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
  34 #ifdef HOST_BSD
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/ioctl.h>
  38 #ifndef __DragonFly__
  39 #include <sys/disk.h>
  40 #endif
  41 #endif
  42
  43 #ifdef _WIN32
  44 #include <windows.h>
  45 #endif
  46
  47 #define SECTOR_BITS 9
  48 #define SECTOR_SIZE (1 << SECTOR_BITS)
  49
  50 static AIOPool vectored_aio_pool;
  51
  52 typedef struct BlockDriverAIOCBSync {
  53     BlockDriverAIOCB common;
  54     QEMUBH *bh;
  55     int ret;
  56 } BlockDriverAIOCBSync;
  57
  58 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  59         int64_t sector_num, uint8_t *buf, int nb_sectors,
  60         BlockDriverCompletionFunc *cb, void *opaque);
  61 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  62         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  63         BlockDriverCompletionFunc *cb, void *opaque);
  64 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  65 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  66                         uint8_t *buf, int nb_sectors);
  67 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  68                          const uint8_t *buf, int nb_sectors);
  69
  70 BlockDriverState *bdrv_first;
  71
  72 static BlockDriver *first_drv;
  73
  74 int path_is_absolute(const char *path)
  75 {
  76     const char *p;
  77 #ifdef _WIN32
  78     /* specific case for names like: "\\.\d:" */
  79     if (*path == '/' || *path == '\\')
  80         return 1;
  81 #endif
  82     p = strchr(path, ':');
  83     if (p)
  84         p++;
  85     else
  86         p = path;
  87 #ifdef _WIN32
  88     return (*p == '/' || *p == '\\');
  89 #else
  90     return (*p == '/');
  91 #endif
  92 }
  93
  94 /* if filename is absolute, just copy it to dest. Otherwise, build a
  95    path to it by considering it is relative to base_path. URL are
  96    supported. */
  97 void path_combine(char *dest, int dest_size,
  98                   const char *base_path,
  99                   const char *filename)
 100 {
 101     const char *p, *p1;
 102     int len;
 103
 104     if (dest_size <= 0)
 105         return;
 106     if (path_is_absolute(filename)) {
 107         pstrcpy(dest, dest_size, filename);
 108     } else {
 109         p = strchr(base_path, ':');
 110         if (p)
 111             p++;
 112         else
 113             p = base_path;
 114         p1 = strrchr(base_path, '/');
 115 #ifdef _WIN32
 116         {
 117             const char *p2;
 118             p2 = strrchr(base_path, '\\');
 119             if (!p1 || p2 > p1)
 120                 p1 = p2;
 121         }
 122 #endif
 123         if (p1)
 124             p1++;
 125         else
 126             p1 = base_path;
 127         if (p1 > p)
 128             p = p1;
 129         len = p - base_path;
 130         if (len > dest_size - 1)
 131             len = dest_size - 1;
 132         memcpy(dest, base_path, len);
 133         dest[len] = '\0';
 134         pstrcat(dest, dest_size, filename);
 135     }
 136 }
 137
 138
 139 static void bdrv_register(BlockDriver *bdrv)
 140 {
 141     if (!bdrv->bdrv_aio_read) {
 142         /* add AIO emulation layer */
 143         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 144         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 145         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 146         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 147     } else if (!bdrv->bdrv_read) {
 148         /* add synchronous IO emulation layer */
 149         bdrv->bdrv_read = bdrv_read_em;
 150         bdrv->bdrv_write = bdrv_write_em;
 151     }
 152     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 153     bdrv->next = first_drv;
 154     first_drv = bdrv;
 155 }
 156
 157 /* create a new block device (by default it is empty) */
 158 BlockDriverState *bdrv_new(const char *device_name)
 159 {
 160     BlockDriverState **pbs, *bs;
 161
 162     bs = qemu_mallocz(sizeof(BlockDriverState));
 163     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 164     if (device_name[0] != '\0') {
 165         /* insert at the end */
 166         pbs = &bdrv_first;
 167         while (*pbs != NULL)
 168             pbs = &(*pbs)->next;
 169         *pbs = bs;
 170     }
 171     return bs;
 172 }
 173
 174 BlockDriver *bdrv_find_format(const char *format_name)
 175 {
 176     BlockDriver *drv1;
 177     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 178         if (!strcmp(drv1->format_name, format_name))
 179             return drv1;
 180     }
 181     return NULL;
 182 }
 183
 184 int bdrv_create(BlockDriver *drv,
 185                 const char *filename, int64_t size_in_sectors,
 186                 const char *backing_file, int flags)
 187 {
 188     if (!drv->bdrv_create)
 189         return -ENOTSUP;
 190     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 191 }
 192
 193 #ifdef _WIN32
 194 void get_tmp_filename(char *filename, int size)
 195 {
 196     char temp_dir[MAX_PATH];
 197
 198     GetTempPath(MAX_PATH, temp_dir);
 199     GetTempFileName(temp_dir, "qem", 0, filename);
 200 }
 201 #else
 202 void get_tmp_filename(char *filename, int size)
 203 {
 204     int fd;
 205     const char *tmpdir;
 206     /* XXX: race condition possible */
 207     tmpdir = getenv("TMPDIR");
 208     if (!tmpdir)
 209         tmpdir = "/tmp";
 210     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 211     fd = mkstemp(filename);
 212     close(fd);
 213 }
 214 #endif
 215
 216 #ifdef _WIN32
 217 static int is_windows_drive_prefix(const char *filename)
 218 {
 219     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 220              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 221             filename[1] == ':');
 222 }
 223
 224 static int is_windows_drive(const char *filename)
 225 {
 226     if (is_windows_drive_prefix(filename) &&
 227         filename[2] == '\0')
 228         return 1;
 229     if (strstart(filename, "\\\\.\\", NULL) ||
 230         strstart(filename, "//./", NULL))
 231         return 1;
 232     return 0;
 233 }
 234 #endif
 235
 236 static BlockDriver *find_protocol(const char *filename)
 237 {
 238     BlockDriver *drv1;
 239     char protocol[128];
 240     int len;
 241     const char *p;
 242
 243 #ifdef _WIN32
 244     if (is_windows_drive(filename) ||
 245         is_windows_drive_prefix(filename))
 246         return &bdrv_raw;
 247 #endif
 248     p = strchr(filename, ':');
 249     if (!p)
 250         return &bdrv_raw;
 251     len = p - filename;
 252     if (len > sizeof(protocol) - 1)
 253         len = sizeof(protocol) - 1;
 254     memcpy(protocol, filename, len);
 255     protocol[len] = '\0';
 256     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 257         if (drv1->protocol_name &&
