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_create(BlockDriver *drv, const char* filename,
 188     QEMUOptionParameter *options)
 189 {
 190     if (!drv->bdrv_create)
 191         return -ENOTSUP;
 192
 193     return drv->bdrv_create(filename, options);
 194 }
 195
 196 #ifdef _WIN32
 197 void get_tmp_filename(char *filename, int size)
 198 {
 199     char temp_dir[MAX_PATH];
 200
 201     GetTempPath(MAX_PATH, temp_dir);
 202     GetTempFileName(temp_dir, "qem", 0, filename);
 203 }
 204 #else
 205 void get_tmp_filename(char *filename, int size)
 206 {
 207     int fd;
 208     const char *tmpdir;
 209     /* XXX: race condition possible */
 210     tmpdir = getenv("TMPDIR");
 211     if (!tmpdir)
 212         tmpdir = "/tmp";
 213     snprintf(filename, size, "%s/vl.XXXXXX", tmpdir);
 214     fd = mkstemp(filename);
 215     close(fd);
 216 }
 217 #endif
 218
 219 #ifdef _WIN32
 220 static int is_windows_drive_prefix(const char *filename)
 221 {
 222     return (((filename[0] >= 'a' && filename[0] <= 'z') ||
 223              (filename[0] >= 'A' && filename[0] <= 'Z')) &&
 224             filename[1] == ':');
 225 }
 226
 227 static int is_windows_drive(const char *filename)
 228 {
 229     if (is_windows_drive_prefix(filename) &&
 230         filename[2] == '\0')
 231         return 1;
 232     if (strstart(filename, "\\\\.\\", NULL) ||
 233         strstart(filename, "//./", NULL))
 234         return 1;
 235     return 0;
 236 }
 237 #endif
 238
 239 static BlockDriver *find_protocol(const char *filename)
 240 {
 241     BlockDriver *drv1;
 242     char protocol[128];
 243     int len;
 244     const char *p;
 245
 246 #ifdef _WIN32
 247     if (is_windows_drive(filename) ||
 248         is_windows_drive_prefix(filename))
 249         return bdrv_find_format("raw");
 250 #endif
 251     p = strchr(filename, ':');
 252     if (!p)
 253         return bdrv_find_format("raw");
 254     len = p - filename;
 255     if (len > sizeof(protocol) - 1)
 256         len = sizeof(protocol) - 1;
 257     memcpy(protocol, filename, len);
 258     protocol[len] = '\0';
 259     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 260         if (drv1->protocol_name &&
 261             !strcmp(drv1->protocol_name, protocol))
 262             return drv1;
 263     }
 264     return NULL;
 265 }
 266
 267 /* XXX: force raw format if block or character device ? It would
 268    simplify the BSD case */
 269 static BlockDriver *find_image_format(const char *filename)
 270 {
 271     int ret, score, score_max;
 272     BlockDriver *drv1, *drv;
 273     uint8_t buf[2048];
 274     BlockDriverState *bs;
 275
 276     /* detect host devices. By convention, /dev/cdrom[N] is always
 277        recognized as a host CDROM */
 278     if (strstart(filename, "/dev/cdrom", NULL))
 279         return bdrv_find_format("host_device");
 280 #ifdef _WIN32
 281     if (is_windows_drive(filename))
 282         return bdrv_find_format("host_device");
 283 #else
 284     {
 285         struct stat st;
 286         if (stat(filename, &st) >= 0 &&
 287             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 288             return bdrv_find_format("host_device");
 289         }
 290     }
 291 #endif
 292
 293     drv = find_protocol(filename);
 294     /* no need to test disk image formats for vvfat */
 295     if (drv && strcmp(drv->format_name, "vvfat") == 0)
 296         return drv;
 297
 298     ret = bdrv_file_open(&bs, filename, BDRV_O_RDONLY);
 299     if (ret < 0)
 300         return NULL;
 301     ret = bdrv_pread(bs, 0, buf, sizeof(buf));
 302     bdrv_delete(bs);
 303     if (ret < 0) {
 304         return NULL;
 305     }
 306
 307     score_max = 0;
 308     for(drv1 = first_drv; drv1 != NULL; drv1 = drv1->next) {
 309         if (drv1->bdrv_probe) {
 310             score = drv1->bdrv_probe(buf, ret, filename);
 311             if (score > score_max) {
 312                 score_max = score;
 313                 drv = drv1;
 314             }
 315         }
 316     }
 317     return drv;
 318 }
 319
 320 int bdrv_file_open(BlockDriverState **pbs, const char *filename, int flags)
 321 {
 322     BlockDriverState *bs;
 323     int ret;
 324
 325     bs = bdrv_new("");
 326     ret = bdrv_open2(bs, filename, flags | BDRV_O_FILE, NULL);
 327     if (ret < 0) {
 328         bdrv_delete(bs);
 329         return ret;
 330     }
 331     bs->growable = 1;
 332     *pbs = bs;
 333     return 0;
 334 }
 335
 336 int bdrv_open(BlockDriverState *bs, const char *filename, int flags)
 337 {
 338     return bdrv_open2(bs, filename, flags, NULL);
 339 }
 340
 341 int bdrv_open2(BlockDriverState *bs, const char *filename, int flags,
 342                BlockDriver *drv)
 343 {
 344     int ret, open_flags;
 345     char tmp_filename[PATH_MAX];
 346     char backing_filename[PATH_MAX];
 347
 348     bs->read_only = 0;
 349     bs->is_temporary = 0;
 350     bs->encrypted = 0;
 351     bs->valid_key = 0;
 352     /* buffer_alignment defaulted to 512, drivers can change this value */
 353     bs->buffer_alignment = 512;
 354
 355     if (flags & BDRV_O_SNAPSHOT) {
 356         BlockDriverState *bs1;
 357         int64_t total_size;
 358         int is_protocol = 0;
 359         BlockDriver *bdrv_qcow2;
 360         QEMUOptionParameter *options;
 361
 362         /* if snapshot, we create a temporary backing file and open it
 363            instead of opening 'filename' directly */
 364
 365         /* if there is a backing file, use it */
 366         bs1 = bdrv_new("");
 367         ret = bdrv_open2(bs1, filename, 0, drv);
 368         if (ret < 0) {
 369             bdrv_delete(bs1);
 370             return ret;
 371         }
 372         total_size = bdrv_getlength(bs1) >> SECTOR_BITS;
 373
 374         if (bs1->drv && bs1->drv->protocol_name)
 375             is_protocol = 1;
 376
 377         bdrv_delete(bs1);
 378
 379         get_tmp_filename(tmp_filename, sizeof(tmp_filename));
 380
 381         /* Real path is meaningless for protocols */
 382         if (is_protocol)
 383             snprintf(backing_filename, sizeof(backing_filename),
 384                      "%s", filename);
 385         else
 386             realpath(filename, backing_filename);
 387
 388         bdrv_qcow2 = bdrv_find_format("qcow2");
 389         options = parse_option_parameters("", bdrv_qcow2->create_options, NULL);
 390
 391         set_option_parameter_int(options, BLOCK_OPT_SIZE, total_size * 512);
 392         set_option_parameter(options, BLOCK_OPT_BACKING_FILE, backing_filename);
