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 _BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "console.h"
  32 #include "block_int.h"
  33 #include "osdep.h"
  34
  35 #ifdef _BSD
  36 #include <sys/types.h>
  37 #include <sys/stat.h>
  38 #include <sys/ioctl.h>
  39 #include <sys/disk.h>
  40 #endif
  41
  42 #define SECTOR_BITS 9
  43 #define SECTOR_SIZE (1 << SECTOR_BITS)
  44
  45 typedef struct BlockDriverAIOCBSync {
  46     BlockDriverAIOCB common;
  47     QEMUBH *bh;
  48     int ret;
  49 } BlockDriverAIOCBSync;
  50
  51 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  52         int64_t sector_num, uint8_t *buf, int nb_sectors,
  53         BlockDriverCompletionFunc *cb, void *opaque);
  54 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  55         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  56         BlockDriverCompletionFunc *cb, void *opaque);
  57 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  58 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  59                         uint8_t *buf, int nb_sectors);
  60 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  61                          const uint8_t *buf, int nb_sectors);
  62
  63 BlockDriverState *bdrv_first;
  64
  65 static BlockDriver *first_drv;
  66
  67 int path_is_absolute(const char *path)
  68 {
  69     const char *p;
  70 #ifdef _WIN32
  71     /* specific case for names like: "\\.\d:" */
  72     if (*path == '/' || *path == '\\')
  73         return 1;
  74 #endif
  75     p = strchr(path, ':');
  76     if (p)
  77         p++;
  78     else
  79         p = path;
  80 #ifdef _WIN32
  81     return (*p == '/' || *p == '\\');
  82 #else
  83     return (*p == '/');
  84 #endif
  85 }
  86
  87 /* if filename is absolute, just copy it to dest. Otherwise, build a
  88    path to it by considering it is relative to base_path. URL are
  89    supported. */
  90 void path_combine(char *dest, int dest_size,
  91                   const char *base_path,
  92                   const char *filename)
  93 {
  94     const char *p, *p1;
  95     int len;
  96
  97     if (dest_size <= 0)
  98         return;
  99     if (path_is_absolute(filename)) {
 100         pstrcpy(dest, dest_size, filename);
 101     } else {
 102         p = strchr(base_path, ':');
 103         if (p)
 104             p++;
 105         else
 106             p = base_path;
 107         p1 = strrchr(base_path, '/');
 108 #ifdef _WIN32
 109         {
 110             const char *p2;
 111             p2 = strrchr(base_path, '\\');
 112             if (!p1 || p2 > p1)
 113                 p1 = p2;
 114         }
 115 #endif
 116         if (p1)
 117             p1++;
 118         else
 119             p1 = base_path;
 120         if (p1 > p)
 121             p = p1;
 122         len = p - base_path;
 123         if (len > dest_size - 1)
 124             len = dest_size - 1;
 125         memcpy(dest, base_path, len);
 126         dest[len] = '\0';
 127         pstrcat(dest, dest_size, filename);
 128     }
 129 }
 130
 131
 132 static void bdrv_register(BlockDriver *bdrv)
 133 {
 134     if (!bdrv->bdrv_aio_read) {
 135         /* add AIO emulation layer */
 136         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 137         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 138         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 139         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 140     } else if (!bdrv->bdrv_read && !bdrv->bdrv_pread) {
 141         /* add synchronous IO emulation layer */
 142         bdrv->bdrv_read = bdrv_read_em;
 143         bdrv->bdrv_write = bdrv_write_em;
 144     }
 145     bdrv->next = first_drv;
 146     first_drv = bdrv;
 147 }
 148
 149 /* create a new block device (by default it is empty) */
 150 BlockDriverState *bdrv_new(const char *device_name)
 151 {
 152     BlockDriverState **pbs, *bs;
 153
 154     bs = qemu_mallocz(sizeof(BlockDriverState));
 155     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 156     if (device_name[0] != '\0') {
 157         /* insert at the end */
 158         pbs = &bdrv_first;
 159         while (*pbs != NULL)
 160             pbs = &(*pbs)->next;
 161         *pbs = bs;
 162     }
 163     return bs;
 164 }
 165
 166 BlockDriver *bdrv_find_format(const char *format_name)
 167 {
 168     BlockDriver *drv1;
 169     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 170         if (!strcmp(drv1->format_name, format_name))
 171             return drv1;
 172     }
 173     return NULL;
 174 }
 175
 176 int bdrv_create(BlockDriver *drv,
 177                 const char *filename, int64_t size_in_sectors,
 178                 const char *backing_file, int flags)
 179 {
 180     if (!drv->bdrv_create)
 181         return -ENOTSUP;
 182     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 183 }
 184
 185 #ifdef _WIN32
 186 void get_tmp_filename(char *filename, int size)
 187 {
 188     char temp_dir[MAX_PATH];
 189
 190     GetTempPath(MAX_PATH, temp_dir);
 191     GetTempFileName(temp_dir, "qem", 0, filename);
 192 }
 193 #else
 194 void get_tmp_filename(char *filename, int size)
 195 {
 196     int fd;
 197     const char *tmpdir;
 198     /* XXX: race condition possible */
 199     tmpdir = getenv("TMPDIR");
 200     if (!tmpdir)
 201         tmpdir = "/tmp";
 202     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 203     fd = mkstemp(filename);
 204     close(fd);
 205 }
 206 #endif
 207
 208 #ifdef _WIN32
 209 static int is_windows_drive_prefix(const char *filename)
 210 {
 211     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 212              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 213             filename[1] == ':');
 214 }
 215
 216 static int is_windows_drive(const char *filename)
 217 {
 218     if (is_windows_drive_prefix(filename) &&
 219         filename[2] == '\0')
 220         return 1;
 221     if (strstart(filename, "\\\\.\\", NULL) ||
 222         strstart(filename, "//./", NULL))
 223         return 1;
 224     return 0;
 225 }
 226 #endif
 227
 228 static BlockDriver *find_protocol(const char *filename)
 229 {
 230     BlockDriver *drv1;
 231     char protocol[128];
 232     int len;
 233     const char *p;
 234
 235 #ifdef _WIN32
