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 "module.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 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
  52         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  53         BlockDriverCompletionFunc *cb, void *opaque);
  54 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
  55         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  56         BlockDriverCompletionFunc *cb, void *opaque);
  57 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  58                         uint8_t *buf, int nb_sectors);
  59 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  60                          const uint8_t *buf, int nb_sectors);
  61
  62 BlockDriverState *bdrv_first;
  63
  64 static BlockDriver *first_drv;
  65
  66 int path_is_absolute(const char *path)
  67 {
  68     const char *p;
  69 #ifdef _WIN32
  70     /* specific case for names like: "\\.\d:" */
  71     if (*path == '/' || *path == '\\')
  72         return 1;
  73 #endif
  74     p = strchr(path, ':');
  75     if (p)
  76         p++;
  77     else
  78         p = path;
  79 #ifdef _WIN32
  80     return (*p == '/' || *p == '\\');
  81 #else
  82     return (*p == '/');
  83 #endif
  84 }
  85
  86 /* if filename is absolute, just copy it to dest. Otherwise, build a
  87    path to it by considering it is relative to base_path. URL are
  88    supported. */
  89 void path_combine(char *dest, int dest_size,
  90                   const char *base_path,
  91                   const char *filename)
  92 {
  93     const char *p, *p1;
  94     int len;
  95
  96     if (dest_size <= 0)
  97         return;
  98     if (path_is_absolute(filename)) {
  99         pstrcpy(dest, dest_size, filename);
 100     } else {
 101         p = strchr(base_path, ':');
 102         if (p)
 103             p++;
 104         else
 105             p = base_path;
 106         p1 = strrchr(base_path, '/');
 107 #ifdef _WIN32
 108         {
 109             const char *p2;
 110             p2 = strrchr(base_path, '\\');
 111             if (!p1 || p2 > p1)
 112                 p1 = p2;
 113         }
 114 #endif
 115         if (p1)
 116             p1++;
 117         else
 118             p1 = base_path;
 119         if (p1 > p)
 120             p = p1;
 121         len = p - base_path;
 122         if (len > dest_size - 1)
 123             len = dest_size - 1;
 124         memcpy(dest, base_path, len);
 125         dest[len] = '\0';
 126         pstrcat(dest, dest_size, filename);
 127     }
 128 }
 129
 130 void bdrv_register(BlockDriver *bdrv)
 131 {
 132     if (!bdrv->bdrv_aio_readv) {
 133         /* add AIO emulation layer */
 134         bdrv->bdrv_aio_readv = bdrv_aio_readv_em;
 135         bdrv->bdrv_aio_writev = bdrv_aio_writev_em;
 136     } else if (!bdrv->bdrv_read) {
 137         /* add synchronous IO emulation layer */
 138         bdrv->bdrv_read = bdrv_read_em;
 139         bdrv->bdrv_write = bdrv_write_em;
 140     }
 141     bdrv->next = first_drv;
 142     first_drv = bdrv;
 143 }
 144
 145 /* create a new block device (by default it is empty) */
 146 BlockDriverState *bdrv_new(const char *device_name)
 147 {
 148     BlockDriverState **pbs, *bs;
 149
 150     bs = qemu_mallocz(sizeof(BlockDriverState));
 151     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 152     if (device_name[0] != '\0') {
 153         /* insert at the end */
 154         pbs = &bdrv_first;
 155         while (*pbs != NULL)
 156             pbs = &(*pbs)->next;
 157         *pbs = bs;
 158     }
 159     return bs;
 160 }
 161
 162 BlockDriver *bdrv_find_format(const char *format_name)
 163 {
 164     BlockDriver *drv1;
 165     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 166         if (!strcmp(drv1->format_name, format_name))
 167             return drv1;
 168     }
 169     return NULL;
 170 }
 171
 172 int bdrv_create(BlockDriver *drv, const char* filename,
 173     QEMUOptionParameter *options)
 174 {
 175     if (!drv->bdrv_create)
 176         return -ENOTSUP;
 177
 178     return drv->bdrv_create(filename, options);
 179 }
 180
 181 #ifdef _WIN32
 182 void get_tmp_filename(char *filename, int size)
 183 {
 184     char temp_dir[MAX_PATH];
 185
 186     GetTempPath(MAX_PATH, temp_dir);
 187     GetTempFileName(temp_dir, "qem", 0, filename);
 188 }
 189 #else
 190 void get_tmp_filename(char *filename, int size)
 191 {
 192     int fd;
 193     const char *tmpdir;
 194     /* XXX: race condition possible */
 195     tmpdir = getenv("TMPDIR");
 196     if (!tmpdir)
 197         tmpdir = "/tmp";
 198     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 199     fd = mkstemp(filename);
 200     close(fd);
 201 }
 202 #endif
 203
 204 #ifdef _WIN32
 205 static int is_windows_drive_prefix(const char *filename)
 206 {
 207     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 208              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 209             filename[1] == ':');
 210 }
 211
 212 int is_windows_drive(const char *filename)
 213 {
 214     if (is_windows_drive_prefix(filename) &&
 215         filename[2] == '\0')
 216         return 1;
 217     if (strstart(filename, "\\\\.\\", NULL) ||
 218         strstart(filename, "//./", NULL))
 219         return 1;
 220     return 0;
 221 }
 222 #endif
 223
 224 static BlockDriver *find_protocol(const char *filename)
 225 {
 226     BlockDriver *drv1;
 227     char protocol[128];
 228     int len;
 229     const char *p;
 230
 231 #ifdef _WIN32
 232     if (is_windows_drive(filename) ||
 233         is_windows_drive_prefix(filename))
 234         return bdrv_find_format("raw");
 235 #endif
 236     p = strchr(filename, ':');
 237     if (!p)
 238         return bdrv_find_format("raw");
 239     len = p - filename;
 240     if (len > sizeof(protocol) - 1)
 241         len = sizeof(protocol) - 1;
 242     memcpy(protocol, filename, len);
 243     protocol[len] = '\0';
 244     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 245         if (drv1->protocol_name &&
 246             !strcmp(drv1->protocol_name, protocol))
 247             return drv1;
 248     }
 249     return NULL;
 250 }
 251
 252 /*
 253  * Detect host devices. By convention, /dev/cdrom[N] is always
 254  * recognized as a host CDROM.
