block.c

   1 /*
   2  * QEMU System Emulator block driver
   3  *
   4  * Copyright (c) 2003 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include "config-host.h"
  25 #ifdef HOST_BSD
  26 /* include native header before sys-queue.h */
  27 #include <sys/queue.h>
  28 #endif
  29
  30 #include "qemu-common.h"
  31 #include "monitor.h"
  32 #include "block_int.h"
  33 #include "module.h"
  34
  35 #ifdef HOST_BSD
  36 #include <sys/types.h>
  37 #include <sys/stat.h>
  38 #include <sys/ioctl.h>
  39 #ifndef __DragonFly__
  40 #include <sys/disk.h>
  41 #endif
  42 #endif
  43
  44 #ifdef _WIN32
  45 #include <windows.h>
  46 #endif
  47
  48 #define SECTOR_BITS 9
  49 #define SECTOR_SIZE (1 << SECTOR_BITS)
  50
  51 typedef struct BlockDriverAIOCBSync {
  52     BlockDriverAIOCB common;
  53     QEMUBH *bh;
  54     int ret;
  55     /* vector translation state */
  56     QEMUIOVector *qiov;
  57     uint8_t *bounce;
  58     int is_write;
  59 } BlockDriverAIOCBSync;
  60
  61 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
  62         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  63         BlockDriverCompletionFunc *cb, void *opaque);
  64 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
  65         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
  66         BlockDriverCompletionFunc *cb, void *opaque);
  67 static void bdrv_aio_cancel_em(BlockDriverAIOCB *acb);
  68 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
  69                         uint8_t *buf, int nb_sectors);
  70 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
  71                          const uint8_t *buf, int nb_sectors);
  72
  73 BlockDriverState *bdrv_first;
  74
  75 static BlockDriver *first_drv;
  76
  77 int path_is_absolute(const char *path)
  78 {
  79     const char *p;
  80 #ifdef _WIN32
  81     /* specific case for names like: "\\.\d:" */
  82     if (*path == '/' || *path == '\\')
  83         return 1;
  84 #endif
  85     p = strchr(path, ':');
  86     if (p)
  87         p++;
  88     else
  89         p = path;
  90 #ifdef _WIN32
  91     return (*p == '/' || *p == '\\');
  92 #else
  93     return (*p == '/');
  94 #endif
  95 }
  96
  97 /* if filename is absolute, just copy it to dest. Otherwise, build a
  98    path to it by considering it is relative to base_path. URL are
  99    supported. */
 100 void path_combine(char *dest, int dest_size,
 101                   const char *base_path,
 102                   const char *filename)
 103 {
 104     const char *p, *p1;
 105     int len;
 106
 107     if (dest_size <= 0)
 108         return;
 109     if (path_is_absolute(filename)) {
 110         pstrcpy(dest, dest_size, filename);
 111     } else {
 112         p = strchr(base_path, ':');
 113         if (p)
 114             p++;
 115         else
 116             p = base_path;
 117         p1 = strrchr(base_path, '/');
 118 #ifdef _WIN32
 119         {
 120             const char *p2;
 121             p2 = strrchr(base_path, '\\');
 122             if (!p1 || p2 > p1)
 123                 p1 = p2;
 124         }
 125 #endif
 126         if (p1)
 127             p1++;
 128         else
 129             p1 = base_path;
 130         if (p1 > p)
 131             p = p1;
 132         len = p - base_path;
 133         if (len > dest_size - 1)
 134             len = dest_size - 1;
 135         memcpy(dest, base_path, len);
 136         dest[len] = '\0';
 137         pstrcat(dest, dest_size, filename);
 138     }
 139 }
 140
 141
 142 void bdrv_register(BlockDriver *bdrv)
 143 {
 144     if (!bdrv->bdrv_aio_readv) {
 145         /* add AIO emulation layer */
 146         bdrv->bdrv_aio_readv = bdrv_aio_readv_em;
 147         bdrv->bdrv_aio_writev = bdrv_aio_writev_em;
 148         bdrv->bdrv_aio_cancel = bdrv_aio_cancel_em;
 149         bdrv->aiocb_size = sizeof(BlockDriverAIOCBSync);
 150     } else if (!bdrv->bdrv_read) {
 151         /* add synchronous IO emulation layer */
 152         bdrv->bdrv_read = bdrv_read_em;
 153         bdrv->bdrv_write = bdrv_write_em;
 154     }
 155     aio_pool_init(&bdrv->aio_pool, bdrv->aiocb_size, bdrv->bdrv_aio_cancel);
 156     bdrv->next = first_drv;
 157     first_drv = bdrv;
 158 }
 159
 160 /* create a new block device (by default it is empty) */
 161 BlockDriverState *bdrv_new(const char *device_name)
 162 {
 163     BlockDriverState **pbs, *bs;
 164
 165     bs = qemu_mallocz(sizeof(BlockDriverState));
 166     pstrcpy(bs->device_name, sizeof(bs->device_name), device_name);
 167     if (device_name[0] != '\0') {
 168         /* insert at the end */
 169         pbs = &bdrv_first;
 170         while (*pbs != NULL)
 171             pbs = &(*pbs)->next;
 172         *pbs = bs;
 173     }
 174     return bs;
 175 }
 176
 177 BlockDriver *bdrv_find_format(const char *format_name)
 178 {
 179     BlockDriver *drv1;
 180     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 181         if (!strcmp(drv1->format_name, format_name))
 182             return drv1;
 183     }
 184     return NULL;
 185 }
 186
 187 int bdrv_create2(BlockDriver *drv,
 188                 const char *filename, int64_t size_in_sectors,
 189                 const char *backing_file, const char *backing_format,
 190                 int flags)
 191 {
 192     if (drv->bdrv_create2)
 193         return drv->bdrv_create2(filename, size_in_sectors, backing_file,
 194                                  backing_format, flags);
 195     if (drv->bdrv_create)
 196         return drv->bdrv_create(filename, size_in_sectors, backing_file,
 197                                 flags);
 198     return -ENOTSUP;
 199 }
 200
 201 int bdrv_create(BlockDriver *drv,
 202                 const char *filename, int64_t size_in_sectors,
 203                 const char *backing_file, int flags)
 204 {
 205     if (!drv->bdrv_create)
 206         return -ENOTSUP;
 207     return drv->bdrv_create(filename, size_in_sectors, backing_file, flags);
 208 }
 209
 210 #ifdef _WIN32
 211 void get_tmp_filename(char *filename, int size)
 212 {