 393         if (drv) {
 394             set_option_parameter(options, BLOCK_OPT_BACKING_FMT,
 395                 drv->format_name);
 396         }
 397
 398         ret = bdrv_create(bdrv_qcow2, tmp_filename, options);
 399         if (ret < 0) {
 400             return ret;
 401         }
 402
 403         filename = tmp_filename;
 404         drv = bdrv_qcow2;
 405         bs->is_temporary = 1;
 406     }
 407
 408     pstrcpy(bs->filename, sizeof(bs->filename), filename);
 409     if (flags & BDRV_O_FILE) {
 410         drv = find_protocol(filename);
 411     } else if (!drv) {
 412         drv = find_image_format(filename);
 413     }
 414     if (!drv) {
 415         ret = -ENOENT;
 416         goto unlink_and_fail;
 417     }
 418     bs->drv = drv;
 419     bs->opaque = qemu_mallocz(drv->instance_size);
 420     /* Note: for compatibility, we open disk image files as RDWR, and
 421        RDONLY as fallback */
 422     if (!(flags & BDRV_O_FILE))
 423         open_flags = BDRV_O_RDWR | (flags & BDRV_O_CACHE_MASK);
 424     else
 425         open_flags = flags & ~(BDRV_O_FILE | BDRV_O_SNAPSHOT);
 426     ret = drv->bdrv_open(bs, filename, open_flags);
 427     if ((ret == -EACCES || ret == -EPERM) && !(flags & BDRV_O_FILE)) {
 428         ret = drv->bdrv_open(bs, filename, open_flags & ~BDRV_O_RDWR);
 429         bs->read_only = 1;
 430     }
 431     if (ret < 0) {
 432         qemu_free(bs->opaque);
 433         bs->opaque = NULL;
 434         bs->drv = NULL;
 435     unlink_and_fail:
 436         if (bs->is_temporary)
 437             unlink(filename);
 438         return ret;
 439     }
 440     if (drv->bdrv_getlength) {
 441         bs->total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 442     }
 443 #ifndef _WIN32
 444     if (bs->is_temporary) {
 445         unlink(filename);
 446     }
 447 #endif
 448     if (bs->backing_file[0] != '\0') {
 449         /* if there is a backing file, use it */
 450         BlockDriver *back_drv = NULL;
 451         bs->backing_hd = bdrv_new("");
 452         path_combine(backing_filename, sizeof(backing_filename),
 453                      filename, bs->backing_file);
 454         if (bs->backing_format[0] != '\0')
 455             back_drv = bdrv_find_format(bs->backing_format);
 456         ret = bdrv_open2(bs->backing_hd, backing_filename, open_flags,
 457                          back_drv);
 458         if (ret < 0) {
 459             bdrv_close(bs);
 460             return ret;
 461         }
 462     }
 463
 464     if (!bdrv_key_required(bs)) {
 465         /* call the change callback */
 466         bs->media_changed = 1;
 467         if (bs->change_cb)
 468             bs->change_cb(bs->change_opaque);
 469     }
 470     return 0;
 471 }
 472
 473 void bdrv_close(BlockDriverState *bs)
 474 {
 475     if (bs->drv) {
 476         if (bs->backing_hd)
 477             bdrv_delete(bs->backing_hd);
 478         bs->drv->bdrv_close(bs);
 479         qemu_free(bs->opaque);
 480 #ifdef _WIN32
 481         if (bs->is_temporary) {
 482             unlink(bs->filename);
 483         }
 484 #endif
 485         bs->opaque = NULL;
 486         bs->drv = NULL;
 487
 488         /* call the change callback */
 489         bs->media_changed = 1;
 490         if (bs->change_cb)
 491             bs->change_cb(bs->change_opaque);
 492     }
 493 }
 494
 495 void bdrv_delete(BlockDriverState *bs)
 496 {
 497     BlockDriverState **pbs;
 498
 499     pbs = &bdrv_first;
 500     while (*pbs != bs && *pbs != NULL)
 501         pbs = &(*pbs)->next;
 502     if (*pbs == bs)
 503         *pbs = bs->next;
 504
 505     bdrv_close(bs);
 506     qemu_free(bs);
 507 }
 508
 509 /*
 510  * Run consistency checks on an image
 511  *
 512  * Returns the number of errors or -errno when an internal error occurs
 513  */
 514 int bdrv_check(BlockDriverState *bs)
 515 {
 516     if (bs->drv->bdrv_check == NULL) {
 517         return -ENOTSUP;
 518     }
 519
 520     return bs->drv->bdrv_check(bs);
 521 }
 522
 523 /* commit COW file into the raw image */
 524 int bdrv_commit(BlockDriverState *bs)
 525 {
 526     BlockDriver *drv = bs->drv;
 527     int64_t i, total_sectors;
 528     int n, j;
 529     unsigned char sector[512];
 530
 531     if (!drv)
 532         return -ENOMEDIUM;
 533
 534     if (bs->read_only) {
 535         return -EACCES;
 536     }
 537
 538     if (!bs->backing_hd) {
 539         return -ENOTSUP;
 540     }
 541
 542     total_sectors = bdrv_getlength(bs) >> SECTOR_BITS;
 543     for (i = 0; i < total_sectors;) {
 544         if (drv->bdrv_is_allocated(bs, i, 65536, &n)) {
 545             for(j = 0; j < n; j++) {
 546                 if (bdrv_read(bs, i, sector, 1) != 0) {
 547                     return -EIO;
 548                 }
 549
 550                 if (bdrv_write(bs->backing_hd, i, sector, 1) != 0) {
 551                     return -EIO;
 552                 }
 553                 i++;
 554             }
 555         } else {
 556             i += n;
 557         }
 558     }
 559
 560     if (drv->bdrv_make_empty)
 561         return drv->bdrv_make_empty(bs);
 562
 563     return 0;
 564 }
 565
 566 static int bdrv_check_byte_request(BlockDriverState *bs, int64_t offset,
 567                                    size_t size)
 568 {
 569     int64_t len;
 570
 571     if (!bdrv_is_inserted(bs))
 572         return -ENOMEDIUM;
 573
 574     if (bs->growable)
 575         return 0;
 576
 577     len = bdrv_getlength(bs);
 578
 579     if (offset < 0)
 580         return -EIO;
 581
 582     if ((offset > len) || (len - offset < size))
 583         return -EIO;
 584
 585     return 0;
 586 }
 587
 588 static int bdrv_check_request(BlockDriverState *bs, int64_t sector_num,
 589                               int nb_sectors)
 590 {
 591     return bdrv_check_byte_request(bs, sector_num * 512, nb_sectors * 512);
 592 }
 593
 594 /* return < 0 if error. See bdrv_write() for the return codes */
 595 int bdrv_read(BlockDriverState *bs, int64_t sector_num,
 596               uint8_t *buf, int nb_sectors)
 597 {
 598     BlockDriver *drv = bs->drv;
 599
 600     if (!drv)
 601         return -ENOMEDIUM;
 602     if (bdrv_check_request(bs, sector_num, nb_sectors))
 603         return -EIO;
 604
 605     return drv->bdrv_read(bs, sector_num, buf, nb_sectors);
 606 }
 607
 608 /* Return < 0 if error. Important errors are:
 609   -EIO         generic I/O error (may happen for all errors)
 610   -ENOMEDIUM   No media inserted.
