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 #if defined(__linux__) && defined(_IO) && !defined(BLKSSZGET)
  62 #define BLKSSZGET  _IO(0x12,104)/* get block device sector size */
  63 #endif
  64
  65 #undef ALIGN_DEBUG
  66
  67 #include "ext2_fs.h"
  68 #include "ext2fs.h"
  69
  70 /*
  71  * For checking structure magic numbers...
  72  */
  73
  74 #define EXT2_CHECK_MAGIC(struct, code) \
  75           if ((struct)->magic != (code)) return (code)
  76
  77 struct unix_cache {
  78         char            *buf;
  79         unsigned long   block;
  80         int             access_time;
  81         unsigned        dirty:1;
  82         unsigned        in_use:1;
  83 };
  84
  85 #define CACHE_SIZE 8
  86 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
  87 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
  88
  89 struct unix_private_data {
  90         int     magic;
  91         int     dev;
  92         int     flags;
  93         int     align;
  94         int     access_time;
  95         ext2_loff_t offset;
  96         struct unix_cache cache[CACHE_SIZE];
  97         void    *bounce;
  98         struct struct_io_stats io_stats;
  99 };
 100
 101 #define IS_ALIGNED(n, align) ((((unsigned long) n) & \
 102                                ((unsigned long) ((align)-1))) == 0)
 103
 104 static errcode_t unix_open(const char *name, int flags, io_channel *channel);
 105 static errcode_t unix_close(io_channel channel);
 106 static errcode_t unix_set_blksize(io_channel channel, int blksize);
 107 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 108                                int count, void *data);
 109 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 110                                 int count, const void *data);
 111 static errcode_t unix_flush(io_channel channel);
 112 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 113                                 int size, const void *data);
 114 static errcode_t unix_set_option(io_channel channel, const char *option,
 115                                  const char *arg);
 116 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 117 ;
 118 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 119                  struct unix_cache *cache, unsigned long long block);
 120 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 121                                int count, void *data);
 122 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 123                                 int count, const void *data);
 124 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 125                               unsigned long long count);
 126
 127 static struct struct_io_manager struct_unix_manager = {
 128         EXT2_ET_MAGIC_IO_MANAGER,
 129         "Unix I/O Manager",
 130         unix_open,
 131         unix_close,
 132         unix_set_blksize,
 133         unix_read_blk,
 134         unix_write_blk,
 135         unix_flush,
 136         unix_write_byte,
 137         unix_set_option,
 138         unix_get_stats,
 139         unix_read_blk64,
 140         unix_write_blk64,
 141         unix_discard,
 142 };
 143
 144 io_manager unix_io_manager = &struct_unix_manager;
 145
 146 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 147 {
 148         errcode_t       retval = 0;
 149
 150         struct unix_private_data *data;
 151
 152         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 153         data = (struct unix_private_data *) channel->private_data;
 154         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 155
 156         if (stats)
 157                 *stats = &data->io_stats;
 158
 159         return retval;
 160 }
 161
 162 /*
 163  * Here are the raw I/O functions
 164  */
 165 static errcode_t raw_read_blk(io_channel channel,
 166                               struct unix_private_data *data,
 167                               unsigned long long block,
 168                               int count, void *bufv)
 169 {
 170         errcode_t       retval;
 171         ssize_t         size;
 172         ext2_loff_t     location;
 173         int             actual = 0;
 174         unsigned char   *buf = bufv;
 175
 176         size = (count < 0) ? -count : count * channel->block_size;
 177         data->io_stats.bytes_read += size;
 178         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 179         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 180                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 181                 goto error_out;
 182         }
 183         if ((data->align == 0) ||
 184             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 185                 actual = read(data->dev, buf, size);
 186                 if (actual != size) {
 187                 short_read:
 188                         if (actual < 0)
 189                                 actual = 0;
 190                         retval = EXT2_ET_SHORT_READ;
 191                         goto error_out;
 192                 }
 193                 return 0;
 194         }
 195
 196 #ifdef ALIGN_DEBUG
 197         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 198                (unsigned long) size);
 199 #endif
 200
 201         /*
 202          * The buffer or size which we're trying to read isn't aligned
 203          * to the O_DIRECT rules, so we need to do this the hard way...
