block/raw-posix.c

   1 /*
   2  * Block driver for RAW files (posix)
   3  *
   4  * Copyright (c) 2006 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 "qemu-common.h"
  25 #include "qemu-timer.h"
  26 #include "qemu-char.h"
  27 #include "block_int.h"
  28 #include "module.h"
  29 #include "compatfd.h"
  30 #include <assert.h>
  31 #ifdef CONFIG_AIO
  32 #include "posix-aio-compat.h"
  33 #endif
  34
  35 #ifdef CONFIG_COCOA
  36 #include <paths.h>
  37 #include <sys/param.h>
  38 #include <IOKit/IOKitLib.h>
  39 #include <IOKit/IOBSD.h>
  40 #include <IOKit/storage/IOMediaBSDClient.h>
  41 #include <IOKit/storage/IOMedia.h>
  42 #include <IOKit/storage/IOCDMedia.h>
  43 //#include <IOKit/storage/IOCDTypes.h>
  44 #include <CoreFoundation/CoreFoundation.h>
  45 #endif
  46
  47 #ifdef __sun__
  48 #define _POSIX_PTHREAD_SEMANTICS 1
  49 #include <signal.h>
  50 #include <sys/dkio.h>
  51 #endif
  52 #ifdef __linux__
  53 #include <sys/ioctl.h>
  54 #include <linux/cdrom.h>
  55 #include <linux/fd.h>
  56 #endif
  57 #ifdef __FreeBSD__
  58 #include <signal.h>
  59 #include <sys/disk.h>
  60 #include <sys/cdio.h>
  61 #endif
  62
  63 #ifdef __OpenBSD__
  64 #include <sys/ioctl.h>
  65 #include <sys/disklabel.h>
  66 #include <sys/dkio.h>
  67 #endif
  68
  69 #ifdef __DragonFly__
  70 #include <sys/ioctl.h>
  71 #include <sys/diskslice.h>
  72 #endif
  73
  74 //#define DEBUG_FLOPPY
  75
  76 //#define DEBUG_BLOCK
  77 #if defined(DEBUG_BLOCK)
  78 #define DEBUG_BLOCK_PRINT(formatCstr, ...) do { if (qemu_log_enabled()) \
  79     { qemu_log(formatCstr, ## __VA_ARGS__); qemu_log_flush(); } } while (0)
  80 #else
  81 #define DEBUG_BLOCK_PRINT(formatCstr, ...)
  82 #endif
  83
  84 /* OS X does not have O_DSYNC */
  85 #ifndef O_DSYNC
  86 #ifdef O_SYNC
  87 #define O_DSYNC O_SYNC
  88 #elif defined(O_FSYNC)
  89 #define O_DSYNC O_FSYNC
  90 #endif
  91 #endif
  92
  93 /* Approximate O_DIRECT with O_DSYNC if O_DIRECT isn't available */
  94 #ifndef O_DIRECT
  95 #define O_DIRECT O_DSYNC
  96 #endif
  97
  98 #define FTYPE_FILE   0
  99 #define FTYPE_CD     1
 100 #define FTYPE_FD     2
 101
 102 #define ALIGNED_BUFFER_SIZE (32 * 512)
 103
 104 /* if the FD is not accessed during that time (in ms), we try to
 105    reopen it to see if the disk has been changed */
 106 #define FD_OPEN_TIMEOUT 1000
 107
 108 typedef struct BDRVRawState {
 109     int fd;
 110     int type;
 111     unsigned int lseek_err_cnt;
 112     int open_flags;
 113 #if defined(__linux__)
 114     /* linux floppy specific */
 115     int64_t fd_open_time;
 116     int64_t fd_error_time;
 117     int fd_got_error;
 118     int fd_media_changed;
 119 #endif
 120     uint8_t* aligned_buf;
 121 } BDRVRawState;
 122
 123 static int posix_aio_init(void);
 124
 125 static int fd_open(BlockDriverState *bs);
 126 static int64_t raw_getlength(BlockDriverState *bs);
 127
 128 #if defined(__FreeBSD__)
 129 static int cdrom_reopen(BlockDriverState *bs);
 130 #endif
 131
 132 static int raw_open_common(BlockDriverState *bs, const char *filename,
 133                            int bdrv_flags, int open_flags)
 134 {
 135     BDRVRawState *s = bs->opaque;
 136     int fd, ret;
 137
 138     posix_aio_init();
 139
 140     s->lseek_err_cnt = 0;
 141
 142     s->open_flags = open_flags | O_BINARY;
 143     s->open_flags &= ~O_ACCMODE;
 144     if ((bdrv_flags & BDRV_O_ACCESS) == BDRV_O_RDWR) {
 145         s->open_flags |= O_RDWR;
 146     } else {
 147         s->open_flags |= O_RDONLY;
 148         bs->read_only = 1;
 149     }
 150
 151     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
 152      * and O_DIRECT for no caching. */
 153     if ((bdrv_flags & BDRV_O_NOCACHE))
 154         s->open_flags |= O_DIRECT;
 155     else if (!(bdrv_flags & BDRV_O_CACHE_WB))
 156         s->open_flags |= O_DSYNC;
 157
 158     s->fd = -1;
 159     fd = open(filename, s->open_flags, 0644);
 160     if (fd < 0) {
 161         ret = -errno;
 162         if (ret == -EROFS)
 163             ret = -EACCES;
 164         return ret;
 165     }
 166     s->fd = fd;
 167     s->aligned_buf = NULL;
 168     if ((bdrv_flags & BDRV_O_NOCACHE)) {
 169         s->aligned_buf = qemu_blockalign(bs, ALIGNED_BUFFER_SIZE);
 170         if (s->aligned_buf == NULL) {
 171             ret = -errno;
 172             close(fd);
 173             return ret;
 174         }
 175     }
 176     return 0;
 177 }
 178
 179 static int raw_open(BlockDriverState *bs, const char *filename, int flags)
 180 {
 181     BDRVRawState *s = bs->opaque;
 182     int open_flags = 0;
 183
 184     s->type = FTYPE_FILE;
 185     if (flags & BDRV_O_CREAT)
 186         open_flags = O_CREAT | O_TRUNC;
 187
 188     return raw_open_common(bs, filename, flags, open_flags);
 189 }
 190
 191 /* XXX: use host sector size if necessary with:
 192 #ifdef DIOCGSECTORSIZE
 193         {
 194             unsigned int sectorsize = 512;
 195             if (!ioctl(fd, DIOCGSECTORSIZE, &sectorsize) &&
 196                 sectorsize > bufsize)
 197                 bufsize = sectorsize;
 198         }
 199 #endif
 200 #ifdef CONFIG_COCOA
 201         u_int32_t   blockSize = 512;
 202         if ( !ioctl( fd, DKIOCGETBLOCKSIZE, &blockSize ) && blockSize > bufsize) {
 203             bufsize = blockSize;
 204         }
 205 #endif
 206 */
 207
 208 /*
 209  * offset and count are in bytes, but must be multiples of 512 for files
 210  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 211  *
 212  * This function may be called without alignment if the caller ensures
 213  * that O_DIRECT is not in effect.
