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 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
  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 #if defined(__linux__)
 111     /* linux floppy specific */
 112     int64_t fd_open_time;
 113     int64_t fd_error_time;
 114     int fd_got_error;
 115     int fd_media_changed;
 116 #endif
 117 #ifdef CONFIG_LINUX_AIO
 118     int use_aio;
 119     void *aio_ctx;
 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__) || defined(__FreeBSD_kernel__)
 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 = qemu_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
 177         /* We're falling back to POSIX AIO in some cases */
 178         paio_init();
 179
 180         s->aio_ctx = laio_init();
 181         if (!s->aio_ctx) {
 182             goto out_free_buf;
 183         }
 184         s->use_aio = 1;
 185     } else
 186 #endif
 187     {
 188         if (paio_init() < 0) {
 189             goto out_free_buf;
 190         }
 191 #ifdef CONFIG_LINUX_AIO
 192         s->use_aio = 0;
 193 #endif
 194     }
 195
 196     return 0;
 197
 198 out_free_buf:
 199     qemu_vfree(s->aligned_buf);
 200 out_close:
 201     close(fd);
 202     return -errno;
 203 }
 204
 205 static int raw_open(BlockDriverState *bs, const char *filename, int flags)
 206 {
 207     BDRVRawState *s = bs->opaque;
 208     int open_flags = 0;
 209
 210     s->type = FTYPE_FILE;
 211     if (flags & BDRV_O_CREAT)
 212         open_flags = O_CREAT | O_TRUNC;
 213
 214     return raw_open_common(bs, filename, flags, open_flags);
 215 }
 216
 217 /* XXX: use host sector size if necessary with:
 218 #ifdef DIOCGSECTORSIZE
 219         {
 220             unsigned int sectorsize = 512;
 221             if (!ioctl(fd, DIOCGSECTORSIZE, &sectorsize) &&
 222                 sectorsize > bufsize)
 223                 bufsize = sectorsize;
 224         }
 225 #endif
 226 #ifdef CONFIG_COCOA
 227         u_int32_t   blockSize = 512;
 228         if ( !ioctl( fd, DKIOCGETBLOCKSIZE, &blockSize ) && blockSize > bufsize) {
 229             bufsize = blockSize;
 230         }
 231 #endif
 232 */
 233
 234 /*
 235  * offset and count are in bytes, but must be multiples of 512 for files
 236  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 237  *
 238  * This function may be called without alignment if the caller ensures
 239  * that O_DIRECT is not in effect.
 240  */
 241 static int raw_pread_aligned(BlockDriverState *bs, int64_t offset,
 242                      uint8_t *buf, int count)
 243 {
 244     BDRVRawState *s = bs->opaque;
 245     int ret;
 246
 247     ret = fd_open(bs);
 248     if (ret < 0)
 249         return ret;
 250
 251     if (offset >= 0 && lseek(s->fd, offset, SEEK_SET) == (off_t)-1) {
 252         ++(s->lseek_err_cnt);
 253         if(s->lseek_err_cnt <= 10) {
 254             DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 255                               "] lseek failed : %d = %s\n",
 256                               s->fd, bs->filename, offset, buf, count,
 257                               bs->total_sectors, errno, strerror(errno));
 258         }
 259         return -1;
 260     }
 261     s->lseek_err_cnt=0;
 262
 263     ret = read(s->fd, buf, count);
 264     if (ret == count)
 265         goto label__raw_read__success;
 266
 267     /* Allow reads beyond the end (needed for pwrite) */
 268     if ((ret == 0) && bs->growable) {
 269         int64_t size = raw_getlength(bs);
 270         if (offset >= size) {
 271             memset(buf, 0, count);
 272             ret = count;
 273             goto label__raw_read__success;
 274         }
 275     }
 276
 277     DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 278                       "] read failed %d : %d = %s\n",
 279                       s->fd, bs->filename, offset, buf, count,
 280                       bs->total_sectors, ret, errno, strerror(errno));
 281
 282     /* Try harder for CDrom. */
 283     if (bs->type == BDRV_TYPE_CDROM) {
 284         lseek(s->fd, offset, SEEK_SET);
 285         ret = read(s->fd, buf, count);
 286         if (ret == count)
 287             goto label__raw_read__success;
 288         lseek(s->fd, offset, SEEK_SET);
 289         ret = read(s->fd, buf, count);
 290         if (ret == count)
 291             goto label__raw_read__success;
 292
 293         DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 294                           "] retry read failed %d : %d = %s\n",
 295                           s->fd, bs->filename, offset, buf, count,
 296                           bs->total_sectors, ret, errno, strerror(errno));
 297     }
 298
 299 label__raw_read__success:
 300
 301     return  (ret < 0) ? -errno : ret;
 302 }
 303
 304 /*
 305  * offset and count are in bytes, but must be multiples of 512 for files
 306  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 307  *
 308  * This function may be called without alignment if the caller ensures
 309  * that O_DIRECT is not in effect.