 213     char temp_dir[MAX_PATH];
 214
 215     GetTempPath(MAX_PATH, temp_dir);
 216     GetTempFileName(temp_dir, "qem", 0, filename);
 217 }
 218 #else
 219 void get_tmp_filename(char *filename, int size)
 220 {
 221     int fd;
 222     const char *tmpdir;
 223     /* XXX: race condition possible */
 224     tmpdir = getenv("TMPDIR");
 225     if (!tmpdir)
 226         tmpdir = "/tmp";
 227     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 228     fd = mkstemp(filename);
 229     close(fd);
 230 }
 231 #endif
 232
 233 #ifdef _WIN32
 234 static int is_windows_drive_prefix(const char *filename)
 235 {
 236     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 237              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 238             filename[1] == ':');
 239 }
 240
 241 static int is_windows_drive(const char *filename)
 242 {
 243     if (is_windows_drive_prefix(filename) &&
 244         filename[2] == '\0')
 245         return 1;
 246     if (strstart(filename, "\\\\.\\", NULL) ||
 247         strstart(filename, "//./", NULL))
 248         return 1;
 249     return 0;
 250 }
 251 #endif
 252
 253 static BlockDriver *find_protocol(const char *filename)
 254 {
 255     BlockDriver *drv1;
 256     char protocol[128];
 257     int len;
 258     const char *p;
 259
 260 #ifdef _WIN32
 261     if (is_windows_drive(filename) ||
 262         is_windows_drive_prefix(filename))
 263         return bdrv_find_format("raw");
 264 #endif
 265     p = strchr(filename, ':');
 266     if (!p)
 267         return bdrv_find_format("raw");
 268     len = p - filename;
 269     if (len > sizeof(protocol) - 1)
 270         len = sizeof(protocol) - 1;
 271     memcpy(protocol, filename, len);
 272     protocol[len] = '\0';
 273     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 274         if (drv1->protocol_name &&
 275             !strcmp(drv1->protocol_name, protocol))
 276             return drv1;
 277     }
 278     return NULL;
 279 }
 280
 281 /* XXX: force raw format if block or character device ? It would
 282    simplify the BSD case */
 283 static BlockDriver *find_image_format(const char *filename)
 284 {
 285     int ret, score, score_max;
 286     BlockDriver *drv1, *drv;
 287     uint8_t buf[2048];
 288     BlockDriverState *bs;
 289
 290     /* detect host devices. By convention, /dev/cdrom[N] is always
 291        recognized as a host CDROM */
 292     if (strstart(filename, "/dev/cdrom", NULL))
 293         return bdrv_find_format("host_device");
 294 #ifdef _WIN32
 295     if (is_windows_drive(filename))
 296         return bdrv_find_format("host_device");
 297 #else
 298     {
 299         struct stat st;
 300         if (stat(filename, &st) >= 0 &&
 301             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 302             return bdrv_find_format("host_device");
 303         }
 304     }
 305 #endif
 306
 307     drv = find_protocol(filename);
 308     /* no need to test disk image formats for vvfat */
 309     if (strcmp(drv->format_name, "vvfat") == 0)
 310         return drv;
 311
 312     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 313     if (ret < 0)
 314         return NULL;
 315     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 316     bdrv_delete(bs);
 317     if (ret < 0) {
 318         return NULL;
 319     }
 320
 321     score_max = 0;
 322     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 323         if (drv1->bdrv_probe) {
 324             score = drv1->bdrv_probe(buf, ret, filename);
 325             if (score > score_max) {
 326                 score_max = score;
 327                 drv = drv1;
 328             }
 329         }
 330     }
 331     return drv;
 332 }
 333
 334 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 335 {
 336     BlockDriverState *bs;
 337     int ret;
 338
 339     bs = bdrv_new("");
 340     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 341     if (ret < 0) {
 342         bdrv_delete(bs);
 343         return ret;
 344     }
 345     bs->growable = 1;
 346     *pbs = bs;
 347     return 0;
 348 }
 349
 350 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 351 {
 352     return bdrv_open2(bs, filename, flags, NULL);
 353 }
 354
 355 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 356                BlockDriver *drv)
 357 {
 358     int ret, open_flags;
 359     char tmp_filename[PATH_MAX];
 360     char backing_filename[PATH_MAX];
 361
 362     bs->read_only = 0;
 363     bs->is_temporary = 0;
 364     bs->encrypted = 0;
 365     bs->valid_key = 0;
 366     /* buffer_alignment defaulted to 512, drivers can change this value */
 367     bs->buffer_alignment = 512;
 368
 369     if (flags & BDRV_O_SNAPSHOT) {
 370         BlockDriverState *bs1;
 371         int64_t total_size;
 372         int is_protocol = 0;
 373
 374         /* if snapshot, we create a temporary backing file and open it
 375            instead of opening 'filename' directly */
 376
 377         /* if there is a backing file, use it */
 378         bs1 = bdrv_new("");
 379         ret = bdrv_open2(bs1, filename, 0, drv);
 380         if (ret < 0) {
 381             bdrv_delete(bs1);
 382             return ret;
 383         }
 384         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 385
 386         if (bs1->drv && bs1->drv->protocol_name)
 387             is_protocol = 1;
 388
 389         bdrv_delete(bs1);
 390
 391         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 392
 393         /* Real path is meaningless for protocols */
 394         if (is_protocol)
 395             snprintf(backing_filename, sizeof(backing_filename),
 396                      "%s", filename);
 397         else
 398             realpath(filename, backing_filename);
 399
 400         ret = bdrv_create2(bdrv_find_format("qcow2"), tmp_filename,
 401                            total_size, backing_filename,
 402                            (drv ? drv->format_name : NULL), 0);
 403         if (ret < 0) {
 404             return ret;
 405         }
 406         filename = tmp_filename;
