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 #if !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
  19 #define _XOPEN_SOURCE 600
  20 #define _DARWIN_C_SOURCE
  21 #define _FILE_OFFSET_BITS 64
  22 #ifndef _LARGEFILE_SOURCE
  23 #define _LARGEFILE_SOURCE
  24 #endif
  25 #ifndef _LARGEFILE64_SOURCE
  26 #define _LARGEFILE64_SOURCE
  27 #endif
  28 #ifndef _GNU_SOURCE
  29 #define _GNU_SOURCE
  30 #endif
  31 #endif
  32
  33 #include "config.h"
  34 #include <stdio.h>
  35 #include <string.h>
  36 #if HAVE_UNISTD_H
  37 #include <unistd.h>
  38 #endif
  39 #if HAVE_ERRNO_H
  40 #include <errno.h>
  41 #endif
  42 #include <fcntl.h>
  43 #include <time.h>
  44 #ifdef __linux__
  45 #include <sys/utsname.h>
  46 #endif
  47 #if HAVE_SYS_TYPES_H
  48 #include <sys/types.h>
  49 #endif
  50 #ifdef HAVE_SYS_IOCTL_H
  51 #include <sys/ioctl.h>
  52 #endif
  53 #ifdef HAVE_SYS_MOUNT_H
  54 #include <sys/mount.h>
  55 #endif
  56 #ifdef HAVE_SYS_PRCTL_H
  57 #include <sys/prctl.h>
  58 #else
  59 #define PR_GET_DUMPABLE 3
  60 #endif
  61 #if HAVE_SYS_STAT_H
  62 #include <sys/stat.h>
  63 #endif
  64 #if HAVE_SYS_RESOURCE_H
  65 #include <sys/resource.h>
  66 #endif
  67 #if HAVE_LINUX_FALLOC_H
  68 #include <linux/falloc.h>
  69 #endif
  70
  71 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
  72 #define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
  73 #endif
  74
  75 #undef ALIGN_DEBUG
  76
  77 #include "ext2_fs.h"
  78 #include "ext2fs.h"
  79
  80 /*
  81  * For checking structure magic numbers...
  82  */
  83
  84 #define EXT2_CHECK_MAGIC(struct, code) \
  85           if ((struct)->magic != (code)) return (code)
  86
  87 struct unix_cache {
  88         char                    *buf;
  89         unsigned long long      block;
  90         int                     access_time;
  91         unsigned                dirty:1;
  92         unsigned                in_use:1;
  93 };
  94
  95 #define CACHE_SIZE 8
  96 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
  97 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
  98
  99 struct unix_private_data {
 100         int     magic;
 101         int     dev;
 102         int     flags;
 103         int     align;
 104         int     access_time;
 105         ext2_loff_t offset;
 106         struct unix_cache cache[CACHE_SIZE];
 107         void    *bounce;
 108         struct struct_io_stats io_stats;
 109 };
 110
 111 #define IS_ALIGNED(n, align) ((((uintptr_t) n) & \
 112                                ((uintptr_t) ((align)-1))) == 0)
 113
 114 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 115 {
 116         errcode_t       retval = 0;
 117
 118         struct unix_private_data *data;
 119
 120         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 121         data = (struct unix_private_data *) channel->private_data;
 122         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 123
 124         if (stats)
 125                 *stats = &data->io_stats;
 126
 127         return retval;
 128 }
 129
 130 static char *safe_getenv(const char *arg)
 131 {
 132         if ((getuid() != geteuid()) || (getgid() != getegid()))
 133                 return NULL;
 134 #ifdef HAVE_PRCTL
 135         if (prctl(PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 136                 return NULL;
 137 #else
 138 #if (defined(linux) && defined(SYS_prctl))
 139         if (syscall(SYS_prctl, PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 140                 return NULL;
 141 #endif
 142 #endif
 143
 144 #if defined(HAVE_SECURE_GETENV)
 145         return secure_getenv(arg);
 146 #elif defined(HAVE___SECURE_GETENV)
 147         return __secure_getenv(arg);
 148 #else
 149         return getenv(arg);
 150 #endif
 151 }
 152
 153 /*
 154  * Here are the raw I/O functions
 155  */
 156 static errcode_t raw_read_blk(io_channel channel,
 157                               struct unix_private_data *data,
 158                               unsigned long long block,
 159                               int count, void *bufv)
 160 {
 161         errcode_t       retval;
 162         ssize_t         size;
 163         ext2_loff_t     location;
 164         int             actual = 0;
 165         unsigned char   *buf = bufv;
 166         ssize_t         really_read = 0;
 167
 168         size = (count < 0) ? -count : count * channel->block_size;
 169         data->io_stats.bytes_read += size;
 170         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 171
 172         if (data->flags & IO_FLAG_FORCE_BOUNCE) {
 173                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 174                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 175                         goto error_out;
 176                 }
 177                 goto bounce_read;
 178         }
 179
 180 #ifdef HAVE_PREAD64
 181         /* Try an aligned pread */
 182         if ((channel->align == 0) ||
 183             (IS_ALIGNED(buf, channel->align) &&
 184              IS_ALIGNED(size, channel->align))) {
 185                 actual = pread64(data->dev, buf, size, location);
 186                 if (actual == size)
 187                         return 0;
 188         }
 189 #elif HAVE_PREAD
 190         /* Try an aligned pread */
 191         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 192             ((channel->align == 0) ||
 193              (IS_ALIGNED(buf, channel->align) &&
 194               IS_ALIGNED(size, channel->align)))) {
 195                 actual = pread(data->dev, buf, size, location);
