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     int open_flags;
 109 #if defined(__linux__)
 110     /* linux floppy specific */
 111     int64_t fd_open_time;
 112     int64_t fd_error_time;
 113     int fd_got_error;
 114     int fd_media_changed;
 115 #endif
 116 #ifdef CONFIG_LINUX_AIO
 117     int use_aio;
 118     void *aio_ctx;
 119 #endif
 120     uint8_t* aligned_buf;
 121 } BDRVRawState;
 122
 123 static int fd_open(BlockDriverState *bs);
 124 static int64_t raw_getlength(BlockDriverState *bs);
 125
 126 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 127 static int cdrom_reopen(BlockDriverState *bs);
 128 #endif
 129
 130 static int raw_open_common(BlockDriverState *bs, const char *filename,
 131                            int bdrv_flags, int open_flags)
 132 {
 133     BDRVRawState *s = bs->opaque;
 134     int fd, ret;
 135
 136     s->open_flags = open_flags | O_BINARY;
 137     s->open_flags &= ~O_ACCMODE;
 138     if (bdrv_flags & BDRV_O_RDWR) {
 139         s->open_flags |= O_RDWR;
 140     } else {
 141         s->open_flags |= O_RDONLY;
 142     }
 143
 144     /* Use O_DSYNC for write-through caching, no flags for write-back caching,
 145      * and O_DIRECT for no caching. */
 146     if ((bdrv_flags & BDRV_O_NOCACHE))
 147         s->open_flags |= O_DIRECT;
 148     else if (!(bdrv_flags & BDRV_O_CACHE_WB))
 149         s->open_flags |= O_DSYNC;
 150
 151     s->fd = -1;
 152     fd = qemu_open(filename, s->open_flags, 0644);
 153     if (fd < 0) {
 154         ret = -errno;
 155         if (ret == -EROFS)
 156             ret = -EACCES;
 157         return ret;
 158     }
 159     s->fd = fd;
 160     s->aligned_buf = NULL;
 161
 162     if ((bdrv_flags & BDRV_O_NOCACHE)) {
 163         s->aligned_buf = qemu_blockalign(bs, ALIGNED_BUFFER_SIZE);
 164         if (s->aligned_buf == NULL) {
 165             goto out_close;
 166         }
 167     }
 168
 169 #ifdef CONFIG_LINUX_AIO
 170     if ((bdrv_flags & (BDRV_O_NOCACHE|BDRV_O_NATIVE_AIO)) ==
 171                       (BDRV_O_NOCACHE|BDRV_O_NATIVE_AIO)) {
 172
 173         /* We're falling back to POSIX AIO in some cases */
 174         paio_init();
 175
 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         if (paio_init() < 0) {
 185             goto out_free_buf;
 186         }
 187 #ifdef CONFIG_LINUX_AIO
 188         s->use_aio = 0;
 189 #endif
 190     }
 191
 192     return 0;
 193
 194 out_free_buf:
 195     qemu_vfree(s->aligned_buf);
 196 out_close:
 197     close(fd);
 198     return -errno;
 199 }
 200
 201 static int raw_open(BlockDriverState *bs, const char *filename, int flags)
 202 {
 203     BDRVRawState *s = bs->opaque;
 204
 205     s->type = FTYPE_FILE;
 206     return raw_open_common(bs, filename, flags, 0);
 207 }
 208
 209 /* XXX: use host sector size if necessary with:
 210 #ifdef DIOCGSECTORSIZE
 211         {
 212             unsigned int sectorsize = 512;
 213             if (!ioctl(fd, DIOCGSECTORSIZE, &sectorsize) &&
 214                 sectorsize > bufsize)
 215                 bufsize = sectorsize;
 216         }
 217 #endif
 218 #ifdef CONFIG_COCOA
 219         u_int32_t   blockSize = 512;
 220         if ( !ioctl( fd, DKIOCGETBLOCKSIZE, &blockSize ) && blockSize > bufsize) {
 221             bufsize = blockSize;
 222         }
 223 #endif
 224 */
 225
 226 /*
 227  * offset and count are in bytes, but must be multiples of 512 for files
 228  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 229  *
 230  * This function may be called without alignment if the caller ensures
 231  * that O_DIRECT is not in effect.
 232  */
 233 static int raw_pread_aligned(BlockDriverState *bs, int64_t offset,
 234                      uint8_t *buf, int count)
 235 {
 236     BDRVRawState *s = bs->opaque;
 237     int ret;
 238
 239     ret = fd_open(bs);
 240     if (ret < 0)
 241         return ret;
 242
 243     ret = pread(s->fd, buf, count, offset);
 244     if (ret == count)
 245         goto label__raw_read__success;
 246
 247     /* Allow reads beyond the end (needed for pwrite) */
 248     if ((ret == 0) && bs->growable) {
 249         int64_t size = raw_getlength(bs);
 250         if (offset >= size) {
 251             memset(buf, 0, count);
 252             ret = count;
 253             goto label__raw_read__success;
 254         }
 255     }
 256
 257     DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 258                       "] read failed %d : %d = %s\n",
 259                       s->fd, bs->filename, offset, buf, count,
 260                       bs->total_sectors, ret, errno, strerror(errno));
 261
 262     /* Try harder for CDrom. */
 263     if (bs->type == BDRV_TYPE_CDROM) {
 264         ret = pread(s->fd, buf, count, offset);
 265         if (ret == count)
 266             goto label__raw_read__success;
 267         ret = pread(s->fd, buf, count, offset);
 268         if (ret == count)
 269             goto label__raw_read__success;
 270
 271         DEBUG_BLOCK_PRINT("raw_pread(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 272                           "] retry read failed %d : %d = %s\n",
 273                           s->fd, bs->filename, offset, buf, count,
 274                           bs->total_sectors, ret, errno, strerror(errno));
 275     }
 276
 277 label__raw_read__success:
 278
 279     return  (ret < 0) ? -errno : ret;
 280 }
 281
 282 /*
 283  * offset and count are in bytes, but must be multiples of 512 for files
 284  * opened with O_DIRECT. buf must be aligned to 512 bytes then.