 255  */
 256 static BlockDriver *find_hdev_driver(const char *filename)
 257 {
 258     int score_max = 0, score;
 259     BlockDriver *drv = NULL, *d;
 260
 261     for (d = first_drv; d; d = d->next) {
 262         if (d->bdrv_probe_device) {
 263             score = d->bdrv_probe_device(filename);
 264             if (score > score_max) {
 265                 score_max = score;
 266                 drv = d;
 267             }
 268         }
 269     }
 270
 271     return drv;
 272 }
 273
 274 static BlockDriver *find_image_format(const char *filename)
 275 {
 276     int ret, score, score_max;
 277     BlockDriver *drv1, *drv;
 278     uint8_t buf[2048];
 279     BlockDriverState *bs;
 280
 281     drv = find_protocol(filename);
 282     /* no need to test disk image formats for vvfat */
 283     if (drv && strcmp(drv->format_name, "vvfat") == 0)
 284         return drv;
 285
 286     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 287     if (ret < 0)
 288         return NULL;
 289     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 290     bdrv_delete(bs);
 291     if (ret < 0) {
 292         return NULL;
 293     }
 294
 295     score_max = 0;
 296     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 297         if (drv1->bdrv_probe) {
 298             score = drv1->bdrv_probe(buf, ret, filename);
 299             if (score > score_max) {
 300                 score_max = score;
 301                 drv = drv1;
 302             }
 303         }
 304     }
 305     return drv;
 306 }
 307
 308 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 309 {
 310     BlockDriverState *bs;
 311     int ret;
 312
 313     bs = bdrv_new("");
 314     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 315     if (ret < 0) {
 316         bdrv_delete(bs);
 317         return ret;
 318     }
 319     bs->growable = 1;
 320     *pbs = bs;
 321     return 0;
 322 }
 323
 324 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 325 {
 326     return bdrv_open2(bs, filename, flags, NULL);
 327 }
 328
 329 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 330                BlockDriver *drv)
 331 {
 332     int ret, open_flags;
 333     char tmp_filename[PATH_MAX];
 334     char backing_filename[PATH_MAX];
 335
 336     bs->read_only = 0;
 337     bs->is_temporary = 0;
 338     bs->encrypted = 0;
 339     bs->valid_key = 0;
 340     /* buffer_alignment defaulted to 512, drivers can change this value */
 341     bs->buffer_alignment = 512;
 342
 343     if (flags & BDRV_O_SNAPSHOT) {
 344         BlockDriverState *bs1;
 345         int64_t total_size;
 346         int is_protocol = 0;
 347         BlockDriver *bdrv_qcow2;
 348         QEMUOptionParameter *options;
 349
 350         /* if snapshot, we create a temporary backing file and open it
 351            instead of opening 'filename' directly */
 352
 353         /* if there is a backing file, use it */
 354         bs1 = bdrv_new("");
 355         ret = bdrv_open2(bs1, filename, 0, drv);
 356         if (ret < 0) {
 357             bdrv_delete(bs1);
 358             return ret;
 359         }
 360         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 361
 362         if (bs1->drv && bs1->drv->protocol_name)
 363             is_protocol = 1;
 364
 365         bdrv_delete(bs1);
 366
 367         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 368
 369         /* Real path is meaningless for protocols */
 370         if (is_protocol)
 371             snprintf(backing_filename, sizeof(backing_filename),
 372                      "%s", filename);
 373         else
 374             realpath(filename, backing_filename);
 375
 376         bdrv_qcow2 = bdrv_find_format("qcow2");
 377         options = parse_option_parameters("", bdrv_qcow2->create_options, NULL);
 378
 379         set_option_parameter_int(options, BLOCK_OPT_SIZE, total_size * 512);
 380         set_option_parameter(options, BLOCK_OPT_BACKING_FILE, backing_filename);
 381         if (drv) {
 382             set_option_parameter(options, BLOCK_OPT_BACKING_FMT,
 383                 drv->format_name);
 384         }
 385
 386         ret = bdrv_create(bdrv_qcow2, tmp_filename, options);
 387         if (ret < 0) {
 388             return ret;
 389         }
 390
 391         filename = tmp_filename;
 392         drv = bdrv_qcow2;
 393         bs->is_temporary = 1;
 394     }
 395
 396     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 397     if (flags & BDRV_O_FILE) {
 398         drv = find_protocol(filename);
 399     } else if (!drv) {
 400         drv = find_hdev_driver(filename);
 401         if (!drv) {
 402             drv = find_image_format(filename);
 403         }
 404     }
 405     if (!drv) {
 406         ret = -ENOENT;
 407         goto unlink_and_fail;
 408     }
 409     bs->drv = drv;
 410     bs->opaque = qemu_mallocz(drv->instance_size);
 411     /* Note: for compatibility, we open disk image files as RDWR, and
 412        RDONLY as fallback */
 413     if (!(flags & BDRV_O_FILE))
 414         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 415     else
 416         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 417     ret = drv->bdrv_open(bs, filename, open_flags);
