vcs.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
  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(BlockDriverState *bs,
  59         int64_t sector_num, uint8_t *buf, int nb_sectors,
  60         BlockDriverCompletionFunc *cb, void *opaque);
  61 static BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs,
  62         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  63         BlockDriverCompletionFunc *cb, void *opaque);
  64 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
  65         int64_t sector_num, uint8_t *buf, int nb_sectors,
  66         BlockDriverCompletionFunc *cb, void *opaque);
  67 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
  68         int64_t sector_num, const uint8_t *buf, int nb_sectors,
  69         BlockDriverCompletionFunc *cb, void *opaque);
  70 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  71 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  72                         uint8_t *buf, int nb_sectors);
  73 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  74                          const uint8_t *buf, int nb_sectors);
  75
  76 BlockDriverState *bdrv_first;
  77
  78 static BlockDriver *first_drv;
  79
  80 int path_is_absolute(const char *path)
  81 {
  82     const char *p;
  83 #ifdef _WIN32
  84     /* specific case for names like: "\\.\d:" */
  85     if (*path == '/' || *path == '\\')
  86         return 1;
  87 #endif
  88     p = strchr(path, ':');
  89     if (p)
  90         p++;
  91     else
  92         p = path;
  93 #ifdef _WIN32
  94     return (*p == '/' || *p == '\\');
  95 #else
  96     return (*p == '/');
  97 #endif
  98 }
  99
 100 /* if filename is absolute, just copy it to dest. Otherwise, build a
 101    path to it by considering it is relative to base_path. URL are
 102    supported. */
 103 void path_combine(char *dest, int dest_size,
 104                   const char *base_path,
 105                   const char *filename)
 106 {
 107     const char *p, *p1;
 108     int len;
 109
 110     if (dest_size <= 0)
 111         return;
 112     if (path_is_absolute(filename)) {
 113         pstrcpy(dest, dest_size, filename);
 114     } else {
 115         p = strchr(base_path, ':');
 116         if (p)
 117             p++;
 118         else
 119             p = base_path;
 120         p1 = strrchr(base_path, '/');
 121 #ifdef _WIN32
 122         {
 123             const char *p2;
 124             p2 = strrchr(base_path, '\\');
 125             if (!p1 || p2 > p1)
 126                 p1 = p2;
 127         }
 128 #endif
 129         if (p1)
 130             p1++;
 131         else
 132             p1 = base_path;
 133         if (p1 > p)
 134             p = p1;
 135         len = p - base_path;
 136         if (len > dest_size - 1)
 137             len = dest_size - 1;
 138         memcpy(dest, base_path, len);
 139         dest[len] = '\0';
 140         pstrcat(dest, dest_size, filename);
 141     }
 142 }
 143
 144
 145 static void bdrv_register(BlockDriver *bdrv)
 146 {
 147     if (!bdrv->bdrv_aio_read) {
 148         /* add AIO emulation layer */
 149         bdrv->bdrv_aio_read = bdrv_aio_read_em;
 150         bdrv->bdrv_aio_write = bdrv_aio_write_em;
 151         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 152         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 153     } else if (!bdrv->bdrv_read) {
 154         /* add synchronous IO emulation layer */
 155         bdrv->bdrv_read = bdrv_read_em;
 156         bdrv->bdrv_write = bdrv_write_em;
 157     }
 158     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 159     bdrv->next = first_drv;
 160     first_drv = bdrv;
 161 }
 162
 163 /* create a new block device (by default it is empty) */
 164 BlockDriverState *bdrv_new(const char *device_name)
 165 {
 166     BlockDriverState **pbs, *bs;
 167
 168     bs = qemu_mallocz(sizeof(BlockDriverState));
 169     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 170     if (device_name[0] != '\0') {
 171         /* insert at the end */
 172         pbs = &bdrv_first;
 173         while (*pbs != NULL)
 174             pbs = &(*pbs)->next;
 175         *pbs = bs;
 176     }
 177     return bs;
 178 }
 179
 180 BlockDriver *bdrv_find_format(const char *format_name)
 181 {
 182     BlockDriver *drv1;
 183     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 184         if (!strcmp(drv1->format_name, format_name))
 185             return drv1;
 186     }
 187     return NULL;
 188 }
 189
 190 int bdrv_create2(BlockDriver *drv,
 191                 const char *filename, int64_t size_in_sectors,
 192                 const char *backing_file, const char *backing_format,
 193                 int flags)
 194 {
 195     if (drv->bdrv_create2)
 196         return drv->bdrv_create2(filename, size_in_sectors, backing_file,
 197                                  backing_format, flags);
 198     if (drv->bdrv_create)
 199         return drv->bdrv_create(filename, size_in_sectors, backing_file,
 200                                 flags);
 201     return -ENOTSUP;
 202 }
 203
 204 int bdrv_create(BlockDriver *drv,
 205                 const char *filename, int64_t size_in_sectors,
 206                 const char *backing_file, int flags)
 207 {
 208     if (!drv->bdrv_create)
 209         return -ENOTSUP;
 210     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 211 }
 212
 213 #ifdef _WIN32
 214 void get_tmp_filename(char *filename, int size)
 215 {
 216     char temp_dir[MAX_PATH];
 217
 218     GetTempPath(MAX_PATH, temp_dir);
 219     GetTempFileName(temp_dir, "qem", 0, filename);
 220 }
 221 #else
 222 void get_tmp_filename(char *filename, int size)
 223 {
 224     int fd;
 225     const char *tmpdir;
 226     /* XXX: race condition possible */
 227     tmpdir = getenv("TMPDIR");
 228     if (!tmpdir)
 229         tmpdir = "/tmp";
 230     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 231     fd = mkstemp(filename);
 232     close(fd);
 233 }
 234 #endif
 235
 236 #ifdef _WIN32
 237 static int is_windows_drive_prefix(const char *filename)
 238 {
 239     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 240              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 241             filename[1] == ':');
 242 }
 243
 244 static int is_windows_drive(const char *filename)
 245 {
 246     if (is_windows_drive_prefix(filename) &&
 247         filename[2] == '\0')
 248         return 1;
 249     if (strstart(filename, "\\\\.\\", NULL) ||
 250         strstart(filename, "//./", NULL))
 251         return 1;
 252     return 0;
 253 }
 254 #endif
 255
 256 static BlockDriver *find_protocol(const char *filename)
 257 {
 258     BlockDriver *drv1;
 259     char protocol[128];
 260     int len;
 261     const char *p;
 262
 263 #ifdef _WIN32
 264     if (is_windows_drive(filename) ||
 265         is_windows_drive_prefix(filename))
 266         return &bdrv_raw;
 267 #endif
 268     p = strchr(filename, ':');
 269     if (!p)