 285  *
 286  * This function may be called without alignment if the caller ensures
 287  * that O_DIRECT is not in effect.
 288  */
 289 static int raw_pwrite_aligned(BlockDriverState *bs, int64_t offset,
 290                       const uint8_t *buf, int count)
 291 {
 292     BDRVRawState *s = bs->opaque;
 293     int ret;
 294
 295     ret = fd_open(bs);
 296     if (ret < 0)
 297         return -errno;
 298
 299     ret = pwrite(s->fd, buf, count, offset);
 300     if (ret == count)
 301         goto label__raw_write__success;
 302
 303     DEBUG_BLOCK_PRINT("raw_pwrite(%d:%s, %" PRId64 ", %p, %d) [%" PRId64
 304                       "] write failed %d : %d = %s\n",
 305                       s->fd, bs->filename, offset, buf, count,
 306                       bs->total_sectors, ret, errno, strerror(errno));
 307
 308 label__raw_write__success:
 309
 310     return  (ret < 0) ? -errno : ret;
 311 }
 312
 313
 314 /*
 315  * offset and count are in bytes and possibly not aligned. For files opened
 316  * with O_DIRECT, necessary alignments are ensured before calling
 317  * raw_pread_aligned to do the actual read.
 318  */
 319 static int raw_pread(BlockDriverState *bs, int64_t offset,
 320                      uint8_t *buf, int count)
 321 {
 322     BDRVRawState *s = bs->opaque;
 323     int size, ret, shift, sum;
 324
 325     sum = 0;
 326
 327     if (s->aligned_buf != NULL)  {
 328
 329         if (offset & 0x1ff) {
 330             /* align offset on a 512 bytes boundary */
 331
 332             shift = offset & 0x1ff;
 333             size = (shift + count + 0x1ff) & ~0x1ff;
 334             if (size > ALIGNED_BUFFER_SIZE)
 335                 size = ALIGNED_BUFFER_SIZE;
 336             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, size);
 337             if (ret < 0)
 338                 return ret;
 339
 340             size = 512 - shift;
 341             if (size > count)
 342                 size = count;
 343             memcpy(buf, s->aligned_buf + shift, size);
 344
 345             buf += size;
 346             offset += size;
 347             count -= size;
 348             sum += size;
 349
 350             if (count == 0)
 351                 return sum;
 352         }
 353         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 354
 355             /* read on aligned buffer */
 356
 357             while (count) {
 358
 359                 size = (count + 0x1ff) & ~0x1ff;
 360                 if (size > ALIGNED_BUFFER_SIZE)
 361                     size = ALIGNED_BUFFER_SIZE;
 362
 363                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, size);
 364                 if (ret < 0) {
 365                     return ret;
 366                 } else if (ret == 0) {
 367                     fprintf(stderr, "raw_pread: read beyond end of file\n");
 368                     abort();
 369                 }
 370
 371                 size = ret;
 372                 if (size > count)
 373                     size = count;
 374
 375                 memcpy(buf, s->aligned_buf, size);
 376
 377                 buf += size;
 378                 offset += size;
 379                 count -= size;
 380                 sum += size;
 381             }
 382
 383             return sum;
 384         }
 385     }
 386
 387     return raw_pread_aligned(bs, offset, buf, count) + sum;
 388 }
 389
 390 static int raw_read(BlockDriverState *bs, int64_t sector_num,
 391                     uint8_t *buf, int nb_sectors)
 392 {
 393     int ret;
 394
 395     ret = raw_pread(bs, sector_num * 512, buf, nb_sectors * 512);
 396     if (ret == (nb_sectors * 512))
 397         ret = 0;
 398     return ret;
 399 }
 400
 401 /*
 402  * offset and count are in bytes and possibly not aligned. For files opened
 403  * with O_DIRECT, necessary alignments are ensured before calling
 404  * raw_pwrite_aligned to do the actual write.