 204          */
 205         while (size > 0) {
 206                 actual = read(data->dev, data->bounce, channel->block_size);
 207                 if (actual != channel->block_size)
 208                         goto short_read;
 209                 actual = size;
 210                 if (size > channel->block_size)
 211                         actual = channel->block_size;
 212                 memcpy(buf, data->bounce, actual);
 213                 size -= actual;
 214                 buf += actual;
 215         }
 216         return 0;
 217
 218 error_out:
 219         memset((char *) buf+actual, 0, size-actual);
 220         if (channel->read_error)
 221                 retval = (channel->read_error)(channel, block, count, buf,
 222                                                size, actual, retval);
 223         return retval;
 224 }
 225
 226 static errcode_t raw_write_blk(io_channel channel,
 227                                struct unix_private_data *data,
 228                                unsigned long long block,
 229                                int count, const void *bufv)
 230 {
 231         ssize_t         size;
 232         ext2_loff_t     location;
 233         int             actual = 0;
 234         errcode_t       retval;
 235         const unsigned char *buf = bufv;
 236
 237         if (count == 1)
 238                 size = channel->block_size;
 239         else {
 240                 if (count < 0)
 241                         size = -count;
 242                 else
 243                         size = count * channel->block_size;
 244         }
 245         data->io_stats.bytes_written += size;
 246
 247         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 248         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 249                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 250                 goto error_out;
 251         }
 252
 253         if ((data->align == 0) ||
 254             ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
 255                 actual = write(data->dev, buf, size);
 256                 if (actual != size) {
 257                 short_write:
 258                         retval = EXT2_ET_SHORT_WRITE;
 259                         goto error_out;
 260                 }
 261                 return 0;
 262         }
 263
 264 #ifdef ALIGN_DEBUG
 265         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 266                (unsigned long) size);
 267 #endif
 268         /*
 269          * The buffer or size which we're trying to write isn't aligned
 270          * to the O_DIRECT rules, so we need to do this the hard way...
 271          */
 272         while (size > 0) {
 273                 if (size < channel->block_size) {
 274                         actual = read(data->dev, data->bounce,
 275                                       channel->block_size);
 276                         if (actual != channel->block_size) {
 277                                 retval = EXT2_ET_SHORT_READ;
 278                                 goto error_out;
 279                         }
 280                 }
 281                 actual = size;
 282                 if (size > channel->block_size)
 283                         actual = channel->block_size;
 284                 memcpy(data->bounce, buf, actual);
 285                 actual = write(data->dev, data->bounce, channel->block_size);
 286                 if (actual != channel->block_size)
 287                         goto short_write;
 288                 size -= actual;
 289                 buf += actual;
 290         }
 291         return 0;
 292
 293 error_out:
 294         if (channel->write_error)
 295                 retval = (channel->write_error)(channel, block, count, buf,
 296                                                 size, actual, retval);
 297         return retval;
 298 }
 299
 300
 301 /*
 302  * Here we implement the cache functions
 303  */
 304
 305 /* Allocate the cache buffers */
 306 static errcode_t alloc_cache(io_channel channel,
 307                              struct unix_private_data *data)
 308 {
 309         errcode_t               retval;
 310         struct unix_cache       *cache;
 311         int                     i;
 312
 313         data->access_time = 0;
 314         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 315                 cache->block = 0;
 316                 cache->access_time = 0;
 317                 cache->dirty = 0;
 318                 cache->in_use = 0;
 319                 if (cache->buf)
 320                         ext2fs_free_mem(&cache->buf);
 321                 retval = ext2fs_get_memalign(channel->block_size,
 322                                              data->align, &cache->buf);
 323                 if (retval)
 324                         return retval;
 325         }
 326         if (data->align) {
 327                 if (data->bounce)
 328                         ext2fs_free_mem(&data->bounce);
 329                 retval = ext2fs_get_memalign(channel->block_size, data->align,
 330                                              &data->bounce);
 331         }
 332         return retval;
 333 }
 334
 335 /* Free the cache buffers */
 336 static void free_cache(struct unix_private_data *data)
 337 {
 338         struct unix_cache       *cache;
 339         int                     i;
 340
 341         data->access_time = 0;
 342         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 343                 cache->block = 0;
 344                 cache->access_time = 0;
 345                 cache->dirty = 0;
 346                 cache->in_use = 0;
 347                 if (cache->buf)
 348                         ext2fs_free_mem(&cache->buf);
 349         }
 350         if (data->bounce)
 351                 ext2fs_free_mem(&data->bounce);
 352 }
 353
 354 #ifndef NO_IO_CACHE
 355 /*
 356  * Try to find a block in the cache.  If the block is not found, and
 357  * eldest is a non-zero pointer, then fill in eldest with the cache
 358  * entry to that should be reused.