 196                 if (actual == size)
 197                         return 0;
 198         }
 199 #endif /* HAVE_PREAD */
 200
 201         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 202                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 203                 goto error_out;
 204         }
 205         if ((channel->align == 0) ||
 206             (IS_ALIGNED(buf, channel->align) &&
 207              IS_ALIGNED(size, channel->align))) {
 208                 actual = read(data->dev, buf, size);
 209                 if (actual != size) {
 210                 short_read:
 211                         if (actual < 0) {
 212                                 retval = errno;
 213                                 actual = 0;
 214                         } else
 215                                 retval = EXT2_ET_SHORT_READ;
 216                         goto error_out;
 217                 }
 218                 return 0;
 219         }
 220
 221 #ifdef ALIGN_DEBUG
 222         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 223                (unsigned long) size);
 224 #endif
 225
 226         /*
 227          * The buffer or size which we're trying to read isn't aligned
 228          * to the O_DIRECT rules, so we need to do this the hard way...
 229          */
 230 bounce_read:
 231         while (size > 0) {
 232                 actual = read(data->dev, data->bounce, channel->block_size);
 233                 if (actual != channel->block_size) {
 234                         actual = really_read;
 235                         buf -= really_read;
 236                         size += really_read;
 237                         goto short_read;
 238                 }
 239                 actual = size;
 240                 if (size > channel->block_size)
 241                         actual = channel->block_size;
 242                 memcpy(buf, data->bounce, actual);
 243                 really_read += actual;
 244                 size -= actual;
 245                 buf += actual;
 246         }
 247         return 0;
 248
 249 error_out:
 250         memset((char *) buf+actual, 0, size-actual);
 251         if (channel->read_error)
 252                 retval = (channel->read_error)(channel, block, count, buf,
 253                                                size, actual, retval);
 254         return retval;
 255 }
 256
 257 static errcode_t raw_write_blk(io_channel channel,
 258                                struct unix_private_data *data,
 259                                unsigned long long block,
 260                                int count, const void *bufv)
 261 {
 262         ssize_t         size;
 263         ext2_loff_t     location;
 264         int             actual = 0;
 265         errcode_t       retval;
 266         const unsigned char *buf = bufv;
 267
 268         if (count == 1)
 269                 size = channel->block_size;
 270         else {
 271                 if (count < 0)
 272                         size = -count;
 273                 else
 274                         size = count * channel->block_size;
 275         }
 276         data->io_stats.bytes_written += size;
 277
 278         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 279
 280         if (data->flags & IO_FLAG_FORCE_BOUNCE) {
 281                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 282                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 283                         goto error_out;
 284                 }
 285                 goto bounce_write;
 286         }
 287
 288 #ifdef HAVE_PWRITE64
 289         /* Try an aligned pwrite */
 290         if ((channel->align == 0) ||
 291             (IS_ALIGNED(buf, channel->align) &&
 292              IS_ALIGNED(size, channel->align))) {
 293                 actual = pwrite64(data->dev, buf, size, location);
 294                 if (actual == size)
 295                         return 0;
 296         }
 297 #elif HAVE_PWRITE
 298         /* Try an aligned pwrite */
 299         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 300             ((channel->align == 0) ||
 301              (IS_ALIGNED(buf, channel->align) &&
 302               IS_ALIGNED(size, channel->align)))) {
 303                 actual = pwrite(data->dev, buf, size, location);
 304                 if (actual == size)
 305                         return 0;
 306         }
 307 #endif /* HAVE_PWRITE */
 308
 309         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 310                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 311                 goto error_out;
 312         }
 313
 314         if ((channel->align == 0) ||
 315             (IS_ALIGNED(buf, channel->align) &&
 316              IS_ALIGNED(size, channel->align))) {
 317                 actual = write(data->dev, buf, size);
 318                 if (actual < 0) {
 319                         retval = errno;
 320                         goto error_out;
 321                 }
 322                 if (actual != size) {
 323                 short_write:
 324                         retval = EXT2_ET_SHORT_WRITE;
 325                         goto error_out;
 326                 }
 327                 return 0;
 328         }
 329
 330 #ifdef ALIGN_DEBUG
 331         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 332                (unsigned long) size);
 333 #endif
 334         /*
 335          * The buffer or size which we're trying to write isn't aligned
 336          * to the O_DIRECT rules, so we need to do this the hard way...
