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 "config.h"
  25 #include <stdio.h>
  26 #include <string.h>
  27 #if HAVE_UNISTD_H
  28 #include <unistd.h>
  29 #endif
  30 #if HAVE_ERRNO_H
  31 #include <errno.h>
  32 #endif
  33 #include <fcntl.h>
  34 #include <time.h>
  35 #ifdef __linux__
  36 #include <sys/utsname.h>
  37 #endif
  38 #ifdef HAVE_SYS_IOCTL_H
  39 #include <sys/ioctl.h>
  40 #endif
  41 #ifdef HAVE_SYS_MOUNT_H
  42 #include <sys/mount.h>
  43 #endif
  44 #if HAVE_SYS_STAT_H
  45 #include <sys/stat.h>
  46 #endif
  47 #if HAVE_SYS_TYPES_H
  48 #include <sys/types.h>
  49 #endif
  50 #if HAVE_SYS_RESOURCE_H
  51 #include <sys/resource.h>
  52 #endif
  53 #if HAVE_LINUX_FALLOC_H
  54 #include <linux/falloc.h>
  55 #endif
  56
  57 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
  58 #define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
  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 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 ((channel->align == 0) ||
 180             (IS_ALIGNED(buf, channel->align) &&
 181              IS_ALIGNED(size, channel->align))) {
 182                 actual = read(data->dev, buf, size);
 183                 if (actual != size) {
 184                 short_read:
 185                         if (actual < 0)
 186                                 actual = 0;
 187                         retval = EXT2_ET_SHORT_READ;
 188                         goto error_out;
 189                 }
 190                 return 0;
 191         }
 192
 193 #ifdef ALIGN_DEBUG
 194         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 195                (unsigned long) size);
 196 #endif
 197
 198         /*
 199          * The buffer or size which we're trying to read isn't aligned
 200          * to the O_DIRECT rules, so we need to do this the hard way...
 201          */
 202         while (size > 0) {
 203                 actual = read(data->dev, data->bounce, channel->block_size);
 204                 if (actual != channel->block_size)
 205                         goto short_read;
 206                 actual = size;
 207                 if (size > channel->block_size)
 208                         actual = channel->block_size;
 209                 memcpy(buf, data->bounce, actual);
 210                 size -= actual;
 211                 buf += actual;
 212         }
 213         return 0;
 214
 215 error_out:
 216         memset((char *) buf+actual, 0, size-actual);
 217         if (channel->read_error)
 218                 retval = (channel->read_error)(channel, block, count, buf,
 219                                                size, actual, retval);
 220         return retval;
 221 }
 222
 223 static errcode_t raw_write_blk(io_channel channel,
 224                                struct unix_private_data *data,
 225                                unsigned long long block,
 226                                int count, const void *bufv)
 227 {
 228         ssize_t         size;
 229         ext2_loff_t     location;
 230         int             actual = 0;
 231         errcode_t       retval;
 232         const unsigned char *buf = bufv;
 233
 234         if (count == 1)
 235                 size = channel->block_size;
 236         else {
 237                 if (count < 0)
 238                         size = -count;
 239                 else
 240                         size = count * channel->block_size;
 241         }
 242         data->io_stats.bytes_written += size;
 243
 244         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 245         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 246                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 247                 goto error_out;
 248         }
 249
 250         if ((channel->align == 0) ||
 251             (IS_ALIGNED(buf, channel->align) &&
 252              IS_ALIGNED(size, channel->align))) {
 253                 actual = write(data->dev, buf, size);
 254                 if (actual != size) {
 255                 short_write:
 256                         retval = EXT2_ET_SHORT_WRITE;
 257                         goto error_out;
 258                 }
 259                 return 0;
 260         }
 261
 262 #ifdef ALIGN_DEBUG
 263         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 264                (unsigned long) size);
 265 #endif
 266         /*
 267          * The buffer or size which we're trying to write isn't aligned
 268          * to the O_DIRECT rules, so we need to do this the hard way...
