block.c

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