 405  */
 406 static int raw_pwrite(BlockDriverState *bs, int64_t offset,
 407                       const uint8_t *buf, int count)
 408 {
 409     BDRVRawState *s = bs->opaque;
 410     int size, ret, shift, sum;
 411
 412     sum = 0;
 413
 414     if (s->aligned_buf != NULL) {
 415
 416         if (offset & 0x1ff) {
 417             /* align offset on a 512 bytes boundary */
 418             shift = offset & 0x1ff;
 419             ret = raw_pread_aligned(bs, offset - shift, s->aligned_buf, 512);
 420             if (ret < 0)
 421                 return ret;
 422
 423             size = 512 - shift;
 424             if (size > count)
 425                 size = count;
 426             memcpy(s->aligned_buf + shift, buf, size);
 427
 428             ret = raw_pwrite_aligned(bs, offset - shift, s->aligned_buf, 512);
 429             if (ret < 0)
 430                 return ret;
 431
 432             buf += size;
 433             offset += size;
 434             count -= size;
 435             sum += size;
 436
 437             if (count == 0)
 438                 return sum;
 439         }
 440         if (count & 0x1ff || (uintptr_t) buf & 0x1ff) {
 441
 442             while ((size = (count & ~0x1ff)) != 0) {
 443
 444                 if (size > ALIGNED_BUFFER_SIZE)
 445                     size = ALIGNED_BUFFER_SIZE;
 446
 447                 memcpy(s->aligned_buf, buf, size);
 448
 449                 ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, size);
 450                 if (ret < 0)
 451                     return ret;
 452
 453                 buf += ret;
 454                 offset += ret;
 455                 count -= ret;
 456                 sum += ret;
 457             }
 458             /* here, count < 512 because (count & ~0x1ff) == 0 */
 459             if (count) {
 460                 ret = raw_pread_aligned(bs, offset, s->aligned_buf, 512);
 461                 if (ret < 0)
 462                     return ret;
 463                  memcpy(s->aligned_buf, buf, count);
 464
 465                  ret = raw_pwrite_aligned(bs, offset, s->aligned_buf, 512);
 466                  if (ret < 0)
 467                      return ret;
 468                  if (count < ret)
 469                      ret = count;
 470
 471                  sum += ret;
 472             }
 473             return sum;
 474         }
 475     }
 476     return raw_pwrite_aligned(bs, offset, buf, count) + sum;
 477 }
 478
 479 static int raw_write(BlockDriverState *bs, int64_t sector_num,
 480                      const uint8_t *buf, int nb_sectors)
 481 {
 482     int ret;
 483     ret = raw_pwrite(bs, sector_num * 512, buf, nb_sectors * 512);
 484     if (ret == (nb_sectors * 512))
 485         ret = 0;
 486     return ret;
 487 }
 488
 489 /*
 490  * Check if all memory in this vector is sector aligned.
 491  */
 492 static int qiov_is_aligned(QEMUIOVector *qiov)
 493 {
 494     int i;
 495
 496     for (i = 0; i < qiov->niov; i++) {
 497         if ((uintptr_t) qiov->iov[i].iov_base % 512) {
 498             return 0;
 499         }
 500     }
 501
 502     return 1;
 503 }
 504
 505 static BlockDriverAIOCB *raw_aio_submit(BlockDriverState *bs,
 506         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 507         BlockDriverCompletionFunc *cb, void *opaque, int type)
 508 {
 509     BDRVRawState *s = bs->opaque;
 510
 511     if (fd_open(bs) < 0)
 512         return NULL;
 513
 514     /*
 515      * If O_DIRECT is used the buffer needs to be aligned on a sector
 516      * boundary.  Check if this is the case or telll the low-level
 517      * driver that it needs to copy the buffer.
 518      */
 519     if (s->aligned_buf) {
 520         if (!qiov_is_aligned(qiov)) {
 521             type |= QEMU_AIO_MISALIGNED;
 522 #ifdef CONFIG_LINUX_AIO
 523         } else if (s->use_aio) {
 524             return laio_submit(bs, s->aio_ctx, s->fd, sector_num, qiov,
 525                                nb_sectors, cb, opaque, type);
 526 #endif
 527         }
 528     }
 529
 530     return paio_submit(bs, s->fd, sector_num, qiov, nb_sectors,
 531                        cb, opaque, type);
 532 }
 533
 534 static BlockDriverAIOCB *raw_aio_readv(BlockDriverState *bs,
 535         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 536         BlockDriverCompletionFunc *cb, void *opaque)
 537 {
 538     return raw_aio_submit(bs, sector_num, qiov, nb_sectors,
 539                           cb, opaque, QEMU_AIO_READ);
 540 }
 541
 542 static BlockDriverAIOCB *raw_aio_writev(BlockDriverState *bs,
 543         int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
 544         BlockDriverCompletionFunc *cb, void *opaque)
 545 {
 546     return raw_aio_submit(bs, sector_num, qiov, nb_sectors,
 547                           cb, opaque, QEMU_AIO_WRITE);
 548 }
 549
 550 static BlockDriverAIOCB *raw_aio_flush(BlockDriverState *bs,
 551         BlockDriverCompletionFunc *cb, void *opaque)
 552 {
 553     BDRVRawState *s = bs->opaque;
 554
 555     if (fd_open(bs) < 0)
 556         return NULL;
 557
 558     return paio_submit(bs, s->fd, 0, NULL, 0, cb, opaque, QEMU_AIO_FLUSH);
 559 }
 560
 561 static void raw_close(BlockDriverState *bs)
 562 {
 563     BDRVRawState *s = bs->opaque;
 564     if (s->fd >= 0) {
 565         close(s->fd);
 566         s->fd = -1;
 567         if (s->aligned_buf != NULL)
 568             qemu_vfree(s->aligned_buf);
 569     }
 570 }
 571
 572 static int raw_truncate(BlockDriverState *bs, int64_t offset)
 573 {
 574     BDRVRawState *s = bs->opaque;
 575     if (s->type != FTYPE_FILE)
 576         return -ENOTSUP;
 577     if (ftruncate(s->fd, offset) < 0)
 578         return -errno;
 579     return 0;
 580 }
 581
 582 #ifdef __OpenBSD__
 583 static int64_t raw_getlength(BlockDriverState *bs)
 584 {
 585     BDRVRawState *s = bs->opaque;
 586     int fd = s->fd;
 587     struct stat st;
 588
 589     if (fstat(fd, &st))
 590         return -1;
 591     if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode)) {
 592         struct disklabel dl;
 593
 594         if (ioctl(fd, DIOCGDINFO, &dl))
 595             return -1;
 596         return (uint64_t)dl.d_secsize *
 597             dl.d_partitions[DISKPART(st.st_rdev)].p_size;
 598     } else
 599         return st.st_size;
 600 }
 601 #elif defined(__sun__)
 602 static int64_t raw_getlength(BlockDriverState *bs)
 603 {
 604     BDRVRawState *s = bs->opaque;
 605     struct dk_minfo minfo;
 606     int ret;
 607
 608     ret = fd_open(bs);
 609     if (ret < 0) {
 610         return ret;
 611     }
 612
 613     /*
 614      * Use the DKIOCGMEDIAINFO ioctl to read the size.