 269          */
 270         while (size > 0) {
 271                 if (size < channel->block_size) {
 272                         actual = read(data->dev, data->bounce,
 273                                       channel->block_size);
 274                         if (actual != channel->block_size) {
 275                                 retval = EXT2_ET_SHORT_READ;
 276                                 goto error_out;
 277                         }
 278                 }
 279                 actual = size;
 280                 if (size > channel->block_size)
 281                         actual = channel->block_size;
 282                 memcpy(data->bounce, buf, actual);
 283                 actual = write(data->dev, data->bounce, channel->block_size);
 284                 if (actual != channel->block_size)
 285                         goto short_write;
 286                 size -= actual;
 287                 buf += actual;
 288         }
 289         return 0;
 290
 291 error_out:
 292         if (channel->write_error)
 293                 retval = (channel->write_error)(channel, block, count, buf,
 294                                                 size, actual, retval);
 295         return retval;
 296 }
 297
 298
 299 /*
 300  * Here we implement the cache functions
 301  */
 302
 303 /* Allocate the cache buffers */
 304 static errcode_t alloc_cache(io_channel channel,
 305                              struct unix_private_data *data)
 306 {
 307         errcode_t               retval;
 308         struct unix_cache       *cache;
 309         int                     i;
 310
 311         data->access_time = 0;
 312         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 313                 cache->block = 0;
 314                 cache->access_time = 0;
 315                 cache->dirty = 0;
 316                 cache->in_use = 0;
 317                 if (cache->buf)
 318                         ext2fs_free_mem(&cache->buf);
 319                 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
 320                 if (retval)
 321                         return retval;
 322         }
 323         if (channel->align) {
 324                 if (data->bounce)
 325                         ext2fs_free_mem(&data->bounce);
 326                 retval = io_channel_alloc_buf(channel, 0, &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 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
 439 {
 440         if (mode)
 441 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 442                 return open64(pathname, flags, mode);
 443         else
 444                 return open64(pathname, flags);
 445 #else
 446                 return open(pathname, flags, mode);
 447         else
 448                 return open(pathname, flags);
 449 #endif
 450 }
 451
 452 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
 453 {
 454 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 455         return stat64(path, buf);
 456 #else
 457         return stat(path, buf);
 458 #endif
 459 }
 460
 461 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
 462 {
 463 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 464         return fstat64(fd, buf);
 465 #else
 466         return fstat(fd, buf);
 467 #endif
 468 }
 469
 470 static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 471 {
 472         io_channel      io = NULL;
 473         struct unix_private_data *data = NULL;
 474         errcode_t       retval;
 475         int             open_flags, zeroes = 0;
 476         int             f_nocache = 0;
 477         ext2fs_struct_stat st;
 478 #ifdef __linux__
 479         struct          utsname ut;
 480 #endif
 481
 482         if (name == 0)
 483                 return EXT2_ET_BAD_DEVICE_NAME;
 484         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 485         if (retval)
 486                 goto cleanup;
 487         memset(io, 0, sizeof(struct struct_io_channel));
 488         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 489         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 490         if (retval)
 491                 goto cleanup;
 492
 493         io->manager = unix_io_manager;
 494         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 495         if (retval)
 496                 goto cleanup;
 497
 498         strcpy(io->name, name);
 499         io->private_data = data;
 500         io->block_size = 1024;
 501         io->read_error = 0;
 502         io->write_error = 0;
 503         io->refcount = 1;
 504
 505         memset(data, 0, sizeof(struct unix_private_data));
 506         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 507         data->io_stats.num_fields = 2;
 508
 509         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 510         if (flags & IO_FLAG_EXCLUSIVE)
 511                 open_flags |= O_EXCL;
 512 #if defined(O_DIRECT)
 513         if (flags & IO_FLAG_DIRECT_IO) {
 514                 open_flags |= O_DIRECT;
 515                 io->align = ext2fs_get_dio_alignment(data->dev);
 516         }
 517 #elif defined(F_NOCACHE)
 518         if (flags & IO_FLAG_DIRECT_IO) {
 519                 f_nocache = F_NOCACHE;
 520                 io->align = 4096;
 521         }
 522 #endif
 523         data->flags = flags;
 524
 525         data->dev = ext2fs_open_file(io->name, open_flags, 0);
 526         if (data->dev < 0) {
 527                 retval = errno;
 528                 goto cleanup;
 529         }
 530         if (f_nocache) {
 531                 if (fcntl(data->dev, f_nocache, 1) < 0) {
 532                         retval = errno;
 533                         goto cleanup;
 534                 }
 535         }
 536
 537         /*
 538          * If the device is really a block device, then set the
 539          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 540          * because we are going to use punch hole instead of discard
 541          * and if it succeed, subsequent read from sparse area returns
 542          * zero.