 214  */
 215 static int raw_pread_aligned(BlockDriverState *bs, int64_t offset,
 216                      uint8_t *buf, int count)
 217 {
 218     BDRVRawState *s = bs->opaque;
 219     int ret;
 220
 221     ret = fd_open(bs);
 222     if (ret < 0)
 223         return ret;
 224
 225     if (offset >= 0 && lseek(s->fd, offset, SEEK_SET) == (off_t)-1) {
 226         ++(s->lseek_err_cnt);
 227         if(s->lseek_err_cnt <= 10) {
 228             DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 229                               "] lseek failed : %d = %s\n",
 230                               s->fd, bs->filename, offset, buf, count,
 231                               bs->total_sectors, errno, strerror(errno));
 232         }
 233         return -1;
 234     }
 235     s->lseek_err_cnt=0;
 236
 237     ret = read(s->fd, buf, count);
 238     if (ret == count)
 239         goto label__raw_read__success;
 240
 241     /* Allow reads beyond the end (needed for pwrite) */
 242     if ((ret == 0) && bs->growable) {
 243         int64_t size = raw_getlength(bs);
 244         if (offset >= size) {
 245             memset(buf, 0, count);
 246             ret = count;
 247             goto label__raw_read__success;
 248         }
 249     }
 250
 251     DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 252                       "] read failed %d : %d = %s\n",
 253                       s->fd, bs->filename, offset, buf, count,
 254                       bs->total_sectors, ret, errno, strerror(errno));
 255
 256     /* Try harder for CDrom. */
 257     if (bs->type == BDRV_TYPE_CDROM) {
 258         lseek(s->fd, offset, SEEK_SET);
 259         ret = read(s->fd, buf, count);
 260         if (ret == count)
 261             goto label__raw_read__success;
 262         lseek(s->fd, offset, SEEK_SET);
 263         ret = read(s->fd, buf, count);
 264         if (ret == count)
 265             goto label__raw_read__success;
 266
 267         DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 268                           "] retry read failed %d : %d = %s\n",
 269                           s->fd, bs->filename, offset, buf, count,
 270                           bs->total_sectors, ret, errno, strerror(errno));
 271     }
 272
 273 label__raw_read__success:
 274
 275     return  (ret < 0) ? -errno : ret;
 276 }
 277
 278 /*
 279  * offset and count are in bytes, but must be multiples of 512 for files
 280  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 281  *
 282  * This function may be called without alignment if the caller ensures
 283  * that O_DIRECT is not in effect.
 284  */
 285 static int raw_pwrite_aligned(BlockDriverState *bs, int64_t offset,
 286                       const uint8_t *buf, int count)
 287 {
 288     BDRVRawState *s = bs->opaque;
 289     int ret;
 290
 291     ret = fd_open(bs);
 292     if (ret < 0)
 293         return -errno;
 294
 295     if (offset >= 0 && lseek(s->fd, offset, SEEK_SET) == (off_t)-1) {
 296         ++(s->lseek_err_cnt);
 297         if(s->lseek_err_cnt) {
 298             DEBUG_BLOCK_PRINT("raw_pwrite(%d:%s, %" PRId64 ", %p, %d) [%"
 299                               PRId64 "] lseek failed : %d = %s\n",
 300                               s->fd, bs->filename, offset, buf, count,
 301                               bs->total_sectors, errno, strerror(errno));
 302         }
 303         return -EIO;
 304     }
 305     s->lseek_err_cnt = 0;
 306
 307     ret = write(s->fd, buf, count);
 308     if (ret == count)
 309         goto label__raw_write__success;
 310
 311     DEBUG_BLOCK_PRINT("raw_pwrite(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 312                       "] write failed %d : %d = %s\n",
 313                       s->fd, bs->filename, offset, buf, count,
 314                       bs->total_sectors, ret, errno, strerror(errno));
 315
 316 label__raw_write__success:
 317
 318     return  (ret < 0) ? -errno : ret;
 319 }
 320
 321
 322 /*
 323  * offset and count are in bytes and possibly not aligned. For files opened
 324  * with O_DIRECT, necessary alignments are ensured before calling
 325  * raw_pread_aligned to do the actual read.
