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