 543          */
 544         if (ext2fs_stat(io->name, &st) == 0) {
 545                 if (S_ISBLK(st.st_mode))
 546                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 547                 else
 548                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 549         }
 550
 551 #ifdef BLKDISCARDZEROES
 552         ioctl(data->dev, BLKDISCARDZEROES, &zeroes);
 553         if (zeroes)
 554                 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 555 #endif
 556
 557 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 558         /*
 559          * Some operating systems require that the buffers be aligned,
 560          * regardless of O_DIRECT
 561          */
 562         if (!io->align)
 563                 io->align = 512;
 564 #endif
 565
 566
 567         if ((retval = alloc_cache(io, data)))
 568                 goto cleanup;
 569
 570 #ifdef BLKROGET
 571         if (flags & IO_FLAG_RW) {
 572                 int error;
 573                 int readonly = 0;
 574
 575                 /* Is the block device actually writable? */
 576                 error = ioctl(data->dev, BLKROGET, &readonly);
 577                 if (!error && readonly) {
 578                         close(data->dev);
 579                         retval = EPERM;
 580                         goto cleanup;
 581                 }
 582         }
 583 #endif
 584
 585 #ifdef __linux__
 586 #undef RLIM_INFINITY
 587 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 588 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 589 #else
 590 #define RLIM_INFINITY  (~0UL)
 591 #endif
 592         /*
 593          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 594          * block devices are wrongly getting hit by the filesize
 595          * limit.  This workaround isn't perfect, since it won't work
 596          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 597          *
 598          */
 599         if ((flags & IO_FLAG_RW) &&
 600             (uname(&ut) == 0) &&
 601             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 602              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 603              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 604              (ut.release[5] < '8')) &&
 605             (ext2fs_stat(io->name, &st) == 0) &&
 606             (S_ISBLK(st.st_mode))) {
 607                 struct rlimit   rlim;
 608
 609                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 610                 setrlimit(RLIMIT_FSIZE, &rlim);
 611                 getrlimit(RLIMIT_FSIZE, &rlim);
 612                 if (((unsigned long) rlim.rlim_cur) <
 613                     ((unsigned long) rlim.rlim_max)) {
 614                         rlim.rlim_cur = rlim.rlim_max;
 615                         setrlimit(RLIMIT_FSIZE, &rlim);
 616                 }
 617         }
 618 #endif
 619         *channel = io;
 620         return 0;
 621
 622 cleanup:
 623         if (data) {
 624                 free_cache(data);
 625                 ext2fs_free_mem(&data);
 626         }
 627         if (io)
 628                 ext2fs_free_mem(&io);
 629         return retval;
 630 }
 631
 632 static errcode_t unix_close(io_channel channel)
 633 {
 634         struct unix_private_data *data;
 635         errcode_t       retval = 0;
 636
 637         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 638         data = (struct unix_private_data *) channel->private_data;
 639         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 640
 641         if (--channel->refcount > 0)
 642                 return 0;
 643
 644 #ifndef NO_IO_CACHE
 645         retval = flush_cached_blocks(channel, data, 0);
 646 #endif
 647
 648         if (close(data->dev) < 0)
 649                 retval = errno;
 650         free_cache(data);
 651
 652         ext2fs_free_mem(&channel->private_data);
 653         if (channel->name)
 654                 ext2fs_free_mem(&channel->name);
 655         ext2fs_free_mem(&channel);
 656         return retval;
 657 }
 658
 659 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 660 {
 661         struct unix_private_data *data;
 662         errcode_t               retval;
 663
 664         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 665         data = (struct unix_private_data *) channel->private_data;
 666         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 667
 668         if (channel->block_size != blksize) {
 669 #ifndef NO_IO_CACHE
 670                 if ((retval = flush_cached_blocks(channel, data, 0)))
 671                         return retval;
 672 #endif
 673
 674                 channel->block_size = blksize;
 675                 free_cache(data);
 676                 if ((retval = alloc_cache(channel, data)))
 677                         return retval;
 678         }
 679         return 0;
 680 }
 681
 682
 683 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 684                                int count, void *buf)
 685 {
 686         struct unix_private_data *data;
 687         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 688         errcode_t       retval;
 689         char            *cp;
 690         int             i, j;
 691
 692         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 693         data = (struct unix_private_data *) channel->private_data;
 694         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 695
 696 #ifdef NO_IO_CACHE
 697         return raw_read_blk(channel, data, block, count, buf);
 698 #else
 699         /*
 700          * If we're doing an odd-sized read or a very large read,
 701          * flush out the cache and then do a direct read.