 310  */
 311 static int raw_pwrite_aligned(BlockDriverState *bs, int64_t offset,
 312                       const uint8_t *buf, int count)
 313 {
 314     BDRVRawState *s = bs->opaque;
 315     int ret;
 316
 317     ret = fd_open(bs);
 318     if (ret < 0)
 319         return -errno;
 320
 321     if (offset >= 0 && lseek(s->fd, offset, SEEK_SET) == (off_t)-1) {
 322         ++(s->lseek_err_cnt);
 323         if(s->lseek_err_cnt) {
 324             DEBUG_BLOCK_PRINT("raw_pwrite(%d:%s, %" PRId64 ", %p, %d) [%"
 325                               PRId64 "] lseek failed : %d = %s\n",
 326                               s->fd, bs->filename, offset, buf, count,
 327                               bs->total_sectors, errno, strerror(errno));
 328         }
 329         return -EIO;
 330     }
 331     s->lseek_err_cnt = 0;
 332
 333     ret = write(s->fd, buf, count);
 334     if (ret == count)
 335         goto label__raw_write__success;
 336
 337     DEBUG_BLOCK_PRINT("raw_pwrite(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 338                       "] write failed %d : %d = %s\n",
 339                       s->fd, bs->filename, offset, buf, count,
 340                       bs->total_sectors, ret, errno, strerror(errno));
 341
 342 label__raw_write__success:
 343
 344     return  (ret < 0) ? -errno : ret;
 345 }
 346
 347
 348 /*
 349  * offset and count are in bytes and possibly not aligned. For files opened
 350  * with O_DIRECT, necessary alignments are ensured before calling
 351  * raw_pread_aligned to do the actual read.
 352  */
 353 static int raw_pread(BlockDriverState *bs, int64_t offset,
 354                      uint8_t *buf, int count)
 355 {
 356     BDRVRawState *s = bs->opaque;
 357     int size, ret, shift, sum;
 358
 359     sum = 0;
 360
 361     if (s->aligned_buf != NULL)  {
 362
 363         if (offset & 0x1ff) {
 364             /* align offset on a 512 bytes boundary */
 365
 366             shift = offset & 0x1ff;
 367             size = (shift + count + 0x1ff) & ~0x1ff;
 368             if (size > ALIGNED_BUFFER_SIZE)
 369                 size = ALIGNED_BUFFER_SIZE;
 370             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, size);
 371             if (ret < 0)
 372                 return ret;
 373
 374             size = 512 - shift;
 375             if (size > count)
 376                 size = count;
 377             memcpy(buf, s->aligned_buf + shift, size);
 378
 379             buf += size;
 380             offset += size;
 381             count -= size;
 382             sum += size;
 383
 384             if (count == 0)
 385                 return sum;
 386         }
 387         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 388
 389             /* read on aligned buffer */
 390
 391             while (count) {
 392
 393                 size = (count + 0x1ff) & ~0x1ff;
 394                 if (size > ALIGNED_BUFFER_SIZE)
 395                     size = ALIGNED_BUFFER_SIZE;
 396
 397                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, size);
 398                 if (ret < 0)
 399                     return ret;
 400
 401                 size = ret;
 402                 if (size > count)
 403                     size = count;
 404
 405                 memcpy(buf, s->aligned_buf, size);
 406
 407                 buf += size;
 408                 offset += size;
 409                 count -= size;
 410                 sum += size;
 411             }
 412
 413             return sum;
 414         }
 415     }
 416
 417     return raw_pread_aligned(bs, offset, buf, count) + sum;
 418 }
 419
 420 static int raw_read(BlockDriverState *bs, int64_t sector_num,
 421                     uint8_t *buf, int nb_sectors)
 422 {
 423     int ret;
 424
 425     ret = raw_pread(bs, sector_num * 512, buf, nb_sectors * 512);
 426     if (ret == (nb_sectors * 512))
 427         ret = 0;
 428     return ret;
 429 }
 430
 431 /*
 432  * offset and count are in bytes and possibly not aligned. For files opened
 433  * with O_DIRECT, necessary alignments are ensured before calling
 434  * raw_pwrite_aligned to do the actual write.
