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 #ifndef _GNU_SOURCE
  21 #define _GNU_SOURCE
  22 #endif
  23
  24 #include <stdio.h>
  25 #include <string.h>
  26 #if HAVE_UNISTD_H
  27 #include <unistd.h>
  28 #endif
  29 #if HAVE_ERRNO_H
  30 #include <errno.h>
  31 #endif
  32 #include <fcntl.h>
  33 #include <time.h>
  34 #ifdef __linux__
  35 #include <sys/utsname.h>
  36 #endif
  37 #ifdef HAVE_SYS_IOCTL_H
  38 #include <sys/ioctl.h>
  39 #endif
  40 #ifdef HAVE_SYS_MOUNT_H
  41 #include <sys/mount.h>
  42 #endif
  43 #if HAVE_SYS_STAT_H
  44 #include <sys/stat.h>
  45 #endif
  46 #if HAVE_SYS_TYPES_H
  47 #include <sys/types.h>
  48 #endif
  49 #if HAVE_SYS_RESOURCE_H
  50 #include <sys/resource.h>
  51 #endif
  52
  53 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
  54 #define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
  55 #endif
  56
  57 #if defined(__linux__) && defined(_IO) && !defined(BLKSSZGET)
  58 #define BLKSSZGET  _IO(0x12,104)/* get block device sector size */
  59 #endif
  60
  61 #undef ALIGN_DEBUG
  62
  63 #include "ext2_fs.h"
  64 #include "ext2fs.h"
  65
  66 /*
  67  * For checking structure magic numbers...
  68  */
  69
  70 #define EXT2_CHECK_MAGIC(struct, code) \
  71           if ((struct)->magic != (code)) return (code)
  72
  73 struct unix_cache {
  74         char            *buf;
  75         unsigned long   block;
  76         int             access_time;
  77         unsigned        dirty:1;
  78         unsigned        in_use:1;
  79 };
  80
  81 #define CACHE_SIZE 8
  82 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
  83 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
  84
  85 struct unix_private_data {
  86         int     magic;
  87         int     dev;
  88         int     flags;
  89         int     align;
  90         int     access_time;
  91         ext2_loff_t offset;
  92         struct unix_cache cache[CACHE_SIZE];
  93         void    *bounce;
  94         struct struct_io_stats io_stats;
  95 };
  96
  97 #define IS_ALIGNED(n, align) ((((unsigned long) n) & \
  98                                ((unsigned long) ((align)-1))) == 0)
  99
 100 static errcode_t unix_open(const char *name, int flags, io_channel *channel);
 101 static errcode_t unix_close(io_channel channel);
 102 static errcode_t unix_set_blksize(io_channel channel, int blksize);
 103 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 104                                int count, void *data);
 105 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 106                                 int count, const void *data);
 107 static errcode_t unix_flush(io_channel channel);
 108 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 109                                 int size, const void *data);
 110 static errcode_t unix_set_option(io_channel channel, const char *option,
 111                                  const char *arg);
 112 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 113 ;
 114 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 115                  struct unix_cache *cache, unsigned long long block);
 116 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 117                                int count, void *data);
 118 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 119                                 int count, const void *data);
 120 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 121                               unsigned long long count);
 122
 123 static struct struct_io_manager struct_unix_manager = {
 124         EXT2_ET_MAGIC_IO_MANAGER,
 125         "Unix I/O Manager",
 126         unix_open,
 127         unix_close,
 128         unix_set_blksize,
 129         unix_read_blk,
 130         unix_write_blk,
 131         unix_flush,
 132         unix_write_byte,
 133         unix_set_option,
 134         unix_get_stats,
 135         unix_read_blk64,
 136         unix_write_blk64,
 137         unix_discard,
 138 };
 139
 140 io_manager unix_io_manager = &struct_unix_manager;
 141
 142 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 143 {
 144         errcode_t       retval = 0;
 145
 146         struct unix_private_data *data;
 147
 148         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 149         data = (struct unix_private_data *) channel->private_data;
 150         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 151
 152         if (stats)
 153                 *stats = &data->io_stats;
 154
 155         return retval;
 156 }
 157
 158 /*
 159  * Here are the raw I/O functions
 160  */
 161 static errcode_t raw_read_blk(io_channel channel,
 162                               struct unix_private_data *data,
 163                               unsigned long long block,
 164                               int count, void *bufv)
 165 {
 166         errcode_t       retval;
 167         ssize_t         size;
 168         ext2_loff_t     location;
 169         int             actual = 0;
 170         unsigned char   *buf = bufv;
 171
 172         size = (count < 0) ? -count : count * channel->block_size;
 173         data->io_stats.bytes_read += size;
 174         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 175         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 176                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 177                 goto error_out;
 178         }
 179         if ((data->align == 0) ||
 180             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 181                 actual = read(data->dev, buf, size);
 182                 if (actual != size) {
 183                 short_read:
 184                         if (actual < 0)
 185                                 actual = 0;
 186                         retval = EXT2_ET_SHORT_READ;
 187                         goto error_out;
 188                 }
 189                 return 0;
 190         }
 191
 192 #ifdef ALIGN_DEBUG
 193         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 194                (unsigned long) size);
 195 #endif
 196
 197         /*
 198          * The buffer or size which we're trying to read isn't aligned
 199          * to the O_DIRECT rules, so we need to do this the hard way...