 615      */
 616     ret = ioctl(s->fd, DKIOCGMEDIAINFO, &minfo);
 617     if (ret != -1) {
 618         return minfo.dki_lbsize * minfo.dki_capacity;
 619     }
 620
 621     /*
 622      * There are reports that lseek on some devices fails, but
 623      * irc discussion said that contingency on contingency was overkill.
 624      */
 625     return lseek(s->fd, 0, SEEK_END);
 626 }
 627 #elif defined(CONFIG_BSD)
 628 static int64_t raw_getlength(BlockDriverState *bs)
 629 {
 630     BDRVRawState *s = bs->opaque;
 631     int fd = s->fd;
 632     int64_t size;
 633     struct stat sb;
 634 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
 635     int reopened = 0;
 636 #endif
 637     int ret;
 638
 639     ret = fd_open(bs);
 640     if (ret < 0)
 641         return ret;
 642
 643 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
 644 again:
 645 #endif
 646     if (!fstat(fd, &sb) && (S_IFCHR & sb.st_mode)) {
 647 #ifdef DIOCGMEDIASIZE
 648         if (ioctl(fd, DIOCGMEDIASIZE, (off_t *)&size))
 649 #elif defined(DIOCGPART)
 650         {
 651                 struct partinfo pi;
 652                 if (ioctl(fd, DIOCGPART, &pi) == 0)
 653                         size = pi.media_size;
 654                 else
 655                         size = 0;
 656         }
 657         if (size == 0)
 658 #endif
 659 #ifdef CONFIG_COCOA
 660         size = LONG_LONG_MAX;
 661 #else
 662         size = lseek(fd, 0LL, SEEK_END);
 663 #endif
 664 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 665         switch(s->type) {
 666         case FTYPE_CD:
 667             /* XXX FreeBSD acd returns UINT_MAX sectors for an empty drive */
 668             if (size == 2048LL * (unsigned)-1)
 669                 size = 0;
 670             /* XXX no disc?  maybe we need to reopen... */
 671             if (size <= 0 && !reopened && cdrom_reopen(bs) >= 0) {
 672                 reopened = 1;
 673                 goto again;
 674             }
 675         }
 676 #endif
 677     } else {
 678         size = lseek(fd, 0, SEEK_END);
 679     }
 680     return size;
 681 }
 682 #else
 683 static int64_t raw_getlength(BlockDriverState *bs)
 684 {
 685     BDRVRawState *s = bs->opaque;
 686     int ret;
 687
 688     ret = fd_open(bs);
 689     if (ret < 0) {
 690         return ret;
 691     }
 692
 693     return lseek(s->fd, 0, SEEK_END);
 694 }
 695 #endif
 696
 697 static int raw_create(const char *filename, QEMUOptionParameter *options)
 698 {
 699     int fd;
 700     int result = 0;
 701     int64_t total_size = 0;
 702
 703     /* Read out options */
 704     while (options && options->name) {
 705         if (!strcmp(options->name, BLOCK_OPT_SIZE)) {
 706             total_size = options->value.n / 512;
 707         }
 708         options++;
 709     }
 710
 711     fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC | O_BINARY,
 712               0644);
 713     if (fd < 0) {
 714         result = -errno;
 715     } else {
 716         if (ftruncate(fd, total_size * 512) != 0) {
 717             result = -errno;
 718         }
 719         if (close(fd) != 0) {
 720             result = -errno;
 721         }
 722     }
 723     return result;
 724 }
 725
 726 static void raw_flush(BlockDriverState *bs)
 727 {
 728     BDRVRawState *s = bs->opaque;
 729     qemu_fdatasync(s->fd);
 730 }
 731
 732
 733 static QEMUOptionParameter raw_create_options[] = {
 734     {
 735         .name = BLOCK_OPT_SIZE,
 736         .type = OPT_SIZE,
 737         .help = "Virtual disk size"
 738     },
 739     { NULL }
 740 };
 741
 742 static BlockDriver bdrv_file = {
 743     .format_name = "file",
 744     .protocol_name = "file",
 745     .instance_size = sizeof(BDRVRawState),
 746     .bdrv_probe = NULL, /* no probe for protocols */
 747     .bdrv_file_open = raw_open,
 748     .bdrv_read = raw_read,
 749     .bdrv_write = raw_write,
 750     .bdrv_close = raw_close,
 751     .bdrv_create = raw_create,
 752     .bdrv_flush = raw_flush,
 753
 754     .bdrv_aio_readv = raw_aio_readv,
 755     .bdrv_aio_writev = raw_aio_writev,
 756     .bdrv_aio_flush = raw_aio_flush,
 757
 758     .bdrv_truncate = raw_truncate,
 759     .bdrv_getlength = raw_getlength,
 760
 761     .create_options = raw_create_options,
 762 };
 763
 764 /***********************************************/
 765 /* host device */
 766
 767 #ifdef CONFIG_COCOA
 768 static kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator );
 769 static kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize );
 770
 771 kern_return_t FindEjectableCDMedia( io_iterator_t *mediaIterator )
 772 {
 773     kern_return_t       kernResult;
 774     mach_port_t     masterPort;
 775     CFMutableDictionaryRef  classesToMatch;
 776