 611   -EINVAL      Invalid sector number or nb_sectors
 612   -EACCES      Trying to write a read-only device
 613 */
 614 int bdrv_write(BlockDriverState *bs, int64_t sector_num,
 615                const uint8_t *buf, int nb_sectors)
 616 {
 617     BlockDriver *drv = bs->drv;
 618     if (!bs->drv)
 619         return -ENOMEDIUM;
 620     if (bs->read_only)
 621         return -EACCES;
 622     if (bdrv_check_request(bs, sector_num, nb_sectors))
 623         return -EIO;
 624
 625     return drv->bdrv_write(bs, sector_num, buf, nb_sectors);
 626 }
 627
 628 int bdrv_pread(BlockDriverState *bs, int64_t offset,
 629                void *buf, int count1)
 630 {
 631     uint8_t tmp_buf[SECTOR_SIZE];
 632     int len, nb_sectors, count;
 633     int64_t sector_num;
 634
 635     count = count1;
 636     /* first read to align to sector start */
 637     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 638     if (len > count)
 639         len = count;
 640     sector_num = offset >> SECTOR_BITS;
 641     if (len > 0) {
 642         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 643             return -EIO;
 644         memcpy(buf, tmp_buf + (offset & (SECTOR_SIZE - 1)), len);
 645         count -= len;
 646         if (count == 0)
 647             return count1;
 648         sector_num++;
 649         buf += len;
 650     }
 651
 652     /* read the sectors "in place" */
 653     nb_sectors = count >> SECTOR_BITS;
 654     if (nb_sectors > 0) {
 655         if (bdrv_read(bs, sector_num, buf, nb_sectors) < 0)
 656             return -EIO;
 657         sector_num += nb_sectors;
 658         len = nb_sectors << SECTOR_BITS;
 659         buf += len;
 660         count -= len;
 661     }
 662
 663     /* add data from the last sector */
 664     if (count > 0) {
 665         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 666             return -EIO;
 667         memcpy(buf, tmp_buf, count);
 668     }
 669     return count1;
 670 }
 671
 672 int bdrv_pwrite(BlockDriverState *bs, int64_t offset,
 673                 const void *buf, int count1)
 674 {
 675     uint8_t tmp_buf[SECTOR_SIZE];
 676     int len, nb_sectors, count;
 677     int64_t sector_num;
 678
 679     count = count1;
 680     /* first write to align to sector start */
 681     len = (SECTOR_SIZE - offset) & (SECTOR_SIZE - 1);
 682     if (len > count)
 683         len = count;
 684     sector_num = offset >> SECTOR_BITS;
 685     if (len > 0) {
 686         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 687             return -EIO;
 688         memcpy(tmp_buf + (offset & (SECTOR_SIZE - 1)), buf, len);
 689         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 690             return -EIO;
 691         count -= len;
 692         if (count == 0)
 693             return count1;
 694         sector_num++;
 695         buf += len;
 696     }
 697
 698     /* write the sectors "in place" */
 699     nb_sectors = count >> SECTOR_BITS;
 700     if (nb_sectors > 0) {
 701         if (bdrv_write(bs, sector_num, buf, nb_sectors) < 0)
 702             return -EIO;
 703         sector_num += nb_sectors;
 704         len = nb_sectors << SECTOR_BITS;
 705         buf += len;
 706         count -= len;
 707     }
 708
 709     /* add data from the last sector */
 710     if (count > 0) {
 711         if (bdrv_read(bs, sector_num, tmp_buf, 1) < 0)
 712             return -EIO;
 713         memcpy(tmp_buf, buf, count);
 714         if (bdrv_write(bs, sector_num, tmp_buf, 1) < 0)
 715             return -EIO;
 716     }
 717     return count1;
 718 }
 719
 720 /**
 721  * Truncate file to 'offset' bytes (needed only for file protocols)
 722  */
 723 int bdrv_truncate(BlockDriverState *bs, int64_t offset)
 724 {
 725     BlockDriver *drv = bs->drv;
 726     if (!drv)
 727         return -ENOMEDIUM;
 728     if (!drv->bdrv_truncate)
 729         return -ENOTSUP;
 730     return drv->bdrv_truncate(bs, offset);
 731 }
 732
 733 /**
 734  * Length of a file in bytes. Return < 0 if error or unknown.