 200          */
 201         while (size > 0) {
 202                 actual = read(data->dev, data->bounce, channel->block_size);
 203                 if (actual != channel->block_size)
 204                         goto short_read;
 205                 actual = size;
 206                 if (size > channel->block_size)
 207                         actual = channel->block_size;
 208                 memcpy(buf, data->bounce, actual);
 209                 size -= actual;
 210                 buf += actual;
 211         }
 212         return 0;
 213
 214 error_out:
 215         memset((char *) buf+actual, 0, size-actual);
 216         if (channel->read_error)
 217                 retval = (channel->read_error)(channel, block, count, buf,
 218                                                size, actual, retval);
 219         return retval;
 220 }
 221
 222 static errcode_t raw_write_blk(io_channel channel,
 223                                struct unix_private_data *data,
 224                                unsigned long long block,
 225                                int count, const void *bufv)
 226 {
 227         ssize_t         size;
 228         ext2_loff_t     location;
 229         int             actual = 0;
 230         errcode_t       retval;
 231         const unsigned char *buf = bufv;
 232
 233         if (count == 1)
 234                 size = channel->block_size;
 235         else {
 236                 if (count < 0)
 237                         size = -count;
 238                 else
 239                         size = count * channel->block_size;
 240         }
 241         data->io_stats.bytes_written += size;
 242
 243         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 244         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 245                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 246                 goto error_out;
 247         }
 248
 249         if ((data->align == 0) ||
 250             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 251                 actual = write(data->dev, buf, size);
 252                 if (actual != size) {
 253                 short_write:
 254                         retval = EXT2_ET_SHORT_WRITE;
 255                         goto error_out;
 256                 }
 257                 return 0;
 258         }
 259
 260 #ifdef ALIGN_DEBUG
 261         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 262                (unsigned long) size);
 263 #endif
 264         /*
 265          * The buffer or size which we're trying to write isn't aligned
 266          * to the O_DIRECT rules, so we need to do this the hard way...