 777     kernResult = IOMasterPort( MACH_PORT_NULL, &masterPort );
 778     if ( KERN_SUCCESS != kernResult ) {
 779         printf( "IOMasterPort returned %d\n", kernResult );
 780     }
 781
 782     classesToMatch = IOServiceMatching( kIOCDMediaClass );
 783     if ( classesToMatch == NULL ) {
 784         printf( "IOServiceMatching returned a NULL dictionary.\n" );
 785     } else {
 786     CFDictionarySetValue( classesToMatch, CFSTR( kIOMediaEjectableKey ), kCFBooleanTrue );
 787     }
 788     kernResult = IOServiceGetMatchingServices( masterPort, classesToMatch, mediaIterator );
 789     if ( KERN_SUCCESS != kernResult )
 790     {
 791         printf( "IOServiceGetMatchingServices returned %d\n", kernResult );
 792     }
 793
 794     return kernResult;
 795 }
 796
 797 kern_return_t GetBSDPath( io_iterator_t mediaIterator, char *bsdPath, CFIndex maxPathSize )
 798 {
 799     io_object_t     nextMedia;
 800     kern_return_t   kernResult = KERN_FAILURE;
 801     *bsdPath = '\0';
 802     nextMedia = IOIteratorNext( mediaIterator );
 803     if ( nextMedia )
 804     {
 805         CFTypeRef   bsdPathAsCFString;
 806     bsdPathAsCFString = IORegistryEntryCreateCFProperty( nextMedia, CFSTR( kIOBSDNameKey ), kCFAllocatorDefault, 0 );
 807         if ( bsdPathAsCFString ) {
 808             size_t devPathLength;
 809             strcpy( bsdPath, _PATH_DEV );
 810             strcat( bsdPath, "r" );
 811             devPathLength = strlen( bsdPath );
 812             if ( CFStringGetCString( bsdPathAsCFString, bsdPath + devPathLength, maxPathSize - devPathLength, kCFStringEncodingASCII ) ) {
 813                 kernResult = KERN_SUCCESS;
 814             }
 815             CFRelease( bsdPathAsCFString );
 816         }
 817         IOObjectRelease( nextMedia );
 818     }
 819
 820     return kernResult;
 821 }
 822
 823 #endif
 824
 825 static int hdev_probe_device(const char *filename)
 826 {
 827     struct stat st;
 828
 829     /* allow a dedicated CD-ROM driver to match with a higher priority */
 830     if (strstart(filename, "/dev/cdrom", NULL))
 831         return 50;
 832
 833     if (stat(filename, &st) >= 0 &&
 834             (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))) {
 835         return 100;
 836     }
 837
 838     return 0;
 839 }
 840
 841 static int hdev_open(BlockDriverState *bs, const char *filename, int flags)
 842 {
 843     BDRVRawState *s = bs->opaque;
 844
 845 #ifdef CONFIG_COCOA
 846     if (strstart(filename, "/dev/cdrom", NULL)) {
 847         kern_return_t kernResult;
 848         io_iterator_t mediaIterator;
 849         char bsdPath[ MAXPATHLEN ];
 850         int fd;
 851
 852         kernResult = FindEjectableCDMedia( &mediaIterator );
 853         kernResult = GetBSDPath( mediaIterator, bsdPath, sizeof( bsdPath ) );
 854
 855         if ( bsdPath[ 0 ] != '\0' ) {
 856             strcat(bsdPath,"s0");
 857             /* some CDs don't have a partition 0 */
 858             fd = open(bsdPath, O_RDONLY | O_BINARY | O_LARGEFILE);
 859             if (fd < 0) {
 860                 bsdPath[strlen(bsdPath)-1] = '1';
 861             } else {
 862                 close(fd);
 863             }
 864             filename = bsdPath;
 865         }
 866
 867         if ( mediaIterator )
 868             IOObjectRelease( mediaIterator );
 869     }
 870 #endif
 871
 872     s->type = FTYPE_FILE;
 873 #if defined(__linux__)
 874     if (strstart(filename, "/dev/sg", NULL)) {
 875         bs->sg = 1;
 876     }
 877 #endif
 878
 879     return raw_open_common(bs, filename, flags, 0);
 880 }
 881
 882 #if defined(__linux__)
 883 /* Note: we do not have a reliable method to detect if the floppy is
 884    present. The current method is to try to open the floppy at every
 885    I/O and to keep it opened during a few hundreds of ms. */
 886 static int fd_open(BlockDriverState *bs)
 887 {
 888     BDRVRawState *s = bs->opaque;
 889     int last_media_present;
 890
 891     if (s->type != FTYPE_FD)
 892         return 0;
 893     last_media_present = (s->fd >= 0);
 894     if (s->fd >= 0 &&
 895         (qemu_get_clock(rt_clock) - s->fd_open_time) >= FD_OPEN_TIMEOUT) {
 896         close(s->fd);
 897         s->fd = -1;
 898 #ifdef DEBUG_FLOPPY
 899         printf("Floppy closed\n");
 900 #endif
 901     }
 902     if (s->fd < 0) {
 903         if (s->fd_got_error &&
 904             (qemu_get_clock(rt_clock) - s->fd_error_time) < FD_OPEN_TIMEOUT) {
 905 #ifdef DEBUG_FLOPPY
 906             printf("No floppy (open delayed)\n");
 907 #endif
 908             return -EIO;
 909         }
 910         s->fd = open(bs->filename, s->open_flags & ~O_NONBLOCK);
 911         if (s->fd < 0) {
 912             s->fd_error_time = qemu_get_clock(rt_clock);
 913             s->fd_got_error = 1;