 735  */
 736 int64_t bdrv_getlength(BlockDriverState *bs)
 737 {
 738     BlockDriver *drv = bs->drv;
 739     if (!drv)
 740         return -ENOMEDIUM;
 741     if (!drv->bdrv_getlength) {
 742         /* legacy mode */
 743         return bs->total_sectors * SECTOR_SIZE;
 744     }
 745     return drv->bdrv_getlength(bs);
 746 }
 747
 748 /* return 0 as number of sectors if no device present or error */
 749 void bdrv_get_geometry(BlockDriverState *bs, uint64_t *nb_sectors_ptr)
 750 {
 751     int64_t length;
 752     length = bdrv_getlength(bs);
 753     if (length < 0)
 754         length = 0;
 755     else
 756         length = length >> SECTOR_BITS;
 757     *nb_sectors_ptr = length;
 758 }
 759
 760 struct partition {
 761         uint8_t boot_ind;           /* 0x80 - active */
 762         uint8_t head;               /* starting head */
 763         uint8_t sector;             /* starting sector */
 764         uint8_t cyl;                /* starting cylinder */
 765         uint8_t sys_ind;            /* What partition type */
 766         uint8_t end_head;           /* end head */
 767         uint8_t end_sector;         /* end sector */
 768         uint8_t end_cyl;            /* end cylinder */
 769         uint32_t start_sect;        /* starting sector counting from 0 */
 770         uint32_t nr_sects;          /* nr of sectors in partition */
 771 } __attribute__((packed));
 772
 773 /* try to guess the disk logical geometry from the MSDOS partition table. Return 0 if OK, -1 if could not guess */
 774 static int guess_disk_lchs(BlockDriverState *bs,
 775                            int *pcylinders, int *pheads, int *psectors)
 776 {
 777     uint8_t buf[512];
 778     int ret, i, heads, sectors, cylinders;
 779     struct partition *p;
 780     uint32_t nr_sects;
 781     uint64_t nb_sectors;
 782
 783     bdrv_get_geometry(bs, &nb_sectors);
 784
 785     ret = bdrv_read(bs, 0, buf, 1);
 786     if (ret < 0)
 787         return -1;
 788     /* test msdos magic */
 789     if (buf[510] != 0x55 || buf[511] != 0xaa)
 790         return -1;
 791     for(i = 0; i < 4; i++) {
 792         p = ((struct partition *)(buf + 0x1be)) + i;
 793         nr_sects = le32_to_cpu(p->nr_sects);
 794         if (nr_sects && p->end_head) {
 795             /* We make the assumption that the partition terminates on
 796                a cylinder boundary */
 797             heads = p->end_head + 1;
 798             sectors = p->end_sector & 63;
 799             if (sectors == 0)
 800                 continue;
 801             cylinders = nb_sectors / (heads * sectors);
 802             if (cylinders < 1 || cylinders > 16383)
 803                 continue;
 804             *pheads = heads;
 805             *psectors = sectors;
 806             *pcylinders = cylinders;
 807 #if 0
 808             printf("guessed geometry: LCHS=%d %d %d\n",
 809                    cylinders, heads, sectors);
 810 #endif
 811             return 0;
 812         }
 813     }
 814     return -1;
 815 }
 816
 817 void bdrv_guess_geometry(BlockDriverState *bs, int *pcyls, int *pheads, int *psecs)
 818 {
 819     int translation, lba_detected = 0;
 820     int cylinders, heads, secs;
 821     uint64_t nb_sectors;
 822
 823     /* if a geometry hint is available, use it */
 824     bdrv_get_geometry(bs, &nb_sectors);
 825     bdrv_get_geometry_hint(bs, &cylinders, &heads, &secs);
 826     translation = bdrv_get_translation_hint(bs);
 827     if (cylinders != 0) {
 828         *pcyls = cylinders;
 829         *pheads = heads;
 830         *psecs = secs;
 831     } else {
 832         if (guess_disk_lchs(bs, &cylinders, &heads, &secs) == 0) {
 833             if (heads > 16) {
 834                 /* if heads > 16, it means that a BIOS LBA
 835                    translation was active, so the default
 836                    hardware geometry is OK */
 837                 lba_detected = 1;
 838                 goto default_geometry;
 839             } else {
 840                 *pcyls = cylinders;
 841                 *pheads = heads;
 842                 *psecs = secs;
 843                 /* disable any translation to be in sync with
 844                    the logical geometry */
 845                 if (translation == BIOS_ATA_TRANSLATION_AUTO) {
 846                     bdrv_set_translation_hint(bs,
 847                                               BIOS_ATA_TRANSLATION_NONE);
 848                 }
 849             }
 850         } else {
 851         default_geometry:
 852             /* if no geometry, use a standard physical disk geometry */
 853             cylinders = nb_sectors / (16 * 63);
 854
 855             if (cylinders > 16383)
 856                 cylinders = 16383;
 857             else if (cylinders < 2)
 858                 cylinders = 2;
 859             *pcyls = cylinders;
 860             *pheads = 16;
 861             *psecs = 63;
 862             if ((lba_detected == 1) && (translation == BIOS_ATA_TRANSLATION_AUTO)) {
 863                 if ((*pcyls * *pheads) <= 131072) {
 864                     bdrv_set_translation_hint(bs,
 865                                               BIOS_ATA_TRANSLATION_LARGE);
 866                 } else {
 867                     bdrv_set_translation_hint(bs,
 868                                               BIOS_ATA_TRANSLATION_LBA);
 869                 }
 870             }
 871         }
 872         bdrv_set_geometry_hint(bs, *pcyls, *pheads, *psecs);
 873     }
 874 }
 875
 876 void bdrv_set_geometry_hint(BlockDriverState *bs,
 877                             int cyls, int heads, int secs)
 878 {
 879     bs->cyls = cyls;
 880     bs->heads = heads;
 881     bs->secs = secs;
 882 }
 883
 884 void bdrv_set_type_hint(BlockDriverState *bs, int type)
 885 {
 886     bs->type = type;