 258             !strcmp(drv1->protocol_name, protocol))
 259             return drv1;
 260     }
 261     return NULL;
 262 }
 263
 264 /* XXX: force raw format if block or character device ? It would
 265    simplify the BSD case */
 266 static BlockDriver *find_image_format(const char *filename)
 267 {
 268     int ret, score, score_max;
 269     BlockDriver *drv1, *drv;
 270     uint8_t buf[2048];
 271     BlockDriverState *bs;
 272
 273     /* detect host devices. By convention, /dev/cdrom[N] is always
 274        recognized as a host CDROM */
 275     if (strstart(filename, "/dev/cdrom", NULL))
 276         return &bdrv_host_device;
 277 #ifdef _WIN32
 278     if (is_windows_drive(filename))
 279         return &bdrv_host_device;
 280 #else
 281     {
 282         struct stat st;
 283         if (stat(filename, &st) >= 0 &&
 284             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 285             return &bdrv_host_device;
 286         }
 287     }
 288 #endif
 289
 290     drv = find_protocol(filename);
 291     /* no need to test disk image formats for vvfat */
 292     if (drv == &bdrv_vvfat)
 293         return drv;
 294
 295     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 296     if (ret < 0)
 297         return NULL;
 298     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 299     bdrv_delete(bs);
 300     if (ret < 0) {
 301         return NULL;
 302     }
 303
 304     score_max = 0;
 305     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 306         if (drv1->bdrv_probe) {
 307             score = drv1->bdrv_probe(buf, ret, filename);
 308             if (score > score_max) {
 309                 score_max = score;
 310                 drv = drv1;
 311             }
 312         }
 313     }
 314     return drv;
 315 }
 316
 317 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 318 {
 319     BlockDriverState *bs;
 320     int ret;
 321
 322     bs = bdrv_new("");
 323     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 324     if (ret < 0) {
 325         bdrv_delete(bs);
 326         return ret;
 327     }
 328     bs->growable = 1;
 329     *pbs = bs;
 330     return 0;
 331 }
 332
 333 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 334 {
 335     return bdrv_open2(bs, filename, flags, NULL);
 336 }
 337
 338 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 339                BlockDriver *drv)
 340 {
 341     int ret, open_flags;
 342     char tmp_filename[PATH_MAX];
 343     char backing_filename[PATH_MAX];
 344
 345     bs->read_only = 0;
 346     bs->is_temporary = 0;
 347     bs->encrypted = 0;
 348     bs->valid_key = 0;
 349
 350     if (flags & BDRV_O_SNAPSHOT) {
 351         BlockDriverState *bs1;
 352         int64_t total_size;
 353         int is_protocol = 0;
 354
 355         /* if snapshot, we create a temporary backing file and open it
 356            instead of opening 'filename' directly */
 357
 358         /* if there is a backing file, use it */
 359         bs1 = bdrv_new("");
 360         ret = bdrv_open(bs1, filename, 0);
 361         if (ret < 0) {
 362             bdrv_delete(bs1);
 363             return ret;
 364         }
 365         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 366
 367         if (bs1->drv && bs1->drv->protocol_name)
 368             is_protocol = 1;
 369
 370         bdrv_delete(bs1);
 371
 372         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 373
 374         /* Real path is meaningless for protocols */
 375         if (is_protocol)
 376             snprintf(backing_filename, sizeof(backing_filename),
 377                      "%s", filename);
 378         else
 379             realpath(filename, backing_filename);
 380
 381         ret = bdrv_create(&bdrv_qcow2, tmp_filename,
 382                           total_size, backing_filename, 0);
 383         if (ret < 0) {
 384             return ret;
 385         }
 386         filename = tmp_filename;
 387         bs->is_temporary = 1;
 388     }
 389
 390     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 391     if (flags & BDRV_O_FILE) {
 392         drv = find_protocol(filename);
 393     } else if (!drv) {
 394         drv = find_image_format(filename);
 395     }
 396     if (!drv) {
 397         ret = -ENOENT;
 398         goto unlink_and_fail;
 399     }
 400     bs->drv = drv;
 401     bs->opaque = qemu_mallocz(drv->instance_size);
 402     /* Note: for compatibility, we open disk image files as RDWR, and
 403        RDONLY as fallback */
 404     if (!(flags & BDRV_O_FILE))
 405         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 406     else
 407         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 408     ret = drv->bdrv_open(bs, filename, open_flags);
 409     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 410         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 411         bs->read_only = 1;
 412     }
 413     if (ret < 0) {
 414         qemu_free(bs->opaque);
 415         bs->opaque = NULL;
 416         bs->drv = NULL;
 417     unlink_and_fail:
 418         if (bs->is_temporary)
 419             unlink(filename);
 420         return ret;
 421     }
 422     if (drv->bdrv_getlength) {
 423         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 424     }
 425 #ifndef _WIN32
 426     if (bs->is_temporary) {
 427         unlink(filename);
 428     }
 429 #endif
 430     if (bs->backing_file[0] != '\0') {
 431         /* if there is a backing file, use it */
 432         bs->backing_hd = bdrv_new("");
 433         path_combine(backing_filename, sizeof(backing_filename),
 434                      filename, bs->backing_file);
 435         ret = bdrv_open(bs->backing_hd, backing_filename, open_flags);
 436         if (ret < 0) {
 437             bdrv_close(bs);
 438             return ret;
 439         }
 440     }
 441
 442     if (!bdrv_key_required(bs)) {
 443         /* call the change callback */
 444         bs->media_changed = 1;
 445         if (bs->change_cb)
 446             bs->change_cb(bs->change_opaque);
 447     }
 448     return 0;
 449 }
 450
 451 void bdrv_close(BlockDriverState *bs)
 452 {
 453     if (bs->drv) {
 454         if (bs->backing_hd)
 455             bdrv_delete(bs->backing_hd);
 456         bs->drv->bdrv_close(bs);
 457         qemu_free(bs->opaque);
 458 #ifdef _WIN32
 459         if (bs->is_temporary) {
 460             unlink(bs->filename);
 461         }
 462 #endif
 463         bs->opaque = NULL;
 464         bs->drv = NULL;
 465
 466         /* call the change callback */
 467         bs->media_changed = 1;
 468         if (bs->change_cb)
 469             bs->change_cb(bs->change_opaque);
 470     }
 471 }
 472
 473 void bdrv_delete(BlockDriverState *bs)
 474 {
 475     BlockDriverState **pbs;
