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