 914             if (last_media_present)
 915                 s->fd_media_changed = 1;
 916 #ifdef DEBUG_FLOPPY
 917             printf("No floppy\n");
 918 #endif
 919             return -EIO;
 920         }
 921 #ifdef DEBUG_FLOPPY
 922         printf("Floppy opened\n");
 923 #endif
 924     }
 925     if (!last_media_present)
 926         s->fd_media_changed = 1;
 927     s->fd_open_time = qemu_get_clock(rt_clock);
 928     s->fd_got_error = 0;
 929     return 0;
 930 }
 931
 932 static int hdev_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 933 {
 934     BDRVRawState *s = bs->opaque;
 935
 936     return ioctl(s->fd, req, buf);
 937 }
 938
 939 static BlockDriverAIOCB *hdev_aio_ioctl(BlockDriverState *bs,
 940         unsigned long int req, void *buf,
 941         BlockDriverCompletionFunc *cb, void *opaque)
 942 {
 943     BDRVRawState *s = bs->opaque;
 944
 945     if (fd_open(bs) < 0)
 946         return NULL;
 947     return paio_ioctl(bs, s->fd, req, buf, cb, opaque);
 948 }
 949
 950 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 951 static int fd_open(BlockDriverState *bs)
 952 {
 953     BDRVRawState *s = bs->opaque;
 954
 955     /* this is just to ensure s->fd is sane (its called by io ops) */
 956     if (s->fd >= 0)
 957         return 0;
 958     return -EIO;
 959 }
 960 #else /* !linux && !FreeBSD */
 961
 962 static int fd_open(BlockDriverState *bs)
 963 {
 964     return 0;
 965 }
 966
 967 #endif /* !linux && !FreeBSD */
 968
 969 static int hdev_create(const char *filename, QEMUOptionParameter *options)
 970 {
 971     int fd;
 972     int ret = 0;
 973     struct stat stat_buf;
 974     int64_t total_size = 0;
 975
 976     /* Read out options */
 977     while (options && options->name) {
 978         if (!strcmp(options->name, "size")) {
 979             total_size = options->value.n / 512;
 980         }
 981         options++;
 982     }
 983
 984     fd = open(filename, O_WRONLY | O_BINARY);
 985     if (fd < 0)
 986         return -errno;
 987
 988     if (fstat(fd, &stat_buf) < 0)
 989         ret = -errno;
 990     else if (!S_ISBLK(stat_buf.st_mode) && !S_ISCHR(stat_buf.st_mode))
 991         ret = -ENODEV;
 992     else if (lseek(fd, 0, SEEK_END) < total_size * 512)
 993         ret = -ENOSPC;
 994
 995     close(fd);
 996     return ret;
 997 }
 998
 999 static BlockDriver bdrv_host_device = {
1000     .format_name        = "host_device",
1001     .protocol_name        = "host_device",
1002     .instance_size      = sizeof(BDRVRawState),
1003     .bdrv_probe_device  = hdev_probe_device,
1004     .bdrv_file_open     = hdev_open,
1005     .bdrv_close         = raw_close,
1006     .bdrv_create        = hdev_create,
1007     .create_options     = raw_create_options,
1008     .no_zero_init       = 1,
1009     .bdrv_flush         = raw_flush,
1010
1011     .bdrv_aio_readv     = raw_aio_readv,
1012     .bdrv_aio_writev    = raw_aio_writev,
1013     .bdrv_aio_flush     = raw_aio_flush,
1014
1015     .bdrv_read          = raw_read,
1016     .bdrv_write         = raw_write,
1017     .bdrv_getlength     = raw_getlength,
1018
1019     /* generic scsi device */
1020 #ifdef __linux__
1021     .bdrv_ioctl         = hdev_ioctl,
1022     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1023 #endif
1024 };
1025
1026 #ifdef __linux__
1027 static int floppy_open(BlockDriverState *bs, const char *filename, int flags)
1028 {
1029     BDRVRawState *s = bs->opaque;
1030     int ret;
1031
1032     s->type = FTYPE_FD;
1033
1034     /* open will not fail even if no floppy is inserted, so add O_NONBLOCK */
1035     ret = raw_open_common(bs, filename, flags, O_NONBLOCK);
1036     if (ret)
1037         return ret;
1038
1039     /* close fd so that we can reopen it as needed */
1040     close(s->fd);
1041     s->fd = -1;
1042     s->fd_media_changed = 1;
1043
1044     return 0;
1045 }
1046
1047 static int floppy_probe_device(const char *filename)
1048 {
1049     int fd, ret;
1050     int prio = 0;
1051     struct floppy_struct fdparam;
1052
1053     if (strstart(filename, "/dev/fd", NULL))
1054         prio = 50;
1055
1056     fd = open(filename, O_RDONLY | O_NONBLOCK);
1057     if (fd < 0) {
1058         goto out;
1059     }
1060
1061     /* Attempt to detect via a floppy specific ioctl */
1062     ret = ioctl(fd, FDGETPRM, &fdparam);
1063     if (ret >= 0)
1064         prio = 100;
1065
1066     close(fd);
1067 out:
1068     return prio;
1069 }
1070
1071
1072 static int floppy_is_inserted(BlockDriverState *bs)
1073 {
1074     return fd_open(bs) >= 0;
1075 }
1076
1077 static int floppy_media_changed(BlockDriverState *bs)
1078 {
1079     BDRVRawState *s = bs->opaque;
1080     int ret;
1081
1082     /*
1083      * XXX: we do not have a true media changed indication.
