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