 236     if (is_windows_drive(filename) ||
 237         is_windows_drive_prefix(filename))
 238         return &bdrv_raw;
 239 #endif
 240     p = strchr(filename, ':');
 241     if (!p)
 242         return &bdrv_raw;
 243     len = p - filename;
 244     if (len > sizeof(protocol) - 1)
 245         len = sizeof(protocol) - 1;
 246     memcpy(protocol, filename, len);
 247     protocol[len] = '\0';
 248     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 249         if (drv1->protocol_name &&
 250             !strcmp(drv1->protocol_name, protocol))
 251             return drv1;
 252     }
 253     return NULL;
 254 }
 255
 256 /* XXX: force raw format if block or character device ? It would
 257    simplify the BSD case */
 258 static BlockDriver *find_image_format(const char *filename)
 259 {
 260     int ret, score, score_max;
 261     BlockDriver *drv1, *drv;
 262     uint8_t buf[2048];
 263     BlockDriverState *bs;
 264
 265     /* detect host devices. By convention, /dev/cdrom[N] is always
 266        recognized as a host CDROM */
 267     if (strstart(filename, "/dev/cdrom", NULL))
 268         return &bdrv_host_device;
 269 #ifdef _WIN32
 270     if (is_windows_drive(filename))
 271         return &bdrv_host_device;
 272 #else
 273     {
 274         struct stat st;
 275         if (stat(filename, &st) >= 0 &&
 276             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 277             return &bdrv_host_device;
 278         }
 279     }
 280 #endif
 281
 282     drv = find_protocol(filename);
 283     /* no need to test disk image formats for vvfat */
 284     if (drv == &bdrv_vvfat)
 285         return drv;
 286
 287     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 288     if (ret < 0)
 289         return NULL;
 290     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 291     bdrv_delete(bs);
 292     if (ret < 0) {
 293         return NULL;
 294     }
 295
 296     score_max = 0;
 297     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 298         if (drv1->bdrv_probe) {
 299             score = drv1->bdrv_probe(buf, ret, filename);
 300             if (score > score_max) {
 301                 score_max = score;
 302                 drv = drv1;
 303             }
 304         }
 305     }
 306     return drv;
 307 }
 308
 309 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 310 {
 311     BlockDriverState *bs;
 312     int ret;
 313
 314     bs = bdrv_new("");
 315     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 316     if (ret < 0) {
 317         bdrv_delete(bs);
 318         return ret;
 319     }
 320     bs->growable = 1;
 321     *pbs = bs;
 322     return 0;
 323 }
 324
 325 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 326 {
 327     return bdrv_open2(bs, filename, flags, NULL);
 328 }
 329
 330 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 331                BlockDriver *drv)
 332 {
 333     int ret, open_flags;
 334     char tmp_filename[PATH_MAX];
 335     char backing_filename[PATH_MAX];
 336
 337     bs->read_only = 0;
 338     bs->is_temporary = 0;
 339     bs->encrypted = 0;
 340     bs->valid_key = 0;
 341
 342     if (flags & BDRV_O_SNAPSHOT) {
 343         BlockDriverState *bs1;
 344         int64_t total_size;
 345         int is_protocol = 0;
 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_open(bs1, filename, 0);
 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         ret = bdrv_create(&bdrv_qcow2, tmp_filename,
 374                           total_size, backing_filename, 0);
 375         if (ret < 0) {
 376             return ret;
 377         }
 378         filename = tmp_filename;
 379         bs->is_temporary = 1;
 380     }
 381
 382     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 383     if (flags & BDRV_O_FILE) {
 384         drv = find_protocol(filename);
 385     } else if (!drv) {
 386         drv = find_image_format(filename);
 387     }
 388     if (!drv) {
 389         ret = -ENOENT;
 390         goto unlink_and_fail;
 391     }
 392     bs->drv = drv;
 393     bs->opaque = qemu_mallocz(drv->instance_size);
 394     /* Note: for compatibility, we open disk image files as RDWR, and
 395        RDONLY as fallback */
 396     if (!(flags & BDRV_O_FILE))
 397         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 398     else
 399         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 400     ret = drv->bdrv_open(bs, filename, open_flags);
 401     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 402         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 403         bs->read_only = 1;
 404     }
 405     if (ret < 0) {
 406         qemu_free(bs->opaque);
 407         bs->opaque = NULL;
 408         bs->drv = NULL;
 409     unlink_and_fail:
 410         if (bs->is_temporary)
 411             unlink(filename);
 412         return ret;
 413     }
 414     if (drv->bdrv_getlength) {
 415         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 416     }
 417 #ifndef _WIN32
 418     if (bs->is_temporary) {
 419         unlink(filename);
 420     }
 421 #endif
 422     if (bs->backing_file[0] != '\0') {
 423         /* if there is a backing file, use it */
 424         bs->backing_hd = bdrv_new("");
 425         path_combine(backing_filename, sizeof(backing_filename),
 426                      filename, bs->backing_file);
 427         ret = bdrv_open(bs->backing_hd, backing_filename, open_flags);
 428         if (ret < 0) {
 429             bdrv_close(bs);
 430             return ret;
 431         }
 432     }
 433
 434     /* call the change callback */
 435     bs->media_changed = 1;
 436     if (bs->change_cb)
 437         bs->change_cb(bs->change_opaque);
 438
 439     return 0;
 440 }
 441
 442 void bdrv_close(BlockDriverState *bs)
 443 {
 444     if (bs->drv) {
 445         if (bs->backing_hd)
 446             bdrv_delete(bs->backing_hd);
 447         bs->drv->bdrv_close(bs);
 448         qemu_free(bs->opaque);
 449 #ifdef _WIN32
 450         if (bs->is_temporary) {
 451             unlink(bs->filename);
 452         }
 453 #endif
 454         bs->opaque = NULL;
 455         bs->drv = NULL;
 456
 457         /* call the change callback */
 458         bs->media_changed = 1;
 459         if (bs->change_cb)
 460             bs->change_cb(bs->change_opaque);