 359  */
 360 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 361                                             unsigned long long block,
 362                                             struct unix_cache **eldest)
 363 {
 364         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 365         int                     i;
 366
 367         unused_cache = oldest_cache = 0;
 368         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 369                 if (!cache->in_use) {
 370                         if (!unused_cache)
 371                                 unused_cache = cache;
 372                         continue;
 373                 }
 374                 if (cache->block == block) {
 375                         cache->access_time = ++data->access_time;
 376                         return cache;
 377                 }
 378                 if (!oldest_cache ||
 379                     (cache->access_time < oldest_cache->access_time))
 380                         oldest_cache = cache;
 381         }
 382         if (eldest)
 383                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 384         return 0;
 385 }
 386
 387 /*
 388  * Reuse a particular cache entry for another block.
 389  */
 390 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 391                  struct unix_cache *cache, unsigned long long block)
 392 {
 393         if (cache->dirty && cache->in_use)
 394                 raw_write_blk(channel, data, cache->block, 1, cache->buf);
 395
 396         cache->in_use = 1;
 397         cache->dirty = 0;
 398         cache->block = block;
 399         cache->access_time = ++data->access_time;
 400 }
 401
 402 /*
 403  * Flush all of the blocks in the cache
 404  */
 405 static errcode_t flush_cached_blocks(io_channel channel,
 406                                      struct unix_private_data *data,
 407                                      int invalidate)
 408
 409 {
 410         struct unix_cache       *cache;
 411         errcode_t               retval, retval2;
 412         int                     i;
 413
 414         retval2 = 0;
 415         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 416                 if (!cache->in_use)
 417                         continue;
 418
 419                 if (invalidate)
 420                         cache->in_use = 0;
 421
 422                 if (!cache->dirty)
 423                         continue;
 424
 425                 retval = raw_write_blk(channel, data,
 426                                        cache->block, 1, cache->buf);
 427                 if (retval)
 428                         retval2 = retval;
 429                 else
 430                         cache->dirty = 0;
 431         }
 432         return retval2;
 433 }
 434 #endif /* NO_IO_CACHE */
 435
 436 #ifdef __linux__
 437 #ifndef BLKDISCARDZEROES
 438 #define BLKDISCARDZEROES _IO(0x12,124)
 439 #endif
 440 #endif
 441
 442 static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 443 {
 444         io_channel      io = NULL;
 445         struct unix_private_data *data = NULL;
 446         errcode_t       retval;
 447         int             open_flags, zeroes = 0;
 448         int             f_nocache = 0;
 449         ext2fs_struct_stat st;
 450 #ifdef __linux__
 451         struct          utsname ut;
 452 #endif
 453
 454         if (name == 0)
 455                 return EXT2_ET_BAD_DEVICE_NAME;
 456         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 457         if (retval)
 458                 goto cleanup;
 459         memset(io, 0, sizeof(struct struct_io_channel));
 460         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 461         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 462         if (retval)
 463                 goto cleanup;
 464
 465         io->manager = unix_io_manager;
 466         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 467         if (retval)
 468                 goto cleanup;
 469
 470         strcpy(io->name, name);
 471         io->private_data = data;
 472         io->block_size = 1024;
 473         io->read_error = 0;
 474         io->write_error = 0;
 475         io->refcount = 1;
 476
 477         memset(data, 0, sizeof(struct unix_private_data));
 478         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 479         data->io_stats.num_fields = 2;
 480
 481         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 482         if (flags & IO_FLAG_EXCLUSIVE)
 483                 open_flags |= O_EXCL;
 484 #if defined(O_DIRECT)
 485         if (flags & IO_FLAG_DIRECT_IO)
 486                 open_flags |= O_DIRECT;
 487 #elif defined(F_NOCACHE)
 488         if (flags & IO_FLAG_DIRECT_IO)
 489                 f_nocache = F_NOCACHE;
 490 #endif
 491         data->flags = flags;
 492
 493         data->dev = ext2fs_open_file(io->name, open_flags, 0);
 494         if (data->dev < 0) {
 495                 retval = errno;
 496                 goto cleanup;
 497         }
 498         if (f_nocache) {
 499                 if (fcntl(data->dev, f_nocache, 1) < 0) {
 500                         retval = errno;
 501                         goto cleanup;
 502                 }
 503         }
 504
 505         /*
 506          * If the device is really a block device, then set the
 507          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 508          * because we are going to use punch hole instead of discard
 509          * and if it succeed, subsequent read from sparse area returns
 510          * zero.
