lib/ext2fs/unix_io.c

   1 /*
   2  * unix_io.c --- This is the Unix (well, really POSIX) implementation
   3  *      of the I/O manager.
   4  *
   5  * Implements a one-block write-through cache.
   6  *
   7  * Includes support for Windows NT support under Cygwin.
   8  *
   9  * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
  10  *      2002 by Theodore Ts'o.
  11  *
  12  * %Begin-Header%
  13  * This file may be redistributed under the terms of the GNU Library
  14  * General Public License, version 2.
  15  * %End-Header%
  16  */
  17
  18 #define _LARGEFILE_SOURCE
  19 #define _LARGEFILE64_SOURCE
  20 #define _GNU_SOURCE
  21
  22 #include <stdio.h>
  23 #include <string.h>
  24 #if HAVE_UNISTD_H
  25 #include <unistd.h>
  26 #endif
  27 #if HAVE_ERRNO_H
  28 #include <errno.h>
  29 #endif
  30 #include <fcntl.h>
  31 #include <time.h>
  32 #ifdef __linux__
  33 #include <sys/utsname.h>
  34 #endif
  35 #ifdef HAVE_SYS_IOCTL_H
  36 #include <sys/ioctl.h>
  37 #endif
  38 #ifdef HAVE_SYS_MOUNT_H
  39 #include <sys/mount.h>
  40 #endif
  41 #if HAVE_SYS_STAT_H
  42 #include <sys/stat.h>
  43 #endif
  44 #if HAVE_SYS_TYPES_H
  45 #include <sys/types.h>
  46 #endif
  47 #if HAVE_SYS_RESOURCE_H
  48 #include <sys/resource.h>
  49 #endif
  50
  51 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
  52 #define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
  53 #endif
  54
  55 #if defined(__linux__) && defined(_IO) && !defined(BLKSSZGET)
  56 #define BLKSSZGET  _IO(0x12,104)/* get block device sector size */
  57 #endif
  58
  59 #undef ALIGN_DEBUG
  60
  61 #include "ext2_fs.h"
  62 #include "ext2fs.h"
  63
  64 /*
  65  * For checking structure magic numbers...
  66  */
  67
  68 #define EXT2_CHECK_MAGIC(struct, code) \
  69           if ((struct)->magic != (code)) return (code)
  70
  71 struct unix_cache {
  72         char            *buf;
  73         unsigned long   block;
  74         int             access_time;
  75         unsigned        dirty:1;
  76         unsigned        in_use:1;
  77 };
  78
  79 #define CACHE_SIZE 8
  80 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
  81 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
  82
  83 struct unix_private_data {
  84         int     magic;
  85         int     dev;
  86         int     flags;
  87         int     align;
  88         int     access_time;
  89         ext2_loff_t offset;
  90         struct unix_cache cache[CACHE_SIZE];
  91         void    *bounce;
  92         struct struct_io_stats io_stats;
  93 };
  94
  95 #define IS_ALIGNED(n, align) ((((unsigned long) n) & \
  96                                ((unsigned long) ((align)-1))) == 0)
  97
  98 static errcode_t unix_open(const char *name, int flags, io_channel *channel);
  99 static errcode_t unix_close(io_channel channel);
 100 static errcode_t unix_set_blksize(io_channel channel, int blksize);
 101 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 102                                int count, void *data);
 103 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 104                                 int count, const void *data);
 105 static errcode_t unix_flush(io_channel channel);
 106 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 107                                 int size, const void *data);
 108 static errcode_t unix_set_option(io_channel channel, const char *option,
 109                                  const char *arg);
 110 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 111 ;
 112 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 113                  struct unix_cache *cache, unsigned long long block);
 114 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 115                                int count, void *data);
 116 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 117                                 int count, const void *data);
 118 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 119                               unsigned long long count);
 120
 121 static struct struct_io_manager struct_unix_manager = {
 122         EXT2_ET_MAGIC_IO_MANAGER,
 123         "Unix I/O Manager",
 124         unix_open,
 125         unix_close,
 126         unix_set_blksize,
 127         unix_read_blk,
 128         unix_write_blk,
 129         unix_flush,
 130         unix_write_byte,
 131         unix_set_option,
 132         unix_get_stats,
 133         unix_read_blk64,
 134         unix_write_blk64,
 135         unix_discard,
 136 };
 137
 138 io_manager unix_io_manager = &struct_unix_manager;
 139
 140 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 141 {
 142         errcode_t       retval = 0;