 435  */
 436 static int raw_pwrite(BlockDriverState *bs, int64_t offset,
 437                       const uint8_t *buf, int count)
 438 {
 439     BDRVRawState *s = bs->opaque;
 440     int size, ret, shift, sum;
 441
 442     sum = 0;
 443
 444     if (s->aligned_buf != NULL) {
 445
 446         if (offset & 0x1ff) {
 447             /* align offset on a 512 bytes boundary */
 448             shift = offset & 0x1ff;
 449             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, 512);
 450             if (ret < 0)
 451                 return ret;
 452
 453             size = 512 - shift;
 454             if (size > count)
 455                 size = count;
 456             memcpy(s->aligned_buf + shift, buf, size);
 457
 458             ret = raw_pwrite_aligned(bs, offset - shift, s->aligned_buf, 512);
 459             if (ret < 0)
 460                 return ret;
 461
 462             buf += size;
 463             offset += size;
 464             count -= size;
 465             sum += size;
 466
 467             if (count == 0)
 468                 return sum;
 469         }
 470         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 471
 472             while ((size = (count & ~0x1ff)) != 0) {
 473
 474                 if (size > ALIGNED_BUFFER_SIZE)
 475                     size = ALIGNED_BUFFER_SIZE;
 476
 477                 memcpy(s->aligned_buf, buf, size);
 478
 479                 ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, size);
 480                 if (ret < 0)
 481                     return ret;
 482
 483                 buf += ret;
 484                 offset += ret;
 485                 count -= ret;
 486                 sum += ret;
 487             }
 488             /* here, count < 512 because (count & ~0x1ff) == 0 */
 489             if (count) {
 490                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, 512);
 491                 if (ret < 0)
 492                     return ret;
 493                  memcpy(s->aligned_buf, buf, count);
 494
 495                  ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, 512);
 496                  if (ret < 0)
 497                      return ret;
 498                  if (count < ret)
 499                      ret = count;
 500
 501                  sum += ret;
 502             }
 503             return sum;
 504         }
 505     }
 506     return raw_pwrite_aligned(bs, offset, buf, count) + sum;
 507 }
 508
 509 static int raw_write(BlockDriverState *bs, int64_t sector_num,
 510                      const uint8_t *buf, int nb_sectors)
 511 {
 512     int ret;
 513     ret = raw_pwrite(bs, sector_num * 512, buf, nb_sectors * 512);
 514     if (ret == (nb_sectors * 512))
 515         ret = 0;
 516     return ret;
 517 }
 518
 519 /*
 520  * Check if all memory in this vector is sector aligned.
 521  */
 522 static int qiov_is_aligned(QEMUIOVector *qiov)
 523 {
 524     int i;
 525
 526     for (i = 0; i < qiov->niov; i++) {
 527         if ((uintptr_t) qiov->iov[i].iov_base % 512) {
 528             return 0;
 529         }
 530     }
 531
 532     return 1;
 533 }
 534
 535 static BlockDriverAIOCB *raw_aio_submit(BlockDriverState *bs,
 536         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 537         BlockDriverCompletionFunc *cb, void *opaque, int type)
 538 {
 539     BDRVRawState *s = bs->opaque;
 540
 541     if (fd_open(bs) < 0)
 542         return NULL;
 543
 544     /*
 545      * If O_DIRECT is used the buffer needs to be aligned on a sector
 546      * boundary.  Check if this is the case or telll the low-level
 547      * driver that it needs to copy the buffer.
 548      */
 549     if (s->aligned_buf) {
 550         if (!qiov_is_aligned(qiov)) {
 551             type |= QEMU_AIO_MISALIGNED;
 552 #ifdef CONFIG_LINUX_AIO
 553         } else if (s->use_aio) {
 554             return laio_submit(bs, s->aio_ctx, s->fd, sector_num, qiov,
 555                                nb_sectors, cb, opaque, type);
 556 #endif
 557         }
 558     }
 559
 560     return paio_submit(bs, s->fd, sector_num, qiov, nb_sectors,
 561                        cb, opaque, type);
 562 }
 563
 564 static BlockDriverAIOCB *raw_aio_readv(BlockDriverState *bs,
 565         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 566         BlockDriverCompletionFunc *cb, void *opaque)
 567 {
 568     return raw_aio_submit(bs, sector_num, qiov, nb_sectors,
 569                           cb, opaque, QEMU_AIO_READ);
 570 }
 571
 572 static BlockDriverAIOCB *raw_aio_writev(BlockDriverState *bs,
 573         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 574         BlockDriverCompletionFunc *cb, void *opaque)
 575 {
 576     return raw_aio_submit(bs, sector_num, qiov, nb_sectors,
 577                           cb, opaque, QEMU_AIO_WRITE);
 578 }
 579
 580 static BlockDriverAIOCB *raw_aio_flush(BlockDriverState *bs,
 581         BlockDriverCompletionFunc *cb, void *opaque)
 582 {
