git.maemo.org Git - qemu/blob - 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 /*
 253  * Detect host devices. By convention, /dev/cdrom[N] is always
 254  * recognized as a host CDROM.
 255  */
 256 #ifdef _WIN32
 257 static BlockDriver *find_hdev_driver(const char *filename)
 258 {
 259     if (strstart(filename, "/dev/cdrom", NULL))
 260         return bdrv_find_format("host_device");
 261     if (is_windows_drive(filename))
 262         return bdrv_find_format("host_device");
 263     return NULL;
 264 }
 265 #else
 266 static BlockDriver *find_hdev_driver(const char *filename)
 267 {
 268     struct stat st;
 269
 270 #ifdef __linux__
 271     if (strstart(filename, "/dev/fd", NULL))
 272         return bdrv_find_format("host_floppy");
 273     if (strstart(filename, "/dev/cd", NULL))
 274         return bdrv_find_format("host_cdrom");
 275 #elif defined(__FreeBSD__)
 276     if (strstart(filename, "/dev/cd", NULL) ||
 277         strstart(filename, "/dev/acd", NULL)) {
 278         return bdrv_find_format("host_cdrom");
 279     }
 280 #else
 281     if (strstart(filename, "/dev/cdrom", NULL))
 282         return bdrv_find_format("host_device");
 283 #endif
 284
 285     if (stat(filename, &st) >= 0 &&
 286             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 287         return bdrv_find_format("host_device");
 288     }
 289
 290     return NULL;
 291 }
 292 #endif
 293
 294 static BlockDriver *find_image_format(const char *filename)
 295 {
 296     int ret, score, score_max;
 297     BlockDriver *drv1, *drv;
 298     uint8_t buf[2048];
 299     BlockDriverState *bs;
 300
 301     drv = find_protocol(filename);
 302     /* no need to test disk image formats for vvfat */
 303     if (drv && strcmp(drv->format_name, "vvfat") == 0)
 304         return drv;
 305
 306     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 307     if (ret < 0)
 308         return NULL;
 309     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 310     bdrv_delete(bs);
 311     if (ret < 0) {
 312         return NULL;
 313     }
 314
 315     score_max = 0;
 316     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 317         if (drv1->bdrv_probe) {
 318             score = drv1->bdrv_probe(buf, ret, filename);
 319             if (score > score_max) {
 320                 score_max = score;
 321                 drv = drv1;
 322             }
 323         }
 324     }
 325     return drv;
 326 }
 327
 328 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 329 {
 330     BlockDriverState *bs;
 331     int ret;
 332
 333     bs = bdrv_new("");
 334     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 335     if (ret < 0) {
 336         bdrv_delete(bs);
 337         return ret;
 338     }
 339     bs->growable = 1;
 340     *pbs = bs;
 341     return 0;
 342 }
 343
 344 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 345 {
 346     return bdrv_open2(bs, filename, flags, NULL);
 347 }
 348
 349 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 350                BlockDriver *drv)
 351 {
 352     int ret, open_flags;
 353     char tmp_filename[PATH_MAX];
 354     char backing_filename[PATH_MAX];
 355
 356     bs->read_only = 0;
 357     bs->is_temporary = 0;
 358     bs->encrypted = 0;
 359     bs->valid_key = 0;
 360     /* buffer_alignment defaulted to 512, drivers can change this value */
 361     bs->buffer_alignment = 512;
 362
 363     if (flags & BDRV_O_SNAPSHOT) {
 364         BlockDriverState *bs1;
 365         int64_t total_size;
 366         int is_protocol = 0;
 367         BlockDriver *bdrv_qcow2;
 368         QEMUOptionParameter *options;
 369
 370         /* if snapshot, we create a temporary backing file and open it
 371            instead of opening 'filename' directly */
 372
 373         /* if there is a backing file, use it */
 374         bs1 = bdrv_new("");
 375         ret = bdrv_open2(bs1, filename, 0, drv);
 376         if (ret < 0) {
 377             bdrv_delete(bs1);
 378             return ret;
 379         }
 380         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 381
 382         if (bs1->drv && bs1->drv->protocol_name)
 383             is_protocol = 1;
 384
 385         bdrv_delete(bs1);
 386
 387         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 388
 389         /* Real path is meaningless for protocols */
 390         if (is_protocol)
 391             snprintf(backing_filename, sizeof(backing_filename),
 392                      "%s", filename);
 393         else
 394             realpath(filename, backing_filename);
 395
 396         bdrv_qcow2 = bdrv_find_format("qcow2");
 397         options = parse_option_parameters("", bdrv_qcow2->create_options, NULL);
 398
 399         set_option_parameter_int(options, BLOCK_OPT_SIZE, total_size * 512);
 400         set_option_parameter(options, BLOCK_OPT_BACKING_FILE, backing_filename);
 401         if (drv) {
 402             set_option_parameter(options, BLOCK_OPT_BACKING_FMT,
 403                 drv->format_name);
 404         }
 405
 406         ret = bdrv_create(bdrv_qcow2, tmp_filename, options);
 407         if (ret < 0) {
 408             return ret;
 409         }
 410
 411         filename = tmp_filename;
 412         drv = bdrv_qcow2;
 413         bs->is_temporary = 1;
 414     }
 415
 416     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 417     if (flags & BDRV_O_FILE) {
 418         drv = find_protocol(filename);
 419     } else if (!drv) {
 420         drv = find_hdev_driver(filename);
 421         if (!drv) {
 422             drv = find_image_format(filename);
 423         }
 424     }
 425     if (!drv) {
 426         ret = -ENOENT;