 511          */
 512         if (ext2fs_stat(io->name, &st) == 0) {
 513                 if (S_ISBLK(st.st_mode))
 514                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 515                 else
 516                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 517         }
 518
 519 #ifdef BLKSSZGET
 520         if (flags & IO_FLAG_DIRECT_IO) {
 521                 if (ioctl(data->dev, BLKSSZGET, &data->align) != 0)
 522                         data->align = io->block_size;
 523         }
 524 #endif
 525
 526 #ifdef BLKDISCARDZEROES
 527         ioctl(data->dev, BLKDISCARDZEROES, &zeroes);
 528         if (zeroes)
 529                 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 530 #endif
 531
 532 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 533         /*
 534          * Some operating systems require that the buffers be aligned,
 535          * regardless of O_DIRECT
 536          */
 537         data->align = 512;
 538 #endif
 539
 540
 541         if ((retval = alloc_cache(io, data)))
 542                 goto cleanup;
 543
 544 #ifdef BLKROGET
 545         if (flags & IO_FLAG_RW) {
 546                 int error;
 547                 int readonly = 0;
 548
 549                 /* Is the block device actually writable? */
 550                 error = ioctl(data->dev, BLKROGET, &readonly);
 551                 if (!error && readonly) {
 552                         close(data->dev);
 553                         retval = EPERM;
 554                         goto cleanup;
 555                 }
 556         }
 557 #endif
 558
 559 #ifdef __linux__
 560 #undef RLIM_INFINITY
 561 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 562 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 563 #else
 564 #define RLIM_INFINITY  (~0UL)
 565 #endif
 566         /*
 567          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 568          * block devices are wrongly getting hit by the filesize
 569          * limit.  This workaround isn't perfect, since it won't work
 570          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 571          *
 572          */
 573         if ((flags & IO_FLAG_RW) &&
 574             (uname(&ut) == 0) &&
 575             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 576              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 577              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 578              (ut.release[5] < '8')) &&
 579             (ext2fs_stat(io->name, &st) == 0) &&
 580             (S_ISBLK(st.st_mode))) {
 581                 struct rlimit   rlim;
 582
 583                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 584                 setrlimit(RLIMIT_FSIZE, &rlim);
 585                 getrlimit(RLIMIT_FSIZE, &rlim);
 586                 if (((unsigned long) rlim.rlim_cur) <
 587                     ((unsigned long) rlim.rlim_max)) {
 588                         rlim.rlim_cur = rlim.rlim_max;
 589                         setrlimit(RLIMIT_FSIZE, &rlim);
 590                 }
 591         }
 592 #endif
 593         *channel = io;
 594         return 0;
 595
 596 cleanup:
 597         if (data) {
 598                 free_cache(data);
 599                 ext2fs_free_mem(&data);
 600         }
 601         if (io)
 602                 ext2fs_free_mem(&io);
 603         return retval;
 604 }
 605
 606 static errcode_t unix_close(io_channel channel)
 607 {
 608         struct unix_private_data *data;
 609         errcode_t       retval = 0;
 610
 611         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 612         data = (struct unix_private_data *) channel->private_data;
 613         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 614
 615         if (--channel->refcount > 0)
 616                 return 0;
 617
 618 #ifndef NO_IO_CACHE
 619         retval = flush_cached_blocks(channel, data, 0);
 620 #endif
 621
 622         if (close(data->dev) < 0)
 623                 retval = errno;
 624         free_cache(data);
 625
 626         ext2fs_free_mem(&channel->private_data);
 627         if (channel->name)
 628                 ext2fs_free_mem(&channel->name);
 629         ext2fs_free_mem(&channel);
 630         return retval;
 631 }
 632
 633 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 634 {
 635         struct unix_private_data *data;
 636         errcode_t               retval;
 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         if (channel->block_size != blksize) {
 643 #ifndef NO_IO_CACHE
 644                 if ((retval = flush_cached_blocks(channel, data, 0)))
 645                         return retval;
 646 #endif
 647
 648                 channel->block_size = blksize;
 649                 free_cache(data);
 650                 if ((retval = alloc_cache(channel, data)))
 651                         return retval;
 652         }
 653         return 0;
 654 }
 655
 656
 657 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 658                                int count, void *buf)
 659 {
 660         struct unix_private_data *data;
 661         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 662         errcode_t       retval;
 663         char            *cp;
 664         int             i, j;
 665
 666         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 667         data = (struct unix_private_data *) channel->private_data;
 668         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 669
 670 #ifdef NO_IO_CACHE
 671         return raw_read_blk(channel, data, block, count, buf);
 672 #else
 673         /*
 674          * If we're doing an odd-sized read or a very large read,
 675          * flush out the cache and then do a direct read.