 143
 144         struct unix_private_data *data;
 145
 146         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 147         data = (struct unix_private_data *) channel->private_data;
 148         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 149
 150         if (stats)
 151                 *stats = &data->io_stats;
 152
 153         return retval;
 154 }
 155
 156 /*
 157  * Here are the raw I/O functions
 158  */
 159 static errcode_t raw_read_blk(io_channel channel,
 160                               struct unix_private_data *data,
 161                               unsigned long long block,
 162                               int count, void *buf)
 163 {
 164         errcode_t       retval;
 165         ssize_t         size;
 166         ext2_loff_t     location;
 167         int             actual = 0;
 168
 169         size = (count < 0) ? -count : count * channel->block_size;
 170         data->io_stats.bytes_read += size;
 171         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 172         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 173                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 174                 goto error_out;
 175         }
 176         if ((data->align == 0) ||
 177             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 178                 actual = read(data->dev, buf, size);
 179                 if (actual != size) {
 180                 short_read:
 181                         if (actual < 0)
 182                                 actual = 0;
 183                         retval = EXT2_ET_SHORT_READ;
 184                         goto error_out;
 185                 }
 186                 return 0;
 187         }
 188
 189 #ifdef ALIGN_DEBUG
 190         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 191                (unsigned long) size);
 192 #endif
 193
 194         /*
 195          * The buffer or size which we're trying to read isn't aligned
 196          * to the O_DIRECT rules, so we need to do this the hard way...
 197          */
 198         while (size > 0) {
 199                 actual = read(data->dev, data->bounce, channel->block_size);
 200                 if (actual != channel->block_size)
 201                         goto short_read;
 202                 actual = size;
 203                 if (size > channel->block_size)
 204                         actual = channel->block_size;
 205                 memcpy(buf, data->bounce, actual);
 206                 size -= actual;
 207                 buf += actual;
 208         }
 209         return 0;
 210
 211 error_out:
 212         memset((char *) buf+actual, 0, size-actual);
 213         if (channel->read_error)
 214                 retval = (channel->read_error)(channel, block, count, buf,
 215                                                size, actual, retval);
 216         return retval;
 217 }
 218
 219 static errcode_t raw_write_blk(io_channel channel,
 220                                struct unix_private_data *data,
 221                                unsigned long long block,
 222                                int count, const void *buf)
 223 {
 224         ssize_t         size;
 225         ext2_loff_t     location;
 226         int             actual = 0;
 227         errcode_t       retval;
 228
 229         if (count == 1)
 230                 size = channel->block_size;
 231         else {
 232                 if (count < 0)
 233                         size = -count;
 234                 else
 235                         size = count * channel->block_size;
 236         }
 237         data->io_stats.bytes_written += size;
 238
 239         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 240         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 241                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 242                 goto error_out;
 243         }
 244
 245         if ((data->align == 0) ||
 246             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 247                 actual = write(data->dev, buf, size);
 248                 if (actual != size) {
 249                 short_write:
 250                         retval = EXT2_ET_SHORT_WRITE;
 251                         goto error_out;
 252                 }
 253                 return 0;
 254         }
 255
 256 #ifdef ALIGN_DEBUG
 257         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 258                (unsigned long) size);
 259 #endif
 260         /*
 261          * The buffer or size which we're trying to write isn't aligned
 262          * to the O_DIRECT rules, so we need to do this the hard way...