1084      * It does not work if the floppy is changed without trying to read it.
1085      */
1086     fd_open(bs);
1087     ret = s->fd_media_changed;
1088     s->fd_media_changed = 0;
1089 #ifdef DEBUG_FLOPPY
1090     printf("Floppy changed=%d\n", ret);
1091 #endif
1092     return ret;
1093 }
1094
1095 static int floppy_eject(BlockDriverState *bs, int eject_flag)
1096 {
1097     BDRVRawState *s = bs->opaque;
1098     int fd;
1099
1100     if (s->fd >= 0) {
1101         close(s->fd);
1102         s->fd = -1;
1103     }
1104     fd = open(bs->filename, s->open_flags | O_NONBLOCK);
1105     if (fd >= 0) {
1106         if (ioctl(fd, FDEJECT, 0) < 0)
1107             perror("FDEJECT");
1108         close(fd);
1109     }
1110
1111     return 0;
1112 }
1113
1114 static BlockDriver bdrv_host_floppy = {
1115     .format_name        = "host_floppy",
1116     .protocol_name      = "host_floppy",
1117     .instance_size      = sizeof(BDRVRawState),
1118     .bdrv_probe_device  = floppy_probe_device,
1119     .bdrv_file_open     = floppy_open,
1120     .bdrv_close         = raw_close,
1121     .bdrv_create        = hdev_create,
1122     .create_options     = raw_create_options,
1123     .no_zero_init       = 1,
1124     .bdrv_flush         = raw_flush,
1125
1126     .bdrv_aio_readv     = raw_aio_readv,
1127     .bdrv_aio_writev    = raw_aio_writev,
1128     .bdrv_aio_flush     = raw_aio_flush,
1129
1130     .bdrv_read          = raw_read,
1131     .bdrv_write         = raw_write,
1132     .bdrv_getlength     = raw_getlength,
1133
1134     /* removable device support */
1135     .bdrv_is_inserted   = floppy_is_inserted,
1136     .bdrv_media_changed = floppy_media_changed,
1137     .bdrv_eject         = floppy_eject,
1138 };
1139
1140 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1141 {
1142     BDRVRawState *s = bs->opaque;
1143
1144     s->type = FTYPE_CD;
1145
1146     /* open will not fail even if no CD is inserted, so add O_NONBLOCK */
1147     return raw_open_common(bs, filename, flags, O_NONBLOCK);
1148 }
1149
1150 static int cdrom_probe_device(const char *filename)
1151 {
1152     int fd, ret;
1153     int prio = 0;
1154
1155     if (strstart(filename, "/dev/cd", NULL))
1156         prio = 50;
1157
1158     fd = open(filename, O_RDONLY | O_NONBLOCK);
1159     if (fd < 0) {
1160         goto out;
1161     }
1162
1163     /* Attempt to detect via a CDROM specific ioctl */
1164     ret = ioctl(fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
1165     if (ret >= 0)
1166         prio = 100;
1167
1168     close(fd);
1169 out:
1170     return prio;
1171 }
1172
1173 static int cdrom_is_inserted(BlockDriverState *bs)
1174 {
1175     BDRVRawState *s = bs->opaque;
1176     int ret;
1177
1178     ret = ioctl(s->fd, CDROM_DRIVE_STATUS, CDSL_CURRENT);
1179     if (ret == CDS_DISC_OK)
1180         return 1;
1181     return 0;
1182 }
1183
1184 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1185 {
1186     BDRVRawState *s = bs->opaque;
1187
1188     if (eject_flag) {
1189         if (ioctl(s->fd, CDROMEJECT, NULL) < 0)
1190             perror("CDROMEJECT");
1191     } else {
1192         if (ioctl(s->fd, CDROMCLOSETRAY, NULL) < 0)
1193             perror("CDROMEJECT");
1194     }
1195
1196     return 0;
1197 }
1198
1199 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1200 {
1201     BDRVRawState *s = bs->opaque;
1202
1203     if (ioctl(s->fd, CDROM_LOCKDOOR, locked) < 0) {
1204         /*
1205          * Note: an error can happen if the distribution automatically
1206          * mounts the CD-ROM
1207          */
1208         /* perror("CDROM_LOCKDOOR"); */
1209     }
1210
1211     return 0;
1212 }
1213
1214 static BlockDriver bdrv_host_cdrom = {
1215     .format_name        = "host_cdrom",
1216     .protocol_name      = "host_cdrom",
1217     .instance_size      = sizeof(BDRVRawState),
1218     .bdrv_probe_device  = cdrom_probe_device,
1219     .bdrv_file_open     = cdrom_open,
1220     .bdrv_close         = raw_close,
1221     .bdrv_create        = hdev_create,
1222     .create_options     = raw_create_options,
1223     .no_zero_init       = 1,
1224     .bdrv_flush         = raw_flush,
1225
1226     .bdrv_aio_readv     = raw_aio_readv,
1227     .bdrv_aio_writev    = raw_aio_writev,
1228     .bdrv_aio_flush     = raw_aio_flush,
1229
1230     .bdrv_read          = raw_read,
1231     .bdrv_write         = raw_write,
