_changed = 0;
1604     return ret;
1605 }
1606
1607 /**
1608  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1609  */
1610 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1611 {
1612     BlockDriver *drv = bs->drv;
1613     int ret;
1614
1615     if (!drv || !drv->bdrv_eject) {
1616         ret = -ENOTSUP;
1617     } else {
1618         ret = drv->bdrv_eject(bs, eject_flag);
1619     }
1620     if (ret == -ENOTSUP) {
1621         if (eject_flag)
1622             bdrv_close(bs);
1623     }
1624 }
1625
1626 int bdrv_is_locked(BlockDriverState *bs)
1627 {
1628     return bs->locked;
1629 }
1630
1631 /**
1632  * Lock or unlock the media (if it is locked, the user won't be able
1633  * to eject it manually).
1634  */
1635 void bdrv_set_locked(BlockDriverState *bs, int locked)
1636 {
1637     BlockDriver *drv = bs->drv;
1638
1639     bs->locked = locked;
1640     if (drv && drv->bdrv_set_locked) {
1641         drv->bdrv_set_locked(bs, locked);
1642     }
1643 }
1644
1645 /* needed for generic scsi interface */
1646
1647 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1648 {
1649     BlockDriver *drv = bs->drv;
1650
1651     if (drv && drv->bdrv_ioctl)
1652         return drv->bdrv_ioctl(bs, req, buf);
1653     return -ENOTSUP;
1654 }
1655
1656 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1657         unsigned long int req, void *buf,
1658         BlockDriverCompletionFunc *cb, void *opaque)
1659 {
1660     BlockDriver *drv = bs->drv;
1661
1662     if (drv && drv->bdrv_aio_ioctl)
1663         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1664     return NULL;
1665 }