 702          */
 703         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 704                 if ((retval = flush_cached_blocks(channel, data, 0)))
 705                         return retval;
 706                 return raw_read_blk(channel, data, block, count, buf);
 707         }
 708
 709         cp = buf;
 710         while (count > 0) {
 711                 /* If it's in the cache, use it! */
 712                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 713 #ifdef DEBUG
 714                         printf("Using cached block %lu\n", block);
 715 #endif
 716                         memcpy(cp, cache->buf, channel->block_size);
 717                         count--;
 718                         block++;
 719                         cp += channel->block_size;
 720                         continue;
 721                 }
 722                 if (count == 1) {
 723                         /*
 724                          * Special case where we read directly into the
 725                          * cache buffer; important in the O_DIRECT case
 726                          */
 727                         cache = reuse[0];
 728                         reuse_cache(channel, data, cache, block);
 729                         if ((retval = raw_read_blk(channel, data, block, 1,
 730                                                    cache->buf))) {
 731                                 cache->in_use = 0;
 732                                 return retval;
 733                         }
 734                         memcpy(cp, cache->buf, channel->block_size);
 735                         return 0;
 736                 }
 737
 738                 /*
 739                  * Find the number of uncached blocks so we can do a
 740                  * single read request
 741                  */
 742                 for (i=1; i < count; i++)
 743                         if (find_cached_block(data, block+i, &reuse[i]))
 744                                 break;
 745 #ifdef DEBUG
 746                 printf("Reading %d blocks starting at %lu\n", i, block);
 747 #endif
 748                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 749                         return retval;
 750
 751                 /* Save the results in the cache */
 752                 for (j=0; j < i; j++) {
 753                         count--;
 754                         cache = reuse[j];
 755                         reuse_cache(channel, data, cache, block++);
 756                         memcpy(cache->buf, cp, channel->block_size);
 757                         cp += channel->block_size;
 758                 }
 759         }
 760         return 0;
 761 #endif /* NO_IO_CACHE */
 762 }
 763
 764 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 765                                int count, void *buf)
 766 {
 767         return unix_read_blk64(channel, block, count, buf);
 768 }
 769
 770 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 771                                 int count, const void *buf)
 772 {
 773         struct unix_private_data *data;
 774         struct unix_cache *cache, *reuse;
 775         errcode_t       retval = 0;
 776         const char      *cp;
 777         int             writethrough;
 778
 779         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 780         data = (struct unix_private_data *) channel->private_data;
 781         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 782
 783 #ifdef NO_IO_CACHE
 784         return raw_write_blk(channel, data, block, count, buf);
 785 #else
 786         /*
 787          * If we're doing an odd-sized write or a very large write,
 788          * flush out the cache completely and then do a direct write.
 789          */
 790         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 791                 if ((retval = flush_cached_blocks(channel, data, 1)))
 792                         return retval;
 793                 return raw_write_blk(channel, data, block, count, buf);
 794         }
 795
 796         /*
 797          * For a moderate-sized multi-block write, first force a write
 798          * if we're in write-through cache mode, and then fill the
 799          * cache with the blocks.