 427         goto unlink_and_fail;
 428     }
 429     bs->drv = drv;
 430     bs->opaque = qemu_mallocz(drv->instance_size);
 431     /* Note: for compatibility, we open disk image files as RDWR, and
 432        RDONLY as fallback */
 433     if (!(flags & BDRV_O_FILE))
 434         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 435     else
 436         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 437     ret = drv->bdrv_open(bs, filename, open_flags);
 438     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 439         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 440         bs->read_only = 1;
 441     }
 442     if (ret < 0) {
 443         qemu_free(bs->opaque);
 444         bs->opaque = NULL;
 445         bs->drv = NULL;
 446     unlink_and_fail:
 447         if (bs->is_temporary)
 448             unlink(filename);
 449         return ret;
 450     }
 451     if (drv->bdrv_getlength) {
 452         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 453     }
 454 #ifndef _WIN32
 455     if (bs->is_temporary) {
 456         unlink(filename);
 457     }
 458 #endif
 459     if (bs->backing_file[0] != '\0') {
 460         /* if there is a backing file, use it */
 461         BlockDriver *back_drv = NULL;
 462         bs->backing_hd = bdrv_new("");
 463         path_combine(backing_filename, sizeof(backing_filename),
 464                      filename, bs->backing_file);
 465         if (bs->backing_format[0] != '\0')
 466             back_drv = bdrv_find_format(bs->backing_format);
 467         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 468                          back_drv);
 469         if (ret < 0) {
 470             bdrv_close(bs);
 471             return ret;
 472         }
 473     }
 474
 475     if (!bdrv_key_required(bs)) {
 476         /* call the change callback */
 477         bs->media_changed = 1;
 478         if (bs->change_cb)
 479             bs->change_cb(bs->change_opaque);
 480     }
 481     return 0;
 482 }
 483
 484 void bdrv_close(BlockDriverState *bs)
 485 {
 486     if (bs->drv) {
 487         if (bs->backing_hd)
 488             bdrv_delete(bs->backing_hd);
 489         bs->drv->bdrv_close(bs);
 490         qemu_free(bs->opaque);
 491 #ifdef _WIN32
 492         if (bs->is_temporary) {
 493             unlink(bs->filename);
 494         }
 495 #endif
 496         bs->opaque = NULL;
 497         bs->drv = NULL;
 498
 499         /* call the change callback */
 500         bs->media_changed = 1;
 501         if (bs->change_cb)
 502             bs->change_cb(bs->change_opaque);
 503     }
 504 }
 505
 506 void bdrv_delete(BlockDriverState *bs)
 507 {
 508     BlockDriverState **pbs;
 509
 510     pbs = &bdrv_first;
 511     while (*pbs != bs && *pbs != NULL)
 512         pbs = &(*pbs)->next;
 513     if (*pbs == bs)
 514         *pbs = bs->next;
 515
 516     bdrv_close(bs);
 517     qemu_free(bs);
 518 }
 519
 520 /*
 521  * Run consistency checks on an image
 522  *
 523  * Returns the number of errors or -errno when an internal error occurs
 524  */
 525 int bdrv_check(BlockDriverState *bs)
 526 {
 527     if (bs->drv->bdrv_check == NULL) {
 528         return -ENOTSUP;
 529     }
 530
 531     return bs->drv->bdrv_check(bs);
 532 }
 533
 534 /* commit COW file into the raw image */
 535 int bdrv_commit(BlockDriverState *bs)
 536 {
 537     BlockDriver *drv = bs->drv;
 538     int64_t i, total_sectors;
 539     int n, j;
 540     unsigned char sector[512];
 541
 542     if (!drv)
 543         return -ENOMEDIUM;
 544
 545     if (bs->read_only) {
 546         return -EACCES;
 547     }
 548
 549     if (!bs->backing_hd) {
 550         return -ENOTSUP;
 551     }
 552
 553     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 554     for (i = 0; i < total_sectors;) {
 555         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 556             for(j = 0; j < n; j++) {
 557                 if (bdrv_read(bs, i, sector, 1) != 0) {
 558                     return -EIO;
 559                 }
 560
 561                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 562                     return -EIO;
 563                 }
 564                 i++;
 565             }
 566         } else {
 567             i += n;
 568         }
 569     }
 570
 571     if (drv->bdrv_make_empty)
 572         return drv->bdrv_make_empty(bs);
 573
 574     return 0;
 575 }
 576
 577 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 578                                    size_t size)
 579 {
 580     int64_t len;
 581
 582     if (!bdrv_is_inserted(bs))
 583         return -ENOMEDIUM;
 584
 585     if (bs->growable)
 586         return 0;
 587
 588     len = bdrv_getlength(bs);
 589
 590     if (offset < 0)
 591         return -EIO;
 592
 593     if ((offset > len) || (len - offset < size))
 594         return -EIO;
 595
 596     return 0;
 597 }
 598
 599 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 600                               int nb_sectors)
 601 {
 602     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 603 }
 604
 605 /* return < 0 if error. See bdrv_write() for the return codes */
 606 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 607               uint8_t *buf, int nb_sectors)
 608 {
 609     BlockDriver *drv = bs->drv;
 610
 611     if (!drv)
 612         return -ENOMEDIUM;
 613     if (bdrv_check_request(bs, sector_num, nb_sectors))
 614         return -EIO;
 615
 616     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 617 }
 618
 619 /* Return < 0 if error. Important errors are:
 620   -EIO         generic I/O error (may happen for all errors)
 621   -ENOMEDIUM   No media inserted.