 887     bs->removable = ((type == BDRV_TYPE_CDROM ||
 888                       type == BDRV_TYPE_FLOPPY));
 889 }
 890
 891 void bdrv_set_translation_hint(BlockDriverState *bs, int translation)
 892 {
 893     bs->translation = translation;
 894 }
 895
 896 void bdrv_get_geometry_hint(BlockDriverState *bs,
 897                             int *pcyls, int *pheads, int *psecs)
 898 {
 899     *pcyls = bs->cyls;
 900     *pheads = bs->heads;
 901     *psecs = bs->secs;
 902 }
 903
 904 int bdrv_get_type_hint(BlockDriverState *bs)
 905 {
 906     return bs->type;
 907 }
 908
 909 int bdrv_get_translation_hint(BlockDriverState *bs)
 910 {
 911     return bs->translation;
 912 }
 913
 914 int bdrv_is_removable(BlockDriverState *bs)
 915 {
 916     return bs->removable;
 917 }
 918
 919 int bdrv_is_read_only(BlockDriverState *bs)
 920 {
 921     return bs->read_only;
 922 }
 923
 924 int bdrv_is_sg(BlockDriverState *bs)
 925 {
 926     return bs->sg;
 927 }
 928
 929 /* XXX: no longer used */
 930 void bdrv_set_change_cb(BlockDriverState *bs,
 931                         void (*change_cb)(void *opaque), void *opaque)
 932 {
 933     bs->change_cb = change_cb;
 934     bs->change_opaque = opaque;
 935 }
 936
 937 int bdrv_is_encrypted(BlockDriverState *bs)
 938 {
 939     if (bs->backing_hd && bs->backing_hd->encrypted)
 940         return 1;
 941     return bs->encrypted;
 942 }
 943
 944 int bdrv_key_required(BlockDriverState *bs)
 945 {
 946     BlockDriverState *backing_hd = bs->backing_hd;
 947
 948     if (backing_hd && backing_hd->encrypted && !backing_hd->valid_key)
 949         return 1;
 950     return (bs->encrypted && !bs->valid_key);
 951 }
 952
 953 int bdrv_set_key(BlockDriverState *bs, const char *key)
 954 {
 955     int ret;
 956     if (bs->backing_hd && bs->backing_hd->encrypted) {
 957         ret = bdrv_set_key(bs->backing_hd, key);
 958         if (ret < 0)
 959             return ret;
 960         if (!bs->encrypted)
 961             return 0;
 962     }
 963     if (!bs->encrypted || !bs->drv || !bs->drv->bdrv_set_key)
 964         return -1;
 965     ret = bs->drv->bdrv_set_key(bs, key);
 966     if (ret < 0) {
 967         bs->valid_key = 0;
 968     } else if (!bs->valid_key) {
 969         bs->valid_key = 1;
 970         /* call the change callback now, we skipped it on open */
 971         bs->media_changed = 1;
 972         if (bs->change_cb)
 973             bs->change_cb(bs->change_opaque);
 974     }
 975     return ret;
 976 }
 977
 978 void bdrv_get_format(BlockDriverState *bs, char *buf, int buf_size)
 979 {
 980     if (!bs->drv) {
 981         buf[0] = '\0';
 982     } else {
 983         pstrcpy(buf, buf_size, bs->drv->format_name);
 984     }
 985 }
 986
 987 void bdrv_iterate_format(void (*it)(void *opaque, const char *name),
 988                          void *opaque)
 989 {
 990     BlockDriver *drv;
 991
 992     for (drv = first_drv; drv != NULL; drv = drv->next) {
 993         it(opaque, drv->format_name);
 994     }
 995 }
 996
 997 BlockDriverState *bdrv_find(const char *name)
 998 {
 999     BlockDriverState *bs;
1000
1001     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1002         if (!strcmp(name, bs->device_name))
1003             return bs;
1004     }
1005     return NULL;
1006 }
1007
1008 void bdrv_iterate(void (*it)(void *opaque, BlockDriverState *bs), void *opaque)
1009 {
1010     BlockDriverState *bs;
1011
1012     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1013         it(opaque, bs);
1014     }
1015 }
1016
1017 const char *bdrv_get_device_name(BlockDriverState *bs)
1018 {
1019     return bs->device_name;
1020 }
1021
1022 void bdrv_flush(BlockDriverState *bs)
1023 {
1024     if (!bs->drv)
1025         return;
1026     if (bs->drv->bdrv_flush)
1027         bs->drv->bdrv_flush(bs);
1028     if (bs->backing_hd)
1029         bdrv_flush(bs->backing_hd);
1030 }
1031
1032 void bdrv_flush_all(void)
1033 {
1034     BlockDriverState *bs;
1035
1036     for (bs = bdrv_first; bs != NULL; bs = bs->next)
1037         if (bs->drv && !bdrv_is_read_only(bs) &&
1038             (!bdrv_is_removable(bs) || bdrv_is_inserted(bs)))
1039             bdrv_flush(bs);
1040 }
1041
1042 /*
1043  * Returns true iff the specified sector is present in the disk image. Drivers
1044  * not implementing the functionality are assumed to not support backing files,
1045  * hence all their sectors are reported as allocated.
1046  *
1047  * 'pnum' is set to the number of sectors (including and immediately following
1048  * the specified sector) that are known to be in the same
1049  * allocated/unallocated state.
1050  *
1051  * 'nb_sectors' is the max value 'pnum' should be set to.
1052  */
1053 int bdrv_is_allocated(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
1054         int *pnum)
1055 {
1056     int64_t n;
1057     if (!bs->drv->bdrv_is_allocated) {
1058         if (sector_num >= bs->total_sectors) {
1059             *pnum = 0;
1060             return 0;
1061         }
1062         n = bs->total_sectors - sector_num;
1063         *pnum = (n < nb_sectors) ? (n) : (nb_sectors);
1064         return 1;
1065     }
1066     return bs->drv->bdrv_is_allocated(bs, sector_num, nb_sectors, pnum);
1067 }
1068
1069 void bdrv_info(Monitor *mon)
1070 {
1071     BlockDriverState *bs;
1072
1073     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1074         monitor_printf(mon, "%s:", bs->device_name);
1075         monitor_printf(mon, " type=");
1076         switch(bs->type) {
1077         case BDRV_TYPE_HD:
1078             monitor_printf(mon, "hd");
1079             break;
1080         case BDRV_TYPE_CDROM:
1081             monitor_printf(mon, "cdrom");
1082             break;
1083         case BDRV_TYPE_FLOPPY:
1084             monitor_printf(mon, "floppy");
1085             break;
1086         }