 326  */
 327 static int raw_pread(BlockDriverState *bs, int64_t offset,
 328                      uint8_t *buf, int count)
 329 {
 330     BDRVRawState *s = bs->opaque;
 331     int size, ret, shift, sum;
 332
 333     sum = 0;
 334
 335     if (s->aligned_buf != NULL)  {
 336
 337         if (offset & 0x1ff) {
 338             /* align offset on a 512 bytes boundary */
 339
 340             shift = offset & 0x1ff;
 341             size = (shift + count + 0x1ff) & ~0x1ff;
 342             if (size > ALIGNED_BUFFER_SIZE)
 343                 size = ALIGNED_BUFFER_SIZE;
 344             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, size);
 345             if (ret < 0)
 346                 return ret;
 347
 348             size = 512 - shift;
 349             if (size > count)
 350                 size = count;
 351             memcpy(buf, s->aligned_buf + shift, size);
 352
 353             buf += size;
 354             offset += size;
 355             count -= size;
 356             sum += size;
 357
 358             if (count == 0)
 359                 return sum;
 360         }
 361         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 362
 363             /* read on aligned buffer */
 364
 365             while (count) {
 366
 367                 size = (count + 0x1ff) & ~0x1ff;
 368                 if (size > ALIGNED_BUFFER_SIZE)
 369                     size = ALIGNED_BUFFER_SIZE;
 370
 371                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, size);
 372                 if (ret < 0)
 373                     return ret;
 374
 375                 size = ret;
 376                 if (size > count)
 377                     size = count;
 378
 379                 memcpy(buf, s->aligned_buf, size);
 380
 381                 buf += size;
 382                 offset += size;
 383                 count -= size;
 384                 sum += size;
 385             }
 386
 387             return sum;
 388         }
 389     }
 390
 391     return raw_pread_aligned(bs, offset, buf, count) + sum;
 392 }
 393
 394 static int raw_read(BlockDriverState *bs, int64_t sector_num,
 395                     uint8_t *buf, int nb_sectors)
 396 {
 397     int ret;
 398
 399     ret = raw_pread(bs, sector_num * 512, buf, nb_sectors * 512);
 400     if (ret == (nb_sectors * 512))
 401         ret = 0;
 402     return ret;
 403 }
 404
 405 /*
 406  * offset and count are in bytes and possibly not aligned. For files opened
 407  * with O_DIRECT, necessary alignments are ensured before calling
 408  * raw_pwrite_aligned to do the actual write.
 409  */
 410 static int raw_pwrite(BlockDriverState *bs, int64_t offset,
 411                       const uint8_t *buf, int count)
 412 {
 413     BDRVRawState *s = bs->opaque;
 414     int size, ret, shift, sum;
 415
 416     sum = 0;
 417
 418     if (s->aligned_buf != NULL) {
 419
 420         if (offset & 0x1ff) {
 421             /* align offset on a 512 bytes boundary */
 422             shift = offset & 0x1ff;
 423             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, 512);
 424             if (ret < 0)
 425                 return ret;
 426
 427             size = 512 - shift;
 428             if (size > count)
 429                 size = count;
 430             memcpy(s->aligned_buf + shift, buf, size);
 431
 432             ret = raw_pwrite_aligned(bs, offset - shift, s->aligned_buf, 512);
 433             if (ret < 0)
 434                 return ret;
 435
 436             buf += size;
 437             offset += size;
 438             count -= size;
 439             sum += size;
 440
 441             if (count == 0)
 442                 return sum;
 443         }
 444         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 445
 446             while ((size = (count & ~0x1ff)) != 0) {
 447
 448                 if (size > ALIGNED_BUFFER_SIZE)
 449                     size = ALIGNED_BUFFER_SIZE;
 450
 451                 memcpy(s->aligned_buf, buf, size);
 452
 453                 ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, size);
 454                 if (ret < 0)
 455                     return ret;
 456
 457                 buf += ret;
 458                 offset += ret;
 459                 count -= ret;
 460                 sum += ret;
 461             }
 462             /* here, count < 512 because (count & ~0x1ff) == 0 */
 463             if (count) {
 464                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, 512);
 465                 if (ret < 0)
 466                     return ret;
 467                  memcpy(s->aligned_buf, buf, count);
 468
 469                  ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, 512);
 470                  if (ret < 0)
 471                      return ret;
 472                  if (count < ret)
 473                      ret = count;
 474
 475                  sum += ret;
 476             }
 477             return sum;
 478         }
 479     }
 480     return raw_pwrite_aligned(bs, offset, buf, count) + sum;
 481 }
 482
 483 static int raw_write(BlockDriverState *bs, int64_t sector_num,
 484                      const uint8_t *buf, int nb_sectors)
 485 {
 486     int ret;
 487     ret = raw_pwrite(bs, sector_num * 512, buf, nb_sectors * 512);
 488     if (ret == (nb_sectors * 512))
 489         ret = 0;
 490     return ret;
 491 }
 492
 493 #ifdef CONFIG_AIO
 494 /***********************************************************/
 495 /* Unix AIO using POSIX AIO */
 496
 497 typedef struct RawAIOCB {
 498     BlockDriverAIOCB common;
 499     struct qemu_paiocb aiocb;
 500     struct RawAIOCB *next;
 501     int ret;
 502 } RawAIOCB;
 503
 504 typedef struct PosixAioState
 505 {
 506     int fd;
 507     RawAIOCB *first_aio;
 508 } PosixAioState;
 509
 510 static void posix_aio_read(void *opaque)
 511 {
 512     PosixAioState *s = opaque;
 513     RawAIOCB *acb, **pacb;