 583     BDRVRawState *s = bs->opaque;
 584
 585     if (fd_open(bs) < 0)
 586         return NULL;
 587
 588     return paio_submit(bs, s->fd, 0, NULL, 0, cb, opaque, QEMU_AIO_FLUSH);
 589 }
 590
 591 static void raw_close(BlockDriverState *bs)
 592 {
 593     BDRVRawState *s = bs->opaque;
 594     if (s->fd >= 0) {
 595         close(s->fd);
 596         s->fd = -1;
 597         if (s->aligned_buf != NULL)
 598             qemu_free(s->aligned_buf);
 599     }
 600 }
 601
 602 static int raw_truncate(BlockDriverState *bs, int64_t offset)
 603 {
 604     BDRVRawState *s = bs->opaque;
 605     if (s->type != FTYPE_FILE)
 606         return -ENOTSUP;
 607     if (ftruncate(s->fd, offset) < 0)
 608         return -errno;
 609     return 0;
 610 }
 611
 612 #ifdef __OpenBSD__
 613 static int64_t raw_getlength(BlockDriverState *bs)
 614 {
 615     BDRVRawState *s = bs->opaque;
 616     int fd = s->fd;
 617     struct stat st;
 618
 619     if (fstat(fd, &st))
 620         return -1;
 621     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
 622         struct disklabel dl;
 623
 624         if (ioctl(fd, DIOCGDINFO, &dl))
 625             return -1;
 626         return (uint64_t)dl.d_secsize *
 627             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
 628     } else
 629         return st.st_size;
 630 }
 631 #else /* !__OpenBSD__ */
 632 static int64_t  raw_getlength(BlockDriverState *bs)
 633 {
 634     BDRVRawState *s = bs->opaque;
 635     int fd = s->fd;
 636     int64_t size;
 637 #ifdef CONFIG_BSD
 638     struct stat sb;
 639 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
 640     int reopened = 0;
 641 #endif
 642 #endif
 643 #ifdef __sun__
 644     struct dk_minfo minfo;
 645     int rv;
 646 #endif
 647     int ret;
 648
 649     ret = fd_open(bs);
 650     if (ret < 0)
 651         return ret;
 652
 653 #ifdef CONFIG_BSD
 654 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
 655 again:
 656 #endif
 657     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
 658 #ifdef DIOCGMEDIASIZE
 659         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size))
 660 #elif defined(DIOCGPART)
 661         {
 662                 struct partinfo pi;
 663                 if (ioctl(fd, DIOCGPART, &pi) == 0)
 664                         size = pi.media_size;
 665                 else
 666                         size = 0;
 667         }
 668         if (size == 0)
 669 #endif
 670 #ifdef CONFIG_COCOA
 671         size = LONG_LONG_MAX;
 672 #else
 673         size = lseek(fd, 0LL, SEEK_END);
 674 #endif
 675 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 676         switch(s->type) {
 677         case FTYPE_CD:
 678             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
 679             if (size == 2048LL * (unsigned)-1)
 680                 size = 0;
 681             /* XXX no disc?  maybe we need to reopen... */
 682             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
 683                 reopened = 1;
 684                 goto again;
 685             }
 686         }
 687 #endif
 688     } else
 689 #endif
 690 #ifdef __sun__
 691     /*
 692      * use the DKIOCGMEDIAINFO ioctl to read the size.
 693      */
 694     rv = ioctl ( fd, DKIOCGMEDIAINFO, &minfo );
 695     if ( rv != -1 ) {
 696         size = minfo.dki_lbsize * minfo.dki_capacity;
 697     } else /* there are reports that lseek on some devices
 698               fails, but irc discussion said that contingency
 699               on contingency was overkill */
 700 #endif
 701     {
 702         size = lseek(fd, 0, SEEK_END);
 703     }
 704     return size;
 705 }
 706 #endif
 707
 708 static int raw_create(const char *filename, QEMUOptionParameter *options)
 709 {
 710     int fd;
 711     int result = 0;
 712     int64_t total_size = 0;
 713
 714     /* Read out options */
 715     while (options && options->name) {
 716         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
 717             total_size = options->value.n / 512;
 718         }
 719         options++;
 720     }
 721
 722     fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC | O_BINARY,
 723               0644);
 724     if (fd < 0) {
 725         result = -errno;
 726     } else {
 727         if (ftruncate(fd, total_size * 512) != 0) {
 728             result = -errno;
 729         }
 730         if (close(fd) != 0) {
 731             result = -errno;
 732         }
 733     }
 734     return result;
 735 }
 736
 737 static void raw_flush(BlockDriverState *bs)
 738 {
 739     BDRVRawState *s = bs->opaque;
 740     qemu_fdatasync(s->fd);
 741 }
 742
 743
 744 static QEMUOptionParameter raw_create_options[] = {
 745     {
 746         .name = BLOCK_OPT_SIZE,
 747         .type = OPT_SIZE,
 748         .help = "Virtual disk size"
 749     },
 750     { NULL }
 751 };
 752
 753 static BlockDriver bdrv_raw = {
 754     .format_name = "raw",
 755     .instance_size = sizeof(BDRVRawState),