 337          */
 338 bounce_write:
 339         while (size > 0) {
 340                 if (size < channel->block_size) {
 341                         actual = read(data->dev, data->bounce,
 342                                       channel->block_size);
 343                         if (actual != channel->block_size) {
 344                                 if (actual < 0) {
 345                                         retval = errno;
 346                                         goto error_out;
 347                                 }
 348                                 memset(data->bounce + actual, 0,
 349                                        channel->block_size - actual);
 350                         }
 351                 }
 352                 actual = size;
 353                 if (size > channel->block_size)
 354                         actual = channel->block_size;
 355                 memcpy(data->bounce, buf, actual);
 356                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 357                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 358                         goto error_out;
 359                 }
 360                 actual = write(data->dev, data->bounce, channel->block_size);
 361                 if (actual < 0) {
 362                         retval = errno;
 363                         goto error_out;
 364                 }
 365                 if (actual != channel->block_size)
 366                         goto short_write;
 367                 size -= actual;
 368                 buf += actual;
 369                 location += actual;
 370         }
 371         return 0;
 372
 373 error_out:
 374         if (channel->write_error)
 375                 retval = (channel->write_error)(channel, block, count, buf,
 376                                                 size, actual, retval);
 377         return retval;
 378 }
 379
 380
 381 /*
 382  * Here we implement the cache functions
 383  */
 384
 385 /* Allocate the cache buffers */
 386 static errcode_t alloc_cache(io_channel channel,
 387                              struct unix_private_data *data)
 388 {
 389         errcode_t               retval;
 390         struct unix_cache       *cache;
 391         int                     i;
 392
 393         data->access_time = 0;
 394         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 395                 cache->block = 0;
 396                 cache->access_time = 0;
 397                 cache->dirty = 0;
 398                 cache->in_use = 0;
 399                 if (cache->buf)
 400                         ext2fs_free_mem(&cache->buf);
 401                 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
 402                 if (retval)
 403                         return retval;
 404         }
 405         if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
 406                 if (data->bounce)
 407                         ext2fs_free_mem(&data->bounce);
 408                 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
 409         }
 410         return retval;
 411 }
 412
 413 /* Free the cache buffers */
 414 static void free_cache(struct unix_private_data *data)
 415 {
 416         struct unix_cache       *cache;
 417         int                     i;
 418
 419         data->access_time = 0;
 420         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 421                 cache->block = 0;
 422                 cache->access_time = 0;
 423                 cache->dirty = 0;
 424                 cache->in_use = 0;
 425                 if (cache->buf)
 426                         ext2fs_free_mem(&cache->buf);
 427         }
 428         if (data->bounce)
 429                 ext2fs_free_mem(&data->bounce);
 430 }
 431
 432 #ifndef NO_IO_CACHE
 433 /*
 434  * Try to find a block in the cache.  If the block is not found, and
 435  * eldest is a non-zero pointer, then fill in eldest with the cache
 436  * entry to that should be reused.
 437  */
 438 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 439                                             unsigned long long block,
 440                                             struct unix_cache **eldest)
 441 {
 442         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 443         int                     i;
 444
 445         unused_cache = oldest_cache = 0;
 446         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 447                 if (!cache->in_use) {
 448                         if (!unused_cache)
 449                                 unused_cache = cache;
 450                         continue;
 451                 }
 452                 if (cache->block == block) {
 453                         cache->access_time = ++data->access_time;
 454                         return cache;
 455                 }
 456                 if (!oldest_cache ||
 457                     (cache->access_time < oldest_cache->access_time))
 458                         oldest_cache = cache;
 459         }
 460         if (eldest)
 461                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 462         return 0;
 463 }
 464
 465 /*
 466  * Reuse a particular cache entry for another block.