 270         return &bdrv_raw;
 271     len = p - filename;
 272     if (len > sizeof(protocol) - 1)
 273         len = sizeof(protocol) - 1;
 274     memcpy(protocol, filename, len);
 275     protocol[len] = '\0';
 276     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 277         if (drv1->protocol_name &&
 278             !strcmp(drv1->protocol_name, protocol))
 279             return drv1;
 280     }
 281     return NULL;
 282 }
 283
 284 /* XXX: force raw format if block or character device ? It would
 285    simplify the BSD case */
 286 static BlockDriver *find_image_format(const char *filename)
 287 {
 288     int ret, score, score_max;
 289     BlockDriver *drv1, *drv;
 290     uint8_t buf[2048];
 291     BlockDriverState *bs;
 292
 293     /* detect host devices. By convention, /dev/cdrom[N] is always
 294        recognized as a host CDROM */
 295     if (strstart(filename, "/dev/cdrom", NULL))
 296         return &bdrv_host_device;
 297 #ifdef _WIN32
 298     if (is_windows_drive(filename))
 299         return &bdrv_host_device;
 300 #else
 301     {
 302         struct stat st;
 303         if (stat(filename, &st) >= 0 &&
 304             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 305             return &bdrv_host_device;
 306         }
 307     }
 308 #endif
 309
 310     drv = find_protocol(filename);
 311     /* no need to test disk image formats for vvfat */
 312     if (drv == &bdrv_vvfat)
 313         return drv;
 314
 315     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 316     if (ret < 0)
 317         return NULL;
 318     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 319     bdrv_delete(bs);
 320     if (ret < 0) {
 321         return NULL;
 322     }
 323
 324     score_max = 0;
 325     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 326         if (drv1->bdrv_probe) {
 327             score = drv1->bdrv_probe(buf, ret, filename);
 328             if (score > score_max) {
 329                 score_max = score;
 330                 drv = drv1;
 331             }
 332         }
 333     }
 334     return drv;
 335 }
 336
 337 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 338 {
 339     BlockDriverState *bs;
 340     int ret;
 341
 342     bs = bdrv_new("");
 343     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 344     if (ret < 0) {
 345         bdrv_delete(bs);
 346         return ret;
 347     }
 348     bs->growable = 1;
 349     *pbs = bs;
 350     return 0;
 351 }
 352
 353 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 354 {
 355     return bdrv_open2(bs, filename, flags, NULL);
 356 }
 357
 358 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 359                BlockDriver *drv)
 360 {
 361     int ret, open_flags;
 362     char tmp_filename[PATH_MAX];
 363     char backing_filename[PATH_MAX];
 364
 365     bs->read_only = 0;
 366     bs->is_temporary = 0;
 367     bs->encrypted = 0;
 368     bs->valid_key = 0;
 369
 370     if (flags & BDRV_O_SNAPSHOT) {
 371         BlockDriverState *bs1;
 372         int64_t total_size;
 373         int is_protocol = 0;
 374
 375         /* if snapshot, we create a temporary backing file and open it
 376            instead of opening 'filename' directly */
 377
 378         /* if there is a backing file, use it */
 379         bs1 = bdrv_new("");
 380         ret = bdrv_open2(bs1, filename, 0, drv);
 381         if (ret < 0) {
 382             bdrv_delete(bs1);
 383             return ret;
 384         }
 385         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 386
 387         if (bs1->drv && bs1->drv->protocol_name)
 388             is_protocol = 1;
 389
 390         bdrv_delete(bs1);
 391
 392         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 393
 394         /* Real path is meaningless for protocols */
 395         if (is_protocol)
 396             snprintf(backing_filename, sizeof(backing_filename),
 397                      "%s", filename);
 398         else
 399             realpath(filename, backing_filename);
 400
 401         ret = bdrv_create2(&bdrv_qcow2, tmp_filename,
 402                            total_size, backing_filename,
 403                            (drv ? drv->format_name : NULL), 0);
 404         if (ret < 0) {
 405             return ret;
 406         }
 407         filename = tmp_filename;
 408         drv = &bdrv_qcow2;
 409         bs->is_temporary = 1;
 410     }
 411
 412     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 413     if (flags & BDRV_O_FILE) {
 414         drv = find_protocol(filename);
 415     } else if (!drv) {
 416         drv = find_image_format(filename);
 417     }
 418     if (!drv) {
 419         ret = -ENOENT;
 420         goto unlink_and_fail;
 421     }
 422     bs->drv = drv;
 423     bs->opaque = qemu_mallocz(drv->instance_size);
 424     /* Note: for compatibility, we open disk image files as RDWR, and
 425        RDONLY as fallback */
 426     if (!(flags & BDRV_O_FILE))
 427         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 428     else
 429         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 430     ret = drv->bdrv_open(bs, filename, open_flags);
 431     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 432         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 433         bs->read_only = 1;
 434     }
 435     if (ret < 0) {
 436         qemu_free(bs->opaque);
 437         bs->opaque = NULL;
 438         bs->drv = NULL;
 439     unlink_and_fail:
 440         if (bs->is_temporary)
 441             unlink(filename);
 442         return ret;
 443     }
 444     if (drv->bdrv_getlength) {
 445         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 446     }
 447 #ifndef _WIN32
 448     if (bs->is_temporary) {
 449         unlink(filename);
 450     }
 451 #endif
 452     if (bs->backing_file[0] != '\0') {
 453         /* if there is a backing file, use it */
 454         BlockDriver *back_drv = NULL;
 455         bs->backing_hd = bdrv_new("");
 456         path_combine(backing_filename, sizeof(backing_filename),
 457                      filename, bs->backing_file);
 458         if (bs->backing_format[0] != '\0')
 459             back_drv = bdrv_find_format(bs->backing_format);
 460         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 461                          back_drv);
 462         if (ret < 0) {
 463             bdrv_close(bs);
 464             return ret;
 465         }
 466     }
 467
 468     if (!bdrv_key_required(bs)) {
 469         /* call the change callback */
 470         bs->media_changed = 1;
 471         if (bs->change_cb)
 472             bs->change_cb(bs->change_opaque);
 473     }
 474     return 0;
 475 }
 476
 477 void bdrv_close(BlockDriverState *bs)
 478 {
 479     if (bs->drv) {
 480         if (bs->backing_hd)
 481             bdrv_delete(bs->backing_hd);
 482         bs->drv->bdrv_close(bs);
 483         qemu_free(bs->opaque);
 484 #ifdef _WIN32
 485         if (bs->is_temporary) {