 514     int ret;
 515     size_t offset;
 516     union {
 517         struct qemu_signalfd_siginfo siginfo;
 518         char buf[128];
 519     } sig;
 520
 521     /* try to read from signalfd, don't freak out if we can't read anything */
 522     offset = 0;
 523     while (offset < 128) {
 524         ssize_t len;
 525
 526         len = read(s->fd, sig.buf + offset, 128 - offset);
 527         if (len == -1 && errno == EINTR)
 528             continue;
 529         if (len == -1 && errno == EAGAIN) {
 530             /* there is no natural reason for this to happen,
 531              * so we'll spin hard until we get everything just
 532              * to be on the safe side. */
 533             if (offset > 0)
 534                 continue;
 535         }
 536
 537         offset += len;
 538     }
 539
 540     for(;;) {
 541         pacb = &s->first_aio;
 542         for(;;) {
 543             acb = *pacb;
 544             if (!acb)
 545                 goto the_end;
 546             ret = qemu_paio_error(&acb->aiocb);
 547             if (ret == ECANCELED) {
 548                 /* remove the request */
 549                 *pacb = acb->next;
 550                 qemu_aio_release(acb);
 551             } else if (ret != EINPROGRESS) {
 552                 /* end of aio */
 553                 if (ret == 0) {
 554                     ret = qemu_paio_return(&acb->aiocb);
 555                     if (ret == acb->aiocb.aio_nbytes)
 556                         ret = 0;
 557                     else
 558                         ret = -EINVAL;
 559                 } else {
 560                     ret = -ret;
 561                 }
 562                 /* remove the request */
 563                 *pacb = acb->next;
 564                 /* call the callback */
 565                 acb->common.cb(acb->common.opaque, ret);
 566                 qemu_aio_release(acb);
 567                 break;
 568             } else {
 569                 pacb = &acb->next;
 570             }
 571         }
 572     }
 573  the_end: ;
 574 }
 575
 576 static int posix_aio_flush(void *opaque)
 577 {
 578     PosixAioState *s = opaque;
 579     return !!s->first_aio;
 580 }
 581
 582 static PosixAioState *posix_aio_state;
 583
 584 static int posix_aio_init(void)
 585 {
 586     sigset_t mask;
 587     PosixAioState *s;
 588     struct qemu_paioinit ai;
 589
 590     if (posix_aio_state)
 591         return 0;
 592
 593     s = qemu_malloc(sizeof(PosixAioState));
 594
 595     /* Make sure to block AIO signal */
 596     sigemptyset(&mask);
 597     sigaddset(&mask, SIGUSR2);
 598     sigprocmask(SIG_BLOCK, &mask, NULL);
 599
 600     s->first_aio = NULL;
 601     s->fd = qemu_signalfd(&mask);
 602     if (s->fd == -1) {
 603         fprintf(stderr, "failed to create signalfd\n");
 604         return -errno;
 605     }
 606
 607     fcntl(s->fd, F_SETFL, O_NONBLOCK);
 608
 609     qemu_aio_set_fd_handler(s->fd, posix_aio_read, NULL, posix_aio_flush, s);
 610
 611     memset(&ai, 0, sizeof(ai));
 612     ai.aio_threads = 64;
 613     ai.aio_num = 64;
 614     qemu_paio_init(&ai);
 615
 616     posix_aio_state = s;
 617
 618     return 0;
 619 }
 620
 621 static void raw_aio_remove(RawAIOCB *acb)
 622 {
 623     RawAIOCB **pacb;
 624
 625     /* remove the callback from the queue */
 626     pacb = &posix_aio_state->first_aio;
 627     for(;;) {
 628         if (*pacb == NULL) {
 629             fprintf(stderr, "raw_aio_remove: aio request not found!\n");
 630             break;
 631         } else if (*pacb == acb) {
 632             *pacb = acb->next;
 633             qemu_aio_release(acb);
 634             break;
 635         }
 636         pacb = &(*pacb)->next;
 637     }
 638 }
 639
 640 static void raw_aio_cancel(BlockDriverAIOCB *blockacb)
 641 {
 642     int ret;
 643     RawAIOCB *acb = (RawAIOCB *)blockacb;
 644
 645     ret = qemu_paio_cancel(acb->aiocb.aio_fildes, &acb->aiocb);
 646     if (ret == QEMU_PAIO_NOTCANCELED) {
 647         /* fail safe: if the aio could not be canceled, we wait for
 648            it */
 649         while (qemu_paio_error(&acb->aiocb) == EINPROGRESS);
 650     }
 651
 652     raw_aio_remove(acb);
 653 }
 654
 655 static AIOPool raw_aio_pool = {
 656     .aiocb_size         = sizeof(RawAIOCB),
 657     .cancel             = raw_aio_cancel,
 658 };
 659
 660 static RawAIOCB *raw_aio_setup(BlockDriverState *bs, int64_t sector_num,
 661         QEMUIOVector *qiov, int nb_sectors,
 662         BlockDriverCompletionFunc *cb, void *opaque)
 663 {
 664     BDRVRawState *s = bs->opaque;
 665     RawAIOCB *acb;
 666
 667     if (fd_open(bs) < 0)
 668         return NULL;
 669
 670     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
 671     if (!acb)
 672         return NULL;
 673     acb->aiocb.aio_fildes = s->fd;
 674     acb->aiocb.ev_signo = SIGUSR2;
 675     acb->aiocb.aio_iov = qiov->iov;
 676     acb->aiocb.aio_niov = qiov->niov;
 677     acb->aiocb.aio_nbytes = nb_sectors * 512;
 678     acb->aiocb.aio_offset = sector_num * 512;
 679     acb->aiocb.aio_flags = 0;
 680
 681     /*
 682      * If O_DIRECT is used the buffer needs to be aligned on a sector
 683      * boundary. Tell the low level code to ensure that in case it's
 684      * not done yet.
 685      */
 686     if (s->aligned_buf)
 687         acb->aiocb.aio_flags |= QEMU_AIO_SECTOR_ALIGNED;
 688
 689     acb->next = posix_aio_state->first_aio;
 690     posix_aio_state->first_aio = acb;
 691     return acb;
 692 }
 693
 694 static BlockDriverAIOCB *raw_aio_readv(BlockDriverState *bs,
 695         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 696         BlockDriverCompletionFunc *cb, void *opaque)
 697 {
 698     RawAIOCB *acb;
 699
 700     acb = raw_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque);
 701     if (!acb)
 702         return NULL;
 703     if (qemu_paio_read(&acb->aiocb) < 0) {