 756     .bdrv_probe = NULL, /* no probe for protocols */
 757     .bdrv_open = raw_open,
 758     .bdrv_read = raw_read,
 759     .bdrv_write = raw_write,
 760     .bdrv_close = raw_close,
 761     .bdrv_create = raw_create,
 762     .bdrv_flush = raw_flush,
 763
 764     .bdrv_aio_readv = raw_aio_readv,
 765     .bdrv_aio_writev = raw_aio_writev,
 766     .bdrv_aio_flush = raw_aio_flush,
 767
 768     .bdrv_truncate = raw_truncate,
 769     .bdrv_getlength = raw_getlength,
 770
 771     .create_options = raw_create_options,
 772 };
 773
 774 /***********************************************/
 775 /* host device */
 776
 777 #ifdef CONFIG_COCOA
 778 static kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator );
 779 static kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize );
 780
 781 kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator )
 782 {
 783     kern_return_t       kernResult;
 784     mach_port_t     masterPort;
 785     CFMutableDictionaryRef  classesToMatch;
 786
 787     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
 788     if ( KERN_SUCCESS != kernResult ) {
 789         printf( "IOMasterPort returned %d\n", kernResult );
 790     }
 791
 792     classesToMatch = IOServiceMatching( kIOCDMediaClass );
 793     if ( classesToMatch == NULL ) {
 794         printf( "IOServiceMatching returned a NULL dictionary.\n" );
 795     } else {
 796     CFDictionarySetValue( classesToMatch, CFSTR( kIOMediaEjectableKey ), kCFBooleanTrue );
 797     }
 798     kernResult = IOServiceGetMatchingServices( masterPort, classesToMatch, mediaIterator );
 799     if ( KERN_SUCCESS != kernResult )
 800     {
 801         printf( "IOServiceGetMatchingServices returned %d\n", kernResult );
 802     }
 803
 804     return kernResult;
 805 }
 806
 807 kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize )
 808 {
 809     io_object_t     nextMedia;
 810     kern_return_t   kernResult = KERN_FAILURE;
 811     *bsdPath = '\0';
 812     nextMedia = IOIteratorNext( mediaIterator );
 813     if ( nextMedia )
 814     {
 815         CFTypeRef   bsdPathAsCFString;
 816     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
 817         if ( bsdPathAsCFString ) {
 818             size_t devPathLength;
 819             strcpy( bsdPath, _PATH_DEV );
 820             strcat( bsdPath, "r" );
 821             devPathLength = strlen( bsdPath );
 822             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
 823                 kernResult = KERN_SUCCESS;
 824             }
 825             CFRelease( bsdPathAsCFString );
 826         }
 827         IOObjectRelease( nextMedia );
 828     }
 829
 830     return kernResult;
 831 }
 832
 833 #endif
 834
 835 static int hdev_probe_device(const char *filename)
 836 {
 837     struct stat st;
 838
 839     /* allow a dedicated CD-ROM driver to match with a higher priority */
 840     if (strstart(filename, "/dev/cdrom", NULL))
 841         return 50;
 842
 843     if (stat(filename, &st) >= 0 &&
 844             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 845         return 100;
 846     }
 847
 848     return 0;
 849 }
 850
 851 static int hdev_open(BlockDriverState *bs, const char *filename, int flags)
 852 {
 853     BDRVRawState *s = bs->opaque;
 854
 855 #ifdef CONFIG_COCOA
 856     if (strstart(filename, "/dev/cdrom", NULL)) {
 857         kern_return_t kernResult;
 858         io_iterator_t mediaIterator;
 859         char bsdPath[ MAXPATHLEN ];
 860         int fd;
 861
 862         kernResult = FindEjectableCDMedia( &mediaIterator );
 863         kernResult = GetBSDPath( mediaIterator, bsdPath, sizeof( bsdPath ) );
 864
 865         if ( bsdPath[ 0 ] != '\0' ) {
 866             strcat(bsdPath,"s0");
 867             /* some CDs don't have a partition 0 */
 868             fd = open(bsdPath, O_RDONLY | O_BINARY | O_LARGEFILE);
 869             if (fd < 0) {
 870                 bsdPath[strlen(bsdPath)-1] = '1';
 871             } else {
 872                 close(fd);
 873             }
 874             filename = bsdPath;
 875         }
 876
 877         if ( mediaIterator )
 878             IOObjectRelease( mediaIterator );
 879     }
 880 #endif
 881
 882     s->type = FTYPE_FILE;
 883 #if defined(__linux__)
 884     if (strstart(filename, "/dev/sg", NULL)) {
 885         bs->sg = 1;
 886     }
 887 #endif
 888
 889     return raw_open_common(bs, filename, flags, 0);
 890 }
 891
 892 #if defined(__linux__)
 893 /* Note: we do not have a reliable method to detect if the floppy is
 894    present. The current method is to try to open the floppy at every
 895    I/O and to keep it opened during a few hundreds of ms. */
 896 static int fd_open(BlockDriverState *bs)
 897 {
 898     BDRVRawState *s = bs->opaque;
 899     int last_media_present;
 900
 901     if (s->type != FTYPE_FD)
 902         return 0;
 903     last_media_present = (s->fd >= 0);
 904     if (s->fd >= 0 &&
 905         (qemu_get_clock(rt_clock) - s->fd_open_time) >= FD_OPEN_TIMEOUT) {