 476
 477     pbs = &bdrv_first;
 478     while (*pbs != bs && *pbs != NULL)
 479         pbs = &(*pbs)->next;
 480     if (*pbs == bs)
 481         *pbs = bs->next;
 482
 483     bdrv_close(bs);
 484     qemu_free(bs);
 485 }
 486
 487 /* commit COW file into the raw image */
 488 int bdrv_commit(BlockDriverState *bs)
 489 {
 490     BlockDriver *drv = bs->drv;
 491     int64_t i, total_sectors;
 492     int n, j;
 493     unsigned char sector[512];
 494
 495     if (!drv)
 496         return -ENOMEDIUM;
 497
 498     if (bs->read_only) {
 499         return -EACCES;
 500     }
 501
 502     if (!bs->backing_hd) {
 503         return -ENOTSUP;
 504     }
 505
 506     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 507     for (i = 0; i < total_sectors;) {
 508         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 509             for(j = 0; j < n; j++) {
 510                 if (bdrv_read(bs, i, sector, 1) != 0) {
 511                     return -EIO;
 512                 }
 513
 514                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 515                     return -EIO;
 516                 }
 517                 i++;
 518             }
 519         } else {
 520             i += n;
 521         }
 522     }
 523
 524     if (drv->bdrv_make_empty)
 525         return drv->bdrv_make_empty(bs);
 526
 527     return 0;
 528 }
 529
 530 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 531                                    size_t size)
 532 {
 533     int64_t len;
 534
 535     if (!bdrv_is_inserted(bs))
 536         return -ENOMEDIUM;
 537
 538     if (bs->growable)
 539         return 0;
 540
 541     len = bdrv_getlength(bs);
 542
 543     if ((offset + size) > len)
 544         return -EIO;
 545
 546     return 0;
 547 }
 548
 549 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 550                               int nb_sectors)
 551 {
 552     int64_t offset;
 553
 554     /* Deal with byte accesses */
 555     if (sector_num < 0)
 556         offset = -sector_num;
 557     else
 558         offset = sector_num * 512;
 559
 560     return bdrv_check_byte_request(bs, offset, nb_sectors * 512);
 561 }
 562
 563 /* return < 0 if error. See bdrv_write() for the return codes */
 564 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 565               uint8_t *buf, int nb_sectors)
 566 {
 567     BlockDriver *drv = bs->drv;
 568
 569     if (!drv)
 570         return -ENOMEDIUM;
 571     if (bdrv_check_request(bs, sector_num, nb_sectors))
 572         return -EIO;
 573
 574     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 575 }
 576
 577 /* Return < 0 if error. Important errors are:
 578   -EIO         generic I/O error (may happen for all errors)
 579   -ENOMEDIUM   No media inserted.
 580   -EINVAL      Invalid sector number or nb_sectors
 581   -EACCES      Trying to write a read-only device
 582 */
 583 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 584                const uint8_t *buf, int nb_sectors)
 585 {
 586     BlockDriver *drv = bs->drv;
 587     if (!bs->drv)
 588         return -ENOMEDIUM;
 589     if (bs->read_only)
 590         return -EACCES;
 591     if (bdrv_check_request(bs, sector_num, nb_sectors))
 592         return -EIO;
 593
 594     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 595 }
 596
 597 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 598                void *buf, int count1)
 599 {
 600     uint8_t tmp_buf[SECTOR_SIZE];
 601     int len, nb_sectors, count;
 602     int64_t sector_num;
 603
 604     count = count1;
 605     /* first read to align to sector start */
 606     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 607     if (len > count)
 608         len = count;
 609     sector_num = offset >> SECTOR_BITS;
 610     if (len > 0) {
 611         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 612             return -EIO;
 613         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 614         count -= len;
 615         if (count == 0)
 616             return count1;
 617         sector_num++;
 618         buf += len;
 619     }
 620
 621     /* read the sectors "in place" */
 622     nb_sectors = count >> SECTOR_BITS;
 623     if (nb_sectors > 0) {
 624         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 625             return -EIO;
 626         sector_num += nb_sectors;
 627         len = nb_sectors << SECTOR_BITS;
 628         buf += len;
 629         count -= len;
 630     }
 631
 632     /* add data from the last sector */
 633     if (count > 0) {
 634         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 635             return -EIO;
 636         memcpy(buf, tmp_buf, count);
 637     }
 638     return count1;
 639 }
 640
 641 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 642                 const void *buf, int count1)
 643 {
 644     uint8_t tmp_buf[SECTOR_SIZE];
 645     int len, nb_sectors, count;
 646     int64_t sector_num;
 647
 648     count = count1;
 649     /* first write to align to sector start */
 650     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 651     if (len > count)
 652         len = count;
 653     sector_num = offset >> SECTOR_BITS;
 654     if (len > 0) {
 655         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 656             return -EIO;
 657         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 658         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 659             return -EIO;
 660         count -= len;
 661         if (count == 0)
 662             return count1;
 663         sector_num++;
 664         buf += len;
 665     }
 666
 667     /* write the sectors "in place" */
 668     nb_sectors = count >> SECTOR_BITS;
 669     if (nb_sectors > 0) {
 670         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 671             return -EIO;
 672         sector_num += nb_sectors;
 673         len = nb_sectors << SECTOR_BITS;
 674         buf += len;
 675         count -= len;
 676     }
 677
 678     /* add data from the last sector */
 679     if (count > 0) {
 680         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 681             return -EIO;
 682         memcpy(tmp_buf, buf, count);
 683         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 684             return -EIO;
 685     }
 686     return count1;
 687 }
 688
 689 /**
 690  * Truncate file to 'offset' bytes (needed only for file protocols)
 691  */
 692 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 693 {
 694     BlockDriver *drv = bs->drv;
 695     if (!drv)
 696         return -ENOMEDIUM;
 697     if (!drv->bdrv_truncate)
 698         return -ENOTSUP;
 699     return drv->bdrv_truncate(bs, offset);
 700 }
 701
 702 /**
 703  * Length of a file in bytes. Return < 0 if error or unknown.