 407         drv = bdrv_find_format("qcow2");
 408         bs->is_temporary = 1;
 409     }
 410
 411     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 412     if (flags & BDRV_O_FILE) {
 413         drv = find_protocol(filename);
 414     } else if (!drv) {
 415         drv = find_image_format(filename);
 416     }
 417     if (!drv) {
 418         ret = -ENOENT;
 419         goto unlink_and_fail;
 420     }
 421     bs->drv = drv;
 422     bs->opaque = qemu_mallocz(drv->instance_size);
 423     /* Note: for compatibility, we open disk image files as RDWR, and
 424        RDONLY as fallback */
 425     if (!(flags & BDRV_O_FILE))
 426         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 427     else
 428         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 429     ret = drv->bdrv_open(bs, filename, open_flags);
 430     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 431         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 432         bs->read_only = 1;
 433     }
 434     if (ret < 0) {
 435         qemu_free(bs->opaque);
 436         bs->opaque = NULL;
 437         bs->drv = NULL;
 438     unlink_and_fail:
 439         if (bs->is_temporary)
 440             unlink(filename);
 441         return ret;
 442     }
 443     if (drv->bdrv_getlength) {
 444         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 445     }
 446 #ifndef _WIN32
 447     if (bs->is_temporary) {
 448         unlink(filename);
 449     }
 450 #endif
 451     if (bs->backing_file[0] != '\0') {
 452         /* if there is a backing file, use it */
 453         BlockDriver *back_drv = NULL;
 454         bs->backing_hd = bdrv_new("");
 455         path_combine(backing_filename, sizeof(backing_filename),
 456                      filename, bs->backing_file);
 457         if (bs->backing_format[0] != '\0')
 458             back_drv = bdrv_find_format(bs->backing_format);
 459         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 460                          back_drv);
 461         if (ret < 0) {
 462             bdrv_close(bs);
 463             return ret;
 464         }
 465     }
 466
 467     if (!bdrv_key_required(bs)) {
 468         /* call the change callback */
 469         bs->media_changed = 1;
 470         if (bs->change_cb)
 471             bs->change_cb(bs->change_opaque);
 472     }
 473     return 0;
 474 }
 475
 476 void bdrv_close(BlockDriverState *bs)
 477 {
 478     if (bs->drv) {
 479         if (bs->backing_hd)
 480             bdrv_delete(bs->backing_hd);
 481         bs->drv->bdrv_close(bs);
 482         qemu_free(bs->opaque);
 483 #ifdef _WIN32
 484         if (bs->is_temporary) {
 485             unlink(bs->filename);
 486         }
 487 #endif
 488         bs->opaque = NULL;
 489         bs->drv = NULL;
 490
 491         /* call the change callback */
 492         bs->media_changed = 1;
 493         if (bs->change_cb)
 494             bs->change_cb(bs->change_opaque);
 495     }
 496 }
 497
 498 void bdrv_delete(BlockDriverState *bs)
 499 {
 500     BlockDriverState **pbs;
 501
 502     pbs = &bdrv_first;
 503     while (*pbs != bs && *pbs != NULL)
 504         pbs = &(*pbs)->next;
 505     if (*pbs == bs)
 506         *pbs = bs->next;
 507
 508     bdrv_close(bs);
 509     qemu_free(bs);
 510 }
 511
 512 /*
 513  * Run consistency checks on an image
 514  *
 515  * Returns the number of errors or -errno when an internal error occurs
 516  */
 517 int bdrv_check(BlockDriverState *bs)
 518 {
 519     if (bs->drv->bdrv_check == NULL) {
 520         return -ENOTSUP;
 521     }
 522
 523     return bs->drv->bdrv_check(bs);
 524 }
 525
 526 /* commit COW file into the raw image */
 527 int bdrv_commit(BlockDriverState *bs)
 528 {
 529     BlockDriver *drv = bs->drv;
 530     int64_t i, total_sectors;
 531     int n, j;
 532     unsigned char sector[512];
 533
 534     if (!drv)
 535         return -ENOMEDIUM;
 536
 537     if (bs->read_only) {
 538         return -EACCES;
 539     }
 540
 541     if (!bs->backing_hd) {
 542         return -ENOTSUP;
 543     }
 544
 545     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 546     for (i = 0; i < total_sectors;) {
 547         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 548             for(j = 0; j < n; j++) {
 549                 if (bdrv_read(bs, i, sector, 1) != 0) {
 550                     return -EIO;
 551                 }
 552
 553                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 554                     return -EIO;
 555                 }
 556                 i++;
 557             }
 558         } else {
 559             i += n;
 560         }
 561     }
 562
 563     if (drv->bdrv_make_empty)
 564         return drv->bdrv_make_empty(bs);
 565
 566     return 0;
 567 }
 568
 569 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 570                                    size_t size)
 571 {
 572     int64_t len;
 573
 574     if (!bdrv_is_inserted(bs))
 575         return -ENOMEDIUM;
 576
 577     if (bs->growable)
 578         return 0;
 579
 580     len = bdrv_getlength(bs);
 581
 582     if (offset < 0)
 583         return -EIO;
 584
 585     if ((offset > len) || (len - offset < size))
 586         return -EIO;
 587
 588     return 0;
 589 }
 590
 591 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 592                               int nb_sectors)
 593 {
 594     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 595 }
 596
 597 /* return < 0 if error. See bdrv_write() for the return codes */
 598 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 599               uint8_t *buf, int nb_sectors)
 600 {
 601     BlockDriver *drv = bs->drv;
 602
 603     if (!drv)
 604         return -ENOMEDIUM;
 605     if (bdrv_check_request(bs, sector_num, nb_sectors))
 606         return -EIO;
 607
 608     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 609 }
 610
 611 /* Return < 0 if error. Important errors are:
 612   -EIO         generic I/O error (may happen for all errors)
 613   -ENOMEDIUM   No media inserted.
