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;
 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         data->dev = -1;
 509
 510         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 511         if (flags & IO_FLAG_EXCLUSIVE)
 512                 open_flags |= O_EXCL;
 513 #if defined(O_DIRECT)
 514         if (flags & IO_FLAG_DIRECT_IO) {
 515                 open_flags |= O_DIRECT;
 516                 io->align = ext2fs_get_dio_alignment(data->dev);
 517         }
 518 #elif defined(F_NOCACHE)
 519         if (flags & IO_FLAG_DIRECT_IO) {
 520                 f_nocache = F_NOCACHE;
 521                 io->align = 4096;
 522         }
 523 #endif
 524         data->flags = flags;
 525
 526         data->dev = ext2fs_open_file(io->name, open_flags, 0);
 527         if (data->dev < 0) {
 528                 retval = errno;
 529                 goto cleanup;
 530         }
 531         if (f_nocache) {
 532                 if (fcntl(data->dev, f_nocache, 1) < 0) {
 533                         retval = errno;
 534                         goto cleanup;
 535                 }
 536         }
 537
 538         /*
 539          * If the device is really a block device, then set the
 540          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 541          * because we are going to use punch hole instead of discard
 542          * and if it succeed, subsequent read from sparse area returns
 543          * zero.
 544          */
 545         if (ext2fs_stat(io->name, &st) == 0) {
 546                 if (S_ISBLK(st.st_mode))
 547                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 548                 else
 549                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 550         }
 551
 552 #ifdef BLKDISCARDZEROES
 553         {
 554                 int zeroes = 0;
 555                 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 556                     zeroes)
 557                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 558         }
 559 #endif
 560
 561 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 562         /*
 563          * Some operating systems require that the buffers be aligned,
 564          * regardless of O_DIRECT
 565          */
 566         if (!io->align)
 567                 io->align = 512;
 568 #endif
 569
 570
 571         if ((retval = alloc_cache(io, data)))
 572                 goto cleanup;
 573
 574 #ifdef BLKROGET
 575         if (flags & IO_FLAG_RW) {
 576                 int error;
 577                 int readonly = 0;
 578
 579                 /* Is the block device actually writable? */
 580                 error = ioctl(data->dev, BLKROGET, &readonly);
 581                 if (!error && readonly) {
 582                         retval = EPERM;
 583                         goto cleanup;
 584                 }
 585         }
 586 #endif
 587
 588 #ifdef __linux__
 589 #undef RLIM_INFINITY
 590 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 591 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 592 #else
 593 #define RLIM_INFINITY  (~0UL)
 594 #endif
 595         /*
 596          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 597          * block devices are wrongly getting hit by the filesize
 598          * limit.  This workaround isn't perfect, since it won't work
 599          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 600          *
 601          */
 602         if ((flags & IO_FLAG_RW) &&
 603             (uname(&ut) == 0) &&
 604             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 605              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 606              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 607              (ut.release[5] < '8')) &&
 608             (ext2fs_stat(io->name, &st) == 0) &&
 609             (S_ISBLK(st.st_mode))) {
 610                 struct rlimit   rlim;
 611
 612                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 613                 setrlimit(RLIMIT_FSIZE, &rlim);
 614                 getrlimit(RLIMIT_FSIZE, &rlim);
 615                 if (((unsigned long) rlim.rlim_cur) <
 616                     ((unsigned long) rlim.rlim_max)) {
 617                         rlim.rlim_cur = rlim.rlim_max;
 618                         setrlimit(RLIMIT_FSIZE, &rlim);
 619                 }
 620         }
 621 #endif
 622         *channel = io;
 623         return 0;
 624
 625 cleanup:
 626         if (data) {
 627                 if (data->dev >= 0)
 628                         close(data->dev);
 629                 free_cache(data);
 630                 ext2fs_free_mem(&data);
 631         }
 632         if (io) {
 633                 if (io->name) {
 634                         ext2fs_free_mem(&io->name);
 635                 }
 636                 ext2fs_free_mem(&io);
 637         }
 638         return retval;
 639 }
 640
 641 static errcode_t unix_close(io_channel channel)
 642 {
 643         struct unix_private_data *data;
 644         errcode_t       retval = 0;
 645
 646         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 647         data = (struct unix_private_data *) channel->private_data;
 648         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 649
 650         if (--channel->refcount > 0)
 651                 return 0;
 652
 653 #ifndef NO_IO_CACHE
 654         retval = flush_cached_blocks(channel, data, 0);
 655 #endif
 656
 657         if (close(data->dev) < 0)
 658                 retval = errno;
 659         free_cache(data);
 660
 661         ext2fs_free_mem(&channel->private_data);
 662         if (channel->name)
 663                 ext2fs_free_mem(&channel->name);
 664         ext2fs_free_mem(&channel);
 665         return retval;
 666 }
 667
 668 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 669 {