1087         monitor_printf(mon, " removable=%d", bs->removable);
1088         if (bs->removable) {
1089             monitor_printf(mon, " locked=%d", bs->locked);
1090         }
1091         if (bs->drv) {
1092             monitor_printf(mon, " file=");
1093             monitor_print_filename(mon, bs->filename);
1094             if (bs->backing_file[0] != '\0') {
1095                 monitor_printf(mon, " backing_file=");
1096                 monitor_print_filename(mon, bs->backing_file);
1097             }
1098             monitor_printf(mon, " ro=%d", bs->read_only);
1099             monitor_printf(mon, " drv=%s", bs->drv->format_name);
1100             monitor_printf(mon, " encrypted=%d", bdrv_is_encrypted(bs));
1101         } else {
1102             monitor_printf(mon, " [not inserted]");
1103         }
1104         monitor_printf(mon, "\n");
1105     }
1106 }
1107
1108 /* The "info blockstats" command. */
1109 void bdrv_info_stats(Monitor *mon)
1110 {
1111     BlockDriverState *bs;
1112
1113     for (bs = bdrv_first; bs != NULL; bs = bs->next) {
1114         monitor_printf(mon, "%s:"
1115                        " rd_bytes=%" PRIu64
1116                        " wr_bytes=%" PRIu64
1117                        " rd_operations=%" PRIu64
1118                        " wr_operations=%" PRIu64
1119                        "\n",
1120                        bs->device_name,
1121                        bs->rd_bytes, bs->wr_bytes,
1122                        bs->rd_ops, bs->wr_ops);
1123     }
1124 }
1125
1126 const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
1127 {
1128     if (bs->backing_hd && bs->backing_hd->encrypted)
1129         return bs->backing_file;
1130     else if (bs->encrypted)
1131         return bs->filename;
1132     else
1133         return NULL;
1134 }
1135
1136 void bdrv_get_backing_filename(BlockDriverState *bs,
1137                                char *filename, int filename_size)
1138 {
1139     if (!bs->backing_hd) {
1140         pstrcpy(filename, filename_size, "");
1141     } else {
1142         pstrcpy(filename, filename_size, bs->backing_file);
1143     }
1144 }
1145
1146 int bdrv_write_compressed(BlockDriverState *bs, int64_t sector_num,
1147                           const uint8_t *buf, int nb_sectors)
1148 {
1149     BlockDriver *drv = bs->drv;
1150     if (!drv)
1151         return -ENOMEDIUM;
1152     if (!drv->bdrv_write_compressed)
1153         return -ENOTSUP;
1154     if (bdrv_check_request(bs, sector_num, nb_sectors))
1155         return -EIO;
1156     return drv->bdrv_write_compressed(bs, sector_num, buf, nb_sectors);
1157 }
1158
1159 int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1160 {
1161     BlockDriver *drv = bs->drv;
1162     if (!drv)
1163         return -ENOMEDIUM;
1164     if (!drv->bdrv_get_info)
1165         return -ENOTSUP;
1166     memset(bdi, 0, sizeof(*bdi));
1167     return drv->bdrv_get_info(bs, bdi);
1168 }
1169
1170 int bdrv_put_buffer(BlockDriverState *bs, const uint8_t *buf, int64_t pos, int size)
1171 {
1172     BlockDriver *drv = bs->drv;
1173     if (!drv)
1174         return -ENOMEDIUM;
1175     if (!drv->bdrv_put_buffer)
1176         return -ENOTSUP;
1177     return drv->bdrv_put_buffer(bs, buf, pos, size);
1178 }
1179
1180 int bdrv_get_buffer(BlockDriverState *bs, uint8_t *buf, int64_t pos, int size)
1181 {
1182     BlockDriver *drv = bs->drv;
1183     if (!drv)
1184         return -ENOMEDIUM;
1185     if (!drv->bdrv_get_buffer)
1186         return -ENOTSUP;
1187     return drv->bdrv_get_buffer(bs, buf, pos, size);
1188 }
1189
1190 /**************************************************************/
1191 /* handling of snapshots */
1192
1193 int bdrv_snapshot_create(BlockDriverState *bs,
1194                          QEMUSnapshotInfo *sn_info)
1195 {
1196     BlockDriver *drv = bs->drv;
1197     if (!drv)
1198         return -ENOMEDIUM;
1199     if (!drv->bdrv_snapshot_create)
1200         return -ENOTSUP;
1201     return drv->bdrv_snapshot_create(bs, sn_info);
1202 }
1203
1204 int bdrv_snapshot_goto(BlockDriverState *bs,
1205                        const char *snapshot_id)
1206 {
1207     BlockDriver *drv = bs->drv;
1208     if (!drv)
1209         return -ENOMEDIUM;
1210     if (!drv->bdrv_snapshot_goto)
1211         return -ENOTSUP;
1212     return drv->bdrv_snapshot_goto(bs, snapshot_id);
1213 }
1214
1215 int bdrv_snapshot_delete(BlockDriverState *bs, const char *snapshot_id)
1216 {
1217     BlockDriver *drv = bs->drv;
1218     if (!drv)
1219         return -ENOMEDIUM;
1220     if (!drv->bdrv_snapshot_delete)
1221         return -ENOTSUP;
1222     return drv->bdrv_snapshot_delete(bs, snapshot_id);
1223 }
1224
1225 int bdrv_snapshot_list(BlockDriverState *bs,
1226                        QEMUSnapshotInfo **psn_info)
1227 {
1228     BlockDriver *drv = bs->drv;
1229     if (!drv)
1230         return -ENOMEDIUM;
1231     if (!drv->bdrv_snapshot_list)
1232         return -ENOTSUP;
1233     return drv->bdrv_snapshot_list(bs, psn_info);
1234 }
1235
1236 #define NB_SUFFIXES 4
1237
1238 char *get_human_readable_size(char *buf, int buf_size, int64_t size)
1239 {
1240     static const char suffixes[NB_SUFFIXES] = "KMGT";
1241     int64_t base;
1242     int i;
1243
1244     if (size <= 999) {
1245         snprintf(buf, buf_size, "%" PRId64, size);
1246     } else {
1247         base = 1024;
1248         for(i = 0; i < NB_SUFFIXES; i++) {
1249             if (size < (10 * base)) {
1250                 snprintf(buf, buf_size, "%0.1f%c",
1251                          (double)size / base,
1252                          suffixes[i]);
1253                 break;
1254             } else if (size < (1000 * base) || i == (NB_SUFFIXES - 1)) {
1255                 snprintf(buf, buf_size, "%" PRId64 "%c",
1256                          ((size + (base >> 1)) / base),
1257                          suffixes[i]);