 614   -EINVAL      Invalid sector number or nb_sectors
 615   -EACCES      Trying to write a read-only device
 616 */
 617 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 618                const uint8_t *buf, int nb_sectors)
 619 {
 620     BlockDriver *drv = bs->drv;
 621     if (!bs->drv)
 622         return -ENOMEDIUM;
 623     if (bs->read_only)
 624         return -EACCES;
 625     if (bdrv_check_request(bs, sector_num, nb_sectors))
 626         return -EIO;
 627
 628     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 629 }
 630
 631 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 632                void *buf, int count1)
 633 {
 634     uint8_t tmp_buf[SECTOR_SIZE];
 635     int len, nb_sectors, count;
 636     int64_t sector_num;
 637
 638     count = count1;
 639     /* first read to align to sector start */
 640     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 641     if (len > count)
 642         len = count;
 643     sector_num = offset >> SECTOR_BITS;
 644     if (len > 0) {
 645         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 646             return -EIO;
 647         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 648         count -= len;
 649         if (count == 0)
 650             return count1;
 651         sector_num++;
 652         buf += len;
 653     }
 654
 655     /* read the sectors "in place" */
 656     nb_sectors = count >> SECTOR_BITS;
 657     if (nb_sectors > 0) {
 658         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 659             return -EIO;
 660         sector_num += nb_sectors;
 661         len = nb_sectors << SECTOR_BITS;
 662         buf += len;
 663         count -= len;
 664     }
 665
 666     /* add data from the last sector */
 667     if (count > 0) {
 668         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 669             return -EIO;
 670         memcpy(buf, tmp_buf, count);
 671     }
 672     return count1;
 673 }
 674
 675 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 676                 const void *buf, int count1)
 677 {
 678     uint8_t tmp_buf[SECTOR_SIZE];
 679     int len, nb_sectors, count;
 680     int64_t sector_num;
 681
 682     count = count1;
 683     /* first write to align to sector start */
 684     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 685     if (len > count)
 686         len = count;
 687     sector_num = offset >> SECTOR_BITS;
 688     if (len > 0) {
 689         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 690             return -EIO;
 691         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 692         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 693             return -EIO;
 694         count -= len;
 695         if (count == 0)
 696             return count1;
 697         sector_num++;
 698         buf += len;
 699     }
 700
 701     /* write the sectors "in place" */
 702     nb_sectors = count >> SECTOR_BITS;
 703     if (nb_sectors > 0) {
 704         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 705             return -EIO;
 706         sector_num += nb_sectors;
 707         len = nb_sectors << SECTOR_BITS;
 708         buf += len;
 709         count -= len;
 710     }
 711
 712     /* add data from the last sector */
 713     if (count > 0) {
 714         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 715             return -EIO;
 716         memcpy(tmp_buf, buf, count);
 717         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 718             return -EIO;
 719     }
 720     return count1;
 721 }
 722
 723 /**
 724  * Truncate file to 'offset' bytes (needed only for file protocols)
 725  */
 726 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 727 {
 728     BlockDriver *drv = bs->drv;
 729     if (!drv)
 730         return -ENOMEDIUM;
 731     if (!drv->bdrv_truncate)
 732         return -ENOTSUP;
 733     return drv->bdrv_truncate(bs, offset);
 734 }
 735
 736 /**
 737  * Length of a file in bytes. Return < 0 if error or unknown.
 738  */
 739 int64_t bdrv_getlength(BlockDriverState *bs)
 740 {
 741     BlockDriver *drv = bs->drv;
 742     if (!drv)
 743         return -ENOMEDIUM;
 744     if (!drv->bdrv_getlength) {
 745         /* legacy mode */
 746         return bs->total_sectors * SECTOR_SIZE;
 747     }
 748     return drv->bdrv_getlength(bs);
 749 }
 750
 751 /* return 0 as number of sectors if no device present or error */
 752 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 753 {
 754     int64_t length;
 755     length = bdrv_getlength(bs);
 756     if (length < 0)
 757         length = 0;
 758     else
 759         length = length >> SECTOR_BITS;
 760     *nb_sectors_ptr = length;
 761 }
 762
 763 struct partition {
 764         uint8_t boot_ind;           /* 0x80 - active */
 765         uint8_t head;               /* starting head */
 766         uint8_t sector;             /* starting sector */
 767         uint8_t cyl;                /* starting cylinder */
 768         uint8_t sys_ind;            /* What partition type */
 769         uint8_t end_head;           /* end head */
 770         uint8_t end_sector;         /* end sector */
 771         uint8_t end_cyl;            /* end cylinder */
 772         uint32_t start_sect;        /* starting sector counting from 0 */
 773         uint32_t nr_sects;          /* nr of sectors in partition */
 774 } __attribute__((packed));
 775
 776 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 777 static int guess_disk_lchs(BlockDriverState *bs,
 778                            int *pcylinders, int *pheads, int *psectors)
 779 {
 780     uint8_t buf[512];
 781     int ret, i, heads, sectors, cylinders;
 782     struct partition *p;
 783     uint32_t nr_sects;
 784     uint64_t nb_sectors;
 785
 786     bdrv_get_geometry(bs, &nb_sectors);
 787
 788     ret = bdrv_read(bs, 0, buf, 1);
 789     if (ret < 0)
 790         return -1;
 791     /* test msdos magic */
 792     if (buf[510] != 0x55 || buf[511] != 0xaa)
 793         return -1;
 794     for(i = 0; i < 4; i++) {
 795         p = ((struct partition *)(buf + 0x1be)) + i;
 796         nr_sects = le32_to_cpu(p->nr_sects);
 797         if (nr_sects && p->end_head) {
 798             /* We make the assumption that the partition terminates on
 799                a cylinder boundary */
 800             heads = p->end_head + 1;
 801             sectors = p->end_sector & 63;
 802             if (sectors == 0)
 803                 continue;
 804             cylinders = nb_sectors / (heads * sectors);
 805             if (cylinders < 1 || cylinders > 16383)
 806                 continue;
 807             *pheads = heads;
 808             *psectors = sectors;
 809             *pcylinders = cylinders;
 810 #if 0
 811             printf("guessed geometry: LCHS=%d %d %d\n",
 812                    cylinders, heads, sectors);
 813 #endif
 814             return 0;
 815         }
 816     }
 817     return -1;
 818 }
 819
 820 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 821 {