 622   -EINVAL      Invalid sector number or nb_sectors
 623   -EACCES      Trying to write a read-only device
 624 */
 625 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 626                const uint8_t *buf, int nb_sectors)
 627 {
 628     BlockDriver *drv = bs->drv;
 629     if (!bs->drv)
 630         return -ENOMEDIUM;
 631     if (bs->read_only)
 632         return -EACCES;
 633     if (bdrv_check_request(bs, sector_num, nb_sectors))
 634         return -EIO;
 635
 636     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 637 }
 638
 639 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 640                void *buf, int count1)
 641 {
 642     uint8_t tmp_buf[SECTOR_SIZE];
 643     int len, nb_sectors, count;
 644     int64_t sector_num;
 645
 646     count = count1;
 647     /* first read to align to sector start */
 648     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 649     if (len > count)
 650         len = count;
 651     sector_num = offset >> SECTOR_BITS;
 652     if (len > 0) {
 653         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 654             return -EIO;
 655         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 656         count -= len;
 657         if (count == 0)
 658             return count1;
 659         sector_num++;
 660         buf += len;
 661     }
 662
 663     /* read the sectors "in place" */
 664     nb_sectors = count >> SECTOR_BITS;
 665     if (nb_sectors > 0) {
 666         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 667             return -EIO;
 668         sector_num += nb_sectors;
 669         len = nb_sectors << SECTOR_BITS;
 670         buf += len;
 671         count -= len;
 672     }
 673
 674     /* add data from the last sector */
 675     if (count > 0) {
 676         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 677             return -EIO;
 678         memcpy(buf, tmp_buf, count);
 679     }
 680     return count1;
 681 }
 682
 683 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 684                 const void *buf, int count1)
 685 {
 686     uint8_t tmp_buf[SECTOR_SIZE];
 687     int len, nb_sectors, count;
 688     int64_t sector_num;
 689
 690     count = count1;
 691     /* first write to align to sector start */
 692     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 693     if (len > count)
 694         len = count;
 695     sector_num = offset >> SECTOR_BITS;
 696     if (len > 0) {
 697         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 698             return -EIO;
 699         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 700         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 701             return -EIO;
 702         count -= len;
 703         if (count == 0)
 704             return count1;
 705         sector_num++;
 706         buf += len;
 707     }
 708
 709     /* write the sectors "in place" */
 710     nb_sectors = count >> SECTOR_BITS;
 711     if (nb_sectors > 0) {
 712         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 713             return -EIO;
 714         sector_num += nb_sectors;
 715         len = nb_sectors << SECTOR_BITS;
 716         buf += len;
 717         count -= len;
 718     }
 719
 720     /* add data from the last sector */
 721     if (count > 0) {
 722         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 723             return -EIO;
 724         memcpy(tmp_buf, buf, count);
 725         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 726             return -EIO;
 727     }
 728     return count1;
 729 }
 730
 731 /**
 732  * Truncate file to 'offset' bytes (needed only for file protocols)
 733  */
 734 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 735 {
 736     BlockDriver *drv = bs->drv;
 737     if (!drv)
 738         return -ENOMEDIUM;
 739     if (!drv->bdrv_truncate)
 740         return -ENOTSUP;
 741     return drv->bdrv_truncate(bs, offset);
 742 }
 743
 744 /**
 745  * Length of a file in bytes. Return < 0 if error or unknown.