 486             unlink(bs->filename);
 487         }
 488 #endif
 489         bs->opaque = NULL;
 490         bs->drv = NULL;
 491
 492         /* call the change callback */
 493         bs->media_changed = 1;
 494         if (bs->change_cb)
 495             bs->change_cb(bs->change_opaque);
 496     }
 497 }
 498
 499 void bdrv_delete(BlockDriverState *bs)
 500 {
 501     BlockDriverState **pbs;
 502
 503     pbs = &bdrv_first;
 504     while (*pbs != bs && *pbs != NULL)
 505         pbs = &(*pbs)->next;
 506     if (*pbs == bs)
 507         *pbs = bs->next;
 508
 509     bdrv_close(bs);
 510     qemu_free(bs);
 511 }
 512
 513 /* commit COW file into the raw image */
 514 int bdrv_commit(BlockDriverState *bs)
 515 {
 516     BlockDriver *drv = bs->drv;
 517     int64_t i, total_sectors;
 518     int n, j;
 519     unsigned char sector[512];
 520
 521     if (!drv)
 522         return -ENOMEDIUM;
 523
 524     if (bs->read_only) {
 525         return -EACCES;
 526     }
 527
 528     if (!bs->backing_hd) {
 529         return -ENOTSUP;
 530     }
 531
 532     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 533     for (i = 0; i < total_sectors;) {
 534         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 535             for(j = 0; j < n; j++) {
 536                 if (bdrv_read(bs, i, sector, 1) != 0) {
 537                     return -EIO;
 538                 }
 539
 540                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 541                     return -EIO;
 542                 }
 543                 i++;
 544             }
 545         } else {
 546             i += n;
 547         }
 548     }
 549
 550     if (drv->bdrv_make_empty)
 551         return drv->bdrv_make_empty(bs);
 552
 553     return 0;
 554 }
 555
 556 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 557                                    size_t size)
 558 {
 559     int64_t len;
 560
 561     if (!bdrv_is_inserted(bs))
 562         return -ENOMEDIUM;
 563
 564     if (bs->growable)
 565         return 0;
 566
 567     len = bdrv_getlength(bs);
 568
 569     if ((offset + size) > len)
 570         return -EIO;
 571
 572     return 0;
 573 }
 574
 575 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 576                               int nb_sectors)
 577 {
 578     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 579 }
 580
 581 /* return < 0 if error. See bdrv_write() for the return codes */
 582 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 583               uint8_t *buf, int nb_sectors)
 584 {
 585     BlockDriver *drv = bs->drv;
 586
 587     if (!drv)
 588         return -ENOMEDIUM;
 589     if (bdrv_check_request(bs, sector_num, nb_sectors))
 590         return -EIO;
 591
 592     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 593 }
 594
 595 /* Return < 0 if error. Important errors are:
 596   -EIO         generic I/O error (may happen for all errors)
 597   -ENOMEDIUM   No media inserted.
 598   -EINVAL      Invalid sector number or nb_sectors
 599   -EACCES      Trying to write a read-only device
 600 */
 601 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 602                const uint8_t *buf, int nb_sectors)
 603 {
 604     BlockDriver *drv = bs->drv;
 605     if (!bs->drv)
 606         return -ENOMEDIUM;
 607     if (bs->read_only)
 608         return -EACCES;
 609     if (bdrv_check_request(bs, sector_num, nb_sectors))
 610         return -EIO;
 611
 612     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 613 }
 614
 615 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 616                void *buf, int count1)
 617 {
 618     uint8_t tmp_buf[SECTOR_SIZE];
 619     int len, nb_sectors, count;
 620     int64_t sector_num;
 621
 622     count = count1;
 623     /* first read to align to sector start */
 624     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 625     if (len > count)
 626         len = count;
 627     sector_num = offset >> SECTOR_BITS;
 628     if (len > 0) {
 629         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 630             return -EIO;
 631         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 632         count -= len;
 633         if (count == 0)
 634             return count1;
 635         sector_num++;
 636         buf += len;
 637     }
 638
 639     /* read the sectors "in place" */
 640     nb_sectors = count >> SECTOR_BITS;
 641     if (nb_sectors > 0) {
 642         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 643             return -EIO;
 644         sector_num += nb_sectors;
 645         len = nb_sectors << SECTOR_BITS;
 646         buf += len;
 647         count -= len;
 648     }
 649
 650     /* add data from the last sector */
 651     if (count > 0) {
 652         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 653             return -EIO;
 654         memcpy(buf, tmp_buf, count);
 655     }
 656     return count1;
 657 }
 658
 659 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 660                 const void *buf, int count1)
 661 {
 662     uint8_t tmp_buf[SECTOR_SIZE];
 663     int len, nb_sectors, count;
 664     int64_t sector_num;
 665
 666     count = count1;
 667     /* first write to align to sector start */
 668     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 669     if (len > count)
 670         len = count;
 671     sector_num = offset >> SECTOR_BITS;
 672     if (len > 0) {
 673         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 674             return -EIO;
 675         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 676         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 677             return -EIO;
 678         count -= len;
 679         if (count == 0)
 680             return count1;
 681         sector_num++;
 682         buf += len;
 683     }
 684
 685     /* write the sectors "in place" */
 686     nb_sectors = count >> SECTOR_BITS;
 687     if (nb_sectors > 0) {
 688         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 689             return -EIO;
 690         sector_num += nb_sectors;
 691         len = nb_sectors << SECTOR_BITS;
 692         buf += len;
 693         count -= len;
 694     }
 695
 696     /* add data from the last sector */
 697     if (count > 0) {
 698         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 699             return -EIO;
 700         memcpy(tmp_buf, buf, count);
 701         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 702             return -EIO;
 703     }
 704     return count1;
 705 }
 706
 707 /**
 708  * Truncate file to 'offset' bytes (needed only for file protocols)
 709  */
 710 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 711 {
 712     BlockDriver *drv = bs->drv;
 713     if (!drv)
 714         return -ENOMEDIUM;
 715     if (!drv->bdrv_truncate)
 716         return -ENOTSUP;
 717     return drv->bdrv_truncate(bs, offset);
 718 }
 719
 720 /**
 721  * Length of a file in bytes. Return < 0 if error or unknown.