 670         struct unix_private_data *data;
 671         errcode_t               retval;
 672
 673         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 674         data = (struct unix_private_data *) channel->private_data;
 675         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 676
 677         if (channel->block_size != blksize) {
 678 #ifndef NO_IO_CACHE
 679                 if ((retval = flush_cached_blocks(channel, data, 0)))
 680                         return retval;
 681 #endif
 682
 683                 channel->block_size = blksize;
 684                 free_cache(data);
 685                 if ((retval = alloc_cache(channel, data)))
 686                         return retval;
 687         }
 688         return 0;
 689 }
 690
 691
 692 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 693                                int count, void *buf)
 694 {
 695         struct unix_private_data *data;
 696         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 697         errcode_t       retval;
 698         char            *cp;
 699         int             i, j;
 700
 701         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 702         data = (struct unix_private_data *) channel->private_data;
 703         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 704
 705 #ifdef NO_IO_CACHE
 706         return raw_read_blk(channel, data, block, count, buf);
 707 #else
 708         /*
 709          * If we're doing an odd-sized read or a very large read,
 710          * flush out the cache and then do a direct read.
 711          */
 712         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 713                 if ((retval = flush_cached_blocks(channel, data, 0)))
 714                         return retval;
 715                 return raw_read_blk(channel, data, block, count, buf);
 716         }
 717
 718         cp = buf;
 719         while (count > 0) {
 720                 /* If it's in the cache, use it! */
 721                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 722 #ifdef DEBUG
 723                         printf("Using cached block %lu\n", block);
 724 #endif
 725                         memcpy(cp, cache->buf, channel->block_size);
 726                         count--;
 727                         block++;
 728                         cp += channel->block_size;
 729                         continue;
 730                 }
 731                 if (count == 1) {
 732                         /*
 733                          * Special case where we read directly into the
 734                          * cache buffer; important in the O_DIRECT case
 735                          */
 736                         cache = reuse[0];
 737                         reuse_cache(channel, data, cache, block);
 738                         if ((retval = raw_read_blk(channel, data, block, 1,
 739                                                    cache->buf))) {
 740                                 cache->in_use = 0;
 741                                 return retval;
 742                         }
 743                         memcpy(cp, cache->buf, channel->block_size);
 744                         return 0;
 745                 }
 746
 747                 /*
 748                  * Find the number of uncached blocks so we can do a
 749                  * single read request
 750                  */
 751                 for (i=1; i < count; i++)
 752                         if (find_cached_block(data, block+i, &reuse[i]))
 753                                 break;
 754 #ifdef DEBUG
 755                 printf("Reading %d blocks starting at %lu\n", i, block);
 756 #endif
 757                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 758                         return retval;
 759
 760                 /* Save the results in the cache */
 761                 for (j=0; j < i; j++) {
 762                         count--;
 763                         cache = reuse[j];
 764                         reuse_cache(channel, data, cache, block++);
 765                         memcpy(cache->buf, cp, channel->block_size);
 766                         cp += channel->block_size;
 767                 }
 768         }
 769         return 0;
 770 #endif /* NO_IO_CACHE */
 771 }
 772
 773 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 774                                int count, void *buf)
 775 {
 776         return unix_read_blk64(channel, block, count, buf);
 777 }
 778
 779 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 780                                 int count, const void *buf)
 781 {
 782         struct unix_private_data *data;
 783         struct unix_cache *cache, *reuse;
 784         errcode_t       retval = 0;
 785         const char      *cp;
 786         int             writethrough;
 787
 788         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 789         data = (struct unix_private_data *) channel->private_data;
 790         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 791
 792 #ifdef NO_IO_CACHE
 793         return raw_write_blk(channel, data, block, count, buf);
 794 #else
 795         /*
 796          * If we're doing an odd-sized write or a very large write,
 797          * flush out the cache completely and then do a direct write.
 798          */
 799         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 800                 if ((retval = flush_cached_blocks(channel, data, 1)))
 801                         return retval;
 802                 return raw_write_blk(channel, data, block, count, buf);
 803         }
 804
 805         /*
 806          * For a moderate-sized multi-block write, first force a write
 807          * if we're in write-through cache mode, and then fill the
 808          * cache with the blocks.