 418     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 419         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 420         bs->read_only = 1;
 421     }
 422     if (ret < 0) {
 423         qemu_free(bs->opaque);
 424         bs->opaque = NULL;
 425         bs->drv = NULL;
 426     unlink_and_fail:
 427         if (bs->is_temporary)
 428             unlink(filename);
 429         return ret;
 430     }
 431     if (drv->bdrv_getlength) {
 432         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 433     }
 434 #ifndef _WIN32
 435     if (bs->is_temporary) {
 436         unlink(filename);
 437     }
 438 #endif
 439     if (bs->backing_file[0] != '\0') {
 440         /* if there is a backing file, use it */
 441         BlockDriver *back_drv = NULL;
 442         bs->backing_hd = bdrv_new("");
 443         path_combine(backing_filename, sizeof(backing_filename),
 444                      filename, bs->backing_file);
 445         if (bs->backing_format[0] != '\0')
 446             back_drv = bdrv_find_format(bs->backing_format);
 447         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 448                          back_drv);
 449         if (ret < 0) {
 450             bdrv_close(bs);
 451             return ret;
 452         }
 453     }
 454
 455     if (!bdrv_key_required(bs)) {
 456         /* call the change callback */
 457         bs->media_changed = 1;
 458         if (bs->change_cb)
 459             bs->change_cb(bs->change_opaque);
 460     }
 461     return 0;
 462 }
 463
 464 void bdrv_close(BlockDriverState *bs)
 465 {
 466     if (bs->drv) {
 467         if (bs->backing_hd)
 468             bdrv_delete(bs->backing_hd);
 469         bs->drv->bdrv_close(bs);
 470         qemu_free(bs->opaque);
 471 #ifdef _WIN32
 472         if (bs->is_temporary) {
 473             unlink(bs->filename);
 474         }
 475 #endif
 476         bs->opaque = NULL;
 477         bs->drv = NULL;
 478
 479         /* call the change callback */
 480         bs->media_changed = 1;
 481         if (bs->change_cb)
 482             bs->change_cb(bs->change_opaque);
 483     }
 484 }
 485
 486 void bdrv_delete(BlockDriverState *bs)
 487 {
 488     BlockDriverState **pbs;
 489
 490     pbs = &bdrv_first;
 491     while (*pbs != bs && *pbs != NULL)
 492         pbs = &(*pbs)->next;
 493     if (*pbs == bs)
 494         *pbs = bs->next;
 495
 496     bdrv_close(bs);
 497     qemu_free(bs);
 498 }
 499
 500 /*
 501  * Run consistency checks on an image
 502  *
 503  * Returns the number of errors or -errno when an internal error occurs
 504  */
 505 int bdrv_check(BlockDriverState *bs)
 506 {
 507     if (bs->drv->bdrv_check == NULL) {
 508         return -ENOTSUP;
 509     }
 510
 511     return bs->drv->bdrv_check(bs);
 512 }
 513
 514 /* commit COW file into the raw image */
 515 int bdrv_commit(BlockDriverState *bs)
 516 {
 517     BlockDriver *drv = bs->drv;
 518     int64_t i, total_sectors;
 519     int n, j;
 520     unsigned char sector[512];
 521
 522     if (!drv)
 523         return -ENOMEDIUM;
 524
 525     if (bs->read_only) {
 526         return -EACCES;
 527     }
 528
 529     if (!bs->backing_hd) {
 530         return -ENOTSUP;
 531     }
 532
 533     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 534     for (i = 0; i < total_sectors;) {
 535         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 536             for(j = 0; j < n; j++) {
 537                 if (bdrv_read(bs, i, sector, 1) != 0) {
 538                     return -EIO;
 539                 }
 540
 541                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 542                     return -EIO;
 543                 }
 544                 i++;
 545             }
 546         } else {
 547             i += n;
 548         }
 549     }
 550
 551     if (drv->bdrv_make_empty)
 552         return drv->bdrv_make_empty(bs);
 553
 554     return 0;
 555 }
 556
 557 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 558                                    size_t size)
 559 {
 560     int64_t len;
 561
 562     if (!bdrv_is_inserted(bs))
 563         return -ENOMEDIUM;
 564
 565     if (bs->growable)
 566         return 0;
 567
 568     len = bdrv_getlength(bs);
 569
 570     if (offset < 0)
 571         return -EIO;
 572
 573     if ((offset > len) || (len - offset < size))
 574         return -EIO;
 575
 576     return 0;
 577 }
 578
 579 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 580                               int nb_sectors)
 581 {
 582     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 583 }
 584
 585 /* return < 0 if error. See bdrv_write() for the return codes */
 586 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 587               uint8_t *buf, int nb_sectors)
 588 {
 589     BlockDriver *drv = bs->drv;
 590
 591     if (!drv)
 592         return -ENOMEDIUM;
 593     if (bdrv_check_request(bs, sector_num, nb_sectors))
 594         return -EIO;
 595
 596     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 597 }
 598
 599 /* Return < 0 if error. Important errors are:
 600   -EIO         generic I/O error (may happen for all errors)
 601   -ENOMEDIUM   No media inserted.
