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 #include "ext2fsP.h"
  80
  81 /*
  82  * For checking structure magic numbers...
  83  */
  84
  85 #define EXT2_CHECK_MAGIC(struct, code) \
  86           if ((struct)->magic != (code)) return (code)
  87
  88 struct unix_cache {
  89         char                    *buf;
  90         unsigned long long      block;
  91         int                     access_time;
  92         unsigned                dirty:1;
  93         unsigned                in_use:1;
  94 };
  95
  96 #define CACHE_SIZE 8
  97 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
  98 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
  99
 100 struct unix_private_data {
 101         int     magic;
 102         int     dev;
 103         int     flags;
 104         int     align;
 105         int     access_time;
 106         ext2_loff_t offset;
 107         struct unix_cache cache[CACHE_SIZE];
 108         void    *bounce;
 109         struct struct_io_stats io_stats;
 110 };
 111
 112 #define IS_ALIGNED(n, align) ((((uintptr_t) n) & \
 113                                ((uintptr_t) ((align)-1))) == 0)
 114
 115 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 116 {
 117         errcode_t       retval = 0;
 118
 119         struct unix_private_data *data;
 120
 121         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 122         data = (struct unix_private_data *) channel->private_data;
 123         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 124
 125         if (stats)
 126                 *stats = &data->io_stats;
 127
 128         return retval;
 129 }
 130
 131 static char *safe_getenv(const char *arg)
 132 {
 133         if ((getuid() != geteuid()) || (getgid() != getegid()))
 134                 return NULL;
 135 #ifdef HAVE_PRCTL
 136         if (prctl(PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 137                 return NULL;
 138 #else
 139 #if (defined(linux) && defined(SYS_prctl))
 140         if (syscall(SYS_prctl, PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 141                 return NULL;
 142 #endif
 143 #endif
 144
 145 #if defined(HAVE_SECURE_GETENV)
 146         return secure_getenv(arg);
 147 #elif defined(HAVE___SECURE_GETENV)
 148         return __secure_getenv(arg);
 149 #else
 150         return getenv(arg);
 151 #endif
 152 }
 153
 154 /*
 155  * Here are the raw I/O functions
 156  */
 157 static errcode_t raw_read_blk(io_channel channel,
 158                               struct unix_private_data *data,
 159                               unsigned long long block,
 160                               int count, void *bufv)
 161 {
 162         errcode_t       retval;
 163         ssize_t         size;
 164         ext2_loff_t     location;
 165         int             actual = 0;
 166         unsigned char   *buf = bufv;
 167         ssize_t         really_read = 0;
 168
 169         size = (count < 0) ? -count : (ext2_loff_t) count * channel->block_size;
 170         data->io_stats.bytes_read += size;
 171         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 172
 173         if (data->flags & IO_FLAG_FORCE_BOUNCE) {
 174                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 175                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 176                         goto error_out;
 177                 }
 178                 goto bounce_read;
 179         }
 180
 181 #ifdef HAVE_PREAD64
 182         /* Try an aligned pread */
 183         if ((channel->align == 0) ||
 184             (IS_ALIGNED(buf, channel->align) &&
 185              IS_ALIGNED(size, channel->align))) {
 186                 actual = pread64(data->dev, buf, size, location);
 187                 if (actual == size)
 188                         return 0;
 189                 actual = 0;
 190         }
 191 #elif HAVE_PREAD
 192         /* Try an aligned pread */
 193         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 194             ((channel->align == 0) ||
 195              (IS_ALIGNED(buf, channel->align) &&
 196               IS_ALIGNED(size, channel->align)))) {
 197                 actual = pread(data->dev, buf, size, location);
 198                 if (actual == size)
 199                         return 0;
 200                 actual = 0;
 201         }
 202 #endif /* HAVE_PREAD */
 203
 204         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 205                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 206                 goto error_out;
 207         }
 208         if ((channel->align == 0) ||
 209             (IS_ALIGNED(buf, channel->align) &&
 210              IS_ALIGNED(size, channel->align))) {
 211                 actual = read(data->dev, buf, size);
 212                 if (actual != size) {
 213                 short_read:
 214                         if (actual < 0) {
 215                                 retval = errno;
 216                                 actual = 0;
 217                         } else
 218                                 retval = EXT2_ET_SHORT_READ;
 219                         goto error_out;
 220                 }
 221                 return 0;
 222         }
 223
 224 #ifdef ALIGN_DEBUG
 225         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 226                (unsigned long) size);
 227 #endif
 228
 229         /*
 230          * The buffer or size which we're trying to read isn't aligned
 231          * to the O_DIRECT rules, so we need to do this the hard way...