 704         raw_aio_remove(acb);
 705         return NULL;
 706     }
 707     return &acb->common;
 708 }
 709
 710 static BlockDriverAIOCB *raw_aio_writev(BlockDriverState *bs,
 711         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 712         BlockDriverCompletionFunc *cb, void *opaque)
 713 {
 714     RawAIOCB *acb;
 715
 716     acb = raw_aio_setup(bs, sector_num, qiov, nb_sectors, cb, opaque);
 717     if (!acb)
 718         return NULL;
 719     if (qemu_paio_write(&acb->aiocb) < 0) {
 720         raw_aio_remove(acb);
 721         return NULL;
 722     }
 723     return &acb->common;
 724 }
 725 #else /* CONFIG_AIO */
 726 static int posix_aio_init(void)
 727 {
 728     return 0;
 729 }
 730 #endif /* CONFIG_AIO */
 731
 732
 733 static void raw_close(BlockDriverState *bs)
 734 {
 735     BDRVRawState *s = bs->opaque;
 736     if (s->fd >= 0) {
 737         close(s->fd);
 738         s->fd = -1;
 739         if (s->aligned_buf != NULL)
 740             qemu_free(s->aligned_buf);
 741     }
 742 }
 743
 744 static int raw_truncate(BlockDriverState *bs, int64_t offset)
 745 {
 746     BDRVRawState *s = bs->opaque;
 747     if (s->type != FTYPE_FILE)
 748         return -ENOTSUP;
 749     if (ftruncate(s->fd, offset) < 0)
 750         return -errno;
 751     return 0;
 752 }
 753
 754 #ifdef __OpenBSD__
 755 static int64_t raw_getlength(BlockDriverState *bs)
 756 {
 757     BDRVRawState *s = bs->opaque;
 758     int fd = s->fd;
 759     struct stat st;
 760
 761     if (fstat(fd, &st))
 762         return -1;
 763     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
 764         struct disklabel dl;
 765
 766         if (ioctl(fd, DIOCGDINFO, &dl))
 767             return -1;
 768         return (uint64_t)dl.d_secsize *
 769             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
 770     } else
 771         return st.st_size;
 772 }
 773 #else /* !__OpenBSD__ */
 774 static int64_t  raw_getlength(BlockDriverState *bs)
 775 {
 776     BDRVRawState *s = bs->opaque;
 777     int fd = s->fd;
 778     int64_t size;
 779 #ifdef HOST_BSD
 780     struct stat sb;
 781 #ifdef __FreeBSD__
 782     int reopened = 0;
 783 #endif
 784 #endif
 785 #ifdef __sun__
 786     struct dk_minfo minfo;
 787     int rv;
 788 #endif
 789     int ret;
 790
 791     ret = fd_open(bs);
 792     if (ret < 0)
 793         return ret;
 794
 795 #ifdef HOST_BSD
 796 #ifdef __FreeBSD__
 797 again:
 798 #endif
 799     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
 800 #ifdef DIOCGMEDIASIZE
 801         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size))
 802 #elif defined(DIOCGPART)
 803         {
 804                 struct partinfo pi;
 805                 if (ioctl(fd, DIOCGPART, &pi) == 0)
 806                         size = pi.media_size;
 807                 else
 808                         size = 0;
 809         }
 810         if (size == 0)
 811 #endif
 812 #ifdef CONFIG_COCOA
 813         size = LONG_LONG_MAX;
 814 #else
 815         size = lseek(fd, 0LL, SEEK_END);
 816 #endif
 817 #ifdef __FreeBSD__
 818         switch(s->type) {
 819         case FTYPE_CD:
 820             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
 821             if (size == 2048LL * (unsigned)-1)
 822                 size = 0;
 823             /* XXX no disc?  maybe we need to reopen... */
 824             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
 825                 reopened = 1;
 826                 goto again;
 827             }
 828         }
 829 #endif
 830     } else
 831 #endif
 832 #ifdef __sun__
 833     /*
 834      * use the DKIOCGMEDIAINFO ioctl to read the size.
 835      */
 836     rv = ioctl ( fd, DKIOCGMEDIAINFO, &minfo );
 837     if ( rv != -1 ) {
 838         size = minfo.dki_lbsize * minfo.dki_capacity;
 839     } else /* there are reports that lseek on some devices
 840               fails, but irc discussion said that contingency
 841               on contingency was overkill */
 842 #endif
 843     {
 844         size = lseek(fd, 0, SEEK_END);
 845     }
 846     return size;
 847 }
 848 #endif
 849
 850 static int raw_create(const char *filename, QEMUOptionParameter *options)
 851 {
 852     int fd;
 853     int64_t total_size = 0;
 854
 855     /* Read out options */
 856     while (options && options->name) {
 857         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
 858             total_size = options->value.n / 512;
 859         }
 860         options++;
 861     }
 862
 863     fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC | O_BINARY,
 864               0644);
 865     if (fd < 0)
 866         return -EIO;
 867     ftruncate(fd, total_size * 512);
 868     close(fd);
 869     return 0;
 870 }
 871
 872 static void raw_flush(BlockDriverState *bs)
 873 {
 874     BDRVRawState *s = bs->opaque;
 875     fsync(s->fd);
 876 }
 877
 878
 879 static QEMUOptionParameter raw_create_options[] = {
 880     {
 881         .name = BLOCK_OPT_SIZE,
 882         .type = OPT_SIZE,
 883         .help = "Virtual disk size"
 884     },
 885     { NULL }
 886 };
 887
 888 static BlockDriver bdrv_raw = {
 889     .format_name = "raw",
 890     .instance_size = sizeof(BDRVRawState),
 891     .bdrv_probe = NULL, /* no probe for protocols */
 892     .bdrv_open = raw_open,
 893     .bdrv_read = raw_read,
 894     .bdrv_write = raw_write,
 895     .bdrv_close = raw_close,
 896     .bdrv_create = raw_create,
 897     .bdrv_flush = raw_flush,
 898
 899 #ifdef CONFIG_AIO
 900     .bdrv_aio_readv = raw_aio_readv,
 901     .bdrv_aio_writev = raw_aio_writev,
 902 #endif
 903
 904     .bdrv_truncate = raw_truncate,
 905     .bdrv_getlength = raw_getlength,
 906
 907     .create_options = raw_create_options,
 908 };
 909
 910 /***********************************************/
 911 /* host device */
 912
 913 #ifdef CONFIG_COCOA
 914 static kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator );
 915 static kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize );
 916
 917 kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator )
 918 {
 919     kern_return_t       kernResult;
 920     mach_port_t     masterPort;
 921     CFMutableDictionaryRef  classesToMatch;
 922
 923     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
 924     if ( KERN_SUCCESS != kernResult ) {
 925         printf( "IOMasterPort returned %d\n", kernResult );
 926     }
 927
 928     classesToMatch = IOServiceMatching( kIOCDMediaClass );
 929     if ( classesToMatch == NULL ) {
 930         printf( "IOServiceMatching returned a NULL dictionary.\n" );
 931     } else {
 932     CFDictionarySetValue( classesToMatch, CFSTR( kIOMediaEjectableKey ), kCFBooleanTrue );
 933     }
 934     kernResult = IOServiceGetMatchingServices( masterPort, classesToMatch, mediaIterator );
 935     if ( KERN_SUCCESS != kernResult )
 936     {
 937         printf( "IOServiceGetMatchingServices returned %d\n", kernResult );
 938     }
 939
 940     return kernResult;
 941 }
 942
 943 kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize )
 944 {
 945     io_object_t     nextMedia;
 946     kern_return_t   kernResult = KERN_FAILURE;
 947     *bsdPath = '\0';
 948     nextMedia = IOIteratorNext( mediaIterator );
 949     if ( nextMedia )
 950     {
 951         CFTypeRef   bsdPathAsCFString;
 952     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
 953         if ( bsdPathAsCFString ) {
 954             size_t devPathLength;
 955             strcpy( bsdPath, _PATH_DEV );
 956             strcat( bsdPath, "r" );
 957             devPathLength = strlen( bsdPath );
 958             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
 959                 kernResult = KERN_SUCCESS;
 960             }
 961             CFRelease( bsdPathAsCFString );
 962         }
 963         IOObjectRelease( nextMedia );
 964     }
 965
 966     return kernResult;
 967 }
 968
 969 #endif
 970
 971 static int hdev_probe_device(const char *filename)
 972 {
 973     struct stat st;
 974
 975     /* allow a dedicated CD-ROM driver to match with a higher priority */
 976     if (strstart(filename, "/dev/cdrom", NULL))
 977         return 50;
 978
 979     if (stat(filename, &st) >= 0 &&
 980             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 981         return 100;
 982     }
 983
 984     return 0;
 985 }
 986
 987 static int hdev_open(BlockDriverState *bs, const char *filename, int flags)
 988 {
 989     BDRVRawState *s = bs->opaque;
 990
 991 #ifdef CONFIG_COCOA
 992     if (strstart(filename, "/dev/cdrom", NULL)) {
 993         kern_return_t kernResult;
 994         io_iterator_t mediaIterator;
 995         char bsdPath[ MAXPATHLEN ];
 996         int fd;
 997
 998         kernResult = FindEjectableCDMedia( &mediaIterator );
 999         kernResult = GetBSDPath( mediaIterator, bsdPath, sizeof( bsdPath ) );
1000
1001         if ( bsdPath[ 0 ] != '\0' ) {
1002             strcat(bsdPath,"s0");
1003             /* some CDs don't have a partition 0 */
1004             fd = open(bsdPath, O_RDONLY | O_BINARY | O_LARGEFILE);
1005             if (fd < 0) {
1006                 bsdPath[strlen(bsdPath)-1] = '1';
1007             } else {
1008                 close(fd);
1009             }
1010             filename = bsdPath;
1011         }
1012
1013         if ( mediaIterator )
1014             IOObjectRelease( mediaIterator );
1015     }
1016 #endif
1017
1018     s->type = FTYPE_FILE;
1019 #if defined(__linux__) && defined(CONFIG_AIO)
1020     if (strstart(filename, "/dev/sg", NULL)) {
1021         bs->sg = 1;
1022     }
1023 #endif
1024
1025     return raw_open_common(bs, filename, flags, 0);
1026 }
1027
1028 #if defined(__linux__)
1029 /* Note: we do not have a reliable method to detect if the floppy is
1030    present. The current method is to try to open the floppy at every
1031    I/O and to keep it opened during a few hundreds of ms. */
1032 static int fd_open(BlockDriverState *bs)
1033 {
1034     BDRVRawState *s = bs->opaque;
1035     int last_media_present;
1036
1037     if (s->type != FTYPE_FD)
1038         return 0;
1039     last_media_present = (s->fd >= 0);
1040     if (s->fd >= 0 &&
1041         (qemu_get_clock(rt_clock) - s->fd_open_time) >= FD_OPEN_TIMEOUT) {
1042         close(s->fd);
1043         s->fd = -1;
1044 #ifdef DEBUG_FLOPPY
1045         printf("Floppy closed\n");
1046 #endif
1047     }
1048     if (s->fd < 0) {
1049         if (s->fd_got_error &&
1050             (qemu_get_clock(rt_clock) - s->fd_error_time) < FD_OPEN_TIMEOUT) {
1051 #ifdef DEBUG_FLOPPY
1052             printf("No floppy (open delayed)\n");
1053 #endif
1054             return -EIO;
1055         }
1056         s->fd = open(bs->filename, s->open_flags & ~O_NONBLOCK);
1057         if (s->fd < 0) {
1058             s->fd_error_time = qemu_get_clock(rt_clock);
1059             s->fd_got_error = 1;
1060             if (last_media_present)
1061                 s->fd_media_changed = 1;
1062 #ifdef DEBUG_FLOPPY
1063             printf("No floppy\n");
1064 #endif
1065             return -EIO;
1066         }
1067 #ifdef DEBUG_FLOPPY
1068         printf("Floppy opened\n");
1069 #endif
1070     }
1071     if (!last_media_present)
1072         s->fd_media_changed = 1;
1073     s->fd_open_time = qemu_get_clock(rt_clock);
1074     s->fd_got_error = 0;
1075     return 0;
1076 }
1077
1078 static int hdev_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
1079 {
1080     BDRVRawState *s = bs->opaque;
1081
1082     return ioctl(s->fd, req, buf);
1083 }
1084
1085 #ifdef CONFIG_AIO
1086 static BlockDriverAIOCB *hdev_aio_ioctl(BlockDriverState *bs,
1087         unsigned long int req, void *buf,
1088         BlockDriverCompletionFunc *cb, void *opaque)
1089 {
1090     BDRVRawState *s = bs->opaque;
1091     RawAIOCB *acb;
1092
1093     if (fd_open(bs) < 0)