 467  */
 468 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 469                  struct unix_cache *cache, unsigned long long block)
 470 {
 471         if (cache->dirty && cache->in_use)
 472                 raw_write_blk(channel, data, cache->block, 1, cache->buf);
 473
 474         cache->in_use = 1;
 475         cache->dirty = 0;
 476         cache->block = block;
 477         cache->access_time = ++data->access_time;
 478 }
 479
 480 /*
 481  * Flush all of the blocks in the cache
 482  */
 483 static errcode_t flush_cached_blocks(io_channel channel,
 484                                      struct unix_private_data *data,
 485                                      int invalidate)
 486
 487 {
 488         struct unix_cache       *cache;
 489         errcode_t               retval, retval2;
 490         int                     i;
 491
 492         retval2 = 0;
 493         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 494                 if (!cache->in_use)
 495                         continue;
 496
 497                 if (invalidate)
 498                         cache->in_use = 0;
 499
 500                 if (!cache->dirty)
 501                         continue;
 502
 503                 retval = raw_write_blk(channel, data,
 504                                        cache->block, 1, cache->buf);
 505                 if (retval)
 506                         retval2 = retval;
 507                 else
 508                         cache->dirty = 0;
 509         }
 510         return retval2;
 511 }
 512 #endif /* NO_IO_CACHE */
 513
 514 #ifdef __linux__
 515 #ifndef BLKDISCARDZEROES
 516 #define BLKDISCARDZEROES _IO(0x12,124)
 517 #endif
 518 #endif
 519
 520 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
 521 {
 522         if (mode)
 523 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 524                 return open64(pathname, flags, mode);
 525         else
 526                 return open64(pathname, flags);
 527 #else
 528                 return open(pathname, flags, mode);
 529         else
 530                 return open(pathname, flags);
 531 #endif
 532 }
 533
 534 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
 535 {
 536 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 537         return stat64(path, buf);
 538 #else
 539         return stat(path, buf);
 540 #endif
 541 }
 542
 543 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
 544 {
 545 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 546         return fstat64(fd, buf);
 547 #else
 548         return fstat(fd, buf);
 549 #endif
 550 }
 551
 552
 553 static errcode_t unix_open_channel(const char *name, int fd,
 554                                    int flags, io_channel *channel,
 555                                    io_manager io_mgr)
 556 {
 557         io_channel      io = NULL;
 558         struct unix_private_data *data = NULL;
 559         errcode_t       retval;
 560         ext2fs_struct_stat st;
 561 #ifdef __linux__
 562         struct          utsname ut;
 563 #endif
 564
 565         if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
 566                 flags |= IO_FLAG_FORCE_BOUNCE;
 567
 568         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 569         if (retval)
 570                 goto cleanup;
 571         memset(io, 0, sizeof(struct struct_io_channel));
 572         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 573         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 574         if (retval)
 575                 goto cleanup;
 576
 577         io->manager = io_mgr;
 578         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 579         if (retval)
 580                 goto cleanup;
 581
 582         strcpy(io->name, name);
 583         io->private_data = data;
 584         io->block_size = 1024;
 585         io->read_error = 0;
 586         io->write_error = 0;
 587         io->refcount = 1;
 588
 589         memset(data, 0, sizeof(struct unix_private_data));
 590         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 591         data->io_stats.num_fields = 2;
 592         data->flags = flags;
 593         data->dev = fd;
 594
 595 #if defined(O_DIRECT)
 596         if (flags & IO_FLAG_DIRECT_IO)
 597                 io->align = ext2fs_get_dio_alignment(data->dev);
 598 #elif defined(F_NOCACHE)
 599         if (flags & IO_FLAG_DIRECT_IO)
 600                 io->align = 4096;
 601 #endif
 602
 603         /*
 604          * If the device is really a block device, then set the
 605          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 606          * because we are going to use punch hole instead of discard
 607          * and if it succeed, subsequent read from sparse area returns
 608          * zero.