 722  */
 723 int64_t bdrv_getlength(BlockDriverState *bs)
 724 {
 725     BlockDriver *drv = bs->drv;
 726     if (!drv)
 727         return -ENOMEDIUM;
 728     if (!drv->bdrv_getlength) {
 729         /* legacy mode */
 730         return bs->total_sectors * SECTOR_SIZE;
 731     }
 732     return drv->bdrv_getlength(bs);
 733 }
 734
 735 /* return 0 as number of sectors if no device present or error */
 736 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 737 {
 738     int64_t length;
 739     length = bdrv_getlength(bs);
 740     if (length < 0)
 741         length = 0;
 742     else
 743         length = length >> SECTOR_BITS;
 744     *nb_sectors_ptr = length;
 745 }
 746
 747 struct partition {
 748         uint8_t boot_ind;           /* 0x80 - active */
 749         uint8_t head;               /* starting head */
 750         uint8_t sector;             /* starting sector */
 751         uint8_t cyl;                /* starting cylinder */
 752         uint8_t sys_ind;            /* What partition type */
 753         uint8_t end_head;           /* end head */
 754         uint8_t end_sector;         /* end sector */
 755         uint8_t end_cyl;            /* end cylinder */
 756         uint32_t start_sect;        /* starting sector counting from 0 */
 757         uint32_t nr_sects;          /* nr of sectors in partition */
 758 } __attribute__((packed));
 759
 760 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 761 static int guess_disk_lchs(BlockDriverState *bs,
 762                            int *pcylinders, int *pheads, int *psectors)
 763 {
 764     uint8_t buf[512];
 765     int ret, i, heads, sectors, cylinders;
 766     struct partition *p;
 767     uint32_t nr_sects;
 768     uint64_t nb_sectors;
 769
 770     bdrv_get_geometry(bs, &nb_sectors);
 771
 772     ret = bdrv_read(bs, 0, buf, 1);
 773     if (ret < 0)
 774         return -1;
 775     /* test msdos magic */
 776     if (buf[510] != 0x55 || buf[511] != 0xaa)
 777         return -1;
 778     for(i = 0; i < 4; i++) {
 779         p = ((struct partition *)(buf + 0x1be)) + i;
 780         nr_sects = le32_to_cpu(p->nr_sects);
 781         if (nr_sects && p->end_head) {
 782             /* We make the assumption that the partition terminates on
 783                a cylinder boundary */
 784             heads = p->end_head + 1;
 785             sectors = p->end_sector & 63;
 786             if (sectors == 0)
 787                 continue;
 788             cylinders = nb_sectors / (heads * sectors);
 789             if (cylinders < 1 || cylinders > 16383)
 790                 continue;
 791             *pheads = heads;
 792             *psectors = sectors;
 793             *pcylinders = cylinders;
 794 #if 0
 795             printf("guessed geometry: LCHS=%d %d %d\n",
 796                    cylinders, heads, sectors);
 797 #endif
 798             return 0;
 799         }
 800     }
 801     return -1;
 802 }
 803
 804 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 805 {
 806     int translation, lba_detected = 0;
 807     int cylinders, heads, secs;
 808     uint64_t nb_sectors;
 809
 810     /* if a geometry hint is available, use it */
 811     bdrv_get_geometry(bs, &nb_sectors);
 812     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 813     translation = bdrv_get_translation_hint(bs);
 814     if (cylinders != 0) {
 815         *pcyls = cylinders;
 816         *pheads = heads;
 817         *psecs = secs;
 818     } else {
 819         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 820             if (heads > 16) {
 821                 /* if heads > 16, it means that a BIOS LBA
 822                    translation was active, so the default
 823                    hardware geometry is OK */
 824                 lba_detected = 1;
 825                 goto default_geometry;
 826             } else {
 827                 *pcyls = cylinders;
 828                 *pheads = heads;
 829                 *psecs = secs;
 830                 /* disable any translation to be in sync with
 831                    the logical geometry */
 832                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 833                     bdrv_set_translation_hint(bs,
 834                                               BIOS_ATA_TRANSLATION_NONE);
 835                 }
 836             }
 837         } else {
 838         default_geometry:
 839             /* if no geometry, use a standard physical disk geometry */
 840             cylinders = nb_sectors / (16 * 63);
 841
 842             if (cylinders > 16383)
 843                 cylinders = 16383;
 844             else if (cylinders < 2)
 845                 cylinders = 2;
 846             *pcyls = cylinders;
 847             *pheads = 16;
 848             *psecs = 63;
 849             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 850                 if ((*pcyls * *pheads) <= 131072) {
 851                     bdrv_set_translation_hint(bs,
 852                                               BIOS_ATA_TRANSLATION_LARGE);
 853                 } else {
 854                     bdrv_set_translation_hint(bs,
 855                                               BIOS_ATA_TRANSLATION_LBA);
 856                 }
 857             }
 858         }
 859         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 860     }
 861 }
 862
 863 void bdrv_set_geometry_hint(BlockDriverState *bs,
 864                             int cyls, int heads, int secs)
 865 {
 866     bs->cyls = cyls;
 867     bs->heads = heads;
 868     bs->secs = secs;
 869 }
 870
 871 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 872 {
 873     bs->type = type;
 874     bs->removable = ((type == BDRV_TYPE_CDROM ||
 875                       type == BDRV_TYPE_FLOPPY));
 876 }
 877
 878 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 879 {
 880     bs->translation = translation;
 881 }
 882
 883 void bdrv_get_geometry_hint(BlockDriverState *bs,
 884                             int *pcyls, int *pheads, int *psecs)
 885 {
 886     *pcyls = bs->cyls;
 887     *pheads = bs->heads;
 888     *psecs = bs->secs;
 889 }
 890