1094         return NULL;
1095
1096     acb = qemu_aio_get(&raw_aio_pool, bs, cb, opaque);
1097     if (!acb)
1098         return NULL;
1099     acb->aiocb.aio_fildes = s->fd;
1100     acb->aiocb.ev_signo = SIGUSR2;
1101     acb->aiocb.aio_offset = 0;
1102     acb->aiocb.aio_flags = 0;
1103
1104     acb->next = posix_aio_state->first_aio;
1105     posix_aio_state->first_aio = acb;
1106
1107     acb->aiocb.aio_ioctl_buf = buf;
1108     acb->aiocb.aio_ioctl_cmd = req;
1109     if (qemu_paio_ioctl(&acb->aiocb) < 0) {
1110         raw_aio_remove(acb);
1111         return NULL;
1112     }
1113
1114     return &acb->common;
1115 }
1116 #endif
1117
1118 #elif defined(__FreeBSD__)
1119 static int fd_open(BlockDriverState *bs)
1120 {
1121     BDRVRawState *s = bs->opaque;
1122
1123     /* this is just to ensure s->fd is sane (its called by io ops) */
1124     if (s->fd >= 0)
1125         return 0;
1126     return -EIO;
1127 }
1128 #else /* !linux && !FreeBSD */
1129
1130 static int fd_open(BlockDriverState *bs)
1131 {
1132     return 0;
1133 }
1134
1135 #endif /* !linux && !FreeBSD */
1136
1137 static int hdev_create(const char *filename, QEMUOptionParameter *options)
1138 {
1139     int fd;
1140     int ret = 0;
1141     struct stat stat_buf;
1142     int64_t total_size = 0;
1143
1144     /* Read out options */
1145     while (options && options->name) {
1146         if (!strcmp(options->name, "size")) {
1147             total_size = options->value.n / 512;
1148         }
1149         options++;
1150     }
1151
1152     fd = open(filename, O_WRONLY | O_BINARY);
1153     if (fd < 0)
1154         return -EIO;
1155
1156     if (fstat(fd, &stat_buf) < 0)
1157         ret = -EIO;
1158     else if (!S_ISBLK(stat_buf.st_mode) && !S_ISCHR(stat_buf.st_mode))
1159         ret = -EIO;
1160     else if (lseek(fd, 0, SEEK_END) < total_size * 512)
1161         ret = -ENOSPC;
1162
1163     close(fd);
1164     return ret;
1165 }
1166
1167 static BlockDriver bdrv_host_device = {
1168     .format_name        = "host_device",
1169     .instance_size      = sizeof(BDRVRawState),
1170     .bdrv_probe_device  = hdev_probe_device,
1171     .bdrv_open          = hdev_open,
1172     .bdrv_close         = raw_close,
1173     .bdrv_create        = hdev_create,
1174     .bdrv_flush         = raw_flush,
1175
1176 #ifdef CONFIG_AIO
1177     .bdrv_aio_readv     = raw_aio_readv,
1178     .bdrv_aio_writev    = raw_aio_writev,
1179 #endif
1180
1181     .bdrv_read          = raw_read,
1182     .bdrv_write         = raw_write,
1183     .bdrv_getlength     = raw_getlength,
1184
1185     /* generic scsi device */
1186 #ifdef __linux__
1187     .bdrv_ioctl         = hdev_ioctl,
1188 #ifdef CONFIG_AIO
1189     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1190 #endif
1191 #endif
1192 };
1193
1194 #ifdef __linux__
1195 static int floppy_open(BlockDriverState *bs, const char *filename, int flags)
1196 {
1197     BDRVRawState *s = bs->opaque;
1198     int ret;
1199
1200     posix_aio_init();
1201
1202     s->type = FTYPE_FD;
1203
1204     /* open will not fail even if no floppy is inserted, so add O_NONBLOCK */
1205     ret = raw_open_common(bs, filename, flags, O_NONBLOCK);
1206     if (ret)
1207         return ret;
1208
1209     /* close fd so that we can reopen it as needed */
1210     close(s->fd);
1211     s->fd = -1;
1212     s->fd_media_changed = 1;
1213
1214     return 0;
1215 }
1216
1217 static int floppy_probe_device(const char *filename)
1218 {
1219     if (strstart(filename, "/dev/fd", NULL))
1220         return 100;
1221     return 0;
1222 }
1223
1224
1225 static int floppy_is_inserted(BlockDriverState *bs)
1226 {
1227     return fd_open(bs) >= 0;
1228 }
1229
1230 static int floppy_media_changed(BlockDriverState *bs)
1231 {
1232     BDRVRawState *s = bs->opaque;
1233     int ret;
1234
1235     /*
1236      * XXX: we do not have a true media changed indication.
1237      * It does not work if the floppy is changed without trying to read it.
1238      */
1239     fd_open(bs);
1240     ret = s->fd_media_changed;
1241     s->fd_media_changed = 0;
1242 #ifdef DEBUG_FLOPPY
1243     printf("Floppy changed=%d\n", ret);
1244 #endif
1245     return ret;
1246 }
1247
1248 static int floppy_eject(BlockDriverState *bs, int eject_flag)
1249 {
1250     BDRVRawState *s = bs->opaque;
1251     int fd;
1252
1253     if (s->fd >= 0) {
1254         close(s->fd);
1255         s->fd = -1;
1256     }
1257     fd = open(bs->filename, s->open_flags | O_NONBLOCK);
1258     if (fd >= 0) {
1259         if (ioctl(fd, FDEJECT, 0) < 0)
1260             perror("FDEJECT");
1261         close(fd);
1262     }
1263
1264     return 0;
1265 }
1266
1267 static BlockDriver bdrv_host_floppy = {
1268     .format_name        = "host_floppy",
1269     .instance_size      = sizeof(BDRVRawState),
1270     .bdrv_probe_device  = floppy_probe_device,
1271     .bdrv_open          = floppy_open,
1272     .bdrv_close         = raw_close,
1273     .bdrv_create        = hdev_create,
1274     .bdrv_flush         = raw_flush,
1275
1276 #ifdef CONFIG_AIO
1277     .bdrv_aio_readv     = raw_aio_readv,
1278     .bdrv_aio_writev    = raw_aio_writev,
1279 #endif
1280
1281     .bdrv_read          = raw_read,
1282     .bdrv_write         = raw_write,
1283     .bdrv_getlength     = raw_getlength,
1284
1285     /* removable device support */
1286     .bdrv_is_inserted   = floppy_is_inserted,
1287     .bdrv_media_changed = floppy_media_changed,
1288     .bdrv_eject         = floppy_eject,
1289 };
1290
1291 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1292 {
1293     BDRVRawState *s = bs->opaque;
1294
1295     s->type = FTYPE_CD;
1296
1297     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
1298     return raw_open_common(bs, filename, flags, O_NONBLOCK);
1299 }
1300
1301 static int cdrom_probe_device(const char *filename)
1302 {
1303     if (strstart(filename, "/dev/cd", NULL))
1304         return 100;