 461     }
 462 }
 463
 464 void bdrv_delete(BlockDriverState *bs)
 465 {
 466     BlockDriverState **pbs;
 467
 468     pbs = &bdrv_first;
 469     while (*pbs != bs && *pbs != NULL)
 470         pbs = &(*pbs)->next;
 471     if (*pbs == bs)
 472         *pbs = bs->next;
 473
 474     bdrv_close(bs);
 475     qemu_free(bs);
 476 }
 477
 478 /* commit COW file into the raw image */
 479 int bdrv_commit(BlockDriverState *bs)
 480 {
 481     BlockDriver *drv = bs->drv;
 482     int64_t i, total_sectors;
 483     int n, j;
 484     unsigned char sector[512];
 485
 486     if (!drv)
 487         return -ENOMEDIUM;
 488
 489     if (bs->read_only) {
 490         return -EACCES;
 491     }
 492
 493     if (!bs->backing_hd) {
 494         return -ENOTSUP;
 495     }
 496
 497     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 498     for (i = 0; i < total_sectors;) {
 499         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 500             for(j = 0; j < n; j++) {
 501                 if (bdrv_read(bs, i, sector, 1) != 0) {
 502                     return -EIO;
 503                 }
 504
 505                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 506                     return -EIO;
 507                 }
 508                 i++;
 509             }
 510         } else {
 511             i += n;
 512         }
 513     }
 514
 515     if (drv->bdrv_make_empty)
 516         return drv->bdrv_make_empty(bs);
 517
 518     return 0;
 519 }
 520
 521 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 522                                    size_t size)
 523 {
 524     int64_t len;
 525
 526     if (!bdrv_is_inserted(bs))
 527         return -ENOMEDIUM;
 528
 529     if (bs->growable)
 530         return 0;
 531
 532     len = bdrv_getlength(bs);
 533
 534     if ((offset + size) > len)
 535         return -EIO;
 536
 537     return 0;
 538 }
 539
 540 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 541                               int nb_sectors)
 542 {
 543     int64_t offset;
 544
 545     /* Deal with byte accesses */
 546     if (sector_num < 0)
 547         offset = -sector_num;
 548     else
 549         offset = sector_num * 512;
 550
 551     return bdrv_check_byte_request(bs, offset, nb_sectors * 512);
 552 }
 553
 554 /* return < 0 if error. See bdrv_write() for the return codes */
 555 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 556               uint8_t *buf, int nb_sectors)
 557 {
 558     BlockDriver *drv = bs->drv;
 559
 560     if (!drv)
 561         return -ENOMEDIUM;
 562     if (bdrv_check_request(bs, sector_num, nb_sectors))
 563         return -EIO;
 564
 565     if (drv->bdrv_pread) {
 566         int ret, len;
 567         len = nb_sectors * 512;
 568         ret = drv->bdrv_pread(bs, sector_num * 512, buf, len);
 569         if (ret < 0)
 570             return ret;
 571         else if (ret != len)
 572             return -EINVAL;
 573         else {
 574             bs->rd_bytes += (unsigned) len;
 575             bs->rd_ops ++;
 576             return 0;
 577         }
 578     } else {
 579         return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 580     }
 581 }
 582
 583 /* Return < 0 if error. Important errors are:
 584   -EIO         generic I/O error (may happen for all errors)
 585   -ENOMEDIUM   No media inserted.
 586   -EINVAL      Invalid sector number or nb_sectors
 587   -EACCES      Trying to write a read-only device
 588 */
 589 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 590                const uint8_t *buf, int nb_sectors)
 591 {
 592     BlockDriver *drv = bs->drv;
 593     if (!bs->drv)
 594         return -ENOMEDIUM;
 595     if (bs->read_only)
 596         return -EACCES;
 597     if (bdrv_check_request(bs, sector_num, nb_sectors))
 598         return -EIO;
 599
 600     if (drv->bdrv_pwrite) {
 601         int ret, len, count = 0;
 602         len = nb_sectors * 512;
 603         do {
 604             ret = drv->bdrv_pwrite(bs, sector_num * 512, buf, len - count);
 605             if (ret < 0) {
 606                 printf("bdrv_write ret=%d\n", ret);
 607                 return ret;
 608             }
 609             count += ret;
 610             buf += ret;
 611         } while (count != len);
 612         bs->wr_bytes += (unsigned) len;
 613         bs->wr_ops ++;
 614         return 0;
 615     }
 616     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 617 }
 618
 619 static int bdrv_pread_em(BlockDriverState *bs, int64_t offset,
 620                          uint8_t *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 static int bdrv_pwrite_em(BlockDriverState *bs, int64_t offset,
 664                           const uint8_t *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  * Read with byte offsets (needed only for file protocols)
 713  */
 714 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 715                void *buf1, int count1)
 716 {
 717     BlockDriver *drv = bs->drv;
 718
 719     if (!drv)
 720         return -ENOMEDIUM;
 721     if (bdrv_check_byte_request(bs, offset, count1))
 722         return -EIO;
 723
 724     if (!drv->bdrv_pread)
 725         return bdrv_pread_em(bs, offset, buf1, count1);
 726     return drv->bdrv_pread(bs, offset, buf1, count1);
 727 }
 728
 729 /**
 730  * Write with byte offsets (needed only for file protocols)
 731  */
 732 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 733                 const void *buf1, int count1)
 734 {
 735     BlockDriver *drv = bs->drv;
 736
 737     if (!drv)
 738         return -ENOMEDIUM;
 739     if (bdrv_check_byte_request(bs, offset, count1))
 740         return -EIO;
 741
 742     if (!drv->bdrv_pwrite)
 743         return bdrv_pwrite_em(bs, offset, buf1, count1);
 744     return drv->bdrv_pwrite(bs, offset, buf1, count1);
 745 }
 746
 747 /**
 748  * Truncate file to 'offset' bytes (needed only for file protocols)
 749  */
 750 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 751 {
 752     BlockDriver *drv = bs->drv;
 753     if (!drv)
 754         return -ENOMEDIUM;
 755     if (!drv->bdrv_truncate)
 756         return -ENOTSUP;
 757     return drv->bdrv_truncate(bs, offset);
 758 }
 759
 760 /**
 761  * Length of a file in bytes. Return < 0 if error or unknown.