1258                 break;
1259             }
1260             base = base * 1024;
1261         }
1262     }
1263     return buf;
1264 }
1265
1266 char *bdrv_snapshot_dump(char *buf, int buf_size, QEMUSnapshotInfo *sn)
1267 {
1268     char buf1[128], date_buf[128], clock_buf[128];
1269 #ifdef _WIN32
1270     struct tm *ptm;
1271 #else
1272     struct tm tm;
1273 #endif
1274     time_t ti;
1275     int64_t secs;
1276
1277     if (!sn) {
1278         snprintf(buf, buf_size,
1279                  "%-10s%-20s%7s%20s%15s",
1280                  "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
1281     } else {
1282         ti = sn->date_sec;
1283 #ifdef _WIN32
1284         ptm = localtime(&ti);
1285         strftime(date_buf, sizeof(date_buf),
1286                  "%Y-%m-%d %H:%M:%S", ptm);
1287 #else
1288         localtime_r(&ti, &tm);
1289         strftime(date_buf, sizeof(date_buf),
1290                  "%Y-%m-%d %H:%M:%S", &tm);
1291 #endif
1292         secs = sn->vm_clock_nsec / 1000000000;
1293         snprintf(clock_buf, sizeof(clock_buf),
1294                  "%02d:%02d:%02d.%03d",
1295                  (int)(secs / 3600),
1296                  (int)((secs / 60) % 60),
1297                  (int)(secs % 60),
1298                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
1299         snprintf(buf, buf_size,
1300                  "%-10s%-20s%7s%20s%15s",
1301                  sn->id_str, sn->name,
1302                  get_human_readable_size(buf1, sizeof(buf1), sn->vm_state_size),
1303                  date_buf,
1304                  clock_buf);
1305     }
1306     return buf;
1307 }
1308
1309
1310 /**************************************************************/
1311 /* async I/Os */
1312
1313 BlockDriverAIOCB *bdrv_aio_readv(BlockDriverState *bs, int64_t sector_num,
1314                                  QEMUIOVector *qiov, int nb_sectors,
1315                                  BlockDriverCompletionFunc *cb, void *opaque)
1316 {
1317     BlockDriver *drv = bs->drv;
1318     BlockDriverAIOCB *ret;
1319
1320     if (!drv)
1321         return NULL;
1322     if (bdrv_check_request(bs, sector_num, nb_sectors))
1323         return NULL;
1324
1325     ret = drv->bdrv_aio_readv(bs, sector_num, qiov, nb_sectors,
1326                               cb, opaque);
1327
1328     if (ret) {
1329         /* Update stats even though technically transfer has not happened. */
1330         bs->rd_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1331         bs->rd_ops ++;
1332     }
1333
1334     return ret;
1335 }
1336
1337 BlockDriverAIOCB *bdrv_aio_writev(BlockDriverState *bs, int64_t sector_num,
1338                                   QEMUIOVector *qiov, int nb_sectors,
1339                                   BlockDriverCompletionFunc *cb, void *opaque)
1340 {
1341     BlockDriver *drv = bs->drv;
1342     BlockDriverAIOCB *ret;
1343
1344     if (!drv)
1345         return NULL;
1346     if (bs->read_only)
1347         return NULL;
1348     if (bdrv_check_request(bs, sector_num, nb_sectors))
1349         return NULL;
1350
1351     ret = drv->bdrv_aio_writev(bs, sector_num, qiov, nb_sectors,
1352                                cb, opaque);
1353
1354     if (ret) {
1355         /* Update stats even though technically transfer has not happened. */
1356         bs->wr_bytes += (unsigned) nb_sectors * SECTOR_SIZE;
1357         bs->wr_ops ++;
1358     }
1359
1360     return ret;
1361 }
1362
1363 void bdrv_aio_cancel(BlockDriverAIOCB *acb)
1364 {
1365     acb->pool->cancel(acb);
1366 }
1367
1368
1369 /**************************************************************/
1370 /* async block device emulation */
1371
1372 static void bdrv_aio_bh_cb(void *opaque)
1373 {
1374     BlockDriverAIOCBSync *acb = opaque;
1375
1376     if (!acb->is_write)
1377         qemu_iovec_from_buffer(acb->qiov, acb->bounce, acb->qiov->size);
1378     qemu_vfree(acb->bounce);
1379     acb->common.cb(acb->common.opaque, acb->ret);
1380
1381     qemu_aio_release(acb);
1382 }
1383
1384 static BlockDriverAIOCB *bdrv_aio_rw_vector(BlockDriverState *bs,
1385                                             int64_t sector_num,
1386                                             QEMUIOVector *qiov,
1387                                             int nb_sectors,
1388                                             BlockDriverCompletionFunc *cb,
1389                                             void *opaque,
1390                                             int is_write)
1391
1392 {
1393     BlockDriverAIOCBSync *acb;
1394
1395     acb = qemu_aio_get(bs, cb, opaque);
1396     acb->is_write = is_write;
1397     acb->qiov = qiov;
1398     acb->bounce = qemu_blockalign(bs, qiov->size);
1399
1400     if (!acb->bh)
1401         acb->bh = qemu_bh_new(bdrv_aio_bh_cb, acb);
1402
1403     if (is_write) {
1404         qemu_iovec_to_buffer(acb->qiov, acb->bounce);
1405         acb->ret = bdrv_write(bs, sector_num, acb->bounce, nb_sectors);
1406     } else {
1407         acb->ret = bdrv_read(bs, sector_num, acb->bounce, nb_sectors);
1408     }
1409
1410     qemu_bh_schedule(acb->bh);
1411
1412     return &acb->common;
1413 }
1414
1415 static BlockDriverAIOCB *bdrv_aio_readv_em(BlockDriverState *bs,
1416         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1417         BlockDriverCompletionFunc *cb, void *opaque)
1418 {
1419     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 0);
1420 }
1421
1422 static BlockDriverAIOCB *bdrv_aio_writev_em(BlockDriverState *bs,
1423         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
1424         BlockDriverCompletionFunc *cb, void *opaque)
1425 {
1426     return bdrv_aio_rw_vector(bs, sector_num, qiov, nb_sectors, cb, opaque, 1);
1427 }
1428
1429
1430 static void bdrv_aio_cancel_em(BlockDriverAIOCB *blockacb)
1431 {