 891 int bdrv_get_type_hint(BlockDriverState *bs)
 892 {
 893     return bs->type;
 894 }
 895
 896 int bdrv_get_translation_hint(BlockDriverState *bs)
 897 {
 898     return bs->translation;
 899 }
 900
 901 int bdrv_is_removable(BlockDriverState *bs)
 902 {
 903     return bs->removable;
 904 }
 905
 906 int bdrv_is_read_only(BlockDriverState *bs)
 907 {
 908     return bs->read_only;
 909 }
 910
 911 int bdrv_is_sg(BlockDriverState *bs)
 912 {
 913     return bs->sg;
 914 }
 915
 916 /* XXX: no longer used */
 917 void bdrv_set_change_cb(BlockDriverState *bs,
 918                         void (*change_cb)(void *opaque), void *opaque)
 919 {
 920     bs->change_cb = change_cb;
 921     bs->change_opaque = opaque;
 922 }
 923
 924 int bdrv_is_encrypted(BlockDriverState *bs)
 925 {
 926     if (bs->backing_hd && bs->backing_hd->encrypted)
 927         return 1;
 928     return bs->encrypted;
 929 }
 930
 931 int bdrv_key_required(BlockDriverState *bs)
 932 {
 933     BlockDriverState *backing_hd = bs->backing_hd;
 934
 935     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 936         return 1;
 937     return (bs->encrypted && !bs->valid_key);
 938 }
 939
 940 int bdrv_set_key(BlockDriverState *bs, const char *key)
 941 {
 942     int ret;
 943     if (bs->backing_hd && bs->backing_hd->encrypted) {
 944         ret = bdrv_set_key(bs->backing_hd, key);
 945         if (ret < 0)
 946             return ret;
 947         if (!bs->encrypted)
 948             return 0;
 949     }
 950     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 951         return -1;
 952     ret = bs->drv->bdrv_set_key(bs, key);
 953     if (ret < 0) {
 954         bs->valid_key = 0;
 955     } else if (!bs->valid_key) {
 956         bs->valid_key = 1;
 957         /* call the change callback now, we skipped it on open */
 958         bs->media_changed = 1;
 959         if (bs->change_cb)
 960             bs->change_cb(bs->change_opaque);
 961     }
 962     return ret;
 963 }
 964
 965 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 966 {
 967     if (!bs->drv) {
 968         buf[0] = '\0';
 969     } else {
 970         pstrcpy(buf, buf_size, bs->drv->format_name);
 971     }
 972 }
 973
 974 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 975                          void *opaque)
 976 {
 977     BlockDriver *drv;
 978
 979     for (drv = first_drv; drv != NULL; drv = drv->next) {
 980         it(opaque, drv->format_name);
 981     }
 982 }
 983
 984 BlockDriverState *bdrv_find(const char *name)
 985 {
 986     BlockDriverState *bs;
 987
 988     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
 989         if (!strcmp(name, bs->device_name))
 990             return bs;
 991     }
 992     return NULL;
 993 }
 994
 995 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
 996 {
 997     BlockDriverState *bs;
 998
 999     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1000         it(opaque, bs);
1001     }
1002 }
1003
1004 const char *bdrv_get_device_name(BlockDriverState *bs)
1005 {
1006     return bs->device_name;
1007 }
1008
1009 void bdrv_flush(BlockDriverState *bs)
1010 {
1011     if (!bs->drv)
1012         return;
1013     if (bs->drv->bdrv_flush)
1014         bs->drv->bdrv_flush(bs);
1015     if (bs->backing_hd)
1016         bdrv_flush(bs->backing_hd);
1017 }
1018
1019 void bdrv_flush_all(void)
1020 {
1021     BlockDriverState *bs;
1022
1023     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1024         if (bs->drv && !bdrv_is_read_only(bs) &&
1025             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1026             bdrv_flush(bs);
1027 }
1028
1029 /*
1030  * Returns true iff the specified sector is present in the disk image. Drivers
1031  * not implementing the functionality are assumed to not support backing files,
1032  * hence all their sectors are reported as allocated.
1033  *
1034  * 'pnum' is set to the number of sectors (including and immediately following
1035  * the specified sector) that are known to be in the same
1036  * allocated/unallocated state.
1037  *
1038  * 'nb_sectors' is the max value 'pnum' should be set to.
1039  */
1040 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1041         int *pnum)
1042 {
1043     int64_t n;
1044     if (!bs->drv->bdrv_is_allocated) {
1045         if (sector_num >= bs->total_sectors) {
1046             *pnum = 0;
1047             return 0;
1048         }
1049         n = bs->total_sectors - sector_num;
1050         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1051         return 1;
1052     }
1053     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1054 }
1055
1056 void bdrv_info(Monitor *mon)
1057 {
1058     BlockDriverState *bs;
1059
1060     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1061         monitor_printf(mon, "%s:", bs->device_name);
1062         monitor_printf(mon, " type=");
1063         switch(bs->type) {
1064         case BDRV_TYPE_HD:
1065             monitor_printf(mon, "hd");
1066             break;
1067         case BDRV_TYPE_CDROM:
1068             monitor_printf(mon, "cdrom");
1069             break;
1070         case BDRV_TYPE_FLOPPY:
1071             monitor_printf(mon, "floppy");
1072             break;
1073         }
1074         monitor_printf(mon, " removable=%d", bs->removable);
1075         if (bs->removable) {
1076             monitor_printf(mon, " locked=%d", bs->locked);
1077         }
1078         if (bs->drv) {
1079             monitor_printf(mon, " file=");
1080             monitor_print_filename(mon, bs->filename);
1081             if (bs->backing_file[0] != '\0') {
1082                 monitor_printf(mon, " backing_file=");
1083                 monitor_print_filename(mon, bs->backing_file);
1084             }
1085             monitor_printf(mon, " ro=%d", bs->read_only);
1086             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1087             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1088         } else {
1089             monitor_printf(mon, " [not inserted]");
1090         }
1091         monitor_printf(mon, "\n");
1092     }
1093 }
1094
1095 /* The "info blockstats" command. */