 602   -EINVAL      Invalid sector number or nb_sectors
 603   -EACCES      Trying to write a read-only device
 604 */
 605 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 606                const uint8_t *buf, int nb_sectors)
 607 {
 608     BlockDriver *drv = bs->drv;
 609     if (!bs->drv)
 610         return -ENOMEDIUM;
 611     if (bs->read_only)
 612         return -EACCES;
 613     if (bdrv_check_request(bs, sector_num, nb_sectors))
 614         return -EIO;
 615
 616     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 617 }
 618
 619 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 620                void *buf, int count1)
 621 {
 622     uint8_t tmp_buf[SECTOR_SIZE];
 623     int len, nb_sectors, count;
 624     int64_t sector_num;
 625
 626     count = count1;
 627     /* first read to align to sector start */
 628     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 629     if (len > count)
 630         len = count;
 631     sector_num = offset >> SECTOR_BITS;
 632     if (len > 0) {
 633         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 634             return -EIO;
 635         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 636         count -= len;
 637         if (count == 0)
 638             return count1;
 639         sector_num++;
 640         buf += len;
 641     }
 642
 643     /* read the sectors "in place" */
 644     nb_sectors = count >> SECTOR_BITS;
 645     if (nb_sectors > 0) {
 646         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 647             return -EIO;
 648         sector_num += nb_sectors;
 649         len = nb_sectors << SECTOR_BITS;
 650         buf += len;
 651         count -= len;
 652     }
 653
 654     /* add data from the last sector */
 655     if (count > 0) {
 656         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 657             return -EIO;
 658         memcpy(buf, tmp_buf, count);
 659     }
 660     return count1;
 661 }
 662
 663 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 664                 const void *buf, int count1)
 665 {
 666     uint8_t tmp_buf[SECTOR_SIZE];
 667     int len, nb_sectors, count;
 668     int64_t sector_num;
 669
 670     count = count1;
 671     /* first write to align to sector start */
 672     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 673     if (len > count)
 674         len = count;
 675     sector_num = offset >> SECTOR_BITS;
 676     if (len > 0) {
 677         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 678             return -EIO;
 679         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 680         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 681             return -EIO;
 682         count -= len;
 683         if (count == 0)
 684             return count1;
 685         sector_num++;
 686         buf += len;
 687     }
 688
 689     /* write the sectors "in place" */
 690     nb_sectors = count >> SECTOR_BITS;
 691     if (nb_sectors > 0) {
 692         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 693             return -EIO;
 694         sector_num += nb_sectors;
 695         len = nb_sectors << SECTOR_BITS;
 696         buf += len;
 697         count -= len;
 698     }
 699
 700     /* add data from the last sector */
 701     if (count > 0) {
 702         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 703             return -EIO;
 704         memcpy(tmp_buf, buf, count);
 705         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 706             return -EIO;
 707     }
 708     return count1;
 709 }
 710
 711 /**
 712  * Truncate file to 'offset' bytes (needed only for file protocols)
 713  */
 714 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 715 {
 716     BlockDriver *drv = bs->drv;
 717     if (!drv)
 718         return -ENOMEDIUM;
 719     if (!drv->bdrv_truncate)
 720         return -ENOTSUP;
 721     return drv->bdrv_truncate(bs, offset);
 722 }
 723
 724 /**
 725  * Length of a file in bytes. Return < 0 if error or unknown.
 726  */
 727 int64_t bdrv_getlength(BlockDriverState *bs)
 728 {
 729     BlockDriver *drv = bs->drv;
 730     if (!drv)
 731         return -ENOMEDIUM;
 732     if (!drv->bdrv_getlength) {
 733         /* legacy mode */
 734         return bs->total_sectors * SECTOR_SIZE;
 735     }
 736     return drv->bdrv_getlength(bs);
 737 }
 738
 739 /* return 0 as number of sectors if no device present or error */
 740 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 741 {
 742     int64_t length;
 743     length = bdrv_getlength(bs);
 744     if (length < 0)
 745         length = 0;
 746     else
 747         length = length >> SECTOR_BITS;
 748     *nb_sectors_ptr = length;
 749 }
 750
 751 struct partition {
 752         uint8_t boot_ind;           /* 0x80 - active */
 753         uint8_t head;               /* starting head */
 754         uint8_t sector;             /* starting sector */
 755         uint8_t cyl;                /* starting cylinder */
 756         uint8_t sys_ind;            /* What partition type */
 757         uint8_t end_head;           /* end head */
 758         uint8_t end_sector;         /* end sector */
 759         uint8_t end_cyl;            /* end cylinder */
 760         uint32_t start_sect;        /* starting sector counting from 0 */
 761         uint32_t nr_sects;          /* nr of sectors in partition */
 762 } __attribute__((packed));
 763
 764 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 765 static int guess_disk_lchs(BlockDriverState *bs,
 766                            int *pcylinders, int *pheads, int *psectors)
 767 {
 768     uint8_t buf[512];
 769     int ret, i, heads, sectors, cylinders;
 770     struct partition *p;
 771     uint32_t nr_sects;
 772     uint64_t nb_sectors;
 773
 774     bdrv_get_geometry(bs, &nb_sectors);
 775
 776     ret = bdrv_read(bs, 0, buf, 1);
 777     if (ret < 0)
 778         return -1;
 779     /* test msdos magic */
 780     if (buf[510] != 0x55 || buf[511] != 0xaa)
 781         return -1;
 782     for(i = 0; i < 4; i++) {
 783         p = ((struct partition *)(buf + 0x1be)) + i;
 784         nr_sects = le32_to_cpu(p->nr_sects);
 785         if (nr_sects && p->end_head) {
 786             /* We make the assumption that the partition terminates on
 787                a cylinder boundary */
 788             heads = p->end_head + 1;
 789             sectors = p->end_sector & 63;
 790             if (sectors == 0)
 791                 continue;
 792             cylinders = nb_sectors / (heads * sectors);
 793             if (cylinders < 1 || cylinders > 16383)
 794                 continue;
 795             *pheads = heads;
 796             *psectors = sectors;
 797             *pcylinders = cylinders;
 798 #if 0
 799             printf("guessed geometry: LCHS=%d %d %d\n",
 800                    cylinders, heads, sectors);
 801 #endif
 802             return 0;
 803         }
 804     }
 805     return -1;
 806 }
 807
 808 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 809 {
 810     int translation, lba_detected = 0;
 811     int cylinders, heads, secs;
 812     uint64_t nb_sectors;