 263          */
 264         while (size > 0) {
 265                 if (size < channel->block_size) {
 266                         actual = read(data->dev, data->bounce,
 267                                       channel->block_size);
 268                         if (actual != channel->block_size) {
 269                                 retval = EXT2_ET_SHORT_READ;
 270                                 goto error_out;
 271                         }
 272                 }
 273                 actual = size;
 274                 if (size > channel->block_size)
 275                         actual = channel->block_size;
 276                 memcpy(data->bounce, buf, actual);
 277                 actual = write(data->dev, data->bounce, channel->block_size);
 278                 if (actual != channel->block_size)
 279                         goto short_write;
 280                 size -= actual;
 281                 buf += actual;
 282         }
 283         return 0;
 284
 285 error_out:
 286         if (channel->write_error)
 287                 retval = (channel->write_error)(channel, block, count, buf,
 288                                                 size, actual, retval);
 289         return retval;
 290 }
 291
 292
 293 /*
 294  * Here we implement the cache functions
 295  */
 296
 297 /* Allocate the cache buffers */
 298 static errcode_t alloc_cache(io_channel channel,
 299                              struct unix_private_data *data)
 300 {
 301         errcode_t               retval;
 302         struct unix_cache       *cache;
 303         int                     i;
 304
 305         data->access_time = 0;
 306         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 307                 cache->block = 0;
 308                 cache->access_time = 0;
 309                 cache->dirty = 0;
 310                 cache->in_use = 0;
 311                 if (cache->buf)
 312                         ext2fs_free_mem(&cache->buf);
 313                 retval = ext2fs_get_memalign(channel->block_size,
 314                                              data->align, &cache->buf);
 315                 if (retval)
 316                         return retval;
 317         }
 318         if (data->align) {
 319                 if (data->bounce)
 320                         ext2fs_free_mem(&data->bounce);
 321                 retval = ext2fs_get_memalign(channel->block_size, data->align,
 322                                              &data->bounce);
 323         }
 324         return retval;
 325 }
 326
 327 /* Free the cache buffers */
 328 static void free_cache(struct unix_private_data *data)
 329 {
 330         struct unix_cache       *cache;
 331         int                     i;
 332
 333         data->access_time = 0;
 334         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 335                 cache->block = 0;
 336                 cache->access_time = 0;
 337                 cache->dirty = 0;
 338                 cache->in_use = 0;
 339                 if (cache->buf)
 340                         ext2fs_free_mem(&cache->buf);
 341         }
 342         if (data->bounce)
 343                 ext2fs_free_mem(&data->bounce);
 344 }
 345
 346 #ifndef NO_IO_CACHE
 347 /*
 348  * Try to find a block in the cache.  If the block is not found, and
 349  * eldest is a non-zero pointer, then fill in eldest with the cache
 350  * entry to that should be reused.
 351  */
 352 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 353                                             unsigned long long block,
 354                                             struct unix_cache **eldest)
 355 {
 356         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 357         int                     i;
 358
 359         unused_cache = oldest_cache = 0;
 360         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 361                 if (!cache->in_use) {
 362                         if (!unused_cache)
 363                                 unused_cache = cache;
 364                         continue;
 365                 }
 366                 if (cache->block == block) {
 367                         cache->access_time = ++data->access_time;
 368                         return cache;
 369                 }
 370                 if (!oldest_cache ||
 371                     (cache->access_time < oldest_cache->access_time))
 372                         oldest_cache = cache;
 373         }
 374         if (eldest)
 375                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 376         return 0;
 377 }
 378
 379 /*
 380  * Reuse a particular cache entry for another block.
 381  */
 382 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 383                  struct unix_cache *cache, unsigned long long block)
 384 {
 385         if (cache->dirty && cache->in_use)
 386                 raw_write_blk(channel, data, cache->block, 1, cache->buf);
 387
 388         cache->in_use = 1;
 389         cache->dirty = 0;
 390         cache->block = block;
 391         cache->access_time = ++data->access_time;
 392 }
 393
 394 /*
 395  * Flush all of the blocks in the cache
 396  */
 397 static errcode_t flush_cached_blocks(io_channel channel,
 398                                      struct unix_private_data *data,
 399                                      int invalidate)
 400
 401 {
 402         struct unix_cache       *cache;
 403         errcode_t               retval, retval2;
 404         int                     i;
 405
 406         retval2 = 0;
 407         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 408                 if (!cache->in_use)
 409                         continue;
 410
 411                 if (invalidate)
 412                         cache->in_use = 0;