1096 void bdrv_info_stats(Monitor *mon)
1097 {
1098     BlockDriverState *bs;
1099
1100     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1101         monitor_printf(mon, "%s:"
1102                        " rd_bytes=%" PRIu64
1103                        " wr_bytes=%" PRIu64
1104                        " rd_operations=%" PRIu64
1105                        " wr_operations=%" PRIu64
1106                        "\n",
1107                        bs->device_name,
1108                        bs->rd_bytes, bs->wr_bytes,
1109                        bs->rd_ops, bs->wr_ops);
1110     }
1111 }
1112
1113 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1114 {
1115     if (bs->backing_hd && bs->backing_hd->encrypted)
1116         return bs->backing_file;
1117     else if (bs->encrypted)
1118         return bs->filename;
1119     else
1120         return NULL;
1121 }
1122
1123 void bdrv_get_backing_filename(BlockDriverState *bs,
1124                                char *filename, int filename_size)
1125 {
1126     if (!bs->backing_hd) {
1127         pstrcpy(filename, filename_size, "");
1128     } else {
1129         pstrcpy(filename, filename_size, bs->backing_file);
1130     }
1131 }
1132
1133 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1134                           const uint8_t *buf, int nb_sectors)
1135 {
1136     BlockDriver *drv = bs->drv;
1137     if (!drv)
1138         return -ENOMEDIUM;
1139     if (!drv->bdrv_write_compressed)
1140         return -ENOTSUP;
1141     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1142 }
1143
1144 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1145 {
1146     BlockDriver *drv = bs->drv;
1147     if (!drv)
1148         return -ENOMEDIUM;
1149     if (!drv->bdrv_get_info)
1150         return -ENOTSUP;
1151     memset(bdi, 0, sizeof(*bdi));
1152     return drv->bdrv_get_info(bs, bdi);
1153 }
1154
1155 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1156 {
1157     BlockDriver *drv = bs->drv;
1158     if (!drv)
1159         return -ENOMEDIUM;
1160     if (!drv->bdrv_put_buffer)
1161         return -ENOTSUP;
1162     return drv->bdrv_put_buffer(bs, buf, pos, size);
1163 }
1164
1165 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1166 {
1167     BlockDriver *drv = bs->drv;
1168     if (!drv)
1169         return -ENOMEDIUM;
1170     if (!drv->bdrv_get_buffer)
1171         return -ENOTSUP;
1172     return drv->bdrv_get_buffer(bs, buf, pos, size);
1173 }
1174
1175 /**************************************************************/
1176 /* handling of snapshots */
1177
1178 int bdrv_snapshot_create(BlockDriverState *bs,
1179                          QEMUSnapshotInfo *sn_info)
1180 {
1181     BlockDriver *drv = bs->drv;
1182     if (!drv)
1183         return -ENOMEDIUM;
1184     if (!drv->bdrv_snapshot_create)
1185         return -ENOTSUP;
1186     return drv->bdrv_snapshot_create(bs, sn_info);
1187 }
1188
1189 int bdrv_snapshot_goto(BlockDriverState *bs,
1190                        const char *snapshot_id)
1191 {
1192     BlockDriver *drv = bs->drv;
1193     if (!drv)
1194         return -ENOMEDIUM;
1195     if (!drv->bdrv_snapshot_goto)
1196         return -ENOTSUP;
1197     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1198 }
1199
1200 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1201 {
1202     BlockDriver *drv = bs->drv;
1203     if (!drv)
1204         return -ENOMEDIUM;
1205     if (!drv->bdrv_snapshot_delete)
1206         return -ENOTSUP;
1207     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1208 }
1209
1210 int bdrv_snapshot_list(BlockDriverState *bs,
1211                        QEMUSnapshotInfo **psn_info)
1212 {
1213     BlockDriver *drv = bs->drv;
1214     if (!drv)
1215         return -ENOMEDIUM;
1216     if (!drv->bdrv_snapshot_list)
1217         return -ENOTSUP;
1218     return drv->bdrv_snapshot_list(bs, psn_info);
1219 }
1220
1221 #define NB_SUFFIXES 4
1222
1223 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1224 {
1225     static const char suffixes[NB_SUFFIXES] = "KMGT";
1226     int64_t base;
1227     int i;
1228
1229     if (size <= 999) {
1230         snprintf(buf, buf_size, "%" PRId64, size);
1231     } else {
1232         base = 1024;
1233         for(i = 0; i < NB_SUFFIXES; i++) {
1234             if (size < (10 * base)) {
1235                 snprintf(buf, buf_size, "%0.1f%c",
1236                          (double)size / base,
1237                          suffixes[i]);
1238                 break;
1239             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1240                 snprintf(buf, buf_size, "%" PRId64 "%c",
1241                          ((size + (base >> 1)) / base),
1242                          suffixes[i]);
1243                 break;
1244             }
1245             base = base * 1024;
1246         }
1247     }
1248     return buf;
1249 }
1250
1251 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1252 {
1253     char buf1[128], date_buf[128], clock_buf[128];
1254 #ifdef _WIN32
1255     struct tm *ptm;
1256 #else
1257     struct tm tm;
1258 #endif
1259     time_t ti;
1260     int64_t secs;
1261
1262     if (!sn) {
1263         snprintf(buf, buf_size,
1264                  "%-10s%-20s%7s%20s%15s",
1265                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1266     } else {
1267         ti = sn->date_sec;
1268 #ifdef _WIN32
1269         ptm = localtime(&ti);
1270         strftime(date_buf, sizeof(date_buf),
1271                  "%Y-%m-%d %H:%M:%S", ptm);
1272 #else
1273         localtime_r(&ti, &tm);
1274         strftime(date_buf, sizeof(date_buf),
1275                  "%Y-%m-%d %H:%M:%S", &tm);
1276 #endif
1277         secs = sn->vm_clock_nsec / 1000000000;
1278         snprintf(clock_buf, sizeof(clock_buf),
1279                  "%02d:%02d:%02d.%03d",
1280                  (int)(secs / 3600),
1281                  (int)((secs / 60) % 60),
1282                  (int)(secs % 60),
1283                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1284         snprintf(buf, buf_size,
1285                  "%-10s%-20s%7s%20s%15s",
1286                  sn->id_str, sn->name,