 822     int translation, lba_detected = 0;
 823     int cylinders, heads, secs;
 824     uint64_t nb_sectors;
 825
 826     /* if a geometry hint is available, use it */
 827     bdrv_get_geometry(bs, &nb_sectors);
 828     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 829     translation = bdrv_get_translation_hint(bs);
 830     if (cylinders != 0) {
 831         *pcyls = cylinders;
 832         *pheads = heads;
 833         *psecs = secs;
 834     } else {
 835         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 836             if (heads > 16) {
 837                 /* if heads > 16, it means that a BIOS LBA
 838                    translation was active, so the default
 839                    hardware geometry is OK */
 840                 lba_detected = 1;
 841                 goto default_geometry;
 842             } else {
 843                 *pcyls = cylinders;
 844                 *pheads = heads;
 845                 *psecs = secs;
 846                 /* disable any translation to be in sync with
 847                    the logical geometry */
 848                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 849                     bdrv_set_translation_hint(bs,
 850                                               BIOS_ATA_TRANSLATION_NONE);
 851                 }
 852             }
 853         } else {
 854         default_geometry:
 855             /* if no geometry, use a standard physical disk geometry */
 856             cylinders = nb_sectors / (16 * 63);
 857
 858             if (cylinders > 16383)
 859                 cylinders = 16383;
 860             else if (cylinders < 2)
 861                 cylinders = 2;
 862             *pcyls = cylinders;
 863             *pheads = 16;
 864             *psecs = 63;
 865             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 866                 if ((*pcyls * *pheads) <= 131072) {
 867                     bdrv_set_translation_hint(bs,
 868                                               BIOS_ATA_TRANSLATION_LARGE);
 869                 } else {
 870                     bdrv_set_translation_hint(bs,
 871                                               BIOS_ATA_TRANSLATION_LBA);
 872                 }
 873             }
 874         }
 875         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 876     }
 877 }
 878
 879 void bdrv_set_geometry_hint(BlockDriverState *bs,
 880                             int cyls, int heads, int secs)
 881 {
 882     bs->cyls = cyls;
 883     bs->heads = heads;
 884     bs->secs = secs;
 885 }
 886
 887 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 888 {
 889     bs->type = type;
 890     bs->removable = ((type == BDRV_TYPE_CDROM ||
 891                       type == BDRV_TYPE_FLOPPY));
 892 }
 893
 894 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 895 {
 896     bs->translation = translation;
 897 }
 898
 899 void bdrv_get_geometry_hint(BlockDriverState *bs,
 900                             int *pcyls, int *pheads, int *psecs)
 901 {
 902     *pcyls = bs->cyls;
 903     *pheads = bs->heads;
 904     *psecs = bs->secs;
 905 }
 906
 907 int bdrv_get_type_hint(BlockDriverState *bs)
 908 {
 909     return bs->type;
 910 }
 911
 912 int bdrv_get_translation_hint(BlockDriverState *bs)
 913 {
 914     return bs->translation;
 915 }
 916
 917 int bdrv_is_removable(BlockDriverState *bs)
 918 {
 919     return bs->removable;
 920 }
 921
 922 int bdrv_is_read_only(BlockDriverState *bs)
 923 {
 924     return bs->read_only;
 925 }
 926
 927 int bdrv_is_sg(BlockDriverState *bs)
 928 {
 929     return bs->sg;
 930 }
 931
 932 /* XXX: no longer used */
 933 void bdrv_set_change_cb(BlockDriverState *bs,
 934                         void (*change_cb)(void *opaque), void *opaque)
 935 {
 936     bs->change_cb = change_cb;
 937     bs->change_opaque = opaque;
 938 }
 939
 940 int bdrv_is_encrypted(BlockDriverState *bs)
 941 {
 942     if (bs->backing_hd && bs->backing_hd->encrypted)
 943         return 1;
 944     return bs->encrypted;
 945 }
 946
 947 int bdrv_key_required(BlockDriverState *bs)
 948 {
 949     BlockDriverState *backing_hd = bs->backing_hd;
 950
 951     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 952         return 1;
 953     return (bs->encrypted && !bs->valid_key);
 954 }
 955
 956 int bdrv_set_key(BlockDriverState *bs, const char *key)
 957 {
 958     int ret;
 959     if (bs->backing_hd && bs->backing_hd->encrypted) {
 960         ret = bdrv_set_key(bs->backing_hd, key);
 961         if (ret < 0)
 962             return ret;
 963         if (!bs->encrypted)
 964             return 0;
 965     }
 966     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 967         return -1;
 968     ret = bs->drv->bdrv_set_key(bs, key);
 969     if (ret < 0) {
 970         bs->valid_key = 0;
 971     } else if (!bs->valid_key) {
 972         bs->valid_key = 1;
 973         /* call the change callback now, we skipped it on open */
 974         bs->media_changed = 1;
 975         if (bs->change_cb)
 976             bs->change_cb(bs->change_opaque);
 977     }
 978     return ret;
 979 }
 980
 981 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 982 {
 983     if (!bs->drv) {
 984         buf[0] = '\0';
 985     } else {
 986         pstrcpy(buf, buf_size, bs->drv->format_name);
 987     }
 988 }
 989
 990 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 991                          void *opaque)
 992 {
 993     BlockDriver *drv;
 994
 995     for (drv = first_drv; drv != NULL; drv = drv->next) {
 996         it(opaque, drv->format_name);
 997     }
 998 }
 999
1000 BlockDriverState *bdrv_find(const char *name)
1001 {
1002     BlockDriverState *bs;
1003
1004     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1005         if (!strcmp(name, bs->device_name))
1006             return bs;
1007     }
1008     return NULL;
1009 }
1010
1011 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1012 {
1013     BlockDriverState *bs;
1014
1015     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1016         it(opaque, bs);
1017     }
1018 }
1019
1020 const char *bdrv_get_device_name(BlockDriverState *bs)
1021 {
1022     return bs->device_name;
1023 }
1024
1025 void bdrv_flush(BlockDriverState *bs)
1026 {
1027     if (!bs->drv)
1028         return;
1029     if (bs->drv->bdrv_flush)
1030         bs->drv->bdrv_flush(bs);
1031     if (bs->backing_hd)
1032         bdrv_flush(bs->backing_hd);
1033 }
1034
1035 void bdrv_flush_all(void)
1036 {
1037     BlockDriverState *bs;
1038
1039     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1040         if (bs->drv && !bdrv_is_read_only(bs) &&
1041             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1042             bdrv_flush(bs);
1043 }
1044
1045 /*
1046  * Returns true iff the specified sector is present in the disk image. Drivers
1047  * not implementing the functionality are assumed to not support backing files,
1048  * hence all their sectors are reported as allocated.