 413
 414                 if (!cache->dirty)
 415                         continue;
 416
 417                 retval = raw_write_blk(channel, data,
 418                                        cache->block, 1, cache->buf);
 419                 if (retval)
 420                         retval2 = retval;
 421                 else
 422                         cache->dirty = 0;
 423         }
 424         return retval2;
 425 }
 426 #endif /* NO_IO_CACHE */
 427
 428 static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 429 {
 430         io_channel      io = NULL;
 431         struct unix_private_data *data = NULL;
 432         errcode_t       retval;
 433         int             open_flags;
 434         struct stat     st;
 435 #ifdef __linux__
 436         struct          utsname ut;
 437 #endif
 438
 439         if (name == 0)
 440                 return EXT2_ET_BAD_DEVICE_NAME;
 441         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 442         if (retval)
 443                 return retval;
 444         memset(io, 0, sizeof(struct struct_io_channel));
 445         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 446         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 447         if (retval)
 448                 goto cleanup;
 449
 450         io->manager = unix_io_manager;
 451         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 452         if (retval)
 453                 goto cleanup;
 454
 455         strcpy(io->name, name);
 456         io->private_data = data;
 457         io->block_size = 1024;
 458         io->read_error = 0;
 459         io->write_error = 0;
 460         io->refcount = 1;
 461
 462         memset(data, 0, sizeof(struct unix_private_data));
 463         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 464         data->io_stats.num_fields = 2;
 465
 466         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 467         if (flags & IO_FLAG_EXCLUSIVE)
 468                 open_flags |= O_EXCL;
 469         if (flags & IO_FLAG_DIRECT_IO)
 470                 open_flags |= O_DIRECT;
 471         data->flags = flags;
 472
 473 #ifdef HAVE_OPEN64
 474         data->dev = open64(io->name, open_flags);
 475 #else
 476         data->dev = open(io->name, open_flags);
 477 #endif
 478         if (data->dev < 0) {
 479                 retval = errno;
 480                 goto cleanup;
 481         }
 482
 483 #ifdef BLKSSZGET
 484         if (flags & IO_FLAG_DIRECT_IO) {
 485                 if (ioctl(data->dev, BLKSSZGET, &data->align) != 0)
 486                         data->align = io->block_size;
 487         }
 488 #endif
 489
 490 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 491         /*
 492          * Some operating systems require that the buffers be aligned,
 493          * regardless of O_DIRECT
 494          */
 495         data->align = 512;
 496 #endif
 497
 498
 499         if ((retval = alloc_cache(io, data)))
 500                 goto cleanup;
 501
 502 #ifdef BLKROGET
 503         if (flags & IO_FLAG_RW) {
 504                 int error;
 505                 int readonly = 0;
 506
 507                 /* Is the block device actually writable? */
 508                 error = ioctl(data->dev, BLKROGET, &readonly);
 509                 if (!error && readonly) {
 510                         close(data->dev);
 511                         retval = EPERM;
 512                         goto cleanup;
 513                 }
 514         }
 515 #endif
 516
 517 #ifdef __linux__
 518 #undef RLIM_INFINITY
 519 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 520 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 521 #else
 522 #define RLIM_INFINITY  (~0UL)
 523 #endif
 524         /*
 525          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 526          * block devices are wrongly getting hit by the filesize
 527          * limit.  This workaround isn't perfect, since it won't work
 528          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 529          *
 530          */
 531         if ((flags & IO_FLAG_RW) &&
 532             (uname(&ut) == 0) &&
 533             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 534              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 535              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 536              (ut.release[5] < '8')) &&
 537             (fstat(data->dev, &st) == 0) &&
 538             (S_ISBLK(st.st_mode))) {
 539                 struct rlimit   rlim;
 540
 541                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 542                 setrlimit(RLIMIT_FSIZE, &rlim);
 543                 getrlimit(RLIMIT_FSIZE, &rlim);
 544                 if (((unsigned long) rlim.rlim_cur) <
 545                     ((unsigned long) rlim.rlim_max)) {
 546                         rlim.rlim_cur = rlim.rlim_max;
 547                         setrlimit(RLIMIT_FSIZE, &rlim);
 548                 }
 549         }
 550 #endif
 551         *channel = io;
 552         return 0;
 553
 554 cleanup:
 555         if (data) {
 556                 free_cache(data);
 557                 ext2fs_free_mem(&data);
 558         }
 559         if (io)
 560                 ext2fs_free_mem(&io);
 561         return retval;
 562 }
 563
 564 static errcode_t unix_close(io_channel channel)
 565 {
 566         struct unix_private_data *data;
 567         errcode_t       retval = 0;
 568
 569         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 570         data = (struct unix_private_data *) channel->private_data;
 571         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 572
 573         if (--channel->refcount > 0)
 574                 return 0;
 575
 576 #ifndef NO_IO_CACHE