 813
 814     /* if a geometry hint is available, use it */
 815     bdrv_get_geometry(bs, &nb_sectors);
 816     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 817     translation = bdrv_get_translation_hint(bs);
 818     if (cylinders != 0) {
 819         *pcyls = cylinders;
 820         *pheads = heads;
 821         *psecs = secs;
 822     } else {
 823         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 824             if (heads > 16) {
 825                 /* if heads > 16, it means that a BIOS LBA
 826                    translation was active, so the default
 827                    hardware geometry is OK */
 828                 lba_detected = 1;
 829                 goto default_geometry;
 830             } else {
 831                 *pcyls = cylinders;
 832                 *pheads = heads;
 833                 *psecs = secs;
 834                 /* disable any translation to be in sync with
 835                    the logical geometry */
 836                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 837                     bdrv_set_translation_hint(bs,
 838                                               BIOS_ATA_TRANSLATION_NONE);
 839                 }
 840             }
 841         } else {
 842         default_geometry:
 843             /* if no geometry, use a standard physical disk geometry */
 844             cylinders = nb_sectors / (16 * 63);
 845
 846             if (cylinders > 16383)
 847                 cylinders = 16383;
 848             else if (cylinders < 2)
 849                 cylinders = 2;
 850             *pcyls = cylinders;
 851             *pheads = 16;
 852             *psecs = 63;
 853             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 854                 if ((*pcyls * *pheads) <= 131072) {
 855                     bdrv_set_translation_hint(bs,
 856                                               BIOS_ATA_TRANSLATION_LARGE);
 857                 } else {
 858                     bdrv_set_translation_hint(bs,
 859                                               BIOS_ATA_TRANSLATION_LBA);
 860                 }
 861             }
 862         }
 863         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 864     }
 865 }
 866
 867 void bdrv_set_geometry_hint(BlockDriverState *bs,
 868                             int cyls, int heads, int secs)
 869 {
 870     bs->cyls = cyls;
 871     bs->heads = heads;
 872     bs->secs = secs;
 873 }
 874
 875 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 876 {
 877     bs->type = type;
 878     bs->removable = ((type == BDRV_TYPE_CDROM ||
 879                       type == BDRV_TYPE_FLOPPY));
 880 }
 881
 882 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 883 {
 884     bs->translation = translation;
 885 }
 886
 887 void bdrv_get_geometry_hint(BlockDriverState *bs,
 888                             int *pcyls, int *pheads, int *psecs)
 889 {
 890     *pcyls = bs->cyls;
 891     *pheads = bs->heads;
 892     *psecs = bs->secs;
 893 }
 894
 895 int bdrv_get_type_hint(BlockDriverState *bs)
 896 {
 897     return bs->type;
 898 }
 899
 900 int bdrv_get_translation_hint(BlockDriverState *bs)
 901 {
 902     return bs->translation;
 903 }
 904
 905 int bdrv_is_removable(BlockDriverState *bs)
 906 {
 907     return bs->removable;
 908 }
 909
 910 int bdrv_is_read_only(BlockDriverState *bs)
 911 {
 912     return bs->read_only;
 913 }
 914
 915 int bdrv_is_sg(BlockDriverState *bs)
 916 {
 917     return bs->sg;
 918 }
 919
 920 /* XXX: no longer used */
 921 void bdrv_set_change_cb(BlockDriverState *bs,
 922                         void (*change_cb)(void *opaque), void *opaque)
 923 {
 924     bs->change_cb = change_cb;
 925     bs->change_opaque = opaque;
 926 }
 927
 928 int bdrv_is_encrypted(BlockDriverState *bs)
 929 {
 930     if (bs->backing_hd && bs->backing_hd->encrypted)
 931         return 1;
 932     return bs->encrypted;
 933 }
 934
 935 int bdrv_key_required(BlockDriverState *bs)
 936 {
 937     BlockDriverState *backing_hd = bs->backing_hd;
 938
 939     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 940         return 1;
 941     return (bs->encrypted && !bs->valid_key);
 942 }
 943
 944 int bdrv_set_key(BlockDriverState *bs, const char *key)
 945 {
 946     int ret;
 947     if (bs->backing_hd && bs->backing_hd->encrypted) {
 948         ret = bdrv_set_key(bs->backing_hd, key);
 949         if (ret < 0)
 950             return ret;
 951         if (!bs->encrypted)
 952             return 0;
 953     }
 954     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 955         return -1;
 956     ret = bs->drv->bdrv_set_key(bs, key);
 957     if (ret < 0) {
 958         bs->valid_key = 0;
 959     } else if (!bs->valid_key) {
 960         bs->valid_key = 1;
 961         /* call the change callback now, we skipped it on open */
 962         bs->media_changed = 1;
 963         if (bs->change_cb)
 964             bs->change_cb(bs->change_opaque);
 965     }
 966     return ret;
 967 }
 968
 969 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 970 {
 971     if (!bs->drv) {
 972         buf[0] = '\0';
 973     } else {
 974         pstrcpy(buf, buf_size, bs->drv->format_name);
 975     }
 976 }
 977
 978 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 979                          void *opaque)
 980 {
 981     BlockDriver *drv;
 982
 983     for (drv = first_drv; drv != NULL; drv = drv->next) {
 984         it(opaque, drv->format_name);
 985     }
 986 }
 987
 988 BlockDriverState *bdrv_find(const char *name)
 989 {
 990     BlockDriverState *bs;
 991
 992     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 993         if (!strcmp(name, bs->device_name))
 994             return bs;
 995     }
 996     return NULL;
 997 }
 998
 999 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1000 {
1001     BlockDriverState *bs;
1002
1003     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1004         it(opaque, bs);
1005     }
1006 }
1007
1008 const char *bdrv_get_device_name(BlockDriverState *bs)
1009 {
1010     return bs->device_name;
1011 }
1012
1013 void bdrv_flush(BlockDriverState *bs)
1014 {
1015     if (!bs->drv)
1016         return;
1017     if (bs->drv->bdrv_flush)
1018         bs->drv->bdrv_flush(bs);
1019     if (bs->backing_hd)
1020         bdrv_flush(bs->backing_hd);
1021 }
1022
1023 void bdrv_flush_all(void)
1024 {
1025     BlockDriverState *bs;
1026
1027     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1028         if (bs->drv && !bdrv_is_read_only(bs) &&
1029             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1030             bdrv_flush(bs);
1031 }
1032
1033 /*
1034  * Returns true iff the specified sector is present in the disk image. Drivers
1035  * not implementing the functionality are assumed to not support backing files,
1036  * hence all their sectors are reported as allocated.
1037  *
1038  * 'pnum' is set to the number of sectors (including and immediately following
1039  * the specified sector) that are known to be in the same
1040  * allocated/unallocated state.
1041  *
1042  * 'nb_sectors' is the max value 'pnum' should be set to.