 762  */
 763 int64_t bdrv_getlength(BlockDriverState *bs)
 764 {
 765     BlockDriver *drv = bs->drv;
 766     if (!drv)
 767         return -ENOMEDIUM;
 768     if (!drv->bdrv_getlength) {
 769         /* legacy mode */
 770         return bs->total_sectors * SECTOR_SIZE;
 771     }
 772     return drv->bdrv_getlength(bs);
 773 }
 774
 775 /* return 0 as number of sectors if no device present or error */
 776 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 777 {
 778     int64_t length;
 779     length = bdrv_getlength(bs);
 780     if (length < 0)
 781         length = 0;
 782     else
 783         length = length >> SECTOR_BITS;
 784     *nb_sectors_ptr = length;
 785 }
 786
 787 struct partition {
 788         uint8_t boot_ind;           /* 0x80 - active */
 789         uint8_t head;               /* starting head */
 790         uint8_t sector;             /* starting sector */
 791         uint8_t cyl;                /* starting cylinder */
 792         uint8_t sys_ind;            /* What partition type */
 793         uint8_t end_head;           /* end head */
 794         uint8_t end_sector;         /* end sector */
 795         uint8_t end_cyl;            /* end cylinder */
 796         uint32_t start_sect;        /* starting sector counting from 0 */
 797         uint32_t nr_sects;          /* nr of sectors in partition */
 798 } __attribute__((packed));
 799
 800 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 801 static int guess_disk_lchs(BlockDriverState *bs,
 802                            int *pcylinders, int *pheads, int *psectors)
 803 {
 804     uint8_t *buf;
 805     int ret, i, heads, sectors, cylinders;
 806     struct partition *p;
 807     uint32_t nr_sects;
 808     uint64_t nb_sectors;
 809
 810     buf = qemu_memalign(512, 512);
 811     if (buf == NULL)
 812         return -1;
 813
 814     bdrv_get_geometry(bs, &nb_sectors);
 815
 816     ret = bdrv_read(bs, 0, buf, 1);
 817     if (ret < 0)
 818         return -1;
 819     /* test msdos magic */
 820     if (buf[510] != 0x55 || buf[511] != 0xaa) {
 821         qemu_free(buf);
 822         return -1;
 823     }
 824     for(i = 0; i < 4; i++) {
 825         p = ((struct partition *)(buf + 0x1be)) + i;
 826         nr_sects = le32_to_cpu(p->nr_sects);
 827         if (nr_sects && p->end_head) {
 828             /* We make the assumption that the partition terminates on
 829                a cylinder boundary */
 830             heads = p->end_head + 1;
 831             sectors = p->end_sector & 63;
 832             if (sectors == 0)
 833                 continue;
 834             cylinders = nb_sectors / (heads * sectors);
 835             if (cylinders < 1 || cylinders > 16383)
 836                 continue;
 837             *pheads = heads;
 838             *psectors = sectors;
 839             *pcylinders = cylinders;
 840 #if 0
 841             printf("guessed geometry: LCHS=%d %d %d\n",
 842                    cylinders, heads, sectors);
 843 #endif
 844             qemu_free(buf);
 845             return 0;
 846         }
 847     }
 848     qemu_free(buf);
 849     return -1;
 850 }
 851
 852 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 853 {
 854     int translation, lba_detected = 0;
 855     int cylinders, heads, secs;
 856     uint64_t nb_sectors;
 857
 858     /* if a geometry hint is available, use it */
 859     bdrv_get_geometry(bs, &nb_sectors);
 860     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 861     translation = bdrv_get_translation_hint(bs);
 862     if (cylinders != 0) {
 863         *pcyls = cylinders;
 864         *pheads = heads;
 865         *psecs = secs;
 866     } else {
 867         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 868             if (heads > 16) {
 869                 /* if heads > 16, it means that a BIOS LBA
 870                    translation was active, so the default
 871                    hardware geometry is OK */
 872                 lba_detected = 1;
 873                 goto default_geometry;
 874             } else {
 875                 *pcyls = cylinders;
 876                 *pheads = heads;
 877                 *psecs = secs;
 878                 /* disable any translation to be in sync with
 879                    the logical geometry */
 880                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 881                     bdrv_set_translation_hint(bs,
 882                                               BIOS_ATA_TRANSLATION_NONE);
 883                 }
 884             }
 885         } else {
 886         default_geometry:
 887             /* if no geometry, use a standard physical disk geometry */
 888             cylinders = nb_sectors / (16 * 63);
 889
 890             if (cylinders > 16383)
 891                 cylinders = 16383;
 892             else if (cylinders < 2)
 893                 cylinders = 2;
 894             *pcyls = cylinders;
 895             *pheads = 16;
 896             *psecs = 63;
 897             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 898                 if ((*pcyls * *pheads) <= 131072) {