 609          */
 610         if (ext2fs_fstat(data->dev, &st) == 0) {
 611                 if (S_ISBLK(st.st_mode))
 612                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 613                 else
 614                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 615         }
 616
 617 #ifdef BLKDISCARDZEROES
 618         {
 619                 int zeroes = 0;
 620                 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 621                     zeroes)
 622                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 623         }
 624 #endif
 625
 626 #if defined(__CYGWIN__)
 627         /*
 628          * Some operating systems require that the buffers be aligned,
 629          * regardless of O_DIRECT
 630          */
 631         if (!io->align)
 632                 io->align = 512;
 633 #endif
 634
 635 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 636         if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 637                 int dio_align = ext2fs_get_dio_alignment(fd);
 638
 639                 if (io->align < dio_align)
 640                         io->align = dio_align;
 641         }
 642 #endif
 643
 644         if ((retval = alloc_cache(io, data)))
 645                 goto cleanup;
 646
 647 #ifdef BLKROGET
 648         if (flags & IO_FLAG_RW) {
 649                 int error;
 650                 int readonly = 0;
 651
 652                 /* Is the block device actually writable? */
 653                 error = ioctl(data->dev, BLKROGET, &readonly);
 654                 if (!error && readonly) {
 655                         retval = EPERM;
 656                         goto cleanup;
 657                 }
 658         }
 659 #endif
 660
 661 #ifdef __linux__
 662 #undef RLIM_INFINITY
 663 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 664 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 665 #else
 666 #define RLIM_INFINITY  (~0UL)
 667 #endif
 668         /*
 669          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 670          * block devices are wrongly getting hit by the filesize
 671          * limit.  This workaround isn't perfect, since it won't work
 672          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 673          *
 674          */
 675         if ((flags & IO_FLAG_RW) &&
 676             (uname(&ut) == 0) &&
 677             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 678              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 679              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 680              (ut.release[5] < '8')) &&
 681             (ext2fs_fstat(data->dev, &st) == 0) &&
 682             (S_ISBLK(st.st_mode))) {
 683                 struct rlimit   rlim;
 684
 685                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 686                 setrlimit(RLIMIT_FSIZE, &rlim);
 687                 getrlimit(RLIMIT_FSIZE, &rlim);
 688                 if (((unsigned long) rlim.rlim_cur) <
 689                     ((unsigned long) rlim.rlim_max)) {
 690                         rlim.rlim_cur = rlim.rlim_max;
 691                         setrlimit(RLIMIT_FSIZE, &rlim);
 692                 }
 693         }
 694 #endif
 695         *channel = io;
 696         return 0;
 697
 698 cleanup:
 699         if (data) {
 700                 if (data->dev >= 0)
 701                         close(data->dev);
 702                 free_cache(data);
 703                 ext2fs_free_mem(&data);
 704         }
 705         if (io) {
 706                 if (io->name) {
 707                         ext2fs_free_mem(&io->name);
 708                 }
 709                 ext2fs_free_mem(&io);
 710         }
 711         return retval;
 712 }
 713
 714 static errcode_t unixfd_open(const char *str_fd, int flags,
 715                              io_channel *channel)
 716 {
 717         int fd;
 718         int fd_flags;
 719
 720         fd = atoi(str_fd);
 721         fd_flags = fcntl(fd, F_GETFD);
 722         if (fd_flags == -1)
 723                 return -EBADF;
 724
 725         flags = 0;
 726         if (fd_flags & O_RDWR)
 727                 flags |= IO_FLAG_RW;
 728         if (fd_flags & O_EXCL)
 729                 flags |= IO_FLAG_EXCLUSIVE;
 730 #if defined(O_DIRECT)
 731         if (fd_flags & O_DIRECT)
 732                 flags |= IO_FLAG_DIRECT_IO;
 733 #endif
 734
 735         return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
 736 }
 737
 738 static errcode_t unix_open(const char *name, int flags,
 739                            io_channel *channel)
 740 {
 741         int fd = -1;
 742         int open_flags;
 743
 744         if (name == 0)
 745                 return EXT2_ET_BAD_DEVICE_NAME;
 746
 747         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 748         if (flags & IO_FLAG_EXCLUSIVE)
 749                 open_flags |= O_EXCL;
 750 #if defined(O_DIRECT)
 751         if (flags & IO_FLAG_DIRECT_IO)
 752                 open_flags |= O_DIRECT;
 753 #endif
 754         fd = ext2fs_open_file(name, open_flags, 0);
 755         if (fd < 0)
 756                 return errno;
 757 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
 758         if (flags & IO_FLAG_DIRECT_IO) {
 759                 if (fcntl(fd, F_NOCACHE, 1) < 0)
 760                         return errno;
 761         }
 762 #endif
 763         return unix_open_channel(name, fd, flags, channel, unix_io_manager);
 764 }
 765
 766 static errcode_t unix_close(io_channel channel)
 767 {
 768         struct unix_private_data *data;
 769         errcode_t       retval = 0;
 770
 771         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 772         data = (struct unix_private_data *) channel->private_data;
 773         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 774
 775         if (--channel->refcount > 0)
 776                 return 0;
 777
 778 #ifndef NO_IO_CACHE
 779         retval = flush_cached_blocks(channel, data, 0);
 780 #endif
 781
 782         if (close(data->dev) < 0)
 783                 retval = errno;
 784         free_cache(data);
 785
 786         ext2fs_free_mem(&channel->private_data);
 787         if (channel->name)
 788                 ext2fs_free_mem(&channel->name);
 789         ext2fs_free_mem(&channel);
 790         return retval;
 791 }
 792
 793 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 794 {
 795         struct unix_private_data *data;
 796         errcode_t               retval;
 797
 798         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 799         data = (struct unix_private_data *) channel->private_data;
 800         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 801
 802         if (channel->block_size != blksize) {
 803 #ifndef NO_IO_CACHE
 804                 if ((retval = flush_cached_blocks(channel, data, 0)))
 805                         return retval;
 806 #endif
 807
 808                 channel->block_size = blksize;
 809                 free_cache(data);
 810                 if ((retval = alloc_cache(channel, data)))
 811                         return retval;
 812         }
 813         return 0;
 814 }
 815
 816 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 817                                int count, void *buf)
 818 {
 819         struct unix_private_data *data;
 820         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 821         errcode_t       retval;
 822         char            *cp;
 823         int             i, j;
 824
 825         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 826         data = (struct unix_private_data *) channel->private_data;
 827         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 828
 829 #ifdef NO_IO_CACHE
 830         return raw_read_blk(channel, data, block, count, buf);
 831 #else
 832         /*
 833          * If we're doing an odd-sized read or a very large read,
 834          * flush out the cache and then do a direct read.