 267          */
 268         while (size > 0) {
 269                 if (size < channel->block_size) {
 270                         actual = read(data->dev, data->bounce,
 271                                       channel->block_size);
 272                         if (actual != channel->block_size) {
 273                                 retval = EXT2_ET_SHORT_READ;
 274                                 goto error_out;
 275                         }
 276                 }
 277                 actual = size;
 278                 if (size > channel->block_size)
 279                         actual = channel->block_size;
 280                 memcpy(data->bounce, buf, actual);
 281                 actual = write(data->dev, data->bounce, channel->block_size);
 282                 if (actual != channel->block_size)
 283                         goto short_write;
 284                 size -= actual;
 285                 buf += actual;
 286         }
 287         return 0;
 288
 289 error_out:
 290         if (channel->write_error)
 291                 retval = (channel->write_error)(channel, block, count, buf,
 292                                                 size, actual, retval);
 293         return retval;
 294 }
 295
 296
 297 /*
 298  * Here we implement the cache functions
 299  */
 300
 301 /* Allocate the cache buffers */
 302 static errcode_t alloc_cache(io_channel channel,
 303                              struct unix_private_data *data)
 304 {
 305         errcode_t               retval;
 306         struct unix_cache       *cache;
 307         int                     i;
 308
 309         data->access_time = 0;
 310         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 311                 cache->block = 0;
 312                 cache->access_time = 0;
 313                 cache->dirty = 0;
 314                 cache->in_use = 0;
 315                 if (cache->buf)
 316                         ext2fs_free_mem(&cache->buf);
 317                 retval = ext2fs_get_memalign(channel->block_size,
 318                                              data->align, &cache->buf);
 319                 if (retval)
 320                         return retval;
 321         }
 322         if (data->align) {
 323                 if (data->bounce)
 324                         ext2fs_free_mem(&data->bounce);
 325                 retval = ext2fs_get_memalign(channel->block_size, data->align,
 326                                              &data->bounce);
 327         }
 328         return retval;
 329 }
 330
 331 /* Free the cache buffers */
 332 static void free_cache(struct unix_private_data *data)
 333 {
 334         struct unix_cache       *cache;
 335         int                     i;
 336
 337         data->access_time = 0;
 338         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 339                 cache->block = 0;
 340                 cache->access_time = 0;
 341                 cache->dirty = 0;
 342                 cache->in_use = 0;
 343                 if (cache->buf)
 344                         ext2fs_free_mem(&cache->buf);
 345         }
 346         if (data->bounce)
 347                 ext2fs_free_mem(&data->bounce);
 348 }
 349
 350 #ifndef NO_IO_CACHE
 351 /*
 352  * Try to find a block in the cache.  If the block is not found, and
 353  * eldest is a non-zero pointer, then fill in eldest with the cache
 354  * entry to that should be reused.
 355  */
 356 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 357                                             unsigned long long block,
 358                                             struct unix_cache **eldest)
 359 {
 360         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 361         int                     i;
 362
 363         unused_cache = oldest_cache = 0;
 364         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 365                 if (!cache->in_use) {
 366                         if (!unused_cache)
 367                                 unused_cache = cache;
 368                         continue;
 369                 }
 370                 if (cache->block == block) {
 371                         cache->access_time = ++data->access_time;
 372                         return cache;
 373                 }
 374                 if (!oldest_cache ||
 375                     (cache->access_time < oldest_cache->access_time))
 376                         oldest_cache = cache;
 377         }
 378         if (eldest)
 379                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 380         return 0;
 381 }
 382
 383 /*
 384  * Reuse a particular cache entry for another block.
 385  */
 386 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 387                  struct unix_cache *cache, unsigned long long block)
 388 {
 389         if (cache->dirty && cache->in_use)
 390                 raw_write_blk(channel, data, cache->block, 1, cache->buf);
 391
 392         cache->in_use = 1;
 393         cache->dirty = 0;
 394         cache->block = block;
 395         cache->access_time = ++data->access_time;
 396 }
 397
 398 /*
 399  * Flush all of the blocks in the cache
 400  */
 401 static errcode_t flush_cached_blocks(io_channel channel,
 402                                      struct unix_private_data *data,
 403                                      int invalidate)
 404
 405 {
 406         struct unix_cache       *cache;
 407         errcode_t               retval, retval2;
 408         int                     i;
 409
 410         retval2 = 0;
 411         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 412                 if (!cache->in_use)
 413                         continue;
 414
 415                 if (invalidate)
 416                         cache->in_use = 0;
 417
 418                 if (!cache->dirty)
 419                         continue;
 420