1049  *
1050  * 'pnum' is set to the number of sectors (including and immediately following
1051  * the specified sector) that are known to be in the same
1052  * allocated/unallocated state.
1053  *
1054  * 'nb_sectors' is the max value 'pnum' should be set to.
1055  */
1056 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1057         int *pnum)
1058 {
1059     int64_t n;
1060     if (!bs->drv->bdrv_is_allocated) {
1061         if (sector_num >= bs->total_sectors) {
1062             *pnum = 0;
1063             return 0;
1064         }
1065         n = bs->total_sectors - sector_num;
1066         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1067         return 1;
1068     }
1069     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1070 }
1071
1072 void bdrv_info(Monitor *mon)
1073 {
1074     BlockDriverState *bs;
1075
1076     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1077         monitor_printf(mon, "%s:", bs->device_name);
1078         monitor_printf(mon, " type=");
1079         switch(bs->type) {
1080         case BDRV_TYPE_HD:
1081             monitor_printf(mon, "hd");
1082             break;
1083         case BDRV_TYPE_CDROM:
1084             monitor_printf(mon, "cdrom");
1085             break;
1086         case BDRV_TYPE_FLOPPY:
1087             monitor_printf(mon, "floppy");
1088             break;
1089         }
1090         monitor_printf(mon, " removable=%d", bs->removable);
1091         if (bs->removable) {
1092             monitor_printf(mon, " locked=%d", bs->locked);
1093         }
1094         if (bs->drv) {
1095             monitor_printf(mon, " file=");
1096             monitor_print_filename(mon, bs->filename);
1097             if (bs->backing_file[0] != '\0') {
1098                 monitor_printf(mon, " backing_file=");
1099                 monitor_print_filename(mon, bs->backing_file);
1100             }
1101             monitor_printf(mon, " ro=%d", bs->read_only);
1102             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1103             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1104         } else {
1105             monitor_printf(mon, " [not inserted]");
1106         }
1107         monitor_printf(mon, "\n");
1108     }
1109 }
1110
1111 /* The "info blockstats" command. */
1112 void bdrv_info_stats(Monitor *mon)
1113 {
1114     BlockDriverState *bs;
1115
1116     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1117         monitor_printf(mon, "%s:"
1118                        " rd_bytes=%" PRIu64
1119                        " wr_bytes=%" PRIu64
1120                        " rd_operations=%" PRIu64
1121                        " wr_operations=%" PRIu64
1122                        "\n",
1123                        bs->device_name,
1124                        bs->rd_bytes, bs->wr_bytes,
1125                        bs->rd_ops, bs->wr_ops);
1126     }
1127 }
1128
1129 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1130 {
1131     if (bs->backing_hd && bs->backing_hd->encrypted)
1132         return bs->backing_file;
1133     else if (bs->encrypted)
1134         return bs->filename;
1135     else
1136         return NULL;
1137 }
1138
1139 void bdrv_get_backing_filename(BlockDriverState *bs,
1140                                char *filename, int filename_size)
1141 {
1142     if (!bs->backing_hd) {
1143         pstrcpy(filename, filename_size, "");
1144     } else {
1145         pstrcpy(filename, filename_size, bs->backing_file);
1146     }
1147 }
1148
1149 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1150                           const uint8_t *buf, int nb_sectors)
1151 {
1152     BlockDriver *drv = bs->drv;
1153     if (!drv)
1154         return -ENOMEDIUM;
1155     if (!drv->bdrv_write_compressed)
1156         return -ENOTSUP;
1157     if (bdrv_check_request(bs, sector_num, nb_sectors))
1158         return -EIO;
1159     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1160 }
1161
1162 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1163 {
1164     BlockDriver *drv = bs->drv;
1165     if (!drv)
1166         return -ENOMEDIUM;
1167     if (!drv->bdrv_get_info)
1168         return -ENOTSUP;
1169     memset(bdi, 0, sizeof(*bdi));
1170     return drv->bdrv_get_info(bs, bdi);
1171 }
1172
1173 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1174 {
1175     BlockDriver *drv = bs->drv;
1176     if (!drv)
1177         return -ENOMEDIUM;
1178     if (!drv->bdrv_put_buffer)
1179         return -ENOTSUP;
1180     return drv->bdrv_put_buffer(bs, buf, pos, size);
1181 }
1182
1183 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1184 {
1185     BlockDriver *drv = bs->drv;
1186     if (!drv)
1187         return -ENOMEDIUM;
1188     if (!drv->bdrv_get_buffer)
1189         return -ENOTSUP;
1190     return drv->bdrv_get_buffer(bs, buf, pos, size);
1191 }
1192
1193 /**************************************************************/
1194 /* handling of snapshots */
1195
1196 int bdrv_snapshot_create(BlockDriverState *bs,
1197                          QEMUSnapshotInfo *sn_info)
1198 {
1199     BlockDriver *drv = bs->drv;
1200     if (!drv)
1201         return -ENOMEDIUM;
1202     if (!drv->bdrv_snapshot_create)
1203         return -ENOTSUP;
1204     return drv->bdrv_snapshot_create(bs, sn_info);
1205 }
1206
1207 int bdrv_snapshot_goto(BlockDriverState *bs,
1208                        const char *snapshot_id)
1209 {
1210     BlockDriver *drv = bs->drv;
1211     if (!drv)
1212         return -ENOMEDIUM;
1213     if (!drv->bdrv_snapshot_goto)
1214         return -ENOTSUP;
1215     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1216 }
1217
1218 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1219 {
1220     BlockDriver *drv = bs->drv;
1221     if (!drv)
1222         return -ENOMEDIUM;
1223     if (!drv->bdrv_snapshot_delete)
1224         return -ENOTSUP;
1225     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1226 }
1227
1228 int bdrv_snapshot_list(BlockDriverState *bs,
1229                        QEMUSnapshotInfo **psn_info)
1230 {
1231     BlockDriver *drv = bs->drv;
1232     if (!drv)
1233         return -ENOMEDIUM;
1234     if (!drv->bdrv_snapshot_list)