 906         close(s->fd);
 907         s->fd = -1;
 908 #ifdef DEBUG_FLOPPY
 909         printf("Floppy closed\n");
 910 #endif
 911     }
 912     if (s->fd < 0) {
 913         if (s->fd_got_error &&
 914             (qemu_get_clock(rt_clock) - s->fd_error_time) < FD_OPEN_TIMEOUT) {
 915 #ifdef DEBUG_FLOPPY
 916             printf("No floppy (open delayed)\n");
 917 #endif
 918             return -EIO;
 919         }
 920         s->fd = open(bs->filename, s->open_flags & ~O_NONBLOCK);
 921         if (s->fd < 0) {
 922             s->fd_error_time = qemu_get_clock(rt_clock);
 923             s->fd_got_error = 1;
 924             if (last_media_present)
 925                 s->fd_media_changed = 1;
 926 #ifdef DEBUG_FLOPPY
 927             printf("No floppy\n");
 928 #endif
 929             return -EIO;
 930         }
 931 #ifdef DEBUG_FLOPPY
 932         printf("Floppy opened\n");
 933 #endif
 934     }
 935     if (!last_media_present)
 936         s->fd_media_changed = 1;
 937     s->fd_open_time = qemu_get_clock(rt_clock);
 938     s->fd_got_error = 0;
 939     return 0;
 940 }
 941
 942 static int hdev_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 943 {
 944     BDRVRawState *s = bs->opaque;
 945
 946     return ioctl(s->fd, req, buf);
 947 }
 948
 949 static BlockDriverAIOCB *hdev_aio_ioctl(BlockDriverState *bs,
 950         unsigned long int req, void *buf,
 951         BlockDriverCompletionFunc *cb, void *opaque)
 952 {
 953     BDRVRawState *s = bs->opaque;
 954
 955     if (fd_open(bs) < 0)
 956         return NULL;
 957     return paio_ioctl(bs, s->fd, req, buf, cb, opaque);
 958 }
 959
 960 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 961 static int fd_open(BlockDriverState *bs)
 962 {
 963     BDRVRawState *s = bs->opaque;
 964
 965     /* this is just to ensure s->fd is sane (its called by io ops) */
 966     if (s->fd >= 0)
 967         return 0;
 968     return -EIO;
 969 }
 970 #else /* !linux && !FreeBSD */
 971
 972 static int fd_open(BlockDriverState *bs)
 973 {
 974     return 0;
 975 }
 976
 977 #endif /* !linux && !FreeBSD */
 978
 979 static int hdev_create(const char *filename, QEMUOptionParameter *options)
 980 {
 981     int fd;
 982     int ret = 0;
 983     struct stat stat_buf;
 984     int64_t total_size = 0;
 985
 986     /* Read out options */
 987     while (options && options->name) {
 988         if (!strcmp(options->name, "size")) {
 989             total_size = options->value.n / 512;
 990         }
 991         options++;
 992     }
 993
 994     fd = open(filename, O_WRONLY | O_BINARY);
 995     if (fd < 0)
 996         return -EIO;
 997
 998     if (fstat(fd, &stat_buf) < 0)
 999         ret = -EIO;
1000     else if (!S_ISBLK(stat_buf.st_mode) && !S_ISCHR(stat_buf.st_mode))
1001         ret = -EIO;
1002     else if (lseek(fd, 0, SEEK_END) < total_size * 512)
1003         ret = -ENOSPC;
1004
1005     close(fd);
1006     return ret;
1007 }
1008
1009 static BlockDriver bdrv_host_device = {
1010     .format_name        = "host_device",
1011     .instance_size      = sizeof(BDRVRawState),
1012     .bdrv_probe_device  = hdev_probe_device,
1013     .bdrv_open          = hdev_open,
1014     .bdrv_close         = raw_close,
1015     .bdrv_create        = hdev_create,
1016     .create_options     = raw_create_options,
1017     .no_zero_init       = 1,
1018     .bdrv_flush         = raw_flush,
1019
1020     .bdrv_aio_readv     = raw_aio_readv,
1021     .bdrv_aio_writev    = raw_aio_writev,
1022     .bdrv_aio_flush     = raw_aio_flush,
1023
1024     .bdrv_read          = raw_read,
1025     .bdrv_write         = raw_write,
1026     .bdrv_getlength     = raw_getlength,
1027
1028     /* generic scsi device */
1029 #ifdef __linux__
1030     .bdrv_ioctl         = hdev_ioctl,
1031     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1032 #endif
1033 };
1034
1035 #ifdef __linux__
1036 static int floppy_open(BlockDriverState *bs, const char *filename, int flags)
1037 {
1038     BDRVRawState *s = bs->opaque;
1039     int ret;
1040
1041     s->type = FTYPE_FD;
1042
1043     /* open will not fail even if no floppy is inserted, so add O_NONBLOCK */
1044     ret = raw_open_common(bs, filename, flags, O_NONBLOCK);
1045     if (ret)
1046         return ret;
1047
1048     /* close fd so that we can reopen it as needed */
1049     close(s->fd);
1050     s->fd = -1;
1051     s->fd_media_changed = 1;
1052
1053     return 0;
1054 }
1055
1056 static int floppy_probe_device(const char *filename)
1057 {
1058     if (strstart(filename, "/dev/fd", NULL))
1059         return 100;
1060     return 0;
1061 }
1062
1063
1064 static int floppy_is_inserted(BlockDriverState *bs)
1065 {
1066     return fd_open(bs) >= 0;
1067 }
1068
1069 static int floppy_media_changed(BlockDriverState *bs)
1070 {
1071     BDRVRawState *s = bs->opaque;
1072     int ret;
1073
1074     /*
1075      * XXX: we do not have a true media changed indication.