1232     .bdrv_getlength     = raw_getlength,
1233
1234     /* removable device support */
1235     .bdrv_is_inserted   = cdrom_is_inserted,
1236     .bdrv_eject         = cdrom_eject,
1237     .bdrv_set_locked    = cdrom_set_locked,
1238
1239     /* generic scsi device */
1240     .bdrv_ioctl         = hdev_ioctl,
1241     .bdrv_aio_ioctl     = hdev_aio_ioctl,
1242 };
1243 #endif /* __linux__ */
1244
1245 #if defined (__FreeBSD__) || defined(__FreeBSD_kernel__)
1246 static int cdrom_open(BlockDriverState *bs, const char *filename, int flags)
1247 {
1248     BDRVRawState *s = bs->opaque;
1249     int ret;
1250
1251     s->type = FTYPE_CD;
1252
1253     ret = raw_open_common(bs, filename, flags, 0);
1254     if (ret)
1255         return ret;
1256
1257     /* make sure the door isnt locked at this time */
1258     ioctl(s->fd, CDIOCALLOW);
1259     return 0;
1260 }
1261
1262 static int cdrom_probe_device(const char *filename)
1263 {
1264     if (strstart(filename, "/dev/cd", NULL) ||
1265             strstart(filename, "/dev/acd", NULL))
1266         return 100;
1267     return 0;
1268 }
1269
1270 static int cdrom_reopen(BlockDriverState *bs)
1271 {
1272     BDRVRawState *s = bs->opaque;
1273     int fd;
1274
1275     /*
1276      * Force reread of possibly changed/newly loaded disc,
1277      * FreeBSD seems to not notice sometimes...
1278      */
1279     if (s->fd >= 0)
1280         close(s->fd);
1281     fd = open(bs->filename, s->open_flags, 0644);
1282     if (fd < 0) {
1283         s->fd = -1;
1284         return -EIO;
1285     }
1286     s->fd = fd;
1287
1288     /* make sure the door isnt locked at this time */
1289     ioctl(s->fd, CDIOCALLOW);
1290     return 0;
1291 }
1292
1293 static int cdrom_is_inserted(BlockDriverState *bs)
1294 {
1295     return raw_getlength(bs) > 0;
1296 }
1297
1298 static int cdrom_eject(BlockDriverState *bs, int eject_flag)
1299 {
1300     BDRVRawState *s = bs->opaque;
1301
1302     if (s->fd < 0)
1303         return -ENOTSUP;
1304
1305     (void) ioctl(s->fd, CDIOCALLOW);
1306
1307     if (eject_flag) {
1308         if (ioctl(s->fd, CDIOCEJECT) < 0)
1309             perror("CDIOCEJECT");
1310     } else {
1311         if (ioctl(s->fd, CDIOCCLOSE) < 0)
1312             perror("CDIOCCLOSE");
1313     }
1314
1315     if (cdrom_reopen(bs) < 0)
1316         return -ENOTSUP;
1317     return 0;
1318 }
1319
1320 static int cdrom_set_locked(BlockDriverState *bs, int locked)
1321 {
1322     BDRVRawState *s = bs->opaque;
1323
1324     if (s->fd < 0)
1325         return -ENOTSUP;
1326     if (ioctl(s->fd, (locked ? CDIOCPREVENT : CDIOCALLOW)) < 0) {
1327         /*
1328          * Note: an error can happen if the distribution automatically
1329          * mounts the CD-ROM
1330          */
1331         /* perror("CDROM_LOCKDOOR"); */
1332     }
1333
1334     return 0;
1335 }
1336
1337 static BlockDriver bdrv_host_cdrom = {
1338     .format_name        = "host_cdrom",
1339     .protocol_name      = "host_cdrom",
1340     .instance_size      = sizeof(BDRVRawState),
1341     .bdrv_probe_device  = cdrom_probe_device,
1342     .bdrv_file_open     = cdrom_open,
1343     .bdrv_close         = raw_close,
1344     .bdrv_create        = hdev_create,
1345     .create_options     = raw_create_options,
1346     .no_zero_init       = 1,
1347     .bdrv_flush         = raw_flush,
1348
1349     .bdrv_aio_readv     = raw_aio_readv,
1350     .bdrv_aio_writev    = raw_aio_writev,
1351     .bdrv_aio_flush     = raw_aio_flush,
1352
1353     .bdrv_read          = raw_read,
1354     .bdrv_write         = raw_write,
1355     .bdrv_getlength     = raw_getlength,
1356
1357     /* removable device support */
1358     .bdrv_is_inserted   = cdrom_is_inserted,
1359     .bdrv_eject         = cdrom_eject,
1360     .bdrv_set_locked    = cdrom_set_locked,
1361 };
1362 #endif /* __FreeBSD__ */
1363
1364 static void bdrv_file_init(void)
1365 {
1366     /*
1367      * Register all the drivers.  Note that order is important, the driver
1368      * registered last will get probed first.
1369      */
1370     bdrv_register(&bdrv_file);
1371     bdrv_register(&bdrv_host_device);
1372 #ifdef __linux__
1373     bdrv_register(&bdrv_host_floppy);
1374     bdrv_register(&bdrv_host_cdrom);
1375 #endif
1376 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
1377     bdrv_register(&bdrv_host_cdrom);
1378 #endif
1379 }
1380
1381 block_init(bdrv_file_init);