1043  */
1044 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1045         int *pnum)
1046 {
1047     int64_t n;
1048     if (!bs->drv->bdrv_is_allocated) {
1049         if (sector_num >= bs->total_sectors) {
1050             *pnum = 0;
1051             return 0;
1052         }
1053         n = bs->total_sectors - sector_num;
1054         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1055         return 1;
1056     }
1057     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1058 }
1059
1060 void bdrv_info(Monitor *mon)
1061 {
1062     BlockDriverState *bs;
1063
1064     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1065         monitor_printf(mon, "%s:", bs->device_name);
1066         monitor_printf(mon, " type=");
1067         switch(bs->type) {
1068         case BDRV_TYPE_HD:
1069             monitor_printf(mon, "hd");
1070             break;
1071         case BDRV_TYPE_CDROM:
1072             monitor_printf(mon, "cdrom");
1073             break;
1074         case BDRV_TYPE_FLOPPY:
1075             monitor_printf(mon, "floppy");
1076             break;
1077         }
1078         monitor_printf(mon, " removable=%d", bs->removable);
1079         if (bs->removable) {
1080             monitor_printf(mon, " locked=%d", bs->locked);
1081         }
1082         if (bs->drv) {
1083             monitor_printf(mon, " file=");
1084             monitor_print_filename(mon, bs->filename);
1085             if (bs->backing_file[0] != '\0') {
1086                 monitor_printf(mon, " backing_file=");
1087                 monitor_print_filename(mon, bs->backing_file);
1088             }
1089             monitor_printf(mon, " ro=%d", bs->read_only);
1090             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1091             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1092         } else {
1093             monitor_printf(mon, " [not inserted]");
1094         }
1095         monitor_printf(mon, "\n");
1096     }
1097 }
1098
1099 /* The "info blockstats" command. */
1100 void bdrv_info_stats(Monitor *mon)
1101 {
1102     BlockDriverState *bs;
1103
1104     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1105         monitor_printf(mon, "%s:"
1106                        " rd_bytes=%" PRIu64
1107                        " wr_bytes=%" PRIu64
1108                        " rd_operations=%" PRIu64
1109                        " wr_operations=%" PRIu64
1110                        "\n",
1111                        bs->device_name,
1112                        bs->rd_bytes, bs->wr_bytes,
1113                        bs->rd_ops, bs->wr_ops);
1114     }
1115 }
1116
1117 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1118 {
1119     if (bs->backing_hd && bs->backing_hd->encrypted)
1120         return bs->backing_file;
1121     else if (bs->encrypted)
1122         return bs->filename;
1123     else
1124         return NULL;
1125 }
1126
1127 void bdrv_get_backing_filename(BlockDriverState *bs,
1128                                char *filename, int filename_size)
1129 {
1130     if (!bs->backing_hd) {
1131         pstrcpy(filename, filename_size, "");
1132     } else {
1133         pstrcpy(filename, filename_size, bs->backing_file);
1134     }
1135 }
1136
1137 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1138                           const uint8_t *buf, int nb_sectors)
1139 {
1140     BlockDriver *drv = bs->drv;
1141     if (!drv)
1142         return -ENOMEDIUM;
1143     if (!drv->bdrv_write_compressed)
1144         return -ENOTSUP;
1145     if (bdrv_check_request(bs, sector_num, nb_sectors))
1146         return -EIO;
1147     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1148 }
1149
1150 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1151 {
1152     BlockDriver *drv = bs->drv;
1153     if (!drv)
1154         return -ENOMEDIUM;
1155     if (!drv->bdrv_get_info)
1156         return -ENOTSUP;
1157     memset(bdi, 0, sizeof(*bdi));
1158     return drv->bdrv_get_info(bs, bdi);
1159 }
1160
1161 int bdrv_save_vmstate(BlockDriverState *bs, const uint8_t *buf,
1162                       int64_t pos, int size)
1163 {
1164     BlockDriver *drv = bs->drv;
1165     if (!drv)
1166         return -ENOMEDIUM;
1167     if (!drv->bdrv_save_vmstate)
1168         return -ENOTSUP;
1169     return drv->bdrv_save_vmstate(bs, buf, pos, size);
1170 }
1171
1172 int bdrv_load_vmstate(BlockDriverState *bs, uint8_t *buf,
1173                       int64_t pos, int size)
1174 {
1175     BlockDriver *drv = bs->drv;
1176     if (!drv)
1177         return -ENOMEDIUM;
1178     if (!drv->bdrv_load_vmstate)
1179         return -ENOTSUP;
1180     return drv->bdrv_load_vmstate(bs, buf, pos, size);
1181 }
1182
1183 /**************************************************************/
1184 /* handling of snapshots */
1185
1186 int bdrv_snapshot_create(BlockDriverState *bs,
1187                          QEMUSnapshotInfo *sn_info)
1188 {
1189     BlockDriver *drv = bs->drv;
1190     if (!drv)
1191         return -ENOMEDIUM;
1192     if (!drv->bdrv_snapshot_create)
1193         return -ENOTSUP;
1194     return drv->bdrv_snapshot_create(bs, sn_info);
1195 }
1196
1197 int bdrv_snapshot_goto(BlockDriverState *bs,
1198                        const char *snapshot_id)
1199 {
1200     BlockDriver *drv = bs->drv;
1201     if (!drv)
1202         return -ENOMEDIUM;
1203     if (!drv->bdrv_snapshot_goto)
1204         return -ENOTSUP;
1205     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1206 }
1207
1208 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1209 {
1210     BlockDriver *drv = bs->drv;
1211     if (!drv)
1212         return -ENOMEDIUM;
1213     if (!drv->bdrv_snapshot_delete)
1214         return -ENOTSUP;
1215     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1216 }
1217
1218 int bdrv_snapshot_list(BlockDriverState *bs,
1219                        QEMUSnapshotInfo **psn_info)
1220 {
1221     BlockDriver *drv = bs->drv;
1222     if (!drv)
1223         return -ENOMEDIUM;
1224     if (!drv->bdrv_snapshot_list)
1225         return -ENOTSUP;
1226     return drv->bdrv_snapshot_list(bs, psn_info);
1227 }
1228
1229 #define NB_SUFFIXES 4
1230