 746  */
 747 int64_t bdrv_getlength(BlockDriverState *bs)
 748 {
 749     BlockDriver *drv = bs->drv;
 750     if (!drv)
 751         return -ENOMEDIUM;
 752     if (!drv->bdrv_getlength) {
 753         /* legacy mode */
 754         return bs->total_sectors * SECTOR_SIZE;
 755     }
 756     return drv->bdrv_getlength(bs);
 757 }
 758
 759 /* return 0 as number of sectors if no device present or error */
 760 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 761 {
 762     int64_t length;
 763     length = bdrv_getlength(bs);
 764     if (length < 0)
 765         length = 0;
 766     else
 767         length = length >> SECTOR_BITS;
 768     *nb_sectors_ptr = length;
 769 }
 770
 771 struct partition {
 772         uint8_t boot_ind;           /* 0x80 - active */
 773         uint8_t head;               /* starting head */
 774         uint8_t sector;             /* starting sector */
 775         uint8_t cyl;                /* starting cylinder */
 776         uint8_t sys_ind;            /* What partition type */
 777         uint8_t end_head;           /* end head */
 778         uint8_t end_sector;         /* end sector */
 779         uint8_t end_cyl;            /* end cylinder */
 780         uint32_t start_sect;        /* starting sector counting from 0 */
 781         uint32_t nr_sects;          /* nr of sectors in partition */
 782 } __attribute__((packed));
 783
 784 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 785 static int guess_disk_lchs(BlockDriverState *bs,
 786                            int *pcylinders, int *pheads, int *psectors)
 787 {
 788     uint8_t buf[512];
 789     int ret, i, heads, sectors, cylinders;
 790     struct partition *p;
 791     uint32_t nr_sects;
 792     uint64_t nb_sectors;
 793
 794     bdrv_get_geometry(bs, &nb_sectors);
 795
 796     ret = bdrv_read(bs, 0, buf, 1);
 797     if (ret < 0)
 798         return -1;
 799     /* test msdos magic */
 800     if (buf[510] != 0x55 || buf[511] != 0xaa)
 801         return -1;
 802     for(i = 0; i < 4; i++) {
 803         p = ((struct partition *)(buf + 0x1be)) + i;
 804         nr_sects = le32_to_cpu(p->nr_sects);
 805         if (nr_sects && p->end_head) {
 806             /* We make the assumption that the partition terminates on
 807                a cylinder boundary */
 808             heads = p->end_head + 1;
 809             sectors = p->end_sector & 63;
 810             if (sectors == 0)
 811                 continue;
 812             cylinders = nb_sectors / (heads * sectors);
 813             if (cylinders < 1 || cylinders > 16383)
 814                 continue;
 815             *pheads = heads;
 816             *psectors = sectors;
 817             *pcylinders = cylinders;
 818 #if 0
 819             printf("guessed geometry: LCHS=%d %d %d\n",
 820                    cylinders, heads, sectors);
 821 #endif
 822             return 0;
 823         }
 824     }
 825     return -1;
 826 }
 827
 828 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 829 {
 830     int translation, lba_detected = 0;
 831     int cylinders, heads, secs;
 832     uint64_t nb_sectors;
 833
 834     /* if a geometry hint is available, use it */
 835     bdrv_get_geometry(bs, &nb_sectors);
 836     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 837     translation = bdrv_get_translation_hint(bs);
 838     if (cylinders != 0) {
 839         *pcyls = cylinders;
 840         *pheads = heads;
 841         *psecs = secs;
 842     } else {
 843         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 844             if (heads > 16) {
 845                 /* if heads > 16, it means that a BIOS LBA
 846                    translation was active, so the default
 847                    hardware geometry is OK */
 848                 lba_detected = 1;
 849                 goto default_geometry;
 850             } else {
 851                 *pcyls = cylinders;
 852                 *pheads = heads;
 853                 *psecs = secs;
 854                 /* disable any translation to be in sync with
 855                    the logical geometry */
 856                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 857                     bdrv_set_translation_hint(bs,
 858                                               BIOS_ATA_TRANSLATION_NONE);
 859                 }
 860             }
 861         } else {
 862         default_geometry:
 863             /* if no geometry, use a standard physical disk geometry */
 864             cylinders = nb_sectors / (16 * 63);
 865
 866             if (cylinders > 16383)
 867                 cylinders = 16383;
 868             else if (cylinders < 2)
 869                 cylinders = 2;
 870             *pcyls = cylinders;
 871             *pheads = 16;
 872             *psecs = 63;
 873             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 874                 if ((*pcyls * *pheads) <= 131072) {
 875                     bdrv_set_translation_hint(bs,
 876                                               BIOS_ATA_TRANSLATION_LARGE);
 877                 } else {
 878                     bdrv_set_translation_hint(bs,
 879                                               BIOS_ATA_TRANSLATION_LBA);
 880                 }
 881             }
 882         }
 883         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 884     }
 885 }
 886
 887 void bdrv_set_geometry_hint(BlockDriverState *bs,
 888                             int cyls, int heads, int secs)
 889 {
 890     bs->cyls = cyls;
 891     bs->heads = heads;
 892     bs->secs = secs;
 893 }
 894
 895 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 896 {
 897     bs->type = type;