 899                     bdrv_set_translation_hint(bs,
 900                                               BIOS_ATA_TRANSLATION_LARGE);
 901                 } else {
 902                     bdrv_set_translation_hint(bs,
 903                                               BIOS_ATA_TRANSLATION_LBA);
 904                 }
 905             }
 906         }
 907         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 908     }
 909 }
 910
 911 void bdrv_set_geometry_hint(BlockDriverState *bs,
 912                             int cyls, int heads, int secs)
 913 {
 914     bs->cyls = cyls;
 915     bs->heads = heads;
 916     bs->secs = secs;
 917 }
 918
 919 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 920 {
 921     bs->type = type;
 922     bs->removable = ((type == BDRV_TYPE_CDROM ||
 923                       type == BDRV_TYPE_FLOPPY));
 924 }
 925
 926 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 927 {
 928     bs->translation = translation;
 929 }
 930
 931 void bdrv_get_geometry_hint(BlockDriverState *bs,
 932                             int *pcyls, int *pheads, int *psecs)
 933 {
 934     *pcyls = bs->cyls;
 935     *pheads = bs->heads;
 936     *psecs = bs->secs;
 937 }
 938
 939 int bdrv_get_type_hint(BlockDriverState *bs)
 940 {
 941     return bs->type;
 942 }
 943
 944 int bdrv_get_translation_hint(BlockDriverState *bs)
 945 {
 946     return bs->translation;
 947 }
 948
 949 int bdrv_is_removable(BlockDriverState *bs)
 950 {
 951     return bs->removable;
 952 }
 953
 954 int bdrv_is_read_only(BlockDriverState *bs)
 955 {
 956     return bs->read_only;
 957 }
 958
 959 int bdrv_is_sg(BlockDriverState *bs)
 960 {
 961     return bs->sg;
 962 }
 963
 964 /* XXX: no longer used */
 965 void bdrv_set_change_cb(BlockDriverState *bs,
 966                         void (*change_cb)(void *opaque), void *opaque)
 967 {
 968     bs->change_cb = change_cb;
 969     bs->change_opaque = opaque;
 970 }
 971
 972 int bdrv_is_encrypted(BlockDriverState *bs)
 973 {
 974     if (bs->backing_hd && bs->backing_hd->encrypted)
 975         return 1;
 976     return bs->encrypted;
 977 }
 978
 979 int bdrv_key_required(BlockDriverState *bs)
 980 {
 981     BlockDriverState *backing_hd = bs->backing_hd;
 982
 983     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 984         return 1;
 985     return (bs->encrypted && !bs->valid_key);
 986 }
 987
 988 int bdrv_set_key(BlockDriverState *bs, const char *key)
 989 {
 990     int ret;
 991     if (bs->backing_hd && bs->backing_hd->encrypted) {
 992         ret = bdrv_set_key(bs->backing_hd, key);
 993         if (ret < 0)
 994             return ret;
 995         if (!bs->encrypted)
 996             return 0;
 997     }
 998     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 999         return -1;
1000     ret = bs->drv->bdrv_set_key(bs, key);
1001     bs->valid_key = (ret == 0);
1002     return ret;
1003 }
1004
1005 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
1006 {
1007     if (!bs->drv) {
1008         buf[0] = '\0';
1009     } else {
1010         pstrcpy(buf, buf_size, bs->drv->format_name);
1011     }
1012 }
1013
1014 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
1015                          void *opaque)
1016 {
1017     BlockDriver *drv;
1018
1019     for (drv = first_drv; drv != NULL; drv = drv->next) {
1020         it(opaque, drv->format_name);
1021     }
1022 }
1023
1024 BlockDriverState *bdrv_find(const char *name)
1025 {
1026     BlockDriverState *bs;
1027
1028     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1029         if (!strcmp(name, bs->device_name))
1030             return bs;
1031     }
1032     return NULL;
1033 }
1034
1035 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1036 {
1037     BlockDriverState *bs;
1038
1039     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1040         it(opaque, bs);
1041     }
1042 }
1043
1044 const char *bdrv_get_device_name(BlockDriverState *bs)
1045 {
1046     return bs->device_name;
1047 }
1048
1049 void bdrv_flush(BlockDriverState *bs)
1050 {
1051     if (bs->drv->bdrv_flush)
1052         bs->drv->bdrv_flush(bs);
1053     if (bs->backing_hd)
1054         bdrv_flush(bs->backing_hd);
1055 }
1056
1057 void bdrv_flush_all(void)
1058 {
1059     BlockDriverState *bs;
1060
1061     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1062         if (bs->drv && !bdrv_is_read_only(bs) &&
1063             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1064             bdrv_flush(bs);
1065 }
1066
1067 /*
1068  * Returns true iff the specified sector is present in the disk image. Drivers
1069  * not implementing the functionality are assumed to not support backing files,
1070  * hence all their sectors are reported as allocated.
1071  *
1072  * 'pnum' is set to the number of sectors (including and immediately following
1073  * the specified sector) that are known to be in the same
1074  * allocated/unallocated state.
1075  *
1076  * 'nb_sectors' is the max value 'pnum' should be set to.