1231 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1232 {
1233     static const char suffixes[NB_SUFFIXES] = "KMGT";
1234     int64_t base;
1235     int i;
1236
1237     if (size <= 999) {
1238         snprintf(buf, buf_size, "%" PRId64, size);
1239     } else {
1240         base = 1024;
1241         for(i = 0; i < NB_SUFFIXES; i++) {
1242             if (size < (10 * base)) {
1243                 snprintf(buf, buf_size, "%0.1f%c",
1244                          (double)size / base,
1245                          suffixes[i]);
1246                 break;
1247             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1248                 snprintf(buf, buf_size, "%" PRId64 "%c",
1249                          ((size + (base >> 1)) / base),
1250                          suffixes[i]);
1251                 break;
1252             }
1253             base = base * 1024;
1254         }
1255     }
1256     return buf;
1257 }
1258
1259 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1260 {
1261     char buf1[128], date_buf[128], clock_buf[128];
1262 #ifdef _WIN32
1263     struct tm *ptm;
1264 #else
1265     struct tm tm;
1266 #endif
1267     time_t ti;
1268     int64_t secs;
1269
1270     if (!sn) {
1271         snprintf(buf, buf_size,
1272                  "%-10s%-20s%7s%20s%15s",
1273                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1274     } else {
1275         ti = sn->date_sec;
1276 #ifdef _WIN32
1277         ptm = localtime(&ti);
1278         strftime(date_buf, sizeof(date_buf),
1279                  "%Y-%m-%d %H:%M:%S", ptm);
1280 #else
1281         localtime_r(&ti, &tm);
1282         strftime(date_buf, sizeof(date_buf),
1283                  "%Y-%m-%d %H:%M:%S", &tm);
1284 #endif
1285         secs = sn->vm_clock_nsec / 1000000000;
1286         snprintf(clock_buf, sizeof(clock_buf),
1287                  "%02d:%02d:%02d.%03d",
1288                  (int)(secs / 3600),
1289                  (int)((secs / 60) % 60),
1290                  (int)(secs % 60),
1291                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1292         snprintf(buf, buf_size,
1293                  "%-10s%-20s%7s%20s%15s",
1294                  sn->id_str, sn->name,
1295                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1296                  date_buf,
1297                  clock_buf);
1298     }
1299     return buf;
1300 }
1301
1302
1303 /**************************************************************/
1304 /* async I/Os */
1305
1306 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1307                                  QEMUIOVector *qiov, int nb_sectors,
1308                                  BlockDriverCompletionFunc *cb, void *opaque)
1309 {
1310     BlockDriver *drv = bs->drv;
1311     BlockDriverAIOCB *ret;
1312
1313     if (!drv)
1314         return NULL;
1315     if (bdrv_check_request(bs, sector_num, nb_sectors))
1316         return NULL;
1317
1318     ret = drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
1319                               cb, opaque);
1320
1321     if (ret) {
1322         /* Update stats even though technically transfer has not happened. */
1323         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1324         bs->rd_ops ++;
1325     }
1326
1327     return ret;
1328 }
1329
1330 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1331                                   QEMUIOVector *qiov, int nb_sectors,
1332                                   BlockDriverCompletionFunc *cb, void *opaque)
1333 {
1334     BlockDriver *drv = bs->drv;
1335     BlockDriverAIOCB *ret;
1336
1337     if (!drv)
1338         return NULL;
1339     if (bs->read_only)
1340         return NULL;
1341     if (bdrv_check_request(bs, sector_num, nb_sectors))
1342         return NULL;
1343
1344     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1345                                cb, opaque);
1346
1347     if (ret) {
1348         /* Update stats even though technically transfer has not happened. */
1349         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1350         bs->wr_ops ++;
1351     }
1352
1353     return ret;
1354 }
1355
1356 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1357 {
1358     acb->pool->cancel(acb);
1359 }
1360
1361
1362 /**************************************************************/
1363 /* async block device emulation */
1364
1365 typedef struct BlockDriverAIOCBSync {
1366     BlockDriverAIOCB common;
1367     QEMUBH *bh;
1368     int ret;
1369     /* vector translation state */
1370     QEMUIOVector *qiov;
1371     uint8_t *bounce;
1372     int is_write;
1373 } BlockDriverAIOCBSync;
1374
1375 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1376 {
1377     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1378     qemu_bh_delete(acb->bh);
1379     acb->bh = NULL;
1380     qemu_aio_release(acb);
1381 }
1382
1383 static AIOPool bdrv_em_aio_pool = {
1384     .aiocb_size         = sizeof(BlockDriverAIOCBSync),
1385     .cancel             = bdrv_aio_cancel_em,
1386 };
1387
1388 static void bdrv_aio_bh_cb(void *opaque)
1389 {
1390     BlockDriverAIOCBSync *acb = opaque;
1391
1392     if (!acb->is_write)
1393         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1394     qemu_vfree(acb->bounce);
1395     acb->common.cb(acb->common.opaque, acb->ret);
1396     qemu_bh_delete(acb->bh);
1397     acb->bh = NULL;
1398     qemu_aio_release(acb);
1399 }
1400
1401 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1402                                             int64_t sector_num,
1403                                             QEMUIOVector *qiov,
1404                                             int nb_sectors,
1405                                             BlockDriverCompletionFunc *cb,
1406                                             void *opaque,
1407                                             int is_write)
1408
1409 {
1410     BlockDriverAIOCBSync *acb;
1411
1412     acb = qemu_aio_get(&bdrv_em_aio_pool, bs, cb, opaque);
1413     acb->is_write = is_write;
1414     acb->qiov = qiov;
1415     acb->bounce = qemu_blockalign(bs, qiov->size);
1416
1417     if (!acb->bh)