1076      * It does not work if the floppy is changed without trying to read it.
1077      */
1078     fd_open(bs);
1079     ret = s->fd_media_changed;
1080     s->fd_media_changed = 0;
1081 #ifdef DEBUG_FLOPPY
1082     printf("Floppy changed=%d\n", ret);
1083 #endif
1084     return ret;
1085 }
1086
1087 static int floppy_eject(BlockDriverState *bs, int eject_flag)
1088 {
1089     BDRVRawState *s = bs->opaque;
1090     int fd;
1091
1092     if (s->fd >= 0) {
1093         close(s->fd);
1094         s->fd = -1;
1095     }
1096     fd = open(bs->filename, s->open_flags | O_NONBLOCK);
1097     if (fd >= 0) {
1098         if (ioctl(fd, FDEJECT, 0) < 0)
1099             perror("FDEJECT");
1100         close(fd);
1101     }
1102
1103     return 0;
1104 }
1105
1106 static BlockDriver bdrv_host_floppy = {
1107     .format_name        = "host_floppy",
1108     .instance_size      = sizeof(BDRVRawState),
1109     .bdrv_probe_device  = floppy_probe_device,
1110     .bdrv_open          = floppy_open,
1111     .bdrv_close         = raw_close,
1112     .bdrv_create        = hdev_create,
1113     .create_options     = raw_create_options,
1114     .no_zero_init       = 1,
1115     .bdrv_flush         = raw_flush,
1116
1117     .bdrv_aio_readv     = raw_aio_readv,
1118     .bdrv_aio_writev    = raw_aio_writev,
1119     .bdrv_aio_flush     = raw_aio_flush,
1120
1121     .bdrv_read          = raw_read,
1122     .bdrv_write         = raw_write,
1123     .bdrv_getlength     = raw_getlength,
1124
1125     /* removable device support */
1126     .bdrv_is_inserted   = floppy_is_inserted,
1127     .bdrv_media_changed = floppy_media_changed,
1128     .bdrv_eject         = floppy_eject,
1129 };
1130
1131 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1132 {
1133     BDRVRawState *s = bs->opaque;
1134
1135     s->type = FTYPE_CD;
1136
1137     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
1138     return raw_open_common(bs, filename, flags, O_NONBLOCK);
1139 }
1140
1141 static int cdrom_probe_device(const char *filename)
1142 {
1143     if (strstart(filename, "/dev/cd", NULL))
1144         return 100;
1145     return 0;
1146 }
1147
1148 static int cdrom_is_inserted(BlockDriverState *bs)
1149 {
1150     BDRVRawState *s = bs->opaque;
1151     int ret;
1152
1153     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
1154     if (ret == CDS_DISC_OK)
1155         return 1;
1156     return 0;
1157 }
1158
1159 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1160 {
1161     BDRVRawState *s = bs->opaque;
1162
1163     if (eject_flag) {
1164         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
1165             perror("CDROMEJECT");
1166     } else {
1167         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
1168             perror("CDROMEJECT");
1169     }
1170
1171     return 0;
1172 }
1173
1174 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1175 {
1176     BDRVRawState *s = bs->opaque;
1177
1178     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
1179         /*
1180          * Note: an error can happen if the distribution automatically
1181          * mounts the CD-ROM
1182          */
1183         /* perror("CDROM_LOCKDOOR"); */
1184     }
1185
1186     return 0;
1187 }
1188
1189 static BlockDriver bdrv_host_cdrom = {
1190     .format_name        = "host_cdrom",
1191     .instance_size      = sizeof(BDRVRawState),
1192     .bdrv_probe_device  = cdrom_probe_device,
1193     .bdrv_open          = cdrom_open,
1194     .bdrv_close         = raw_close,
1195     .bdrv_create        = hdev_create,
1196     .create_options     = raw_create_options,
1197     .no_zero_init       = 1,
1198     .bdrv_flush         = raw_flush,
1199
1200     .bdrv_aio_readv     = raw_aio_readv,
1201     .bdrv_aio_writev    = raw_aio_writev,
1202     .bdrv_aio_flush     = raw_aio_flush,
1203
1204     .bdrv_read          = raw_read,
1205     .bdrv_write         = raw_write,
1206     .bdrv_getlength     = raw_getlength,
1207
1208     /* removable device support */
1209     .bdrv_is_inserted   = cdrom_is_inserted,
1210     .bdrv_eject         = cdrom_eject,