1432     BlockDriverAIOCBSync *acb = (BlockDriverAIOCBSync *)blockacb;
1433     qemu_bh_cancel(acb->bh);
1434     qemu_aio_release(acb);
1435 }
1436
1437 /**************************************************************/
1438 /* sync block device emulation */
1439
1440 static void bdrv_rw_em_cb(void *opaque, int ret)
1441 {
1442     *(int *)opaque = ret;
1443 }
1444
1445 #define NOT_DONE 0x7fffffff
1446
1447 static int bdrv_read_em(BlockDriverState *bs, int64_t sector_num,
1448                         uint8_t *buf, int nb_sectors)
1449 {
1450     int async_ret;
1451     BlockDriverAIOCB *acb;
1452     struct iovec iov;
1453     QEMUIOVector qiov;
1454
1455     async_ret = NOT_DONE;
1456     iov.iov_base = (void *)buf;
1457     iov.iov_len = nb_sectors * 512;
1458     qemu_iovec_init_external(&qiov, &iov, 1);
1459     acb = bdrv_aio_readv(bs, sector_num, &qiov, nb_sectors,
1460         bdrv_rw_em_cb, &async_ret);
1461     if (acb == NULL)
1462         return -1;
1463
1464     while (async_ret == NOT_DONE) {
1465         qemu_aio_wait();
1466     }
1467
1468     return async_ret;
1469 }
1470
1471 static int bdrv_write_em(BlockDriverState *bs, int64_t sector_num,
1472                          const uint8_t *buf, int nb_sectors)
1473 {
1474     int async_ret;
1475     BlockDriverAIOCB *acb;
1476     struct iovec iov;
1477     QEMUIOVector qiov;
1478
1479     async_ret = NOT_DONE;
1480     iov.iov_base = (void *)buf;
1481     iov.iov_len = nb_sectors * 512;
1482     qemu_iovec_init_external(&qiov, &iov, 1);
1483     acb = bdrv_aio_writev(bs, sector_num, &qiov, nb_sectors,
1484         bdrv_rw_em_cb, &async_ret);
1485     if (acb == NULL)
1486         return -1;
1487     while (async_ret == NOT_DONE) {
1488         qemu_aio_wait();
1489     }
1490     return async_ret;
1491 }
1492
1493 void bdrv_init(void)
1494 {
1495     module_call_init(MODULE_INIT_BLOCK);
1496 }
1497
1498 void aio_pool_init(AIOPool *pool, int aiocb_size,
1499                    void (*cancel)(BlockDriverAIOCB *acb))
1500 {
1501     pool->aiocb_size = aiocb_size;
1502     pool->cancel = cancel;
1503     pool->free_aiocb = NULL;
1504 }
1505
1506 void *qemu_aio_get_pool(AIOPool *pool, BlockDriverState *bs,
1507                         BlockDriverCompletionFunc *cb, void *opaque)
1508 {
1509     BlockDriverAIOCB *acb;
1510
1511     if (pool->free_aiocb) {
1512         acb = pool->free_aiocb;
1513         pool->free_aiocb = acb->next;
1514     } else {
1515         acb = qemu_mallocz(pool->aiocb_size);
1516         acb->pool = pool;
1517     }
1518     acb->bs = bs;
1519     acb->cb = cb;
1520     acb->opaque = opaque;
1521     return acb;
1522 }
1523
1524 void *qemu_aio_get(BlockDriverState *bs, BlockDriverCompletionFunc *cb,
1525                    void *opaque)
1526 {
1527     return qemu_aio_get_pool(&bs->drv->aio_pool, bs, cb, opaque);
1528 }
1529
1530 void qemu_aio_release(void *p)
1531 {
1532     BlockDriverAIOCB *acb = (BlockDriverAIOCB *)p;
1533     AIOPool *pool = acb->pool;
1534     acb->next = pool->free_aiocb;
1535     pool->free_aiocb = acb;
1536 }
1537
1538 /**************************************************************/
1539 /* removable device support */
1540
1541 /**
1542  * Return TRUE if the media is present
1543  */
1544 int bdrv_is_inserted(BlockDriverState *bs)
1545 {
1546     BlockDriver *drv = bs->drv;
1547     int ret;
1548     if (!drv)
1549         return 0;
1550     if (!drv->bdrv_is_inserted)
1551         return 1;
1552     ret = drv->bdrv_is_inserted(bs);
1553     return ret;
1554 }
1555
1556 /**
1557  * Return TRUE if the media changed since the last call to this
1558  * function. It is currently only used for floppy disks
1559  */
1560 int bdrv_media_changed(BlockDriverState *bs)
1561 {
1562     BlockDriver *drv = bs->drv;
1563     int ret;
1564
1565     if (!drv || !drv->bdrv_media_changed)
1566         ret = -ENOTSUP;
1567     else
1568         ret = drv->bdrv_media_changed(bs);
1569     if (ret == -ENOTSUP)
1570         ret = bs->media_changed;
1571     bs->media_changed = 0;
1572     return ret;
1573 }
1574
1575 /**
1576  * If eject_flag is TRUE, eject the media. Otherwise, close the tray
1577  */
1578 void bdrv_eject(BlockDriverState *bs, int eject_flag)
1579 {
1580     BlockDriver *drv = bs->drv;
1581     int ret;
1582
1583     if (!drv || !drv->bdrv_eject) {
1584         ret = -ENOTSUP;
1585     } else {
1586         ret = drv->bdrv_eject(bs, eject_flag);
1587     }
1588     if (ret == -ENOTSUP) {
1589         if (eject_flag)
1590             bdrv_close(bs);
1591     }
1592 }
1593
1594 int bdrv_is_locked(BlockDriverState *bs)
1595 {
1596     return bs->locked;
1597 }
1598
1599 /**
1600  * Lock or unlock the media (if it is locked, the user won't be able
1601  * to eject it manually).
1602  */
1603 void bdrv_set_locked(BlockDriverState *bs, int locked)
1604 {
1605     BlockDriver *drv = bs->drv;
1606
1607     bs->locked = locked;
1608     if (drv && drv->bdrv_set_locked) {
1609         drv->bdrv_set_locked(bs, locked);
1610     }
1611 }
1612
1613 /* needed for generic scsi interface */
1614
1615 int bdrv_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1616 {
1617     BlockDriver *drv = bs->drv;
1618
1619     if (drv && drv->bdrv_ioctl)
1620         return drv->bdrv_ioctl(bs, req, buf);
1621     return -ENOTSUP;
1622 }
1623
1624 BlockDriverAIOCB *bdrv_aio_ioctl(BlockDriverState *bs,
1625         unsigned long int req, void *buf,
1626         BlockDriverCompletionFunc *cb, void *opaque)
1627 {
1628     BlockDriver *drv = bs->drv;
1629
1630     if (drv && drv->bdrv_aio_ioctl)
1631         return drv->bdrv_aio_ioctl(bs, req, buf, cb, opaque);
1632     return NULL;
1633 }
1634
1635 void *qemu_blockalign(BlockDriverState *bs, size_t size)
1636 {
1637     return qemu_memalign((bs && bs->buffer_alignment) ? bs->buffer_alignment : 512, size);
1638 }