 421                 retval = raw_write_blk(channel, data,
 422                                        cache->block, 1, cache->buf);
 423                 if (retval)
 424                         retval2 = retval;
 425                 else
 426                         cache->dirty = 0;
 427         }
 428         return retval2;
 429 }
 430 #endif /* NO_IO_CACHE */
 431
 432 #ifdef __linux__
 433 #ifndef BLKDISCARDZEROES
 434 #define BLKDISCARDZEROES _IO(0x12,124)
 435 #endif
 436 #endif
 437
 438 static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 439 {
 440         io_channel      io = NULL;
 441         struct unix_private_data *data = NULL;
 442         errcode_t       retval;
 443         int             open_flags, zeroes = 0;
 444         ext2fs_struct_stat st;
 445 #ifdef __linux__
 446         struct          utsname ut;
 447 #endif
 448
 449         if (name == 0)
 450                 return EXT2_ET_BAD_DEVICE_NAME;
 451         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 452         if (retval)
 453                 return retval;
 454         memset(io, 0, sizeof(struct struct_io_channel));
 455         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 456         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 457         if (retval)
 458                 goto cleanup;
 459
 460         io->manager = unix_io_manager;
 461         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 462         if (retval)
 463                 goto cleanup;
 464
 465         strcpy(io->name, name);
 466         io->private_data = data;
 467         io->block_size = 1024;
 468         io->read_error = 0;
 469         io->write_error = 0;
 470         io->refcount = 1;
 471
 472         memset(data, 0, sizeof(struct unix_private_data));
 473         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 474         data->io_stats.num_fields = 2;
 475
 476         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 477         if (flags & IO_FLAG_EXCLUSIVE)
 478                 open_flags |= O_EXCL;
 479 #ifdef O_DIRECT
 480         if (flags & IO_FLAG_DIRECT_IO)
 481                 open_flags |= O_DIRECT;
 482 #endif
 483         data->flags = flags;
 484
 485         data->dev = ext2fs_open_file(io->name, open_flags);
 486         if (data->dev < 0) {
 487                 retval = errno;
 488                 goto cleanup;
 489         }
 490
 491 #ifdef BLKSSZGET
 492         if (flags & IO_FLAG_DIRECT_IO) {
 493                 if (ioctl(data->dev, BLKSSZGET, &data->align) != 0)
 494                         data->align = io->block_size;
 495         }
 496 #endif
 497
 498 #ifdef BLKDISCARDZEROES
 499         ioctl(data->dev, BLKDISCARDZEROES, &zeroes);
 500         if (zeroes)
 501                 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 502 #endif
 503
 504 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 505         /*
 506          * Some operating systems require that the buffers be aligned,
 507          * regardless of O_DIRECT
 508          */
 509         data->align = 512;
 510 #endif
 511
 512
 513         if ((retval = alloc_cache(io, data)))
 514                 goto cleanup;
 515
 516 #ifdef BLKROGET
 517         if (flags & IO_FLAG_RW) {
 518                 int error;
 519                 int readonly = 0;
 520
 521                 /* Is the block device actually writable? */
 522                 error = ioctl(data->dev, BLKROGET, &readonly);
 523                 if (!error && readonly) {
 524                         close(data->dev);
 525                         retval = EPERM;
 526                         goto cleanup;
 527                 }
 528         }
 529 #endif
 530
 531 #ifdef __linux__
 532 #undef RLIM_INFINITY
 533 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 534 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 535 #else
 536 #define RLIM_INFINITY  (~0UL)
 537 #endif
 538         /*
 539          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 540          * block devices are wrongly getting hit by the filesize
 541          * limit.  This workaround isn't perfect, since it won't work
 542          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 543          *
 544          */
 545         if ((flags & IO_FLAG_RW) &&
 546             (uname(&ut) == 0) &&
 547             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 548              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 549              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 550              (ut.release[5] < '8')) &&
 551             (ext2fs_stat(io->name, &st) == 0) &&
 552             (S_ISBLK(st.st_mode))) {
 553                 struct rlimit   rlim;
 554
 555                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 556                 setrlimit(RLIMIT_FSIZE, &rlim);
 557                 getrlimit(RLIMIT_FSIZE, &rlim);
 558                 if (((unsigned long) rlim.rlim_cur) <
 559                     ((unsigned long) rlim.rlim_max)) {
 560                         rlim.rlim_cur = rlim.rlim_max;
 561                         setrlimit(RLIMIT_FSIZE, &rlim);
 562                 }
 563         }
 564 #endif
 565         *channel = io;
 566         return 0;
 567
 568 cleanup:
 569         if (data) {
 570                 free_cache(data);
 571                 ext2fs_free_mem(&data);
 572         }
 573         if (io)
 574                 ext2fs_free_mem(&io);
 575         return retval;
 576 }
 577
 578 static errcode_t unix_close(io_channel channel)
 579 {
 580         struct unix_private_data *data;
 581         errcode_t       retval = 0;
 582
 583         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 584         data = (struct unix_private_data *) channel->private_data;
 585         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 586
 587         if (--channel->refcount > 0)