1235         return -ENOTSUP;
1236     return drv->bdrv_snapshot_list(bs, psn_info);
1237 }
1238
1239 #define NB_SUFFIXES 4
1240
1241 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1242 {
1243     static const char suffixes[NB_SUFFIXES] = "KMGT";
1244     int64_t base;
1245     int i;
1246
1247     if (size <= 999) {
1248         snprintf(buf, buf_size, "%" PRId64, size);
1249     } else {
1250         base = 1024;
1251         for(i = 0; i < NB_SUFFIXES; i++) {
1252             if (size < (10 * base)) {
1253                 snprintf(buf, buf_size, "%0.1f%c",
1254                          (double)size / base,
1255                          suffixes[i]);
1256                 break;
1257             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1258                 snprintf(buf, buf_size, "%" PRId64 "%c",
1259                          ((size + (base >> 1)) / base),
1260                          suffixes[i]);
1261                 break;
1262             }
1263             base = base * 1024;
1264         }
1265     }
1266     return buf;
1267 }
1268
1269 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1270 {
1271     char buf1[128], date_buf[128], clock_buf[128];
1272 #ifdef _WIN32
1273     struct tm *ptm;
1274 #else
1275     struct tm tm;
1276 #endif
1277     time_t ti;
1278     int64_t secs;
1279
1280     if (!sn) {
1281         snprintf(buf, buf_size,
1282                  "%-10s%-20s%7s%20s%15s",
1283                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1284     } else {
1285         ti = sn->date_sec;
1286 #ifdef _WIN32
1287         ptm = localtime(&ti);
1288         strftime(date_buf, sizeof(date_buf),
1289                  "%Y-%m-%d %H:%M:%S", ptm);
1290 #else
1291         localtime_r(&ti, &tm);
1292         strftime(date_buf, sizeof(date_buf),
1293                  "%Y-%m-%d %H:%M:%S", &tm);
1294 #endif
1295         secs = sn->vm_clock_nsec / 1000000000;
1296         snprintf(clock_buf, sizeof(clock_buf),
1297                  "%02d:%02d:%02d.%03d",
1298                  (int)(secs / 3600),
1299                  (int)((secs / 60) % 60),
1300                  (int)(secs % 60),
1301                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1302         snprintf(buf, buf_size,
1303                  "%-10s%-20s%7s%20s%15s",
1304                  sn->id_str, sn->name,
1305                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1306                  date_buf,
1307                  clock_buf);
1308     }
1309     return buf;
1310 }
1311
1312
1313 /**************************************************************/
1314 /* async I/Os */
1315
1316 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1317                                  QEMUIOVector *qiov, int nb_sectors,
1318                                  BlockDriverCompletionFunc *cb, void *opaque)
1319 {
1320     BlockDriver *drv = bs->drv;
1321     BlockDriverAIOCB *ret;
1322
1323     if (!drv)
1324         return NULL;
1325     if (bdrv_check_request(bs, sector_num, nb_sectors))
1326         return NULL;
1327
1328     ret = drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
1329                               cb, opaque);
1330
1331     if (ret) {
1332         /* Update stats even though technically transfer has not happened. */
1333         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1334         bs->rd_ops ++;
1335     }
1336
1337     return ret;
1338 }
1339
1340 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1341                                   QEMUIOVector *qiov, int nb_sectors,
1342                                   BlockDriverCompletionFunc *cb, void *opaque)
1343 {
1344     BlockDriver *drv = bs->drv;
1345     BlockDriverAIOCB *ret;
1346
1347     if (!drv)
1348         return NULL;
1349     if (bs->read_only)
1350         return NULL;
1351     if (bdrv_check_request(bs, sector_num, nb_sectors))
1352         return NULL;
1353
1354     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1355                                cb, opaque);
1356
1357     if (ret) {
1358         /* Update stats even though technically transfer has not happened. */
1359         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1360         bs->wr_ops ++;
1361     }
1362
1363     return ret;
1364 }
1365
1366 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1367 {
1368     acb->pool->cancel(acb);
1369 }
1370
1371
1372 /**************************************************************/
1373 /* async block device emulation */
1374
1375 static void bdrv_aio_bh_cb(void *opaque)
1376 {
1377     BlockDriverAIOCBSync *acb = opaque;
1378
1379     if (!acb->is_write)
1380         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1381     qemu_vfree(acb->bounce);
1382     acb->common.cb(acb->common.opaque, acb->ret);
1383
1384     qemu_aio_release(acb);
1385 }
1386
1387 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1388                                             int64_t sector_num,
1389                                             QEMUIOVector *qiov,
1390                                             int nb_sectors,
1391                                             BlockDriverCompletionFunc *cb,
1392                                             void *opaque,
1393                                             int is_write)
1394
1395 {
1396     BlockDriverAIOCBSync *acb;
1397
1398     acb = qemu_aio_get(bs, cb, opaque);
1399     acb->is_write = is_write;
1400     acb->qiov = qiov;
1401     acb->bounce = qemu_blockalign(bs, qiov->size);
1402
1403     if (!acb->bh)
1404         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1405
1406     if (is_write) {
1407         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1408         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1409     } else {
1410         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1411     }
1412
1413     qemu_bh_schedule(acb->bh);
1414
1415     return &acb->common;
1416 }
1417
1418 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1419         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1420         BlockDriverCompletionFunc *cb, void *opaque)