 577         retval = flush_cached_blocks(channel, data, 0);
 578 #endif
 579
 580         if (close(data->dev) < 0)
 581                 retval = errno;
 582         free_cache(data);
 583
 584         ext2fs_free_mem(&channel->private_data);
 585         if (channel->name)
 586                 ext2fs_free_mem(&channel->name);
 587         ext2fs_free_mem(&channel);
 588         return retval;
 589 }
 590
 591 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 592 {
 593         struct unix_private_data *data;
 594         errcode_t               retval;
 595
 596         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 597         data = (struct unix_private_data *) channel->private_data;
 598         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 599
 600         if (channel->block_size != blksize) {
 601 #ifndef NO_IO_CACHE
 602                 if ((retval = flush_cached_blocks(channel, data, 0)))
 603                         return retval;
 604 #endif
 605
 606                 channel->block_size = blksize;
 607                 free_cache(data);
 608                 if ((retval = alloc_cache(channel, data)))
 609                         return retval;
 610         }
 611         return 0;
 612 }
 613
 614
 615 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 616                                int count, void *buf)
 617 {
 618         struct unix_private_data *data;
 619         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 620         errcode_t       retval;
 621         char            *cp;
 622         int             i, j;
 623
 624         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 625         data = (struct unix_private_data *) channel->private_data;
 626         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 627
 628 #ifdef NO_IO_CACHE
 629         return raw_read_blk(channel, data, block, count, buf);
 630 #else
 631         /*
 632          * If we're doing an odd-sized read or a very large read,
 633          * flush out the cache and then do a direct read.
 634          */
 635         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 636                 if ((retval = flush_cached_blocks(channel, data, 0)))
 637                         return retval;
 638                 return raw_read_blk(channel, data, block, count, buf);
 639         }
 640
 641         cp = buf;
 642         while (count > 0) {
 643                 /* If it's in the cache, use it! */
 644                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 645 #ifdef DEBUG
 646                         printf("Using cached block %lu\n", block);
 647 #endif
 648                         memcpy(cp, cache->buf, channel->block_size);
 649                         count--;
 650                         block++;
 651                         cp += channel->block_size;
 652                         continue;
 653                 }
 654                 if (count == 1) {
 655                         /*
 656                          * Special case where we read directly into the
 657                          * cache buffer; important in the O_DIRECT case
 658                          */
 659                         cache = reuse[0];
 660                         reuse_cache(channel, data, cache, block);
 661                         if ((retval = raw_read_blk(channel, data, block, 1,
 662                                                    cache->buf))) {
 663                                 cache->in_use = 0;
 664                                 return retval;
 665                         }
 666                         memcpy(cp, cache->buf, channel->block_size);
 667                         return 0;
 668                 }
 669
 670                 /*
 671                  * Find the number of uncached blocks so we can do a
 672                  * single read request
 673                  */
 674                 for (i=1; i < count; i++)
 675                         if (find_cached_block(data, block+i, &reuse[i]))
 676                                 break;
 677 #ifdef DEBUG
 678                 printf("Reading %d blocks starting at %lu\n", i, block);
 679 #endif
 680                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 681                         return retval;
 682
 683                 /* Save the results in the cache */
 684                 for (j=0; j < i; j++) {
 685                         count--;
 686                         cache = reuse[j];
 687                         reuse_cache(channel, data, cache, block++);
 688                         memcpy(cache->buf, cp, channel->block_size);
 689                         cp += channel->block_size;
 690                 }
 691         }
 692         return 0;
 693 #endif /* NO_IO_CACHE */
 694 }
 695
 696 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 697                                int count, void *buf)
 698 {
 699         return unix_read_blk64(channel, block, count, buf);
 700 }
 701
 702 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 703                                 int count, const void *buf)
 704 {
 705         struct unix_private_data *data;
 706         struct unix_cache *cache, *reuse;
 707         errcode_t       retval = 0;
 708         const char      *cp;
 709         int             writethrough;
 710
 711         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 712         data = (struct unix_private_data *) channel->private_data;
 713         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 714
 715 #ifdef NO_IO_CACHE
 716         return raw_write_blk(channel, data, block, count, buf);
 717 #else
 718         /*
 719          * If we're doing an odd-sized write or a very large write,
 720          * flush out the cache completely and then do a direct write.