 588                 return 0;
 589
 590 #ifndef NO_IO_CACHE
 591         retval = flush_cached_blocks(channel, data, 0);
 592 #endif
 593
 594         if (close(data->dev) < 0)
 595                 retval = errno;
 596         free_cache(data);
 597
 598         ext2fs_free_mem(&channel->private_data);
 599         if (channel->name)
 600                 ext2fs_free_mem(&channel->name);
 601         ext2fs_free_mem(&channel);
 602         return retval;
 603 }
 604
 605 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 606 {
 607         struct unix_private_data *data;
 608         errcode_t               retval;
 609
 610         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 611         data = (struct unix_private_data *) channel->private_data;
 612         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 613
 614         if (channel->block_size != blksize) {
 615 #ifndef NO_IO_CACHE
 616                 if ((retval = flush_cached_blocks(channel, data, 0)))
 617                         return retval;
 618 #endif
 619
 620                 channel->block_size = blksize;
 621                 free_cache(data);
 622                 if ((retval = alloc_cache(channel, data)))
 623                         return retval;
 624         }
 625         return 0;
 626 }
 627
 628
 629 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 630                                int count, void *buf)
 631 {
 632         struct unix_private_data *data;
 633         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 634         errcode_t       retval;
 635         char            *cp;
 636         int             i, j;
 637
 638         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 639         data = (struct unix_private_data *) channel->private_data;
 640         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 641
 642 #ifdef NO_IO_CACHE
 643         return raw_read_blk(channel, data, block, count, buf);
 644 #else
 645         /*
 646          * If we're doing an odd-sized read or a very large read,
 647          * flush out the cache and then do a direct read.
 648          */
 649         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 650                 if ((retval = flush_cached_blocks(channel, data, 0)))
 651                         return retval;
 652                 return raw_read_blk(channel, data, block, count, buf);
 653         }
 654
 655         cp = buf;
 656         while (count > 0) {
 657                 /* If it's in the cache, use it! */
 658                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 659 #ifdef DEBUG
 660                         printf("Using cached block %lu\n", block);
 661 #endif
 662                         memcpy(cp, cache->buf, channel->block_size);
 663                         count--;
 664                         block++;
 665                         cp += channel->block_size;
 666                         continue;
 667                 }
 668                 if (count == 1) {
 669                         /*
 670                          * Special case where we read directly into the
 671                          * cache buffer; important in the O_DIRECT case
 672                          */
 673                         cache = reuse[0];
 674                         reuse_cache(channel, data, cache, block);
 675                         if ((retval = raw_read_blk(channel, data, block, 1,
 676                                                    cache->buf))) {
 677                                 cache->in_use = 0;
 678                                 return retval;
 679                         }
 680                         memcpy(cp, cache->buf, channel->block_size);
 681                         return 0;
 682                 }
 683
 684                 /*
 685                  * Find the number of uncached blocks so we can do a
 686                  * single read request
 687                  */
 688                 for (i=1; i < count; i++)
 689                         if (find_cached_block(data, block+i, &reuse[i]))
 690                                 break;
 691 #ifdef DEBUG
 692                 printf("Reading %d blocks starting at %lu\n", i, block);
 693 #endif
 694                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 695                         return retval;
 696
 697                 /* Save the results in the cache */
 698                 for (j=0; j < i; j++) {
 699                         count--;
 700                         cache = reuse[j];
 701                         reuse_cache(channel, data, cache, block++);
 702                         memcpy(cache->buf, cp, channel->block_size);
 703                         cp += channel->block_size;
 704                 }
 705         }
 706         return 0;
 707 #endif /* NO_IO_CACHE */
 708 }
 709
 710 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 711                                int count, void *buf)
 712 {
 713         return unix_read_blk64(channel, block, count, buf);
 714 }
 715
 716 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 717                                 int count, const void *buf)
 718 {
 719         struct unix_private_data *data;
 720         struct unix_cache *cache, *reuse;
 721         errcode_t       retval = 0;
 722         const char      *cp;
 723         int             writethrough;
 724
 725         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 726         data = (struct unix_private_data *) channel->private_data;
 727         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 728
 729 #ifdef NO_IO_CACHE
 730         return raw_write_blk(channel, data, block, count, buf);
 731 #else
 732         /*
 733          * If we're doing an odd-sized write or a very large write,
 734          * flush out the cache completely and then do a direct write.