1305     return 0;
1306 }
1307
1308 static int cdrom_is_inserted(BlockDriverState *bs)
1309 {
1310     BDRVRawState *s = bs->opaque;
1311     int ret;
1312
1313     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
1314     if (ret == CDS_DISC_OK)
1315         return 1;
1316     return 0;
1317 }
1318
1319 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1320 {
1321     BDRVRawState *s = bs->opaque;
1322
1323     if (eject_flag) {
1324         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
1325             perror("CDROMEJECT");
1326     } else {
1327         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
1328             perror("CDROMEJECT");
1329     }
1330
1331     return 0;
1332 }
1333
1334 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1335 {
1336     BDRVRawState *s = bs->opaque;
1337
1338     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
1339         /*
1340          * Note: an error can happen if the distribution automatically
1341          * mounts the CD-ROM
1342          */
1343         /* perror("CDROM_LOCKDOOR"); */
1344     }
1345
1346     return 0;
1347 }
1348
1349 static BlockDriver bdrv_host_cdrom = {
1350     .format_name        = "host_cdrom",
1351     .instance_size      = sizeof(BDRVRawState),
1352     .bdrv_probe_device  = cdrom_probe_device,
1353     .bdrv_open          = cdrom_open,
1354     .bdrv_close         = raw_close,
1355     .bdrv_create        = hdev_create,
1356     .bdrv_flush         = raw_flush,
1357
1358 #ifdef CONFIG_AIO
1359     .bdrv_aio_readv     = raw_aio_readv,
1360     .bdrv_aio_writev    = raw_aio_writev,
1361 #endif
1362
1363     .bdrv_read          = raw_read,
1364     .bdrv_write         = raw_write,
1365     .bdrv_getlength     = raw_getlength,
1366
1367     /* removable device support */
1368     .bdrv_is_inserted   = cdrom_is_inserted,
1369     .bdrv_eject         = cdrom_eject,
1370     .bdrv_set_locked    = cdrom_set_locked,
1371
1372     /* generic scsi device */
1373     .bdrv_ioctl         = hdev_ioctl,
1374 #ifdef CONFIG_AIO
1375     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1376 #endif
1377 };
1378 #endif /* __linux__ */
1379
1380 #ifdef __FreeBSD__
1381 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1382 {
1383     BDRVRawState *s = bs->opaque;
1384     int ret;
1385
1386     s->type = FTYPE_CD;
1387
1388     ret = raw_open_common(bs, filename, flags, 0);
1389     if (ret)
1390         return ret;
1391
1392     /* make sure the door isnt locked at this time */
1393     ioctl(s->fd, CDIOCALLOW);
1394     return 0;
1395 }
1396
1397 static int cdrom_probe_device(const char *filename)
1398 {
1399     if (strstart(filename, "/dev/cd", NULL) ||
1400             strstart(filename, "/dev/acd", NULL))
1401         return 100;
1402     return 0;
1403 }
1404
1405 static int cdrom_reopen(BlockDriverState *bs)
1406 {
1407     BDRVRawState *s = bs->opaque;
1408     int fd;
1409
1410     /*
1411      * Force reread of possibly changed/newly loaded disc,
1412      * FreeBSD seems to not notice sometimes...
1413      */
1414     if (s->fd >= 0)
1415         close(s->fd);
1416     fd = open(bs->filename, s->open_flags, 0644);
1417     if (fd < 0) {
1418         s->fd = -1;
1419         return -EIO;
1420     }
1421     s->fd = fd;
1422
1423     /* make sure the door isnt locked at this time */
1424     ioctl(s->fd, CDIOCALLOW);
1425     return 0;
1426 }
1427
1428 static int cdrom_is_inserted(BlockDriverState *bs)
1429 {
1430     return raw_getlength(bs) > 0;
1431 }
1432
1433 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1434 {
1435     BDRVRawState *s = bs->opaque;
1436
1437     if (s->fd < 0)
1438         return -ENOTSUP;
1439
1440     (void) ioctl(s->fd, CDIOCALLOW);
1441
1442     if (eject_flag) {
1443         if (ioctl(s->fd, CDIOCEJECT) < 0)
1444             perror("CDIOCEJECT");
1445     } else {
1446         if (ioctl(s->fd, CDIOCCLOSE) < 0)
1447             perror("CDIOCCLOSE");
1448     }
1449
1450     if (cdrom_reopen(bs) < 0)
1451         return -ENOTSUP;
1452     return 0;
1453 }
1454
1455 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1456 {
1457     BDRVRawState *s = bs->opaque;
1458
1459     if (s->fd < 0)
1460         return -ENOTSUP;
1461     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
1462         /*
1463          * Note: an error can happen if the distribution automatically
1464          * mounts the CD-ROM
1465          */
1466         /* perror("CDROM_LOCKDOOR"); */
1467     }
1468
1469     return 0;
1470 }
1471
1472 static BlockDriver bdrv_host_cdrom = {
1473     .format_name        = "host_cdrom",
1474     .instance_size      = sizeof(BDRVRawState),
1475     .bdrv_probe_device  = cdrom_probe_device,
1476     .bdrv_open          = cdrom_open,
1477     .bdrv_close         = raw_close,
1478     .bdrv_create        = hdev_create,
1479     .bdrv_flush         = raw_flush,
1480
1481 #ifdef CONFIG_AIO
1482     .bdrv_aio_readv     = raw_aio_readv,
1483     .bdrv_aio_writev    = raw_aio_writev,
1484 #endif
1485
1486     .bdrv_read          = raw_read,
1487     .bdrv_write         = raw_write,
1488     .bdrv_getlength     = raw_getlength,
1489
1490     /* removable device support */
1491     .bdrv_is_inserted   = cdrom_is_inserted,
1492     .bdrv_eject         = cdrom_eject,
1493     .bdrv_set_locked    = cdrom_set_locked,
1494 };
1495 #endif /* __FreeBSD__ */
1496
1497 static void bdrv_raw_init(void)
1498 {
1499     /*
1500      * Register all the drivers.  Note that order is important, the driver
1501      * registered last will get probed first.
1502      */
1503     bdrv_register(&bdrv_raw);
1504     bdrv_register(&bdrv_host_device);
1505 #ifdef __linux__
1506     bdrv_register(&bdrv_host_floppy);
1507     bdrv_register(&bdrv_host_cdrom);
1508 #endif
1509 #ifdef __FreeBSD__
1510     bdrv_register(&bdrv_host_cdrom);
1511 #endif
1512 }
1513
1514 block_init(bdrv_raw_init);