1421 {
1422     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1423 }
1424
1425 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1426         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1427         BlockDriverCompletionFunc *cb, void *opaque)
1428 {
1429     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1430 }
1431
1432
1433 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1434 {
1435     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1436     qemu_bh_cancel(acb->bh);
1437     qemu_aio_release(acb);
1438 }
1439
1440 /**************************************************************/
1441 /* sync block device emulation */
1442
1443 static void bdrv_rw_em_cb(void *opaque, int ret)
1444 {
1445     *(int *)opaque = ret;
1446 }
1447
1448 #define NOT_DONE 0x7fffffff
1449
1450 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1451                         uint8_t *buf, int nb_sectors)
1452 {
1453     int async_ret;
1454     BlockDriverAIOCB *acb;
1455     struct iovec iov;
1456     QEMUIOVector qiov;
1457
1458     async_ret = NOT_DONE;
1459     iov.iov_base = (void *)buf;
1460     iov.iov_len = nb_sectors * 512;
1461     qemu_iovec_init_external(&qiov, &iov, 1);
1462     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1463         bdrv_rw_em_cb, &async_ret);
1464     if (acb == NULL)
1465         return -1;
1466
1467     while (async_ret == NOT_DONE) {
1468         qemu_aio_wait();
1469     }
1470
1471     return async_ret;
1472 }
1473
1474 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1475                          const uint8_t *buf, int nb_sectors)
1476 {
1477     int async_ret;
1478     BlockDriverAIOCB *acb;
1479     struct iovec iov;
1480     QEMUIOVector qiov;
1481
1482     async_ret = NOT_DONE;
1483     iov.iov_base = (void *)buf;
1484     iov.iov_len = nb_sectors * 512;
1485     qemu_iovec_init_external(&qiov, &iov, 1);
1486     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1487         bdrv_rw_em_cb, &async_ret);
1488     if (acb == NULL)
1489         return -1;
1490     while (async_ret == NOT_DONE) {
1491         qemu_aio_wait();
1492     }
1493     return async_ret;
1494 }
1495
1496 void bdrv_init(void)
1497 {
1498     module_call_init(MODULE_INIT_BLOCK);
1499 }
1500
1501 void aio_pool_init(AIOPool *pool, int aiocb_size,
1502                    void (*cancel)(BlockDriverAIOCB *acb))
1503 {
1504     pool->aiocb_size = aiocb_size;
1505     pool->cancel = cancel;
1506     pool->free_aiocb = NULL;
1507 }
1508
1509 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1510                         BlockDriverCompletionFunc *cb, void *opaque)
1511 {
1512     BlockDriverAIOCB *acb;
1513
1514     if (pool->free_aiocb) {
1515         acb = pool->free_aiocb;
1516         pool->free_aiocb = acb->next;
1517     } else {
1518         acb = qemu_mallocz(pool->aiocb_size);
1519         acb->pool = pool;
1520     }
1521     acb->bs = bs;
1522     acb->cb = cb;
1523     acb->opaque = opaque;
1524     return acb;
1525 }
1526
1527 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1528                    void *opaque)
1529 {
1530     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1531 }
1532
1533 void qemu_aio_release(void *p)
1534 {
1535     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1536     AIOPool *pool = acb->pool;
1537     acb->next = pool->free_aiocb;
1538     pool->free_aiocb = acb;
1539 }
1540
1541 /**************************************************************/
1542 /* removable device support */
1543
1544 /**
1545  * Return TRUE if the media is present
1546  */
1547 int bdrv_is_inserted(BlockDriverState *bs)
1548 {
1549     BlockDriver *drv = bs->drv;
1550     int ret;
1551     if (!drv)
1552         return 0;
1553     if (!drv->bdrv_is_inserted)
1554         return 1;
1555     ret = drv->bdrv_is_inserted(bs);
1556     return ret;
1557 }
1558
1559 /**
1560  * Return TRUE if the media changed since the last call to this
1561  * function. It is currently only used for floppy disks
1562  */
1563 int bdrv_media_changed(BlockDriverState *bs)
1564 {
1565     BlockDriver *drv = bs->drv;
1566     int ret;
1567
1568     if (!drv || !drv->bdrv_media_changed)
1569         ret = -ENOTSUP;
1570     else
1571         ret = drv->bdrv_media_changed(bs);
1572     if (ret == -ENOTSUP)
1573         ret = bs->media_changed;
1574     bs->media_changed = 0;
1575     return ret;
1576 }
1577
1578 /**
1579  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1580  */
1581 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1582 {
1583     BlockDriver *drv = bs->drv;
1584     int ret;
1585
1586     if (!drv || !drv->bdrv_eject) {
1587         ret = -ENOTSUP;
1588     } else {
1589         ret = drv->bdrv_eject(bs, eject_flag);
1590     }
1591     if (ret == -ENOTSUP) {
1592         if (eject_flag)
1593             bdrv_close(bs);
1594     }
1595 }
1596
1597 int bdrv_is_locked(BlockDriverState *bs)
1598 {
1599     return bs->locked;
1600 }
1601
1602 /**
1603  * Lock or unlock the media (if it is locked, the user won't be able
1604  * to eject it manually).
1605  */
1606 void bdrv_set_locked(BlockDriverState *bs, int locked)
1607 {
1608     BlockDriver *drv = bs->drv;
1609
1610     bs->locked = locked;
1611     if (drv && drv->bdrv_set_locked) {
1612         drv->bdrv_set_locked(bs, locked);
1613     }
1614 }
1615
1616 /* needed for generic scsi interface */
1617
1618 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1619 {
1620     BlockDriver *drv = bs->drv;
1621
1622     if (drv && drv->bdrv_ioctl)
1623         return drv->bdrv_ioctl(bs, req, buf);
1624     return -ENOTSUP;
1625 }
1626
1627 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1628         unsigned long int req, void *buf,
1629         BlockDriverCompletionFunc *cb, void *opaque)
1630 {
1631     BlockDriver *drv = bs->drv;
1632
1633     if (drv && drv->bdrv_aio_ioctl)
1634         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1635     return NULL;
1636 }
1637
1638 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1639 {
1640     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1641 }