 735          */
 736         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 737                 if ((retval = flush_cached_blocks(channel, data, 1)))
 738                         return retval;
 739                 return raw_write_blk(channel, data, block, count, buf);
 740         }
 741
 742         /*
 743          * For a moderate-sized multi-block write, first force a write
 744          * if we're in write-through cache mode, and then fill the
 745          * cache with the blocks.
 746          */
 747         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 748         if (writethrough)
 749                 retval = raw_write_blk(channel, data, block, count, buf);
 750
 751         cp = buf;
 752         while (count > 0) {
 753                 cache = find_cached_block(data, block, &reuse);
 754                 if (!cache) {
 755                         cache = reuse;
 756                         reuse_cache(channel, data, cache, block);
 757                 }
 758                 memcpy(cache->buf, cp, channel->block_size);
 759                 cache->dirty = !writethrough;
 760                 count--;
 761                 block++;
 762                 cp += channel->block_size;
 763         }
 764         return retval;
 765 #endif /* NO_IO_CACHE */
 766 }
 767
 768 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 769                                 int count, const void *buf)
 770 {
 771         return unix_write_blk64(channel, block, count, buf);
 772 }
 773
 774 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 775                                  int size, const void *buf)
 776 {
 777         struct unix_private_data *data;
 778         errcode_t       retval = 0;
 779         ssize_t         actual;
 780
 781         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 782         data = (struct unix_private_data *) channel->private_data;
 783         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 784
 785         if (data->align != 0) {
 786 #ifdef ALIGN_DEBUG
 787                 printf("unix_write_byte: O_DIRECT fallback\n");
 788 #endif
 789                 return EXT2_ET_UNIMPLEMENTED;
 790         }
 791
 792 #ifndef NO_IO_CACHE
 793         /*
 794          * Flush out the cache completely
 795          */
 796         if ((retval = flush_cached_blocks(channel, data, 1)))
 797                 return retval;
 798 #endif
 799
 800         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
 801                 return errno;
 802
 803         actual = write(data->dev, buf, size);
 804         if (actual != size)
 805                 return EXT2_ET_SHORT_WRITE;
 806
 807         return 0;
 808 }
 809
 810 /*
 811  * Flush data buffers to disk.
 812  */
 813 static errcode_t unix_flush(io_channel channel)
 814 {
 815         struct unix_private_data *data;
 816         errcode_t retval = 0;
 817
 818         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 819         data = (struct unix_private_data *) channel->private_data;
 820         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 821
 822 #ifndef NO_IO_CACHE
 823         retval = flush_cached_blocks(channel, data, 0);
 824 #endif
 825         fsync(data->dev);
 826         return retval;
 827 }
 828
 829 static errcode_t unix_set_option(io_channel channel, const char *option,
 830                                  const char *arg)
 831 {
 832         struct unix_private_data *data;
 833         unsigned long long tmp;
 834         char *end;
 835
 836         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 837         data = (struct unix_private_data *) channel->private_data;
 838         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 839
 840         if (!strcmp(option, "offset")) {
 841                 if (!arg)
 842                         return EXT2_ET_INVALID_ARGUMENT;
 843
 844                 tmp = strtoull(arg, &end, 0);
 845                 if (*end)
 846                         return EXT2_ET_INVALID_ARGUMENT;
 847                 data->offset = tmp;
 848                 if (data->offset < 0)
 849                         return EXT2_ET_INVALID_ARGUMENT;
 850                 return 0;
 851         }
 852         return EXT2_ET_INVALID_ARGUMENT;
 853 }
 854
 855 #if defined(__linux__) && !defined(BLKDISCARD)
 856 #define BLKDISCARD      _IO(0x12,119)
 857 #endif
 858
 859 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 860                               unsigned long long count)
 861 {
 862 #ifdef BLKDISCARD
 863         struct unix_private_data *data;
 864         __uint64_t      range[2];
 865         int             ret;
 866
 867         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 868         data = (struct unix_private_data *) channel->private_data;
 869         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 870
 871         range[0] = (__uint64_t)(block) * channel->block_size;
 872         range[1] = (__uint64_t)(count) * channel->block_size;
 873
 874         ret = ioctl(data->dev, BLKDISCARD, &range);
 875         if (ret < 0)
 876                 return errno;
 877         return 0;
 878 #else
 879         return EXT2_ET_UNIMPLEMENTED;
 880 #endif
 881 }