 898     bs->removable = ((type == BDRV_TYPE_CDROM ||
 899                       type == BDRV_TYPE_FLOPPY));
 900 }
 901
 902 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 903 {
 904     bs->translation = translation;
 905 }
 906
 907 void bdrv_get_geometry_hint(BlockDriverState *bs,
 908                             int *pcyls, int *pheads, int *psecs)
 909 {
 910     *pcyls = bs->cyls;
 911     *pheads = bs->heads;
 912     *psecs = bs->secs;
 913 }
 914
 915 int bdrv_get_type_hint(BlockDriverState *bs)
 916 {
 917     return bs->type;
 918 }
 919
 920 int bdrv_get_translation_hint(BlockDriverState *bs)
 921 {
 922     return bs->translation;
 923 }
 924
 925 int bdrv_is_removable(BlockDriverState *bs)
 926 {
 927     return bs->removable;
 928 }
 929
 930 int bdrv_is_read_only(BlockDriverState *bs)
 931 {
 932     return bs->read_only;
 933 }
 934
 935 int bdrv_is_sg(BlockDriverState *bs)
 936 {
 937     return bs->sg;
 938 }
 939
 940 /* XXX: no longer used */
 941 void bdrv_set_change_cb(BlockDriverState *bs,
 942                         void (*change_cb)(void *opaque), void *opaque)
 943 {
 944     bs->change_cb = change_cb;
 945     bs->change_opaque = opaque;
 946 }
 947
 948 int bdrv_is_encrypted(BlockDriverState *bs)
 949 {
 950     if (bs->backing_hd && bs->backing_hd->encrypted)
 951         return 1;
 952     return bs->encrypted;
 953 }
 954
 955 int bdrv_key_required(BlockDriverState *bs)
 956 {
 957     BlockDriverState *backing_hd = bs->backing_hd;
 958
 959     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 960         return 1;
 961     return (bs->encrypted && !bs->valid_key);
 962 }
 963
 964 int bdrv_set_key(BlockDriverState *bs, const char *key)
 965 {
 966     int ret;
 967     if (bs->backing_hd && bs->backing_hd->encrypted) {
 968         ret = bdrv_set_key(bs->backing_hd, key);
 969         if (ret < 0)
 970             return ret;
 971         if (!bs->encrypted)
 972             return 0;
 973     }
 974     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 975         return -1;
 976     ret = bs->drv->bdrv_set_key(bs, key);
 977     if (ret < 0) {
 978         bs->valid_key = 0;
 979     } else if (!bs->valid_key) {
 980         bs->valid_key = 1;
 981         /* call the change callback now, we skipped it on open */
 982         bs->media_changed = 1;
 983         if (bs->change_cb)
 984             bs->change_cb(bs->change_opaque);
 985     }
 986     return ret;
 987 }
 988
 989 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 990 {
 991     if (!bs->drv) {
 992         buf[0] = '\0';
 993     } else {
 994         pstrcpy(buf, buf_size, bs->drv->format_name);
 995     }
 996 }
 997
 998 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 999                          void *opaque)
1000 {
1001     BlockDriver *drv;
1002
1003     for (drv = first_drv; drv != NULL; drv = drv->next) {
1004         it(opaque, drv->format_name);
1005     }
1006 }
1007
1008 BlockDriverState *bdrv_find(const char *name)
1009 {
1010     BlockDriverState *bs;
1011
1012     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1013         if (!strcmp(name, bs->device_name))
1014             return bs;
1015     }
1016     return NULL;
1017 }
1018
1019 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1020 {
1021     BlockDriverState *bs;
1022
1023     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1024         it(opaque, bs);
1025     }
1026 }
1027
1028 const char *bdrv_get_device_name(BlockDriverState *bs)
1029 {
1030     return bs->device_name;
1031 }
1032
1033 void bdrv_flush(BlockDriverState *bs)
1034 {
1035     if (!bs->drv)
1036         return;
1037     if (bs->drv->bdrv_flush)
1038         bs->drv->bdrv_flush(bs);
1039     if (bs->backing_hd)
1040         bdrv_flush(bs->backing_hd);
1041 }
1042
1043 void bdrv_flush_all(void)
1044 {
1045     BlockDriverState *bs;
1046
1047     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1048         if (bs->drv && !bdrv_is_read_only(bs) &&
1049             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1050             bdrv_flush(bs);
1051 }
1052
1053 /*
1054  * Returns true iff the specified sector is present in the disk image. Drivers
1055  * not implementing the functionality are assumed to not support backing files,
1056  * hence all their sectors are reported as allocated.
1057  *
1058  * 'pnum' is set to the number of sectors (including and immediately following
1059  * the specified sector) that are known to be in the same
1060  * allocated/unallocated state.
1061  *
1062  * 'nb_sectors' is the max value 'pnum' should be set to.
1063  */
1064 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1065         int *pnum)
1066 {
1067     int64_t n;
1068     if (!bs->drv->bdrv_is_allocated) {
1069         if (sector_num >= bs->total_sectors) {
1070             *pnum = 0;
1071             return 0;
1072         }
1073         n = bs->total_sectors - sector_num;
1074         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1075         return 1;
1076     }
1077     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1078 }
1079
1080 void bdrv_info(Monitor *mon)
1081 {
1082     BlockDriverState *bs;
1083
1084     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1085         monitor_printf(mon, "%s:", bs->device_name);
1086         monitor_printf(mon, " type=");
1087         switch(bs->type) {
1088         case BDRV_TYPE_HD:
1089             monitor_printf(mon, "hd");
1090             break;
1091         case BDRV_TYPE_CDROM:
1092             monitor_printf(mon, "cdrom");
1093             break;
1094         case BDRV_TYPE_FLOPPY:
1095             monitor_printf(mon, "floppy");