 676          */
 677         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 678                 if ((retval = flush_cached_blocks(channel, data, 0)))
 679                         return retval;
 680                 return raw_read_blk(channel, data, block, count, buf);
 681         }
 682
 683         cp = buf;
 684         while (count > 0) {
 685                 /* If it's in the cache, use it! */
 686                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 687 #ifdef DEBUG
 688                         printf("Using cached block %lu\n", block);
 689 #endif
 690                         memcpy(cp, cache->buf, channel->block_size);
 691                         count--;
 692                         block++;
 693                         cp += channel->block_size;
 694                         continue;
 695                 }
 696                 if (count == 1) {
 697                         /*
 698                          * Special case where we read directly into the
 699                          * cache buffer; important in the O_DIRECT case
 700                          */
 701                         cache = reuse[0];
 702                         reuse_cache(channel, data, cache, block);
 703                         if ((retval = raw_read_blk(channel, data, block, 1,
 704                                                    cache->buf))) {
 705                                 cache->in_use = 0;
 706                                 return retval;
 707                         }
 708                         memcpy(cp, cache->buf, channel->block_size);
 709                         return 0;
 710                 }
 711
 712                 /*
 713                  * Find the number of uncached blocks so we can do a
 714                  * single read request
 715                  */
 716                 for (i=1; i < count; i++)
 717                         if (find_cached_block(data, block+i, &reuse[i]))
 718                                 break;
 719 #ifdef DEBUG
 720                 printf("Reading %d blocks starting at %lu\n", i, block);
 721 #endif
 722                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 723                         return retval;
 724
 725                 /* Save the results in the cache */
 726                 for (j=0; j < i; j++) {
 727                         count--;
 728                         cache = reuse[j];
 729                         reuse_cache(channel, data, cache, block++);
 730                         memcpy(cache->buf, cp, channel->block_size);
 731                         cp += channel->block_size;
 732                 }
 733         }
 734         return 0;
 735 #endif /* NO_IO_CACHE */
 736 }
 737
 738 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 739                                int count, void *buf)
 740 {
 741         return unix_read_blk64(channel, block, count, buf);
 742 }
 743
 744 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 745                                 int count, const void *buf)
 746 {
 747         struct unix_private_data *data;
 748         struct unix_cache *cache, *reuse;
 749         errcode_t       retval = 0;
 750         const char      *cp;
 751         int             writethrough;
 752
 753         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 754         data = (struct unix_private_data *) channel->private_data;
 755         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 756
 757 #ifdef NO_IO_CACHE
 758         return raw_write_blk(channel, data, block, count, buf);
 759 #else
 760         /*
 761          * If we're doing an odd-sized write or a very large write,
 762          * flush out the cache completely and then do a direct write.
 763          */
 764         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 765                 if ((retval = flush_cached_blocks(channel, data, 1)))
 766                         return retval;
 767                 return raw_write_blk(channel, data, block, count, buf);
 768         }
 769
 770         /*
 771          * For a moderate-sized multi-block write, first force a write
 772          * if we're in write-through cache mode, and then fill the
 773          * cache with the blocks.