1287                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1288                  date_buf,
1289                  clock_buf);
1290     }
1291     return buf;
1292 }
1293
1294
1295 /**************************************************************/
1296 /* async I/Os */
1297
1298 typedef struct VectorTranslationAIOCB {
1299     BlockDriverAIOCB common;
1300     QEMUIOVector *iov;
1301     uint8_t *bounce;
1302     int is_write;
1303     BlockDriverAIOCB *aiocb;
1304 } VectorTranslationAIOCB;
1305
1306 static void bdrv_aio_cancel_vector(BlockDriverAIOCB *_acb)
1307 {
1308     VectorTranslationAIOCB *acb
1309         = container_of(_acb, VectorTranslationAIOCB, common);
1310
1311     bdrv_aio_cancel(acb->aiocb);
1312 }
1313
1314 static void bdrv_aio_rw_vector_cb(void *opaque, int ret)
1315 {
1316     VectorTranslationAIOCB *s = (VectorTranslationAIOCB *)opaque;
1317
1318     if (!s->is_write) {
1319         qemu_iovec_from_buffer(s->iov, s->bounce, s->iov->size);
1320     }
1321     qemu_vfree(s->bounce);
1322     s->common.cb(s->common.opaque, ret);
1323     qemu_aio_release(s);
1324 }
1325
1326 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1327                                             int64_t sector_num,
1328                                             QEMUIOVector *iov,
1329                                             int nb_sectors,
1330                                             BlockDriverCompletionFunc *cb,
1331                                             void *opaque,
1332                                             int is_write)
1333
1334 {
1335     VectorTranslationAIOCB *s = qemu_aio_get_pool(&vectored_aio_pool, bs,
1336                                                   cb, opaque);
1337
1338     s->iov = iov;
1339     s->bounce = qemu_memalign(512, nb_sectors * 512);
1340     s->is_write = is_write;
1341     if (is_write) {
1342         qemu_iovec_to_buffer(s->iov, s->bounce);
1343         s->aiocb = bdrv_aio_write(bs, sector_num, s->bounce, nb_sectors,
1344                                   bdrv_aio_rw_vector_cb, s);
1345     } else {
1346         s->aiocb = bdrv_aio_read(bs, sector_num, s->bounce, nb_sectors,
1347                                  bdrv_aio_rw_vector_cb, s);
1348     }
1349     if (!s->aiocb) {
1350         qemu_vfree(s->bounce);
1351         qemu_aio_release(s);
1352         return NULL;
1353     }
1354     return &s->common;
1355 }
1356
1357 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1358                                  QEMUIOVector *iov, int nb_sectors,
1359                                  BlockDriverCompletionFunc *cb, void *opaque)
1360 {
1361     if (bdrv_check_request(bs, sector_num, nb_sectors))
1362         return NULL;
1363
1364     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1365                               cb, opaque, 0);
1366 }
1367
1368 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1369                                   QEMUIOVector *iov, int nb_sectors,
1370                                   BlockDriverCompletionFunc *cb, void *opaque)
1371 {
1372     if (bdrv_check_request(bs, sector_num, nb_sectors))
1373         return NULL;
1374
1375     return bdrv_aio_rw_vector(bs, sector_num, iov, nb_sectors,
1376                               cb, opaque, 1);
1377 }
1378
1379 static BlockDriverAIOCB *bdrv_aio_read(BlockDriverState *bs, int64_t sector_num,
1380                                 uint8_t *buf, int nb_sectors,
1381                                 BlockDriverCompletionFunc *cb, void *opaque)
1382 {
1383     BlockDriver *drv = bs->drv;
1384     BlockDriverAIOCB *ret;
1385
1386     if (!drv)
1387         return NULL;
1388     if (bdrv_check_request(bs, sector_num, nb_sectors))
1389         return NULL;
1390
1391     ret = drv->bdrv_aio_read(bs, sector_num, buf, nb_sectors, cb, opaque);
1392
1393     if (ret) {
1394         /* Update stats even though technically transfer has not happened. */
1395         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1396         bs->rd_ops ++;
1397     }
1398
1399     return ret;
1400 }
1401
1402 static BlockDriverAIOCB *bdrv_aio_write(BlockDriverState *bs, int64_t sector_num,
1403                                  const uint8_t *buf, int nb_sectors,
1404                                  BlockDriverCompletionFunc *cb, void *opaque)
1405 {
1406     BlockDriver *drv = bs->drv;
1407     BlockDriverAIOCB *ret;
1408
1409     if (!drv)
1410         return NULL;
1411     if (bs->read_only)
1412         return NULL;
1413     if (bdrv_check_request(bs, sector_num, nb_sectors))
1414         return NULL;
1415
1416     ret = drv->bdrv_aio_write(bs, sector_num, buf, nb_sectors, cb, opaque);
1417
1418     if (ret) {
1419         /* Update stats even though technically transfer has not happened. */
1420         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1421         bs->wr_ops ++;
1422     }
1423
1424     return ret;
1425 }
1426
1427 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1428 {
1429     acb->pool->cancel(acb);
1430 }
1431
1432
1433 /**************************************************************/
1434 /* async block device emulation */
1435
1436 static void bdrv_aio_bh_cb(void *opaque)
1437 {
1438     BlockDriverAIOCBSync *acb = opaque;
1439     acb->common.cb(acb->common.opaque, acb->ret);
1440     qemu_aio_release(acb);
1441 }
1442
1443 static BlockDriverAIOCB *bdrv_aio_read_em(BlockDriverState *bs,
1444         int64_t sector_num, uint8_t *buf, int nb_sectors,
1445         BlockDriverCompletionFunc *cb, void *opaque)
1446 {
1447     BlockDriverAIOCBSync *acb;
1448     int ret;
1449
1450     acb = qemu_aio_get(bs, cb, opaque);
1451     if (!acb->bh)
1452         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1453     ret = bdrv_read(bs, sector_num, buf, nb_sectors);
1454     acb->ret = ret;
1455     qemu_bh_schedule(acb->bh);
1456     return &acb->common;
1457 }
1458
1459 static BlockDriverAIOCB *bdrv_aio_write_em(BlockDriverState *bs,
1460         int64_t sector_num, const uint8_t *buf, int nb_sectors,
1461         BlockDriverCompletionFunc *cb, void *opaque)
1462 {
1463     BlockDriverAIOCBSync *acb;
1464     int ret;
1465
1466     acb = qemu_aio_get(bs, cb, opaque);