1418         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1419
1420     if (is_write) {
1421         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1422         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1423     } else {
1424         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1425     }
1426
1427     qemu_bh_schedule(acb->bh);
1428
1429     return &acb->common;
1430 }
1431
1432 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1433         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1434         BlockDriverCompletionFunc *cb, void *opaque)
1435 {
1436     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1437 }
1438
1439 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1440         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1441         BlockDriverCompletionFunc *cb, void *opaque)
1442 {
1443     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1444 }
1445
1446 /**************************************************************/
1447 /* sync block device emulation */
1448
1449 static void bdrv_rw_em_cb(void *opaque, int ret)
1450 {
1451     *(int *)opaque = ret;
1452 }
1453
1454 #define NOT_DONE 0x7fffffff
1455
1456 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1457                         uint8_t *buf, int nb_sectors)
1458 {
1459     int async_ret;
1460     BlockDriverAIOCB *acb;
1461     struct iovec iov;
1462     QEMUIOVector qiov;
1463
1464     async_ret = NOT_DONE;
1465     iov.iov_base = (void *)buf;
1466     iov.iov_len = nb_sectors * 512;
1467     qemu_iovec_init_external(&qiov, &iov, 1);
1468     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1469         bdrv_rw_em_cb, &async_ret);
1470     if (acb == NULL)
1471         return -1;
1472
1473     while (async_ret == NOT_DONE) {
1474         qemu_aio_wait();
1475     }
1476
1477     return async_ret;
1478 }
1479
1480 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1481                          const uint8_t *buf, int nb_sectors)
1482 {
1483     int async_ret;
1484     BlockDriverAIOCB *acb;
1485     struct iovec iov;
1486     QEMUIOVector qiov;
1487
1488     async_ret = NOT_DONE;
1489     iov.iov_base = (void *)buf;
1490     iov.iov_len = nb_sectors * 512;
1491     qemu_iovec_init_external(&qiov, &iov, 1);
1492     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1493         bdrv_rw_em_cb, &async_ret);
1494     if (acb == NULL)
1495         return -1;
1496     while (async_ret == NOT_DONE) {
1497         qemu_aio_wait();
1498     }
1499     return async_ret;
1500 }
1501
1502 void bdrv_init(void)
1503 {
1504     module_call_init(MODULE_INIT_BLOCK);
1505 }
1506
1507 void *qemu_aio_get(AIOPool *pool, BlockDriverState *bs,
1508                    BlockDriverCompletionFunc *cb, void *opaque)
1509 {
1510     BlockDriverAIOCB *acb;
1511
1512     if (pool->free_aiocb) {
1513         acb = pool->free_aiocb;
1514         pool->free_aiocb = acb->next;
1515     } else {
1516         acb = qemu_mallocz(pool->aiocb_size);
1517         acb->pool = pool;
1518     }
1519     acb->bs = bs;
1520     acb->cb = cb;
1521     acb->opaque = opaque;
1522     return acb;
1523 }
1524
1525 void qemu_aio_release(void *p)
1526 {
1527     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1528     AIOPool *pool = acb->pool;
1529     acb->next = pool->free_aiocb;
1530     pool->free_aiocb = acb;
1531 }
1532
1533 /**************************************************************/
1534 /* removable device support */
1535
1536 /**
1537  * Return TRUE if the media is present
1538  */
1539 int bdrv_is_inserted(BlockDriverState *bs)
1540 {
1541     BlockDriver *drv = bs->drv;
1542     int ret;
1543     if (!drv)
1544         return 0;
1545     if (!drv->bdrv_is_inserted)
1546         return 1;
1547     ret = drv->bdrv_is_inserted(bs);
1548     return ret;
1549 }
1550
1551 /**
1552  * Return TRUE if the media changed since the last call to this
1553  * function. It is currently only used for floppy disks
1554  */
1555 int bdrv_media_changed(BlockDriverState *bs)
1556 {
1557     BlockDriver *drv = bs->drv;
1558     int ret;
1559
1560     if (!drv || !drv->bdrv_media_changed)
1561         ret = -ENOTSUP;
1562     else
1563         ret = drv->bdrv_media_changed(bs);
1564     if (ret == -ENOTSUP)
1565         ret = bs->media_changed;
1566     bs->media_changed = 0;
1567     return ret;
1568 }
1569
1570 /**
1571  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1572  */
1573 int bdrv_eject(BlockDriverState *bs, int eject_flag)
1574 {
1575     BlockDriver *drv = bs->drv;
1576     int ret;
1577
1578     if (bs->locked) {
1579         return -EBUSY;
1580     }
1581
1582     if (!drv || !drv->bdrv_eject) {
1583         ret = -ENOTSUP;
1584     } else {
1585         ret = drv->bdrv_eject(bs, eject_flag);
1586     }
1587     if (ret == -ENOTSUP) {
1588         if (eject_flag)
1589             bdrv_close(bs);
1590         ret = 0;
1591     }
1592
1593     return ret;
1594 }
1595
1596 int bdrv_is_locked(BlockDriverState *bs)
1597 {
1598     return bs->locked;
1599 }
1600
1601 /**
1602  * Lock or unlock the media (if it is locked, the user won't be able
1603  * to eject it manually).
1604  */
1605 void bdrv_set_locked(BlockDriverState *bs, int locked)
1606 {
1607     BlockDriver *drv = bs->drv;
1608
1609     bs->locked = locked;
1610     if (drv && drv->bdrv_set_locked) {
1611         drv->bdrv_set_locked(bs, locked);
1612     }
1613 }
1614
1615 /* needed for generic scsi interface */
1616
1617 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1618 {
1619     BlockDriver *drv = bs->drv;
1620
1621     if (drv && drv->bdrv_ioctl)
1622         return drv->bdrv_ioctl(bs, req, buf);
1623     return -ENOTSUP;
1624 }
1625
1626 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1627         unsigned long int req, void *buf,
1628         BlockDriverCompletionFunc *cb, void *opaque)
1629 {
1630     BlockDriver *drv = bs->drv;
1631
1632     if (drv && drv->bdrv_aio_ioctl)
1633         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1634     return NULL;
1635 }
1636
1637 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1638 {
1639     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1640 }