 835          */
 836         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 837                 if ((retval = flush_cached_blocks(channel, data, 0)))
 838                         return retval;
 839                 return raw_read_blk(channel, data, block, count, buf);
 840         }
 841
 842         cp = buf;
 843         while (count > 0) {
 844                 /* If it's in the cache, use it! */
 845                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 846 #ifdef DEBUG
 847                         printf("Using cached block %lu\n", block);
 848 #endif
 849                         memcpy(cp, cache->buf, channel->block_size);
 850                         count--;
 851                         block++;
 852                         cp += channel->block_size;
 853                         continue;
 854                 }
 855                 if (count == 1) {
 856                         /*
 857                          * Special case where we read directly into the
 858                          * cache buffer; important in the O_DIRECT case
 859                          */
 860                         cache = reuse[0];
 861                         reuse_cache(channel, data, cache, block);
 862                         if ((retval = raw_read_blk(channel, data, block, 1,
 863                                                    cache->buf))) {
 864                                 cache->in_use = 0;
 865                                 return retval;
 866                         }
 867                         memcpy(cp, cache->buf, channel->block_size);
 868                         return 0;
 869                 }
 870
 871                 /*
 872                  * Find the number of uncached blocks so we can do a
 873                  * single read request
 874                  */
 875                 for (i=1; i < count; i++)
 876                         if (find_cached_block(data, block+i, &reuse[i]))
 877                                 break;
 878 #ifdef DEBUG
 879                 printf("Reading %d blocks starting at %lu\n", i, block);
 880 #endif
 881                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 882                         return retval;
 883
 884                 /* Save the results in the cache */
 885                 for (j=0; j < i; j++) {
 886                         count--;
 887                         cache = reuse[j];
 888                         reuse_cache(channel, data, cache, block++);
 889                         memcpy(cache->buf, cp, channel->block_size);
 890                         cp += channel->block_size;
 891                 }
 892         }
 893         return 0;
 894 #endif /* NO_IO_CACHE */
 895 }
 896
 897 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 898                                int count, void *buf)
 899 {
 900         return unix_read_blk64(channel, block, count, buf);
 901 }
 902
 903 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 904                                 int count, const void *buf)
 905 {
 906         struct unix_private_data *data;
 907         struct unix_cache *cache, *reuse;
 908         errcode_t       retval = 0;
 909         const char      *cp;
 910         int             writethrough;
 911
 912         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 913         data = (struct unix_private_data *) channel->private_data;
 914         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 915
 916 #ifdef NO_IO_CACHE
 917         return raw_write_blk(channel, data, block, count, buf);
 918 #else
 919         /*
 920          * If we're doing an odd-sized write or a very large write,
 921          * flush out the cache completely and then do a direct write.
 922          */
 923         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 924                 if ((retval = flush_cached_blocks(channel, data, 1)))
 925                         return retval;
 926                 return raw_write_blk(channel, data, block, count, buf);
 927         }
 928
 929         /*
 930          * For a moderate-sized multi-block write, first force a write
 931          * if we're in write-through cache mode, and then fill the
 932          * cache with the blocks.