 800          */
 801         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 802         if (writethrough)
 803                 retval = raw_write_blk(channel, data, block, count, buf);
 804
 805         cp = buf;
 806         while (count > 0) {
 807                 cache = find_cached_block(data, block, &reuse);
 808                 if (!cache) {
 809                         cache = reuse;
 810                         reuse_cache(channel, data, cache, block);
 811                 }
 812                 memcpy(cache->buf, cp, channel->block_size);
 813                 cache->dirty = !writethrough;
 814                 count--;
 815                 block++;
 816                 cp += channel->block_size;
 817         }
 818         return retval;
 819 #endif /* NO_IO_CACHE */
 820 }
 821
 822 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 823                                 int count, const void *buf)
 824 {
 825         return unix_write_blk64(channel, block, count, buf);
 826 }
 827
 828 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 829                                  int size, const void *buf)
 830 {
 831         struct unix_private_data *data;
 832         errcode_t       retval = 0;
 833         ssize_t         actual;
 834
 835         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 836         data = (struct unix_private_data *) channel->private_data;
 837         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 838
 839         if (channel->align != 0) {
 840 #ifdef ALIGN_DEBUG
 841                 printf("unix_write_byte: O_DIRECT fallback\n");
 842 #endif
 843                 return EXT2_ET_UNIMPLEMENTED;
 844         }
 845
 846 #ifndef NO_IO_CACHE
 847         /*
 848          * Flush out the cache completely
 849          */
 850         if ((retval = flush_cached_blocks(channel, data, 1)))
 851                 return retval;
 852 #endif
 853
 854         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
 855                 return errno;
 856
 857         actual = write(data->dev, buf, size);
 858         if (actual != size)
 859                 return EXT2_ET_SHORT_WRITE;
 860
 861         return 0;
 862 }
 863
 864 /*
 865  * Flush data buffers to disk.
 866  */
 867 static errcode_t unix_flush(io_channel channel)
 868 {
 869         struct unix_private_data *data;
 870         errcode_t retval = 0;
 871
 872         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 873         data = (struct unix_private_data *) channel->private_data;
 874         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 875
 876 #ifndef NO_IO_CACHE
 877         retval = flush_cached_blocks(channel, data, 0);
 878 #endif
 879         fsync(data->dev);
 880         return retval;
 881 }
 882
 883 static errcode_t unix_set_option(io_channel channel, const char *option,
 884                                  const char *arg)
 885 {
 886         struct unix_private_data *data;
 887         unsigned long long tmp;
 888         char *end;
 889
 890         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 891         data = (struct unix_private_data *) channel->private_data;
 892         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 893
 894         if (!strcmp(option, "offset")) {
 895                 if (!arg)
 896                         return EXT2_ET_INVALID_ARGUMENT;
 897
 898                 tmp = strtoull(arg, &end, 0);
 899                 if (*end)
 900                         return EXT2_ET_INVALID_ARGUMENT;
 901                 data->offset = tmp;
 902                 if (data->offset < 0)
 903                         return EXT2_ET_INVALID_ARGUMENT;
 904                 return 0;
 905         }
 906         return EXT2_ET_INVALID_ARGUMENT;
 907 }
 908
 909 #if defined(__linux__) && !defined(BLKDISCARD)
 910 #define BLKDISCARD              _IO(0x12,119)
 911 #endif
 912
 913 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 914                               unsigned long long count)
 915 {
 916         struct unix_private_data *data;
 917         __uint64_t      range[2];
 918         int             ret;
 919
 920         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 921         data = (struct unix_private_data *) channel->private_data;
 922         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 923
 924         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 925 #ifdef BLKDISCARD
 926                 range[0] = (__uint64_t)(block) * channel->block_size;
 927                 range[1] = (__uint64_t)(count) * channel->block_size;
 928
 929                 ret = ioctl(data->dev, BLKDISCARD, &range);
 930 #else
 931                 goto unimplemented;
 932 #endif
 933         } else {
 934 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
 935                 /*
 936                  * If we are not on block device, try to use punch hole
 937                  * to reclaim free space.
 938                  */
 939                 ret = fallocate(data->dev,
 940                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 941                                 (off_t)(block) * channel->block_size,
 942                                 (off_t)(count) * channel->block_size);
 943 #else
 944                 goto unimplemented;
 945 #endif
 946         }
 947         if (ret < 0) {
 948                 if (errno == EOPNOTSUPP)
 949                         goto unimplemented;
 950                 return errno;
 951         }
 952         return 0;
 953 unimplemented:
 954         return EXT2_ET_UNIMPLEMENTED;
 955 }