 704  */
 705 int64_t bdrv_getlength(BlockDriverState *bs)
 706 {
 707     BlockDriver *drv = bs->drv;
 708     if (!drv)
 709         return -ENOMEDIUM;
 710     if (!drv->bdrv_getlength) {
 711         /* legacy mode */
 712         return bs->total_sectors * SECTOR_SIZE;
 713     }
 714     return drv->bdrv_getlength(bs);
 715 }
 716
 717 /* return 0 as number of sectors if no device present or error */
 718 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 719 {
 720     int64_t length;
 721     length = bdrv_getlength(bs);
 722     if (length < 0)
 723         length = 0;
 724     else
 725         length = length >> SECTOR_BITS;
 726     *nb_sectors_ptr = length;
 727 }
 728
 729 struct partition {
 730         uint8_t boot_ind;           /* 0x80 - active */
 731         uint8_t head;               /* starting head */
 732         uint8_t sector;             /* starting sector */
 733         uint8_t cyl;                /* starting cylinder */
 734         uint8_t sys_ind;            /* What partition type */
 735         uint8_t end_head;           /* end head */
 736         uint8_t end_sector;         /* end sector */
 737         uint8_t end_cyl;            /* end cylinder */
 738         uint32_t start_sect;        /* starting sector counting from 0 */
 739         uint32_t nr_sects;          /* nr of sectors in partition */
 740 } __attribute__((packed));
 741
 742 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 743 static int guess_disk_lchs(BlockDriverState *bs,
 744                            int *pcylinders, int *pheads, int *psectors)
 745 {
 746     uint8_t buf[512];
 747     int ret, i, heads, sectors, cylinders;
 748     struct partition *p;
 749     uint32_t nr_sects;
 750     uint64_t nb_sectors;
 751
 752     bdrv_get_geometry(bs, &nb_sectors);
 753
 754     ret = bdrv_read(bs, 0, buf, 1);
 755     if (ret < 0)
 756         return -1;
 757     /* test msdos magic */
 758     if (buf[510] != 0x55 || buf[511] != 0xaa)
 759         return -1;
 760     for(i = 0; i < 4; i++) {
 761         p = ((struct partition *)(buf + 0x1be)) + i;
 762         nr_sects = le32_to_cpu(p->nr_sects);
 763         if (nr_sects && p->end_head) {
 764             /* We make the assumption that the partition terminates on
 765                a cylinder boundary */
 766             heads = p->end_head + 1;
 767             sectors = p->end_sector & 63;
 768             if (sectors == 0)
 769                 continue;
 770             cylinders = nb_sectors / (heads * sectors);
 771             if (cylinders < 1 || cylinders > 16383)
 772                 continue;
 773             *pheads = heads;
 774             *psectors = sectors;
 775             *pcylinders = cylinders;
 776 #if 0
 777             printf("guessed geometry: LCHS=%d %d %d\n",
 778                    cylinders, heads, sectors);
 779 #endif
 780             return 0;
 781         }
 782     }
 783     return -1;
 784 }
 785
 786 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 787 {
 788     int translation, lba_detected = 0;
 789     int cylinders, heads, secs;
 790     uint64_t nb_sectors;
 791
 792     /* if a geometry hint is available, use it */
 793     bdrv_get_geometry(bs, &nb_sectors);
 794     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 795     translation = bdrv_get_translation_hint(bs);
 796     if (cylinders != 0) {
 797         *pcyls = cylinders;
 798         *pheads = heads;
 799         *psecs = secs;
 800     } else {
 801         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 802             if (heads > 16) {
 803                 /* if heads > 16, it means that a BIOS LBA
 804                    translation was active, so the default
 805                    hardware geometry is OK */
 806                 lba_detected = 1;
 807                 goto default_geometry;
 808             } else {
 809                 *pcyls = cylinders;
 810                 *pheads = heads;
 811                 *psecs = secs;
 812                 /* disable any translation to be in sync with
 813                    the logical geometry */
 814                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 815                     bdrv_set_translation_hint(bs,
 816                                               BIOS_ATA_TRANSLATION_NONE);
 817                 }
 818             }
 819         } else {
 820         default_geometry:
 821             /* if no geometry, use a standard physical disk geometry */
 822             cylinders = nb_sectors / (16 * 63);
 823
 824             if (cylinders > 16383)
 825                 cylinders = 16383;
 826             else if (cylinders < 2)
 827                 cylinders = 2;
 828             *pcyls = cylinders;
 829             *pheads = 16;
 830             *psecs = 63;
 831             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 832                 if ((*pcyls * *pheads) <= 131072) {
 833                     bdrv_set_translation_hint(bs,
 834                                               BIOS_ATA_TRANSLATION_LARGE);
 835                 } else {
 836                     bdrv_set_translation_hint(bs,