 933          */
 934         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 935         if (writethrough)
 936                 retval = raw_write_blk(channel, data, block, count, buf);
 937
 938         cp = buf;
 939         while (count > 0) {
 940                 cache = find_cached_block(data, block, &reuse);
 941                 if (!cache) {
 942                         cache = reuse;
 943                         reuse_cache(channel, data, cache, block);
 944                 }
 945                 if (cache->buf != cp)
 946                         memcpy(cache->buf, cp, channel->block_size);
 947                 cache->dirty = !writethrough;
 948                 count--;
 949                 block++;
 950                 cp += channel->block_size;
 951         }
 952         return retval;
 953 #endif /* NO_IO_CACHE */
 954 }
 955
 956 static errcode_t unix_cache_readahead(io_channel channel,
 957                                       unsigned long long block,
 958                                       unsigned long long count)
 959 {
 960 #ifdef POSIX_FADV_WILLNEED
 961         struct unix_private_data *data;
 962
 963         data = (struct unix_private_data *)channel->private_data;
 964         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 965         return posix_fadvise(data->dev,
 966                              (ext2_loff_t)block * channel->block_size + data->offset,
 967                              (ext2_loff_t)count * channel->block_size,
 968                              POSIX_FADV_WILLNEED);
 969 #else
 970         return EXT2_ET_OP_NOT_SUPPORTED;
 971 #endif
 972 }
 973
 974 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 975                                 int count, const void *buf)
 976 {
 977         return unix_write_blk64(channel, block, count, buf);
 978 }
 979
 980 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 981                                  int size, const void *buf)
 982 {
 983         struct unix_private_data *data;
 984         errcode_t       retval = 0;
 985         ssize_t         actual;
 986
 987         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 988         data = (struct unix_private_data *) channel->private_data;
 989         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 990
 991         if (channel->align != 0) {
 992 #ifdef ALIGN_DEBUG
 993                 printf("unix_write_byte: O_DIRECT fallback\n");
 994 #endif
 995                 return EXT2_ET_UNIMPLEMENTED;
 996         }
 997
 998 #ifndef NO_IO_CACHE
 999         /*
1000          * Flush out the cache completely
1001          */
1002         if ((retval = flush_cached_blocks(channel, data, 1)))
1003                 return retval;
1004 #endif
1005
1006         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1007                 return errno;
1008
1009         actual = write(data->dev, buf, size);
1010         if (actual < 0)
1011                 return errno;
1012         if (actual != size)
1013                 return EXT2_ET_SHORT_WRITE;
1014
1015         return 0;
1016 }
1017
1018 /*
1019  * Flush data buffers to disk.
1020  */
1021 static errcode_t unix_flush(io_channel channel)
1022 {
1023         struct unix_private_data *data;
1024         errcode_t retval = 0;
1025
1026         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1027         data = (struct unix_private_data *) channel->private_data;
1028         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1029
1030 #ifndef NO_IO_CACHE
1031         retval = flush_cached_blocks(channel, data, 0);
1032 #endif
1033         fsync(data->dev);
1034         return retval;
1035 }
1036
1037 static errcode_t unix_set_option(io_channel channel, const char *option,
1038                                  const char *arg)
1039 {
1040         struct unix_private_data *data;
1041         unsigned long long tmp;
1042         char *end;
1043
1044         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1045         data = (struct unix_private_data *) channel->private_data;
1046         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1047
1048         if (!strcmp(option, "offset")) {
1049                 if (!arg)
1050                         return EXT2_ET_INVALID_ARGUMENT;
1051
1052                 tmp = strtoull(arg, &end, 0);
1053                 if (*end)
1054                         return EXT2_ET_INVALID_ARGUMENT;
1055                 data->offset = tmp;
1056                 if (data->offset < 0)
1057                         return EXT2_ET_INVALID_ARGUMENT;
1058                 return 0;
1059         }
1060         return EXT2_ET_INVALID_ARGUMENT;
1061 }
1062
1063 #if defined(__linux__) && !defined(BLKDISCARD)
1064 #define BLKDISCARD              _IO(0x12,119)
1065 #endif
1066
1067 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1068                               unsigned long long count)
1069 {
1070         struct unix_private_data *data;
1071         int             ret;
1072
1073         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1074         data = (struct unix_private_data *) channel->private_data;
1075         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1076
1077         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1078 #ifdef BLKDISCARD
1079                 __u64 range[2];
1080
1081                 range[0] = (__u64)(block) * channel->block_size + data->offset;
1082                 range[1] = (__u64)(count) * channel->block_size;
1083
1084                 ret = ioctl(data->dev, BLKDISCARD, &range);
1085 #else
1086                 goto unimplemented;
1087 #endif
1088         } else {
1089 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1090                 /*
1091                  * If we are not on block device, try to use punch hole
1092                  * to reclaim free space.