 809          */
 810         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 811         if (writethrough)
 812                 retval = raw_write_blk(channel, data, block, count, buf);
 813
 814         cp = buf;
 815         while (count > 0) {
 816                 cache = find_cached_block(data, block, &reuse);
 817                 if (!cache) {
 818                         cache = reuse;
 819                         reuse_cache(channel, data, cache, block);
 820                 }
 821                 memcpy(cache->buf, cp, channel->block_size);
 822                 cache->dirty = !writethrough;
 823                 count--;
 824                 block++;
 825                 cp += channel->block_size;
 826         }
 827         return retval;
 828 #endif /* NO_IO_CACHE */
 829 }
 830
 831 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 832                                 int count, const void *buf)
 833 {
 834         return unix_write_blk64(channel, block, count, buf);
 835 }
 836
 837 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 838                                  int size, const void *buf)
 839 {
 840         struct unix_private_data *data;
 841         errcode_t       retval = 0;
 842         ssize_t         actual;
 843
 844         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 845         data = (struct unix_private_data *) channel->private_data;
 846         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 847
 848         if (channel->align != 0) {
 849 #ifdef ALIGN_DEBUG
 850                 printf("unix_write_byte: O_DIRECT fallback\n");
 851 #endif
 852                 return EXT2_ET_UNIMPLEMENTED;
 853         }
 854
 855 #ifndef NO_IO_CACHE
 856         /*
 857          * Flush out the cache completely
 858          */
 859         if ((retval = flush_cached_blocks(channel, data, 1)))
 860                 return retval;
 861 #endif
 862
 863         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
 864                 return errno;
 865
 866         actual = write(data->dev, buf, size);
 867         if (actual != size)
 868                 return EXT2_ET_SHORT_WRITE;
 869
 870         return 0;
 871 }
 872
 873 /*
 874  * Flush data buffers to disk.
 875  */
 876 static errcode_t unix_flush(io_channel channel)
 877 {
 878         struct unix_private_data *data;
 879         errcode_t retval = 0;
 880
 881         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 882         data = (struct unix_private_data *) channel->private_data;
 883         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 884
 885 #ifndef NO_IO_CACHE
 886         retval = flush_cached_blocks(channel, data, 0);
 887 #endif
 888         fsync(data->dev);
 889         return retval;
 890 }
 891
 892 static errcode_t unix_set_option(io_channel channel, const char *option,
 893                                  const char *arg)
 894 {
 895         struct unix_private_data *data;
 896         unsigned long long tmp;
 897         char *end;
 898
 899         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 900         data = (struct unix_private_data *) channel->private_data;
 901         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 902
 903         if (!strcmp(option, "offset")) {
 904                 if (!arg)
 905                         return EXT2_ET_INVALID_ARGUMENT;
 906
 907                 tmp = strtoull(arg, &end, 0);
 908                 if (*end)
 909                         return EXT2_ET_INVALID_ARGUMENT;
 910                 data->offset = tmp;
 911                 if (data->offset < 0)
 912                         return EXT2_ET_INVALID_ARGUMENT;
 913                 return 0;
 914         }
 915         return EXT2_ET_INVALID_ARGUMENT;
 916 }
 917
 918 #if defined(__linux__) && !defined(BLKDISCARD)
 919 #define BLKDISCARD              _IO(0x12,119)
 920 #endif
 921
 922 static errcode_t unix_discard(io_channel channel, unsigned long long block,
 923                               unsigned long long count)
 924 {
 925         struct unix_private_data *data;
 926         int             ret;
 927
 928         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 929         data = (struct unix_private_data *) channel->private_data;
 930         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 931
 932         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 933 #ifdef BLKDISCARD
 934                 __uint64_t range[2];
 935
 936                 range[0] = (__uint64_t)(block) * channel->block_size;
 937                 range[1] = (__uint64_t)(count) * channel->block_size;
 938
 939                 ret = ioctl(data->dev, BLKDISCARD, &range);
 940 #else
 941                 goto unimplemented;
 942 #endif
 943         } else {
 944 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
 945                 /*
 946                  * If we are not on block device, try to use punch hole
 947                  * to reclaim free space.
 948                  */
 949                 ret = fallocate(data->dev,
 950                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 951                                 (off_t)(block) * channel->block_size,
 952                                 (off_t)(count) * channel->block_size);
 953 #else
 954                 goto unimplemented;
 955 #endif
 956         }
 957         if (ret < 0) {
 958                 if (errno == EOPNOTSUPP)
 959                         goto unimplemented;
 960                 return errno;
 961         }
 962         return 0;
 963 unimplemented:
 964         return EXT2_ET_UNIMPLEMENTED;
 965 }