 837                                               BIOS_ATA_TRANSLATION_LBA);
 838                 }
 839             }
 840         }
 841         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 842     }
 843 }
 844
 845 void bdrv_set_geometry_hint(BlockDriverState *bs,
 846                             int cyls, int heads, int secs)
 847 {
 848     bs->cyls = cyls;
 849     bs->heads = heads;
 850     bs->secs = secs;
 851 }
 852
 853 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 854 {
 855     bs->type = type;
 856     bs->removable = ((type == BDRV_TYPE_CDROM ||
 857                       type == BDRV_TYPE_FLOPPY));
 858 }
 859
 860 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 861 {
 862     bs->translation = translation;
 863 }
 864
 865 void bdrv_get_geometry_hint(BlockDriverState *bs,
 866                             int *pcyls, int *pheads, int *psecs)
 867 {
 868     *pcyls = bs->cyls;
 869     *pheads = bs->heads;
 870     *psecs = bs->secs;
 871 }
 872
 873 int bdrv_get_type_hint(BlockDriverState *bs)
 874 {
 875     return bs->type;
 876 }
 877
 878 int bdrv_get_translation_hint(BlockDriverState *bs)
 879 {
 880     return bs->translation;
 881 }
 882
 883 int bdrv_is_removable(BlockDriverState *bs)
 884 {
 885     return bs->removable;
 886 }
 887
 888 int bdrv_is_read_only(BlockDriverState *bs)
 889 {
 890     return bs->read_only;
 891 }
 892
 893 int bdrv_is_sg(BlockDriverState *bs)
 894 {
 895     return bs->sg;
 896 }
 897
 898 /* XXX: no longer used */
 899 void bdrv_set_change_cb(BlockDriverState *bs,
 900                         void (*change_cb)(void *opaque), void *opaque)
 901 {
 902     bs->change_cb = change_cb;
 903     bs->change_opaque = opaque;
 904 }
 905
 906 int bdrv_is_encrypted(BlockDriverState *bs)
 907 {
 908     if (bs->backing_hd && bs->backing_hd->encrypted)
 909         return 1;
 910     return bs->encrypted;
 911 }
 912
 913 int bdrv_key_required(BlockDriverState *bs)
 914 {
 915     BlockDriverState *backing_hd = bs->backing_hd;
 916
 917     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 918         return 1;
 919     return (bs->encrypted && !bs->valid_key);
 920 }
 921
 922 int bdrv_set_key(BlockDriverState *bs, const char *key)
 923 {
 924     int ret;
 925     if (bs->backing_hd && bs->backing_hd->encrypted) {
 926         ret = bdrv_set_key(bs->backing_hd, key);
 927         if (ret < 0)
 928             return ret;
 929         if (!bs->encrypted)
 930             return 0;
 931     }
 932     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 933         return -1;
 934     ret = bs->drv->bdrv_set_key(bs, key);
 935     if (ret < 0) {
 936         bs->valid_key = 0;
 937     } else if (!bs->valid_key) {
 938         bs->valid_key = 1;
 939         /* call the change callback now, we skipped it on open */
 940         bs->media_changed = 1;
 941         if (bs->change_cb)
 942             bs->change_cb(bs->change_opaque);
 943     }
 944     return ret;
 945 }
 946
 947 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 948 {
 949     if (!bs->drv) {
 950         buf[0] = '\0';
 951     } else {
 952         pstrcpy(buf, buf_size, bs->drv->format_name);
 953     }
 954 }
 955
 956 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 957                          void *opaque)
 958 {
 959     BlockDriver *drv;
 960
 961     for (drv = first_drv; drv != NULL; drv = drv->next) {
 962         it(opaque, drv->format_name);
 963     }
 964 }
 965
 966 BlockDriverState *bdrv_find(const char *name)
 967 {
 968     BlockDriverState *bs;
 969
 970     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 971         if (!strcmp(name, bs->device_name))
 972             return bs;
 973     }
 974     return NULL;
 975 }
 976
 977 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
 978 {
 979     BlockDriverState *bs;
 980
 981     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 982         it(opaque, bs);
 983     }
 984 }
 985
 986 const char *bdrv_get_device_name(BlockDriverState *bs)
 987 {
 988     return bs->device_name;
 989 }
 990
 991 void bdrv_flush(BlockDriverState *bs)
 992 {
 993     if (bs->drv->bdrv_flush)
 994         bs->drv->bdrv_flush(bs);
 995     if (bs->backing_hd)
 996         bdrv_flush(bs->backing_hd);
 997 }
 998
 999 void bdrv_flush_all(void)
1000 {
1001     BlockDriverState *bs;
1002
1003     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1004         if (bs->drv && !bdrv_is_read_only(bs) &&
1005             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1006             bdrv_flush(bs);
1007 }
1008
1009 /*
1010  * Returns true iff the specified sector is present in the disk image. Drivers
1011  * not implementing the functionality are assumed to not support backing files,
1012  * hence all their sectors are reported as allocated.
1013  *
1014  * 'pnum' is set to the number of sectors (including and immediately following
1015  * the specified sector) that are known to be in the same
1016  * allocated/unallocated state.
1017  *
1018  * 'nb_sectors' is the max value 'pnum' should be set to.
