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