1096             break;
1097         }
1098         monitor_printf(mon, " removable=%d", bs->removable);
1099         if (bs->removable) {
1100             monitor_printf(mon, " locked=%d", bs->locked);
1101         }
1102         if (bs->drv) {
1103             monitor_printf(mon, " file=");
1104             monitor_print_filename(mon, bs->filename);
1105             if (bs->backing_file[0] != '\0') {
1106                 monitor_printf(mon, " backing_file=");
1107                 monitor_print_filename(mon, bs->backing_file);
1108             }
1109             monitor_printf(mon, " ro=%d", bs->read_only);
1110             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1111             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1112         } else {
1113             monitor_printf(mon, " [not inserted]");
1114         }
1115         monitor_printf(mon, "\n");
1116     }
1117 }
1118
1119 /* The "info blockstats" command. */
1120 void bdrv_info_stats(Monitor *mon)
1121 {
1122     BlockDriverState *bs;
1123
1124     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1125         monitor_printf(mon, "%s:"
1126                        " rd_bytes=%" PRIu64
1127                        " wr_bytes=%" PRIu64
1128                        " rd_operations=%" PRIu64
1129                        " wr_operations=%" PRIu64
1130                        "\n",
1131                        bs->device_name,
1132                        bs->rd_bytes, bs->wr_bytes,
1133                        bs->rd_ops, bs->wr_ops);
1134     }
1135 }
1136
1137 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1138 {
1139     if (bs->backing_hd && bs->backing_hd->encrypted)
1140         return bs->backing_file;
1141     else if (bs->encrypted)
1142         return bs->filename;
1143     else
1144         return NULL;
1145 }
1146
1147 void bdrv_get_backing_filename(BlockDriverState *bs,
1148                                char *filename, int filename_size)
1149 {
1150     if (!bs->backing_hd) {
1151         pstrcpy(filename, filename_size, "");
1152     } else {
1153         pstrcpy(filename, filename_size, bs->backing_file);
1154     }
1155 }
1156
1157 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1158                           const uint8_t *buf, int nb_sectors)
1159 {
1160     BlockDriver *drv = bs->drv;
1161     if (!drv)
1162         return -ENOMEDIUM;
1163     if (!drv->bdrv_write_compressed)
1164         return -ENOTSUP;
1165     if (bdrv_check_request(bs, sector_num, nb_sectors))
1166         return -EIO;
1167     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1168 }
1169
1170 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1171 {
1172     BlockDriver *drv = bs->drv;
1173     if (!drv)
1174         return -ENOMEDIUM;
1175     if (!drv->bdrv_get_info)
1176         return -ENOTSUP;
1177     memset(bdi, 0, sizeof(*bdi));
1178     return drv->bdrv_get_info(bs, bdi);
1179 }
1180
1181 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1182 {
1183     BlockDriver *drv = bs->drv;
1184     if (!drv)
1185         return -ENOMEDIUM;
1186     if (!drv->bdrv_put_buffer)
1187         return -ENOTSUP;
1188     return drv->bdrv_put_buffer(bs, buf, pos, size);
1189 }
1190
1191 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1192 {
1193     BlockDriver *drv = bs->drv;
1194     if (!drv)
1195         return -ENOMEDIUM;
1196     if (!drv->bdrv_get_buffer)
1197         return -ENOTSUP;
1198     return drv->bdrv_get_buffer(bs, buf, pos, size);
1199 }
1200
1201 /**************************************************************/
1202 /* handling of snapshots */
1203
1204 int bdrv_snapshot_create(BlockDriverState *bs,
1205                          QEMUSnapshotInfo *sn_info)
1206 {
1207     BlockDriver *drv = bs->drv;
1208     if (!drv)
1209         return -ENOMEDIUM;
1210     if (!drv->bdrv_snapshot_create)
1211         return -ENOTSUP;
1212     return drv->bdrv_snapshot_create(bs, sn_info);
1213 }
1214
1215 int bdrv_snapshot_goto(BlockDriverState *bs,
1216                        const char *snapshot_id)
1217 {
1218     BlockDriver *drv = bs->drv;
1219     if (!drv)
1220         return -ENOMEDIUM;
1221     if (!drv->bdrv_snapshot_goto)
1222         return -ENOTSUP;
1223     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1224 }
1225
1226 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1227 {
1228     BlockDriver *drv = bs->drv;
1229     if (!drv)
1230         return -ENOMEDIUM;
1231     if (!drv->bdrv_snapshot_delete)
1232         return -ENOTSUP;
1233     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1234 }
1235
1236 int bdrv_snapshot_list(BlockDriverState *bs,
1237                        QEMUSnapshotInfo **psn_info)
1238 {
1239     BlockDriver *drv = bs->drv;
1240     if (!drv)
1241         return -ENOMEDIUM;
1242     if (!drv->bdrv_snapshot_list)
1243         return -ENOTSUP;
1244     return drv->bdrv_snapshot_list(bs, psn_info);
1245 }
1246
1247 #define NB_SUFFIXES 4
1248
1249 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1250 {
1251     static const char suffixes[NB_SUFFIXES] = "KMGT";
1252     int64_t base;
1253     int i;
1254
1255     if (size <= 999) {
1256         snprintf(buf, buf_size, "%" PRId64, size);
1257     } else {
1258         base = 1024;
1259         for(i = 0; i < NB_SUFFIXES; i++) {
1260             if (size < (10 * base)) {
1261                 snprintf(buf, buf_size, "%0.1f%c",
1262                          (double)size / base,
1263                          suffixes[i]);
1264                 break;
1265             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1266                 snprintf(buf, buf_size, "%" PRId64 "%c",