1019  */
1020 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1021         int *pnum)
1022 {
1023     int64_t n;
1024     if (!bs->drv->bdrv_is_allocated) {
1025         if (sector_num >= bs->total_sectors) {
1026             *pnum = 0;
1027             return 0;
1028         }
1029         n = bs->total_sectors - sector_num;
1030         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1031         return 1;
1032     }
1033     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1034 }
1035
1036 void bdrv_info(Monitor *mon)
1037 {
1038     BlockDriverState *bs;
1039
1040     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1041         monitor_printf(mon, "%s:", bs->device_name);
1042         monitor_printf(mon, " type=");
1043         switch(bs->type) {
1044         case BDRV_TYPE_HD:
1045             monitor_printf(mon, "hd");
1046             break;
1047         case BDRV_TYPE_CDROM:
1048             monitor_printf(mon, "cdrom");
1049             break;
1050         case BDRV_TYPE_FLOPPY:
1051             monitor_printf(mon, "floppy");
1052             break;
1053         }
1054         monitor_printf(mon, " removable=%d", bs->removable);
1055         if (bs->removable) {
1056             monitor_printf(mon, " locked=%d", bs->locked);
1057         }
1058         if (bs->drv) {
1059             monitor_printf(mon, " file=");
1060             monitor_print_filename(mon, bs->filename);
1061             if (bs->backing_file[0] != '\0') {
1062                 monitor_printf(mon, " backing_file=");
1063                 monitor_print_filename(mon, bs->backing_file);
1064             }
1065             monitor_printf(mon, " ro=%d", bs->read_only);
1066             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1067             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1068         } else {
1069             monitor_printf(mon, " [not inserted]");
1070         }
1071         monitor_printf(mon, "\n");
1072     }
1073 }
1074
1075 /* The "info blockstats" command. */
1076 void bdrv_info_stats(Monitor *mon)
1077 {
1078     BlockDriverState *bs;
1079
1080     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1081         monitor_printf(mon, "%s:"
1082                        " rd_bytes=%" PRIu64
1083                        " wr_bytes=%" PRIu64
1084                        " rd_operations=%" PRIu64
1085                        " wr_operations=%" PRIu64
1086                        "\n",
1087                        bs->device_name,
1088                        bs->rd_bytes, bs->wr_bytes,
1089                        bs->rd_ops, bs->wr_ops);
1090     }
1091 }
1092
1093 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1094 {
1095     if (bs->backing_hd && bs->backing_hd->encrypted)
1096         return bs->backing_file;
1097     else if (bs->encrypted)
1098         return bs->filename;
1099     else
1100         return NULL;
1101 }
1102
1103 void bdrv_get_backing_filename(BlockDriverState *bs,
1104                                char *filename, int filename_size)
1105 {
1106     if (!bs->backing_hd) {
1107         pstrcpy(filename, filename_size, "");
1108     } else {
1109         pstrcpy(filename, filename_size, bs->backing_file);
1110     }
1111 }
1112
1113 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1114                           const uint8_t *buf, int nb_sectors)
1115 {
1116     BlockDriver *drv = bs->drv;
1117     if (!drv)
1118         return -ENOMEDIUM;
1119     if (!drv->bdrv_write_compressed)
1120         return -ENOTSUP;
1121     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1122 }
1123
1124 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1125 {
1126     BlockDriver *drv = bs->drv;
1127     if (!drv)
1128         return -ENOMEDIUM;
1129     if (!drv->bdrv_get_info)
1130         return -ENOTSUP;
1131     memset(bdi, 0, sizeof(*bdi));
1132     return drv->bdrv_get_info(bs, bdi);
1133 }
1134
1135 /**************************************************************/
1136 /* handling of snapshots */
1137
1138 int bdrv_snapshot_create(BlockDriverState *bs,
1139                          QEMUSnapshotInfo *sn_info)
1140 {
1141     BlockDriver *drv = bs->drv;
1142     if (!drv)
1143         return -ENOMEDIUM;
1144     if (!drv->bdrv_snapshot_create)
1145         return -ENOTSUP;
1146     return drv->bdrv_snapshot_create(bs, sn_info);
1147 }
1148
1149 int bdrv_snapshot_goto(BlockDriverState *bs,
1150                        const char *snapshot_id)
1151 {
1152     BlockDriver *drv = bs->drv;
1153     if (!drv)
1154         return -ENOMEDIUM;
1155     if (!drv->bdrv_snapshot_goto)
1156         return -ENOTSUP;
1157     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1158 }
1159
1160 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1161 {
1162     BlockDriver *drv = bs->drv;
1163     if (!drv)
1164         return -ENOMEDIUM;
1165     if (!drv->bdrv_snapshot_delete)
1166         return -ENOTSUP;
1167     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1168 }
1169
1170 int bdrv_snapshot_list(BlockDriverState *bs,
1171                        QEMUSnapshotInfo **psn_info)
1172 {
1173     BlockDriver *drv = bs->drv;
1174     if (!drv)
1175         return -ENOMEDIUM;
1176     if (!drv->bdrv_snapshot_list)
1177         return -ENOTSUP;
1178     return drv->bdrv_snapshot_list(bs, psn_info);
1179 }
1180
1181 #define NB_SUFFIXES 4
1182
1183 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1184 {
1185     static const char suffixes[NB_SUFFIXES] = "KMGT";
1186     int64_t base;
1187     int i;
1188
1189     if (size <= 999) {
1190         snprintf(buf, buf_size, "%" PRId64, size);
1191     } else {
1192         base = 1024;
1193         for(i = 0; i < NB_SUFFIXES; i++) {
1194             if (size < (10 * base)) {
1195                 snprintf(buf, buf_size, "%0.1f%c",
1196                          (double)size / base,
1197                          suffixes[i]);
1198                 break;
1199             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1200                 snprintf(buf, buf_size, "%" PRId64 "%c",
1201                          ((size + (base >> 1)) / base),
1202                          suffixes[i]);
1203                 break;
1204             }
1205             base = base * 1024;
1206         }
1207     }
1208     return buf;
1209 }
1210
1211 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1212 {
1213     char buf1[128], date_buf[128], clock_buf[128];
1214 #ifdef _WIN32
1215     struct tm *ptm;
1216 #else
1217     struct tm tm;
1218 #endif
1219     time_t ti;
1220     int64_t secs;
1221
1222     if (!sn) {
1223         snprintf(buf, buf_size,
1224                  "%-10s%-20s%7s%20s%15s",
1225                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1226     } else {