1467     if (!acb->bh)
1468         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1469     ret = bdrv_write(bs, sector_num, buf, nb_sectors);
1470     acb->ret = ret;
1471     qemu_bh_schedule(acb->bh);
1472     return &acb->common;
1473 }
1474
1475 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1476 {
1477     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1478     qemu_bh_cancel(acb->bh);
1479     qemu_aio_release(acb);
1480 }
1481
1482 /**************************************************************/
1483 /* sync block device emulation */
1484
1485 static void bdrv_rw_em_cb(void *opaque, int ret)
1486 {
1487     *(int *)opaque = ret;
1488 }
1489
1490 #define NOT_DONE 0x7fffffff
1491
1492 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1493                         uint8_t *buf, int nb_sectors)
1494 {
1495     int async_ret;
1496     BlockDriverAIOCB *acb;
1497
1498     async_ret = NOT_DONE;
1499     acb = bdrv_aio_read(bs, sector_num, buf, nb_sectors,
1500                         bdrv_rw_em_cb, &async_ret);
1501     if (acb == NULL)
1502         return -1;
1503
1504     while (async_ret == NOT_DONE) {
1505         qemu_aio_wait();
1506     }
1507
1508     return async_ret;
1509 }
1510
1511 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1512                          const uint8_t *buf, int nb_sectors)
1513 {
1514     int async_ret;
1515     BlockDriverAIOCB *acb;
1516
1517     async_ret = NOT_DONE;
1518     acb = bdrv_aio_write(bs, sector_num, buf, nb_sectors,
1519                          bdrv_rw_em_cb, &async_ret);
1520     if (acb == NULL)
1521         return -1;
1522     while (async_ret == NOT_DONE) {
1523         qemu_aio_wait();
1524     }
1525     return async_ret;
1526 }
1527
1528 void bdrv_init(void)
1529 {
1530     aio_pool_init(&vectored_aio_pool, sizeof(VectorTranslationAIOCB),
1531                   bdrv_aio_cancel_vector);
1532
1533     bdrv_register(&bdrv_raw);
1534     bdrv_register(&bdrv_host_device);
1535 #ifndef _WIN32
1536     bdrv_register(&bdrv_cow);
1537 #endif
1538     bdrv_register(&bdrv_qcow);
1539     bdrv_register(&bdrv_vmdk);
1540     bdrv_register(&bdrv_cloop);
1541     bdrv_register(&bdrv_dmg);
1542     bdrv_register(&bdrv_bochs);
1543     bdrv_register(&bdrv_vpc);
1544     bdrv_register(&bdrv_vvfat);
1545     bdrv_register(&bdrv_qcow2);
1546     bdrv_register(&bdrv_parallels);
1547     bdrv_register(&bdrv_nbd);
1548 }
1549
1550 void aio_pool_init(AIOPool *pool, int aiocb_size,
1551                    void (*cancel)(BlockDriverAIOCB *acb))
1552 {
1553     pool->aiocb_size = aiocb_size;
1554     pool->cancel = cancel;
1555     pool->free_aiocb = NULL;
1556 }
1557
1558 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1559                         BlockDriverCompletionFunc *cb, void *opaque)
1560 {
1561     BlockDriverAIOCB *acb;
1562
1563     if (pool->free_aiocb) {
1564         acb = pool->free_aiocb;
1565         pool->free_aiocb = acb->next;
1566     } else {
1567         acb = qemu_mallocz(pool->aiocb_size);
1568         acb->pool = pool;
1569     }
1570     acb->bs = bs;
1571     acb->cb = cb;
1572     acb->opaque = opaque;
1573     return acb;
1574 }
1575
1576 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1577                    void *opaque)
1578 {
1579     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1580 }
1581
1582 void qemu_aio_release(void *p)
1583 {
1584     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1585     AIOPool *pool = acb->pool;
1586     acb->next = pool->free_aiocb;
1587     pool->free_aiocb = acb;
1588 }
1589
1590 /**************************************************************/
1591 /* removable device support */
1592
1593 /**
1594  * Return TRUE if the media is present
1595  */
1596 int bdrv_is_inserted(BlockDriverState *bs)
1597 {
1598     BlockDriver *drv = bs->drv;
1599     int ret;
1600     if (!drv)
1601         return 0;
1602     if (!drv->bdrv_is_inserted)
1603         return 1;
1604     ret = drv->bdrv_is_inserted(bs);
1605     return ret;
1606 }
1607
1608 /**
1609  * Return TRUE if the media changed since the last call to this
1610  * function. It is currently only used for floppy disks
1611  */
1612 int bdrv_media_changed(BlockDriverState *bs)
1613 {
1614     BlockDriver *drv = bs->drv;
1615     int ret;
1616
1617     if (!drv || !drv->bdrv_media_changed)
1618         ret = -ENOTSUP;
1619     else
1620         ret = drv->bdrv_media_changed(bs);
1621     if (ret == -ENOTSUP)
1622         ret = bs->media_changed;
1623     bs->media_changed = 0;
1624     return ret;
1625 }
1626
1627 /**
1628  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1629  */
1630 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1631 {
1632     BlockDriver *drv = bs->drv;
1633     int ret;
1634
1635     if (!drv || !drv->bdrv_eject) {
1636         ret = -ENOTSUP;
1637     } else {
1638         ret = drv->bdrv_eject(bs, eject_flag);
1639     }
1640     if (ret == -ENOTSUP) {
1641         if (eject_flag)
1642             bdrv_close(bs);
1643     }
1644 }
1645
1646 int bdrv_is_locked(BlockDriverState *bs)
1647 {
1648     return bs->locked;
1649 }
1650
1651 /**
1652  * Lock or unlock the media (if it is locked, the user won't be able
1653  * to eject it manually).
1654  */
1655 void bdrv_set_locked(BlockDriverState *bs, int locked)
1656 {
1657     BlockDriver *drv = bs->drv;
1658
1659     bs->locked = locked;
1660     if (drv && drv->bdrv_set_locked) {
1661         drv->bdrv_set_locked(bs, locked);
1662     }
1663 }
1664
1665 /* needed for generic scsi interface */
1666
1667 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1668 {
1669     BlockDriver *drv = bs->drv;
1670
1671     if (drv && drv->bdrv_ioctl)
1672         return drv->bdrv_ioctl(bs, req, buf);
1673     return -ENOTSUP;
1674 }
1675
1676 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1677         unsigned long int req, void *buf,
1678         BlockDriverCompletionFunc *cb, void *opaque)
1679 {
1680     BlockDriver *drv = bs->drv;
1681
1682     if (drv && drv->bdrv_aio_ioctl)
1683         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1684     return NULL;
1685 }