1093                  */
1094                 ret = fallocate(data->dev,
1095                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1096                                 (off_t)(block) * channel->block_size + data->offset,
1097                                 (off_t)(count) * channel->block_size);
1098 #else
1099                 goto unimplemented;
1100 #endif
1101         }
1102         if (ret < 0) {
1103                 if (errno == EOPNOTSUPP)
1104                         goto unimplemented;
1105                 return errno;
1106         }
1107         return 0;
1108 unimplemented:
1109         return EXT2_ET_UNIMPLEMENTED;
1110 }
1111
1112 /* parameters might not be used if OS doesn't support zeroout */
1113 #pragma GCC diagnostic push
1114 #pragma GCC diagnostic ignored "-Wunused-parameter"
1115 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1116                               unsigned long long count)
1117 {
1118         struct unix_private_data *data;
1119         int             ret;
1120
1121         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1122         data = (struct unix_private_data *) channel->private_data;
1123         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1124
1125         if (safe_getenv("UNIX_IO_NOZEROOUT"))
1126                 goto unimplemented;
1127
1128         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1129                 /* Not implemented until the BLKZEROOUT mess is fixed */
1130                 goto unimplemented;
1131         } else {
1132                 /* Regular file, try to use truncate/punch/zero. */
1133                 struct stat statbuf;
1134
1135                 if (count == 0)
1136                         return 0;
1137                 /*
1138                  * If we're trying to zero a range past the end of the file,
1139                  * extend the file size, then truncate everything.
1140                  */
1141                 ret = fstat(data->dev, &statbuf);
1142                 if (ret)
1143                         goto err;
1144                 if ((unsigned long long) statbuf.st_size <
1145                         (block + count) * channel->block_size + data->offset) {
1146                         ret = ftruncate(data->dev,
1147                                         (block + count) * channel->block_size + data->offset);
1148                         if (ret)
1149                                 goto err;
1150                 }
1151 #if defined(HAVE_FALLOCATE) && (defined(FALLOC_FL_ZERO_RANGE) || \
1152         (defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)))
1153 #if defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1154                 ret = fallocate(data->dev,
1155                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1156                                 (off_t)(block) * channel->block_size + data->offset,
1157                                 (off_t)(count) * channel->block_size);
1158                 if (ret == 0)
1159                         goto err;
1160 #endif
1161 #ifdef FALLOC_FL_ZERO_RANGE
1162                 ret = fallocate(data->dev,
1163                                 FALLOC_FL_ZERO_RANGE,
1164                                 (off_t)(block) * channel->block_size + data->offset,
1165                                 (off_t)(count) * channel->block_size);
1166 #endif
1167 #else
1168                 goto unimplemented;
1169 #endif /* HAVE_FALLOCATE && (ZERO_RANGE || (PUNCH_HOLE && KEEP_SIZE)) */
1170         }
1171 err:
1172         if (ret < 0) {
1173                 if (errno == EOPNOTSUPP)
1174                         goto unimplemented;
1175                 return errno;
1176         }
1177         return 0;
1178 unimplemented:
1179         return EXT2_ET_UNIMPLEMENTED;
1180 }
1181 #pragma GCC diagnostic pop
1182
1183 static struct struct_io_manager struct_unix_manager = {
1184         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1185         .name           = "Unix I/O Manager",
1186         .open           = unix_open,
1187         .close          = unix_close,
1188         .set_blksize    = unix_set_blksize,
1189         .read_blk       = unix_read_blk,
1190         .write_blk      = unix_write_blk,
1191         .flush          = unix_flush,
1192         .write_byte     = unix_write_byte,
1193         .set_option     = unix_set_option,
1194         .get_stats      = unix_get_stats,
1195         .read_blk64     = unix_read_blk64,
1196         .write_blk64    = unix_write_blk64,
1197         .discard        = unix_discard,
1198         .cache_readahead        = unix_cache_readahead,
1199         .zeroout        = unix_zeroout,
1200 };
1201
1202 io_manager unix_io_manager = &struct_unix_manager;
1203
1204 static struct struct_io_manager struct_unixfd_manager = {
1205         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1206         .name           = "Unix fd I/O Manager",
1207         .open           = unixfd_open,
1208         .close          = unix_close,
1209         .set_blksize    = unix_set_blksize,
1210         .read_blk       = unix_read_blk,
1211         .write_blk      = unix_write_blk,
1212         .flush          = unix_flush,
1213         .write_byte     = unix_write_byte,
1214         .set_option     = unix_set_option,
1215         .get_stats      = unix_get_stats,
1216         .read_blk64     = unix_read_blk64,
1217         .write_blk64    = unix_write_blk64,
1218         .discard        = unix_discard,
1219         .cache_readahead        = unix_cache_readahead,
1220         .zeroout        = unix_zeroout,
1221 };
1222
1223 io_manager unixfd_io_manager = &struct_unixfd_manager;