1077  */
1078 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1079         int *pnum)
1080 {
1081     int64_t n;
1082     if (!bs->drv->bdrv_is_allocated) {
1083         if (sector_num >= bs->total_sectors) {
1084             *pnum = 0;
1085             return 0;
1086         }
1087         n = bs->total_sectors - sector_num;
1088         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1089         return 1;
1090     }
1091     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1092 }
1093
1094 void bdrv_info(void)
1095 {
1096     BlockDriverState *bs;
1097
1098     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1099         term_printf("%s:", bs->device_name);
1100         term_printf(" type=");
1101         switch(bs->type) {
1102         case BDRV_TYPE_HD:
1103             term_printf("hd");
1104             break;
1105         case BDRV_TYPE_CDROM:
1106             term_printf("cdrom");
1107             break;
1108         case BDRV_TYPE_FLOPPY:
1109             term_printf("floppy");
1110             break;
1111         }
1112         term_printf(" removable=%d", bs->removable);
1113         if (bs->removable) {
1114             term_printf(" locked=%d", bs->locked);
1115         }
1116         if (bs->drv) {
1117             term_printf(" file=");
1118             term_print_filename(bs->filename);
1119             if (bs->backing_file[0] != '\0') {
1120                 term_printf(" backing_file=");
1121                 term_print_filename(bs->backing_file);
1122             }
1123             term_printf(" ro=%d", bs->read_only);
1124             term_printf(" drv=%s", bs->drv->format_name);
1125             term_printf(" encrypted=%d", bdrv_is_encrypted(bs));
1126         } else {
1127             term_printf(" [not inserted]");
1128         }
1129         term_printf("\n");
1130     }
1131 }
1132
1133 /* The "info blockstats" command. */
1134 void bdrv_info_stats (void)
1135 {
1136     BlockDriverState *bs;
1137
1138     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1139         term_printf ("%s:"
1140                      " rd_bytes=%" PRIu64
1141                      " wr_bytes=%" PRIu64
1142                      " rd_operations=%" PRIu64
1143                      " wr_operations=%" PRIu64
1144                      "\n",
1145                      bs->device_name,
1146                      bs->rd_bytes, bs->wr_bytes,
1147                      bs->rd_ops, bs->wr_ops);
1148     }
1149 }
1150
1151 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1152 {
1153     if (bs->backing_hd && bs->backing_hd->encrypted)
1154         return bs->backing_file;
1155     else if (bs->encrypted)
1156         return bs->filename;
1157     else
1158         return NULL;
1159 }
1160
1161 void bdrv_get_backing_filename(BlockDriverState *bs,
1162                                char *filename, int filename_size)
1163 {
1164     if (!bs->backing_hd) {
1165         pstrcpy(filename, filename_size, "");
1166     } else {
1167         pstrcpy(filename, filename_size, bs->backing_file);
1168     }
1169 }
1170
1171 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1172                           const uint8_t *buf, int nb_sectors)
1173 {
1174     BlockDriver *drv = bs->drv;
1175     if (!drv)
1176         return -ENOMEDIUM;
1177     if (!drv->bdrv_write_compressed)
1178         return -ENOTSUP;
1179     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1180 }
1181
1182 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1183 {
1184     BlockDriver *drv = bs->drv;
1185     if (!drv)
1186         return -ENOMEDIUM;
1187     if (!drv->bdrv_get_info)
1188         return -ENOTSUP;
1189     memset(bdi, 0, sizeof(*bdi));
1190     return drv->bdrv_get_info(bs, bdi);
1191 }
1192
1193 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1194 {
1195     BlockDriver *drv = bs->drv;
1196     if (!drv)
1197         return -ENOMEDIUM;
1198     if (!drv->bdrv_put_buffer)
1199         return -ENOTSUP;
1200     return drv->bdrv_put_buffer(bs, buf, pos, size);
1201 }
1202
1203 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1204 {
1205     BlockDriver *drv = bs->drv;
1206     if (!drv)
1207         return -ENOMEDIUM;
1208     if (!drv->bdrv_get_buffer)
1209         return -ENOTSUP;
1210     return drv->bdrv_get_buffer(bs, buf, pos, size);
1211 }
1212
1213 /**************************************************************/
1214 /* handling of snapshots */
1215
1216 int bdrv_snapshot_create(BlockDriverState *bs,
1217                          QEMUSnapshotInfo *sn_info)
1218 {
1219     BlockDriver *drv = bs->drv;
1220     if (!drv)
1221         return -ENOMEDIUM;
1222     if (!drv->bdrv_snapshot_create)
1223         return -ENOTSUP;
1224     return drv->bdrv_snapshot_create(bs, sn_info);
1225 }
1226
1227 int bdrv_snapshot_goto(BlockDriverState *bs,
1228                        const char *snapshot_id)
1229 {
1230     BlockDriver *drv = bs->drv;
1231     if (!drv)
1232         return -ENOMEDIUM;
1233     if (!drv->bdrv_snapshot_goto)
1234         return -ENOTSUP;
1235     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1236 }
1237
1238 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1239 {
1240     BlockDriver *drv = bs->drv;
1241     if (!drv)
1242         return -ENOMEDIUM;
1243     if (!drv->bdrv_snapshot_delete)
1244         return -ENOTSUP;
1245     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1246 }
1247
1248 int bdrv_snapshot_list(BlockDriverState *bs,
1249                        QEMUSnapshotInfo **psn_info)
1250 {
1251     BlockDriver *drv = bs->drv;
1252     if (!drv)
1253         return -ENOMEDIUM;
1254     if (!drv->bdrv_snapshot_list)
1255         return -ENOTSUP;
1256     return drv->bdrv_snapshot_list(bs, psn_info);
1257 }
1258
1259 #define NB_SUFFIXES 4
1260
1261 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1262 {
1263     static const char suffixes[NB_SUFFIXES] = "KMGT";
1264     int64_t base;
1265     int i;
1266
1267     if (size <= 999) {
1268         snprintf(buf, buf_size, "%" PRId64, size);
1269     } else {
1270         base = 1024;
1271         for(i = 0; i < NB_SUFFIXES; i++) {
1272             if (size < (10 * base)) {
1273                 snprintf(buf, buf_size, "%0.1f%c",
1274                          (double)size / base,
1275                          suffixes[i]);
1276                 break;