1267                          ((size + (base >> 1)) / base),
1268                          suffixes[i]);
1269                 break;
1270             }
1271             base = base * 1024;
1272         }
1273     }
1274     return buf;
1275 }
1276
1277 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1278 {
1279     char buf1[128], date_buf[128], clock_buf[128];
1280 #ifdef _WIN32
1281     struct tm *ptm;
1282 #else
1283     struct tm tm;
1284 #endif
1285     time_t ti;
1286     int64_t secs;
1287
1288     if (!sn) {
1289         snprintf(buf, buf_size,
1290                  "%-10s%-20s%7s%20s%15s",
1291                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1292     } else {
1293         ti = sn->date_sec;
1294 #ifdef _WIN32
1295         ptm = localtime(&ti);
1296         strftime(date_buf, sizeof(date_buf),
1297                  "%Y-%m-%d %H:%M:%S", ptm);
1298 #else
1299         localtime_r(&ti, &tm);
1300         strftime(date_buf, sizeof(date_buf),
1301                  "%Y-%m-%d %H:%M:%S", &tm);
1302 #endif
1303         secs = sn->vm_clock_nsec / 1000000000;
1304         snprintf(clock_buf, sizeof(clock_buf),
1305                  "%02d:%02d:%02d.%03d",
1306                  (int)(secs / 3600),
1307                  (int)((secs / 60) % 60),
1308                  (int)(secs % 60),
1309                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1310         snprintf(buf, buf_size,
1311                  "%-10s%-20s%7s%20s%15s",
1312                  sn->id_str, sn->name,
1313                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1314                  date_buf,
1315                  clock_buf);
1316     }
1317     return buf;
1318 }
1319
1320
1321 /**************************************************************/
1322 /* async I/Os */
1323
1324 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1325                                  QEMUIOVector *qiov, int nb_sectors,
1326                                  BlockDriverCompletionFunc *cb, void *opaque)
1327 {
1328     BlockDriver *drv = bs->drv;
1329     BlockDriverAIOCB *ret;
1330
1331     if (!drv)
1332         return NULL;
1333     if (bdrv_check_request(bs, sector_num, nb_sectors))
1334         return NULL;
1335
1336     ret = drv->bdrv_aio_readv(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->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1342         bs->rd_ops ++;
1343     }
1344
1345     return ret;
1346 }
1347
1348 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1349                                   QEMUIOVector *qiov, int nb_sectors,
1350                                   BlockDriverCompletionFunc *cb, void *opaque)
1351 {
1352     BlockDriver *drv = bs->drv;
1353     BlockDriverAIOCB *ret;
1354
1355     if (!drv)
1356         return NULL;
1357     if (bs->read_only)
1358         return NULL;
1359     if (bdrv_check_request(bs, sector_num, nb_sectors))
1360         return NULL;
1361
1362     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1363                                cb, opaque);
1364
1365     if (ret) {
1366         /* Update stats even though technically transfer has not happened. */
1367         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1368         bs->wr_ops ++;
1369     }
1370
1371     return ret;
1372 }
1373
1374 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1375 {
1376     acb->pool->cancel(acb);
1377 }
1378
1379
1380 /**************************************************************/
1381 /* async block device emulation */
1382
1383 typedef struct BlockDriverAIOCBSync {
1384     BlockDriverAIOCB common;
1385     QEMUBH *bh;
1386     int ret;
1387     /* vector translation state */
1388     QEMUIOVector *qiov;
1389     uint8_t *bounce;
1390     int is_write;
1391 } BlockDriverAIOCBSync;
1392
1393 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1394 {
1395     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1396     qemu_bh_cancel(acb->bh);
1397     qemu_aio_release(acb);
1398 }
1399
1400 static AIOPool bdrv_em_aio_pool = {
1401     .aiocb_size         = sizeof(BlockDriverAIOCBSync),
1402     .cancel             = bdrv_aio_cancel_em,
1403 };
1404
1405 static void bdrv_aio_bh_cb(void *opaque)
1406 {
1407     BlockDriverAIOCBSync *acb = opaque;
1408
1409     if (!acb->is_write)
1410         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1411     qemu_vfree(acb->bounce);
1412     acb->common.cb(acb->common.opaque, acb->ret);
1413
1414     qemu_aio_release(acb);
1415 }
1416
1417 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1418                                             int64_t sector_num,
1419                                             QEMUIOVector *qiov,
1420                                             int nb_sectors,
1421                                             BlockDriverCompletionFunc *cb,
1422                                             void *opaque,
1423                                             int is_write)
1424
1425 {
1426     BlockDriverAIOCBSync *acb;
1427
1428     acb = qemu_aio_get(&bdrv_em_aio_pool, bs, cb, opaque);
1429     acb->is_write = is_write;
1430     acb->qiov = qiov;
1431     acb->bounce = qemu_blockalign(bs, qiov->size);
1432
1433     if (!acb->bh)
1434         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1435
1436     if (is_write) {
1437         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1438         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1439     } else {
1440         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1441     }
1442
1443     qemu_bh_schedule(acb->bh);
1444
1445     return &acb->common;
1446 }
1447