1227         ti = sn->date_sec;
1228 #ifdef _WIN32
1229         ptm = localtime(&ti);
1230         strftime(date_buf, sizeof(date_buf),
1231                  "%Y-%m-%d %H:%M:%S", ptm);
1232 #else
1233         localtime_r(&ti, &tm);
1234         strftime(date_buf, sizeof(date_buf),
1235                  "%Y-%m-%d %H:%M:%S", &tm);
1236 #endif
1237         secs = sn->vm_clock_nsec / 1000000000;
1238         snprintf(clock_buf, sizeof(clock_buf),
1239                  "%02d:%02d:%02d.%03d",
1240                  (int)(secs / 3600),
1241                  (int)((secs / 60) % 60),
1242                  (int)(secs % 60),
1243                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1244         snprintf(buf, buf_size,
1245                  "%-10s%-20s%7s%20s%15s",
1246                  sn->id_str, sn->name,
1247                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1248                  date_buf,
1249                  clock_buf);
1250     }
1251     return buf;
1252 }
1253
1254
1255 /**************************************************************/
1256 /* async I/Os */
1257
1258 typedef struct VectorTranslationAIOCB {
1259     BlockDriverAIOCB common;
1260     QEMUIOVector *iov;
1261     uint8_t *bounce;
1262     int is_write;
1263     BlockDriverAIOCB *aiocb;
1264 } VectorTranslationAIOCB;
1265
1266 static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
1267 {
1268     VectorTranslationAIOCB *acb
1269         = container_of(_acb, VectorTranslationAIOCB, common);
1270
1271     bdrv_aio_cancel(acb->aiocb);
1272 }
1273
1274 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1275 {
1276     VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
1277
1278     if (!s->is_write) {
1279         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1280     }
1281     qemu_vfree(s->bounce);
1282     s->common.cb(s->common.opaque, ret);
1283     qemu_aio_release(s);
1284 }
1285
1286 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1287                                             int64_t sector_num,
1288                                             QEMUIOVector *iov,
1289                                             int nb_sectors,
1290                                             BlockDriverCompletionFunc *cb,
1291                                             void *opaque,
1292                                             int is_write)
1293
1294 {
1295     VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
1296                                                   cb, opaque);
1297
1298     s->iov = iov;
1299     s->bounce = qemu_memalign(512, nb_sectors * 512);
1300     s->is_write = is_write;
1301     if (is_write) {
1302         qemu_iovec_to_buffer(s->iov, s->bounce);
1303         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1304                                   bdrv_aio_rw_vector_cb, s);
1305     } else {
1306         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1307                                  bdrv_aio_rw_vector_cb, s);
1308     }
1309     if (!s->aiocb) {
1310         qemu_vfree(s->bounce);
1311         qemu_aio_release(s);
1312         return NULL;
1313     }
1314     return &s->common;
1315 }
1316
1317 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1318                                  QEMUIOVector *iov, int nb_sectors,
1319                                  BlockDriverCompletionFunc *cb, void *opaque)
1320 {
1321     if (bdrv_check_request(bs, sector_num, nb_sectors))
1322         return NULL;
1323
1324     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1325                               cb, opaque, 0);
1326 }
1327
1328 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1329                                   QEMUIOVector *iov, int nb_sectors,
1330                                   BlockDriverCompletionFunc *cb, void *opaque)
1331 {
1332     if (bdrv_check_request(bs, sector_num, nb_sectors))
1333         return NULL;
1334
1335     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1336                               cb, opaque, 1);
1337 }
1338
1339 BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1340                                 uint8_t *buf, int nb_sectors,
1341                                 BlockDriverCompletionFunc *cb, void *opaque)
1342 {
1343     BlockDriver *drv = bs->drv;
1344     BlockDriverAIOCB *ret;
1345
1346     if (!drv)
1347         return NULL;
1348     if (bdrv_check_request(bs, sector_num, nb_sectors))
1349         return NULL;
1350
1351     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1352
1353     if (ret) {
1354         /* Update stats even though technically transfer has not happened. */
1355         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1356         bs->rd_ops ++;
1357     }
1358
1359     return ret;
1360 }
1361
1362 BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1363                                  const uint8_t *buf, int nb_sectors,
1364                                  BlockDriverCompletionFunc *cb, void *opaque)
1365 {
1366     BlockDriver *drv = bs->drv;
1367     BlockDriverAIOCB *ret;
1368
1369     if (!drv)
1370         return NULL;
1371     if (bs->read_only)
1372         return NULL;
1373     if (bdrv_check_request(bs, sector_num, nb_sectors))
1374         return NULL;
1375
1376     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1377
1378     if (ret) {
1379         /* Update stats even though technically transfer has not happened. */
1380         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1381         bs->wr_ops ++;
1382     }
1383
1384     return ret;
1385 }
1386
1387 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1388 {
1389     acb->pool->cancel(acb);
1390 }
1391
1392
1393 /**************************************************************/
1394 /* async block device emulation */
1395
1396 static void bdrv_aio_bh_cb(void *opaque)
1397 {
1398     BlockDriverAIOCBSync *acb = opaque;
1399     acb->common.cb(acb->common.opaque, acb->ret);
1400     qemu_aio_release(acb);
1401 }
1402
1403 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1404         int64_t sector_num, uint8_t *buf, int nb_sectors,
1405         BlockDriverCompletionFunc *cb, void *opaque)
1406 {
1407     BlockDriverAIOCBSync *acb;
1408     int ret;
1409
1410     acb = qemu_aio_get(bs, cb, opaque);
1411     if (!acb->bh)
1412         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1413     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1414     acb->ret = ret;
1415     qemu_bh_schedule(acb->bh);
1416     return &acb->common;
1417 }
1418
1419 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1420         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1421         BlockDriverCompletionFunc *cb, void *opaque)