1277             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1278                 snprintf(buf, buf_size, "%" PRId64 "%c",
1279                          ((size + (base >> 1)) / base),
1280                          suffixes[i]);
1281                 break;
1282             }
1283             base = base * 1024;
1284         }
1285     }
1286     return buf;
1287 }
1288
1289 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1290 {
1291     char buf1[128], date_buf[128], clock_buf[128];
1292 #ifdef _WIN32
1293     struct tm *ptm;
1294 #else
1295     struct tm tm;
1296 #endif
1297     time_t ti;
1298     int64_t secs;
1299
1300     if (!sn) {
1301         snprintf(buf, buf_size,
1302                  "%-10s%-20s%7s%20s%15s",
1303                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1304     } else {
1305         ti = sn->date_sec;
1306 #ifdef _WIN32
1307         ptm = localtime(&ti);
1308         strftime(date_buf, sizeof(date_buf),
1309                  "%Y-%m-%d %H:%M:%S", ptm);
1310 #else
1311         localtime_r(&ti, &tm);
1312         strftime(date_buf, sizeof(date_buf),
1313                  "%Y-%m-%d %H:%M:%S", &tm);
1314 #endif
1315         secs = sn->vm_clock_nsec / 1000000000;
1316         snprintf(clock_buf, sizeof(clock_buf),
1317                  "%02d:%02d:%02d.%03d",
1318                  (int)(secs / 3600),
1319                  (int)((secs / 60) % 60),
1320                  (int)(secs % 60),
1321                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1322         snprintf(buf, buf_size,
1323                  "%-10s%-20s%7s%20s%15s",
1324                  sn->id_str, sn->name,
1325                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1326                  date_buf,
1327                  clock_buf);
1328     }
1329     return buf;
1330 }
1331
1332
1333 /**************************************************************/
1334 /* async I/Os */
1335
1336 typedef struct VectorTranslationState {
1337     QEMUIOVector *iov;
1338     uint8_t *bounce;
1339     int is_write;
1340     BlockDriverAIOCB *aiocb;
1341     BlockDriverAIOCB *this_aiocb;
1342 } VectorTranslationState;
1343
1344 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1345 {
1346     VectorTranslationState *s = opaque;
1347
1348     if (!s->is_write) {
1349         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1350     }
1351     qemu_vfree(s->bounce);
1352     s->this_aiocb->cb(s->this_aiocb->opaque, ret);
1353     qemu_aio_release(s->this_aiocb);
1354 }
1355
1356 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1357                                             int64_t sector_num,
1358                                             QEMUIOVector *iov,
1359                                             int nb_sectors,
1360                                             BlockDriverCompletionFunc *cb,
1361                                             void *opaque,
1362                                             int is_write)
1363
1364 {
1365     VectorTranslationState *s = qemu_mallocz(sizeof(*s));
1366     BlockDriverAIOCB *aiocb = qemu_aio_get(bs, cb, opaque);
1367
1368     s->this_aiocb = aiocb;
1369     s->iov = iov;
1370     s->bounce = qemu_memalign(512, nb_sectors * 512);
1371     s->is_write = is_write;
1372     if (is_write) {
1373         qemu_iovec_to_buffer(s->iov, s->bounce);
1374         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1375                                   bdrv_aio_rw_vector_cb, s);
1376     } else {
1377         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1378                                  bdrv_aio_rw_vector_cb, s);
1379     }
1380     return aiocb;
1381 }
1382
1383 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1384                                  QEMUIOVector *iov, int nb_sectors,
1385                                  BlockDriverCompletionFunc *cb, void *opaque)
1386 {
1387     if (bdrv_check_request(bs, sector_num, nb_sectors))
1388         return NULL;
1389
1390     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1391                               cb, opaque, 0);
1392 }
1393
1394 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1395                                   QEMUIOVector *iov, int nb_sectors,
1396                                   BlockDriverCompletionFunc *cb, void *opaque)
1397 {
1398     if (bdrv_check_request(bs, sector_num, nb_sectors))
1399         return NULL;
1400
1401     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1402                               cb, opaque, 1);
1403 }
1404
1405 BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1406                                 uint8_t *buf, int nb_sectors,
1407                                 BlockDriverCompletionFunc *cb, void *opaque)
1408 {
1409     BlockDriver *drv = bs->drv;
1410     BlockDriverAIOCB *ret;
1411
1412     if (!drv)
1413         return NULL;
1414     if (bdrv_check_request(bs, sector_num, nb_sectors))
1415         return NULL;
1416
1417     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1418
1419     if (ret) {
1420         /* Update stats even though technically transfer has not happened. */
1421         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1422         bs->rd_ops ++;
1423     }
1424
1425     return ret;
1426 }
1427
1428 BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1429                                  const uint8_t *buf, int nb_sectors,
1430                                  BlockDriverCompletionFunc *cb, void *opaque)
1431 {
1432     BlockDriver *drv = bs->drv;
1433     BlockDriverAIOCB *ret;
1434
1435     if (!drv)
1436         return NULL;
1437     if (bs->read_only)
1438         return NULL;
1439     if (bdrv_check_request(bs, sector_num, nb_sectors))
1440         return NULL;
1441
1442     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1443
1444     if (ret) {
1445         /* Update stats even though technically transfer has not happened. */
1446         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1447         bs->wr_ops ++;
1448     }
1449
1450     return ret;
1451 }
1452
1453 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1454 {
1455     BlockDriver *drv = acb->bs->drv;
1456
1457     if (acb->cb == bdrv_aio_rw_vector_cb) {
1458         VectorTranslationState *s = acb->opaque;
1459         acb = s->aiocb;
1460     }
1461
1462     drv->bdrv_aio_cancel(acb);