1448 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1449         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1450         BlockDriverCompletionFunc *cb, void *opaque)
1451 {
1452     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1453 }
1454
1455 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1456         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1457         BlockDriverCompletionFunc *cb, void *opaque)
1458 {
1459     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1460 }
1461
1462 /**************************************************************/
1463 /* sync block device emulation */
1464
1465 static void bdrv_rw_em_cb(void *opaque, int ret)
1466 {
1467     *(int *)opaque = ret;
1468 }
1469
1470 #define NOT_DONE 0x7fffffff
1471
1472 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1473                         uint8_t *buf, int nb_sectors)
1474 {
1475     int async_ret;
1476     BlockDriverAIOCB *acb;
1477     struct iovec iov;
1478     QEMUIOVector qiov;
1479
1480     async_ret = NOT_DONE;
1481     iov.iov_base = (void *)buf;
1482     iov.iov_len = nb_sectors * 512;
1483     qemu_iovec_init_external(&qiov, &iov, 1);
1484     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1485         bdrv_rw_em_cb, &async_ret);
1486     if (acb == NULL)
1487         return -1;
1488
1489     while (async_ret == NOT_DONE) {
1490         qemu_aio_wait();
1491     }
1492
1493     return async_ret;
1494 }
1495
1496 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1497                          const uint8_t *buf, int nb_sectors)
1498 {
1499     int async_ret;
1500     BlockDriverAIOCB *acb;
1501     struct iovec iov;
1502     QEMUIOVector qiov;
1503
1504     async_ret = NOT_DONE;
1505     iov.iov_base = (void *)buf;
1506     iov.iov_len = nb_sectors * 512;
1507     qemu_iovec_init_external(&qiov, &iov, 1);
1508     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1509         bdrv_rw_em_cb, &async_ret);
1510     if (acb == NULL)
1511         return -1;
1512     while (async_ret == NOT_DONE) {
1513         qemu_aio_wait();
1514     }
1515     return async_ret;
1516 }
1517
1518 void bdrv_init(void)
1519 {
1520     module_call_init(MODULE_INIT_BLOCK);
1521 }
1522
1523 void *qemu_aio_get(AIOPool *pool, BlockDriverState *bs,
1524                    BlockDriverCompletionFunc *cb, void *opaque)
1525 {
1526     BlockDriverAIOCB *acb;
1527
1528     if (pool->free_aiocb) {
1529         acb = pool->free_aiocb;
1530         pool->free_aiocb = acb->next;
1531     } else {
1532         acb = qemu_mallocz(pool->aiocb_size);
1533         acb->pool = pool;
1534     }
1535     acb->bs = bs;
1536     acb->cb = cb;
1537     acb->opaque = opaque;
1538     return acb;
1539 }
1540
1541 void qemu_aio_release(void *p)
1542 {
1543     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1544     AIOPool *pool = acb->pool;
1545     acb->next = pool->free_aiocb;
1546     pool->free_aiocb = acb;
1547 }
1548
1549 /**************************************************************/
1550 /* removable device support */
1551
1552 /**
1553  * Return TRUE if the media is present
1554  */
1555 int bdrv_is_inserted(BlockDriverState *bs)
1556 {
1557     BlockDriver *drv = bs->drv;
1558     int ret;
1559     if (!drv)
1560         return 0;
1561     if (!drv->bdrv_is_inserted)
1562         return 1;
1563     ret = drv->bdrv_is_inserted(bs);
1564     return ret;
1565 }
1566
1567 /**
1568  * Return TRUE if the media changed since the last call to this
1569  * function. It is currently only used for floppy disks
1570  */
1571 int bdrv_media_changed(BlockDriverState *bs)
1572 {
1573     BlockDriver *drv = bs->drv;
1574     int ret;
1575
1576     if (!drv || !drv->bdrv_media_changed)
1577         ret = -ENOTSUP;
1578     else
1579         ret = drv->bdrv_media_changed(bs);
1580     if (ret == -ENOTSUP)
1581         ret = bs->media_changed;
1582     bs->media_changed = 0;
1583     return ret;
1584 }
1585
1586 /**
1587  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1588  */
1589 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1590 {
1591     BlockDriver *drv = bs->drv;
1592     int ret;
1593
1594     if (!drv || !drv->bdrv_eject) {
1595         ret = -ENOTSUP;
1596     } else {
1597         ret = drv->bdrv_eject(bs, eject_flag);
1598     }
1599     if (ret == -ENOTSUP) {
1600         if (eject_flag)
1601             bdrv_close(bs);
1602     }
1603 }
1604
1605 int bdrv_is_locked(BlockDriverState *bs)
1606 {
1607     return bs->locked;
1608 }
1609
1610 /**
1611  * Lock or unlock the media (if it is locked, the user won't be able
1612  * to eject it manually).
1613  */
1614 void bdrv_set_locked(BlockDriverState *bs, int locked)
1615 {
1616     BlockDriver *drv = bs->drv;
1617
1618     bs->locked = locked;
1619     if (drv && drv->bdrv_set_locked) {
1620         drv->bdrv_set_locked(bs, locked);
1621     }
1622 }
1623
1624 /* needed for generic scsi interface */
1625
1626 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1627 {
1628     BlockDriver *drv = bs->drv;
1629
1630     if (drv && drv->bdrv_ioctl)
1631         return drv->bdrv_ioctl(bs, req, buf);
1632     return -ENOTSUP;
1633 }
1634
1635 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1636         unsigned long int req, void *buf,
1637         BlockDriverCompletionFunc *cb, void *opaque)
1638 {
1639     BlockDriver *drv = bs->drv;
1640
1641     if (drv && drv->bdrv_aio_ioctl)
1642         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1643     return NULL;
1644 }
1645
1646 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1647 {
1648     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1649 }