1422 {
1423     BlockDriverAIOCBSync *acb;
1424     int ret;
1425
1426     acb = qemu_aio_get(bs, cb, opaque);
1427     if (!acb->bh)
1428         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1429     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1430     acb->ret = ret;
1431     qemu_bh_schedule(acb->bh);
1432     return &acb->common;
1433 }
1434
1435 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1436 {
1437     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1438     qemu_bh_cancel(acb->bh);
1439     qemu_aio_release(acb);
1440 }
1441
1442 /**************************************************************/
1443 /* sync block device emulation */
1444
1445 static void bdrv_rw_em_cb(void *opaque, int ret)
1446 {
1447     *(int *)opaque = ret;
1448 }
1449
1450 #define NOT_DONE 0x7fffffff
1451
1452 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1453                         uint8_t *buf, int nb_sectors)
1454 {
1455     int async_ret;
1456     BlockDriverAIOCB *acb;
1457
1458     async_ret = NOT_DONE;
1459     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1460                         bdrv_rw_em_cb, &async_ret);
1461     if (acb == NULL)
1462         return -1;
1463
1464     while (async_ret == NOT_DONE) {
1465         qemu_aio_wait();
1466     }
1467
1468     return async_ret;
1469 }
1470
1471 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1472                          const uint8_t *buf, int nb_sectors)
1473 {
1474     int async_ret;
1475     BlockDriverAIOCB *acb;
1476
1477     async_ret = NOT_DONE;
1478     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1479                          bdrv_rw_em_cb, &async_ret);
1480     if (acb == NULL)
1481         return -1;
1482     while (async_ret == NOT_DONE) {
1483         qemu_aio_wait();
1484     }
1485     return async_ret;
1486 }
1487
1488 void bdrv_init(void)
1489 {
1490     aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
1491                   bdrv_aio_cancel_vector);
1492
1493     bdrv_register(&bdrv_raw);
1494     bdrv_register(&bdrv_host_device);
1495 #ifndef _WIN32
1496     bdrv_register(&bdrv_cow);
1497 #endif
1498     bdrv_register(&bdrv_qcow);
1499     bdrv_register(&bdrv_vmdk);
1500     bdrv_register(&bdrv_cloop);
1501     bdrv_register(&bdrv_dmg);
1502     bdrv_register(&bdrv_bochs);
1503     bdrv_register(&bdrv_vpc);
1504     bdrv_register(&bdrv_vvfat);
1505     bdrv_register(&bdrv_qcow2);
1506     bdrv_register(&bdrv_parallels);
1507     bdrv_register(&bdrv_nbd);
1508 }
1509
1510 void aio_pool_init(AIOPool *pool, int aiocb_size,
1511                    void (*cancel)(BlockDriverAIOCB *acb))
1512 {
1513     pool->aiocb_size = aiocb_size;
1514     pool->cancel = cancel;
1515     pool->free_aiocb = NULL;
1516 }
1517
1518 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1519                         BlockDriverCompletionFunc *cb, void *opaque)
1520 {
1521     BlockDriverAIOCB *acb;
1522
1523     if (pool->free_aiocb) {
1524         acb = pool->free_aiocb;
1525         pool->free_aiocb = acb->next;
1526     } else {
1527         acb = qemu_mallocz(pool->aiocb_size);
1528         acb->pool = pool;
1529     }
1530     acb->bs = bs;
1531     acb->cb = cb;
1532     acb->opaque = opaque;
1533     return acb;
1534 }
1535
1536 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1537                    void *opaque)
1538 {
1539     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1540 }
1541
1542 void qemu_aio_release(void *p)
1543 {
1544     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1545     AIOPool *pool = acb->pool;
1546     acb->next = pool->free_aiocb;
1547     pool->free_aiocb = acb;
1548 }
1549
1550 /**************************************************************/
1551 /* removable device support */
1552
1553 /**
1554  * Return TRUE if the media is present
1555  */
1556 int bdrv_is_inserted(BlockDriverState *bs)
1557 {
1558     BlockDriver *drv = bs->drv;
1559     int ret;
1560     if (!drv)
1561         return 0;
1562     if (!drv->bdrv_is_inserted)
1563         return 1;
1564     ret = drv->bdrv_is_inserted(bs);
1565     return ret;
1566 }
1567
1568 /**
1569  * Return TRUE if the media changed since the last call to this
1570  * function. It is currently only used for floppy disks
1571  */
1572 int bdrv_media_changed(BlockDriverState *bs)
1573 {
1574     BlockDriver *drv = bs->drv;
1575     int ret;
1576
1577     if (!drv || !drv->bdrv_media_changed)
1578         ret = -ENOTSUP;
1579     else
1580         ret = drv->bdrv_media_changed(bs);
1581     if (ret == -ENOTSUP)
1582         ret = bs->media_changed;
1583     bs->media_changed = 0;
1584     return ret;
1585 }
1586
1587 /**
1588  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1589  */
1590 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1591 {
1592     BlockDriver *drv = bs->drv;
1593     int ret;
1594
1595     if (!drv || !drv->bdrv_eject) {
1596         ret = -ENOTSUP;
1597     } else {
1598         ret = drv->bdrv_eject(bs, eject_flag);
1599     }
1600     if (ret == -ENOTSUP) {
1601         if (eject_flag)
1602             bdrv_close(bs);
1603     }
1604 }
1605
1606 int bdrv_is_locked(BlockDriverState *bs)
1607 {
1608     return bs->locked;
1609 }
1610
1611 /**
1612  * Lock or unlock the media (if it is locked, the user won't be able
1613  * to eject it manually).
1614  */
1615 void bdrv_set_locked(BlockDriverState *bs, int locked)
1616 {
1617     BlockDriver *drv = bs->drv;
1618
1619     bs->locked = locked;
1620     if (drv && drv->bdrv_set_locked) {
1621         drv->bdrv_set_locked(bs, locked);
1622     }
1623 }
1624
1625 /* needed for generic scsi interface */
1626
1627 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1628 {
1629     BlockDriver *drv = bs->drv;
1630
1631     if (drv && drv->bdrv_ioctl)
1632         return drv->bdrv_ioctl(bs, req, buf);
1633     return -ENOTSUP;
1634 }
1635
1636 int bdrv_sg_send_command(BlockDriverState *bs, void *buf, int count)
1637 {
1638     return bs->drv->bdrv_sg_send_command(bs, buf, count);
1639 }
1640
1641 int bdrv_sg_recv_response(BlockDriverState *bs, void *buf, int count)
1642 {
1643     return bs->drv->bdrv_sg_recv_response(bs, buf, count);
1644 }
1645
1646 BlockDriverAIOCB *bdrv_sg_aio_read(BlockDriverState *bs, void *buf, int count,
1647                                    BlockDriverCompletionFunc *cb, void *opaque)
1648 {
1649     return bs->drv->bdrv_sg_aio_read(bs, buf, count, cb, opaque);
1650 }
1651
1652 BlockDriverAIOCB *bdrv_sg_aio_write(BlockDriverState *bs, void *buf, int count,
1653                                     BlockDriverCompletionFunc *cb, void *opaque)
1654 {
1655     return bs->drv->bdrv_sg_aio_write(bs, buf, count, cb, opaque);
1656 }