 774          */
 775         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 776         if (writethrough)
 777                 retval = raw_write_blk(channel, data, block, count, buf);
 778
 779         cp = buf;
 780         while (count > 0) {
 781                 cache = find_cached_block(data, block, &reuse);
 782                 if (!cache) {
 783                         cache = reuse;
 784                         reuse_cache(channel, data, cache, block);
 785                 }
 786                 memcpy(cache->buf, cp, channel->block_size);
 787                 cache->dirty = !writethrough;
 788                 count--;
 789                 block++;
 790                 cp += channel->block_size;
 791         }
 792         return retval;
 793 #endif /* NO_IO_CACHE */
 794 }
 795
 796 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 797                                 int count, const void *buf)
 798 {
 799         return unix_write_blk64(channel, block, count, buf);
 800 }
 801
 802 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 803                                  int size, const void *buf)
 804 {
 805         struct unix_private_data *data;
 806         errcode_t       retval = 0;
 807         ssize_t         actual;
 808
 809         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 810         data = (struct unix_private_data *) channel->private_data;
 811         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 812
 813         if (data->align != 0) {
 814 #ifdef ALIGN_DEBUG
 815                 printf("unix_write_byte: O_DIRECT fallback\n");
 816 #endif
 817                 return EXT2_ET_UNIMPLEMENTED;
 818         }
 819
 820 #ifndef NO_IO_CACHE
 821         /*
 822          * Flush out the cache completely
 823          */
 824         if ((retval = flush_cached_blocks(channel, data, 1)))
 825                 return retval;
 826 #endif
 827
 828         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
 829                 return errno;
 830
 831         actual = write(data->dev, buf, size);
 832         if (actual != size)
 833                 return EXT2_ET_SHORT_WRITE;
 834
 835         return 0;
 836 }
 837
 838 /*
 839  * Flush data buffers to disk.
 840  */
 841 static errcode_t unix_flush(io_channel channel)
 842 {
 843         struct unix_private_data *data;
 844         errcode_t retval = 0;
 845
 846         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 847         data = (struct unix_private_data *) channel->private_data;
 848         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 849
 850 #ifndef NO_IO_CACHE
 851         retval = flush_cached_blocks(channel, data, 0);
 852 #endif
 853         fsync(data->dev);
 854         return retval;
 855 }
 856
 857 static errcode_t unix_set_option(io_channel channel, const char *option,
 858                                  const char *arg)
 859 {
 860         struct unix_private_data *data;
 861         unsigned long long tmp;
 862         char *end;
 863
 864         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 865         data = (struct unix_private_data *) channel->private_data;
 866         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 867
 868         if (!strcmp(option, "offset")) {
 869                 if (!arg)
 870                         return EXT2_ET_INVALID_ARGUMENT;
 871
 872                 tmp = strtoull(arg, &end, 0);
 873                 if (*end)
 874                         return EXT2_ET_INVALID_ARGUMENT;
 875                 data->offset = tmp;
 876                 if (data->offset < 0)
 877                         return EXT2_ET_INVALID_ARGUMENT;
 878                 return 0;
 879         }
 880         return EXT2_ET_INVALID_ARGUMENT;
 881 }
 882
 883 #if defined(__linux__) && !defined(BLKDISCARD)
 884 #define BLKDISCARD              _IO(0x12,119)
 885 #endif
 886
 887 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 888                               unsigned long long count)
 889 {
 890         struct unix_private_data *data;
 891         __uint64_t      range[2];
 892         int             ret;
 893
 894         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 895         data = (struct unix_private_data *) channel->private_data;
 896         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 897
 898         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 899 #ifdef BLKDISCARD
 900                 range[0] = (__uint64_t)(block) * channel->block_size;
 901                 range[1] = (__uint64_t)(count) * channel->block_size;
 902
 903                 ret = ioctl(data->dev, BLKDISCARD, &range);
 904 #else
 905                 goto unimplemented;
 906 #endif
 907         } else {
 908 #ifdef FALLOC_FL_PUNCH_HOLE
 909                 /*
 910                  * If we are not on block device, try to use punch hole
 911                  * to reclaim free space.
 912                  */
 913                 ret = fallocate(data->dev,
 914                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 915                                 (off_t)(block) * channel->block_size,
 916                                 (off_t)(count) * channel->block_size);
 917 #else
 918                 goto unimplemented;
 919 #endif
 920         }
 921         if (ret < 0) {
 922                 if (errno == EOPNOTSUPP)
 923                         goto unimplemented;
 924                 return errno;
 925         }
 926         return 0;
 927 unimplemented:
 928         return EXT2_ET_UNIMPLEMENTED;
 929 }