 721          */
 722         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 723                 if ((retval = flush_cached_blocks(channel, data, 1)))
 724                         return retval;
 725                 return raw_write_blk(channel, data, block, count, buf);
 726         }
 727
 728         /*
 729          * For a moderate-sized multi-block write, first force a write
 730          * if we're in write-through cache mode, and then fill the
 731          * cache with the blocks.
 732          */
 733         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 734         if (writethrough)
 735                 retval = raw_write_blk(channel, data, block, count, buf);
 736
 737         cp = buf;
 738         while (count > 0) {
 739                 cache = find_cached_block(data, block, &reuse);
 740                 if (!cache) {
 741                         cache = reuse;
 742                         reuse_cache(channel, data, cache, block);
 743                 }
 744                 memcpy(cache->buf, cp, channel->block_size);
 745                 cache->dirty = !writethrough;
 746                 count--;
 747                 block++;
 748                 cp += channel->block_size;
 749         }
 750         return retval;
 751 #endif /* NO_IO_CACHE */
 752 }
 753
 754 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 755                                 int count, const void *buf)
 756 {
 757         return unix_write_blk64(channel, block, count, buf);
 758 }
 759
 760 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 761                                  int size, const void *buf)
 762 {
 763         struct unix_private_data *data;
 764         errcode_t       retval = 0;
 765         ssize_t         actual;
 766
 767         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 768         data = (struct unix_private_data *) channel->private_data;
 769         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 770
 771         if (data->align != 0) {
 772 #ifdef ALIGN_DEBUG
 773                 printf("unix_write_byte: O_DIRECT fallback\n");
 774 #endif
 775                 return EXT2_ET_UNIMPLEMENTED;
 776         }
 777
 778 #ifndef NO_IO_CACHE
 779         /*
 780          * Flush out the cache completely
 781          */
 782         if ((retval = flush_cached_blocks(channel, data, 1)))
 783                 return retval;
 784 #endif
 785
 786         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
 787                 return errno;
 788
 789         actual = write(data->dev, buf, size);
 790         if (actual != size)
 791                 return EXT2_ET_SHORT_WRITE;
 792
 793         return 0;
 794 }
 795
 796 /*
 797  * Flush data buffers to disk.
 798  */
 799 static errcode_t unix_flush(io_channel channel)
 800 {
 801         struct unix_private_data *data;
 802         errcode_t retval = 0;
 803
 804         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 805         data = (struct unix_private_data *) channel->private_data;
 806         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 807
 808 #ifndef NO_IO_CACHE
 809         retval = flush_cached_blocks(channel, data, 0);
 810 #endif
 811         fsync(data->dev);
 812         return retval;
 813 }
 814
 815 static errcode_t unix_set_option(io_channel channel, const char *option,
 816                                  const char *arg)
 817 {
 818         struct unix_private_data *data;
 819         unsigned long long tmp;
 820         char *end;
 821
 822         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 823         data = (struct unix_private_data *) channel->private_data;
 824         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 825
 826         if (!strcmp(option, "offset")) {
 827                 if (!arg)
 828                         return EXT2_ET_INVALID_ARGUMENT;
 829
 830                 tmp = strtoull(arg, &end, 0);
 831                 if (*end)
 832                         return EXT2_ET_INVALID_ARGUMENT;
 833                 data->offset = tmp;
 834                 if (data->offset < 0)
 835                         return EXT2_ET_INVALID_ARGUMENT;
 836                 return 0;
 837         }
 838         return EXT2_ET_INVALID_ARGUMENT;
 839 }
 840
 841 #if defined(__linux__) && !defined(BLKDISCARD)
 842 #define BLKDISCARD      _IO(0x12,119)
 843 #endif
 844
 845 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 846                               unsigned long long count)
 847 {
 848 #ifdef BLKDISCARD
 849         struct unix_private_data *data;
 850         __uint64_t      range[2];
 851         int             ret;
 852
 853         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 854         data = (struct unix_private_data *) channel->private_data;
 855         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 856
 857         range[0] = (__uint64_t)(block) * channel->block_size;
 858         range[1] = (__uint64_t)(count) * channel->block_size;
 859
 860         ret = ioctl(data->dev, BLKDISCARD, &range);
 861         if (ret < 0)
 862                 return errno;
 863         return 0;
 864 #else
 865         return EXT2_ET_UNIMPLEMENTED;
 866 #endif
 867 }