1463 }
1464
1465
1466 /**************************************************************/
1467 /* async block device emulation */
1468
1469 static void bdrv_aio_bh_cb(void *opaque)
1470 {
1471     BlockDriverAIOCBSync *acb = opaque;
1472     acb->common.cb(acb->common.opaque, acb->ret);
1473     qemu_aio_release(acb);
1474 }
1475
1476 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1477         int64_t sector_num, uint8_t *buf, int nb_sectors,
1478         BlockDriverCompletionFunc *cb, void *opaque)
1479 {
1480     BlockDriverAIOCBSync *acb;
1481     int ret;
1482
1483     acb = qemu_aio_get(bs, cb, opaque);
1484     if (!acb->bh)
1485         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1486     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1487     acb->ret = ret;
1488     qemu_bh_schedule(acb->bh);
1489     return &acb->common;
1490 }
1491
1492 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1493         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1494         BlockDriverCompletionFunc *cb, void *opaque)
1495 {
1496     BlockDriverAIOCBSync *acb;
1497     int ret;
1498
1499     acb = qemu_aio_get(bs, cb, opaque);
1500     if (!acb->bh)
1501         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1502     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1503     acb->ret = ret;
1504     qemu_bh_schedule(acb->bh);
1505     return &acb->common;
1506 }
1507
1508 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1509 {
1510     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1511     qemu_bh_cancel(acb->bh);
1512     qemu_aio_release(acb);
1513 }
1514
1515 /**************************************************************/
1516 /* sync block device emulation */
1517
1518 static void bdrv_rw_em_cb(void *opaque, int ret)
1519 {
1520     *(int *)opaque = ret;
1521 }
1522
1523 #define NOT_DONE 0x7fffffff
1524
1525 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1526                         uint8_t *buf, int nb_sectors)
1527 {
1528     int async_ret;
1529     BlockDriverAIOCB *acb;
1530
1531     async_ret = NOT_DONE;
1532     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1533                         bdrv_rw_em_cb, &async_ret);
1534     if (acb == NULL)
1535         return -1;
1536
1537     while (async_ret == NOT_DONE) {
1538         qemu_aio_wait();
1539     }
1540
1541     return async_ret;
1542 }
1543
1544 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1545                          const uint8_t *buf, int nb_sectors)
1546 {
1547     int async_ret;
1548     BlockDriverAIOCB *acb;
1549
1550     async_ret = NOT_DONE;
1551     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1552                          bdrv_rw_em_cb, &async_ret);
1553     if (acb == NULL)
1554         return -1;
1555     while (async_ret == NOT_DONE) {
1556         qemu_aio_wait();
1557     }
1558     return async_ret;
1559 }
1560
1561 void bdrv_init(void)
1562 {
1563     bdrv_register(&bdrv_raw);
1564     bdrv_register(&bdrv_host_device);
1565 #ifndef _WIN32
1566     bdrv_register(&bdrv_cow);
1567 #endif
1568     bdrv_register(&bdrv_qcow);
1569     bdrv_register(&bdrv_vmdk);
1570     bdrv_register(&bdrv_cloop);
1571     bdrv_register(&bdrv_dmg);
1572     bdrv_register(&bdrv_bochs);
1573     bdrv_register(&bdrv_vpc);
1574     bdrv_register(&bdrv_vvfat);
1575     bdrv_register(&bdrv_qcow2);
1576     bdrv_register(&bdrv_parallels);
1577     bdrv_register(&bdrv_nbd);
1578 }
1579
1580 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1581                    void *opaque)
1582 {
1583     BlockDriver *drv;
1584     BlockDriverAIOCB *acb;
1585
1586     drv = bs->drv;
1587     if (drv->free_aiocb) {
1588         acb = drv->free_aiocb;
1589         drv->free_aiocb = acb->next;
1590     } else {
1591         acb = qemu_mallocz(drv->aiocb_size);
1592     }
1593     acb->bs = bs;
1594     acb->cb = cb;
1595     acb->opaque = opaque;
1596     return acb;
1597 }
1598
1599 void qemu_aio_release(void *p)
1600 {
1601     BlockDriverAIOCB *acb = p;
1602     BlockDriver *drv = acb->bs->drv;
1603     acb->next = drv->free_aiocb;
1604     drv->free_aiocb = acb;
1605 }
1606
1607 /**************************************************************/
1608 /* removable device support */
1609
1610 /**
1611  * Return TRUE if the media is present
1612  */
1613 int bdrv_is_inserted(BlockDriverState *bs)
1614 {
1615     BlockDriver *drv = bs->drv;
1616     int ret;
1617     if (!drv)
1618         return 0;
1619     if (!drv->bdrv_is_inserted)
1620         return 1;
1621     ret = drv->bdrv_is_inserted(bs);
1622     return ret;
1623 }
1624
1625 /**
1626  * Return TRUE if the media changed since the last call to this
1627  * function. It is currently only used for floppy disks
1628  */
1629 int bdrv_media_changed(BlockDriverState *bs)
1630 {
1631     BlockDriver *drv = bs->drv;
1632     int ret;
1633
1634     if (!drv || !drv->bdrv_media_changed)
1635         ret = -ENOTSUP;
1636     else
1637         ret = drv->bdrv_media_changed(bs);
1638     if (ret == -ENOTSUP)
1639         ret = bs->media_changed;
1640     bs->media_changed = 0;
1641     return ret;
1642 }
1643
1644 /**
1645  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1646  */
1647 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1648 {
1649     BlockDriver *drv = bs->drv;
1650     int ret;
1651
1652     if (!drv || !drv->bdrv_eject) {
1653         ret = -ENOTSUP;
1654     } else {
1655         ret = drv->bdrv_eject(bs, eject_flag);
1656     }
1657     if (ret == -ENOTSUP) {
1658         if (eject_flag)
1659             bdrv_close(bs);
1660     }
1661 }
1662
1663 int bdrv_is_locked(BlockDriverState *bs)
1664 {
1665     return bs->locked;
1666 }
1667
1668 /**
1669  * Lock or unlock the media (if it is locked, the user won't be able
1670  * to eject it manually).
1671  */
1672 void bdrv_set_locked(BlockDriverState *bs, int locked)
1673 {
1674     BlockDriver *drv = bs->drv;
1675
1676     bs->locked = locked;
1677     if (drv && drv->bdrv_set_locked) {
1678         drv->bdrv_set_locked(bs, locked);
1679     }
1680 }
1681
1682 /* needed for generic scsi interface */
1683
1684 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1685 {
1686     BlockDriver *drv = bs->drv;
1687
1688     if (drv && drv->bdrv_ioctl)
1689         return drv->bdrv_ioctl(bs, req, buf);
1690     return -ENOTSUP;
1691 }