1211     .bdrv_set_locked    = cdrom_set_locked,
1212
1213     /* generic scsi device */
1214     .bdrv_ioctl         = hdev_ioctl,
1215     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1216 };
1217 #endif /* __linux__ */
1218
1219 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
1220 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1221 {
1222     BDRVRawState *s = bs->opaque;
1223     int ret;
1224
1225     s->type = FTYPE_CD;
1226
1227     ret = raw_open_common(bs, filename, flags, 0);
1228     if (ret)
1229         return ret;
1230
1231     /* make sure the door isnt locked at this time */
1232     ioctl(s->fd, CDIOCALLOW);
1233     return 0;
1234 }
1235
1236 static int cdrom_probe_device(const char *filename)
1237 {
1238     if (strstart(filename, "/dev/cd", NULL) ||
1239             strstart(filename, "/dev/acd", NULL))
1240         return 100;
1241     return 0;
1242 }
1243
1244 static int cdrom_reopen(BlockDriverState *bs)
1245 {
1246     BDRVRawState *s = bs->opaque;
1247     int fd;
1248
1249     /*
1250      * Force reread of possibly changed/newly loaded disc,
1251      * FreeBSD seems to not notice sometimes...
1252      */
1253     if (s->fd >= 0)
1254         close(s->fd);
1255     fd = open(bs->filename, s->open_flags, 0644);
1256     if (fd < 0) {
1257         s->fd = -1;
1258         return -EIO;
1259     }
1260     s->fd = fd;
1261
1262     /* make sure the door isnt locked at this time */
1263     ioctl(s->fd, CDIOCALLOW);
1264     return 0;
1265 }
1266
1267 static int cdrom_is_inserted(BlockDriverState *bs)
1268 {
1269     return raw_getlength(bs) > 0;
1270 }
1271
1272 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1273 {
1274     BDRVRawState *s = bs->opaque;
1275
1276     if (s->fd < 0)
1277         return -ENOTSUP;
1278
1279     (void) ioctl(s->fd, CDIOCALLOW);
1280
1281     if (eject_flag) {
1282         if (ioctl(s->fd, CDIOCEJECT) < 0)
1283             perror("CDIOCEJECT");
1284     } else {
1285         if (ioctl(s->fd, CDIOCCLOSE) < 0)
1286             perror("CDIOCCLOSE");
1287     }
1288
1289     if (cdrom_reopen(bs) < 0)
1290         return -ENOTSUP;
1291     return 0;
1292 }
1293
1294 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1295 {
1296     BDRVRawState *s = bs->opaque;
1297
1298     if (s->fd < 0)
1299         return -ENOTSUP;
1300     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
1301         /*
1302          * Note: an error can happen if the distribution automatically
1303          * mounts the CD-ROM
1304          */
1305         /* perror("CDROM_LOCKDOOR"); */
1306     }
1307
1308     return 0;
1309 }
1310
1311 static BlockDriver bdrv_host_cdrom = {
1312     .format_name        = "host_cdrom",
1313     .instance_size      = sizeof(BDRVRawState),
1314     .bdrv_probe_device  = cdrom_probe_device,
1315     .bdrv_open          = cdrom_open,
1316     .bdrv_close         = raw_close,
1317     .bdrv_create        = hdev_create,
1318     .create_options     = raw_create_options,
1319     .no_zero_init       = 1,
1320     .bdrv_flush         = raw_flush,
1321
1322     .bdrv_aio_readv     = raw_aio_readv,
1323     .bdrv_aio_writev    = raw_aio_writev,
1324     .bdrv_aio_flush     = raw_aio_flush,
1325
1326     .bdrv_read          = raw_read,
1327     .bdrv_write         = raw_write,
1328     .bdrv_getlength     = raw_getlength,
1329
1330     /* removable device support */
1331     .bdrv_is_inserted   = cdrom_is_inserted,
1332     .bdrv_eject         = cdrom_eject,
1333     .bdrv_set_locked    = cdrom_set_locked,
1334 };
1335 #endif /* __FreeBSD__ */
1336
1337 static void bdrv_raw_init(void)
1338 {
1339     /*
1340      * Register all the drivers.  Note that order is important, the driver
1341      * registered last will get probed first.
1342      */
1343     bdrv_register(&bdrv_raw);
1344     bdrv_register(&bdrv_host_device);
1345 #ifdef __linux__
1346     bdrv_register(&bdrv_host_floppy);
1347     bdrv_register(&bdrv_host_cdrom);
1348 #endif
1349 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
1350     bdrv_register(&bdrv_host_cdrom);
1351 #endif
1352 }
1353
1354 block_init(bdrv_raw_init);