 232          */
 233 bounce_read:
 234         while (size > 0) {
 235                 actual = read(data->dev, data->bounce, channel->block_size);
 236                 if (actual != channel->block_size) {
 237                         actual = really_read;
 238                         buf -= really_read;
 239                         size += really_read;
 240                         goto short_read;
 241                 }
 242                 actual = size;
 243                 if (size > channel->block_size)
 244                         actual = channel->block_size;
 245                 memcpy(buf, data->bounce, actual);
 246                 really_read += actual;
 247                 size -= actual;
 248                 buf += actual;
 249         }
 250         return 0;
 251
 252 error_out:
 253         if (actual >= 0 && actual < size)
 254                 memset((char *) buf+actual, 0, size-actual);
 255         if (channel->read_error)
 256                 retval = (channel->read_error)(channel, block, count, buf,
 257                                                size, actual, retval);
 258         return retval;
 259 }
 260
 261 static errcode_t raw_write_blk(io_channel channel,
 262                                struct unix_private_data *data,
 263                                unsigned long long block,
 264                                int count, const void *bufv)
 265 {
 266         ssize_t         size;
 267         ext2_loff_t     location;
 268         int             actual = 0;
 269         errcode_t       retval;
 270         const unsigned char *buf = bufv;
 271
 272         if (count == 1)
 273                 size = channel->block_size;
 274         else {
 275                 if (count < 0)
 276                         size = -count;
 277                 else
 278                         size = (ext2_loff_t) count * channel->block_size;
 279         }
 280         data->io_stats.bytes_written += size;
 281
 282         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 283
 284         if (data->flags & IO_FLAG_FORCE_BOUNCE) {
 285                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 286                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 287                         goto error_out;
 288                 }
 289                 goto bounce_write;
 290         }
 291
 292 #ifdef HAVE_PWRITE64
 293         /* Try an aligned pwrite */
 294         if ((channel->align == 0) ||
 295             (IS_ALIGNED(buf, channel->align) &&
 296              IS_ALIGNED(size, channel->align))) {
 297                 actual = pwrite64(data->dev, buf, size, location);
 298                 if (actual == size)
 299                         return 0;
 300         }
 301 #elif HAVE_PWRITE
 302         /* Try an aligned pwrite */
 303         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 304             ((channel->align == 0) ||
 305              (IS_ALIGNED(buf, channel->align) &&
 306               IS_ALIGNED(size, channel->align)))) {
 307                 actual = pwrite(data->dev, buf, size, location);
 308                 if (actual == size)
 309                         return 0;
 310         }
 311 #endif /* HAVE_PWRITE */
 312
 313         if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 314                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 315                 goto error_out;
 316         }
 317
 318         if ((channel->align == 0) ||
 319             (IS_ALIGNED(buf, channel->align) &&
 320              IS_ALIGNED(size, channel->align))) {
 321                 actual = write(data->dev, buf, size);
 322                 if (actual < 0) {
 323                         retval = errno;
 324                         goto error_out;
 325                 }
 326                 if (actual != size) {
 327                 short_write:
 328                         retval = EXT2_ET_SHORT_WRITE;
 329                         goto error_out;
 330                 }
 331                 return 0;
 332         }
 333
 334 #ifdef ALIGN_DEBUG
 335         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 336                (unsigned long) size);
 337 #endif
 338         /*
 339          * The buffer or size which we're trying to write isn't aligned
 340          * to the O_DIRECT rules, so we need to do this the hard way...
 341          */
 342 bounce_write:
 343         while (size > 0) {
 344                 if (size < channel->block_size) {
 345                         actual = read(data->dev, data->bounce,
 346                                       channel->block_size);
 347                         if (actual != channel->block_size) {
 348                                 if (actual < 0) {
 349                                         retval = errno;
 350                                         goto error_out;
 351                                 }
 352                                 memset((char *) data->bounce + actual, 0,
 353                                        channel->block_size - actual);
 354                         }
 355                 }
 356                 actual = size;
 357                 if (size > channel->block_size)
 358                         actual = channel->block_size;
 359                 memcpy(data->bounce, buf, actual);
 360                 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
 361                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 362                         goto error_out;
 363                 }
 364                 actual = write(data->dev, data->bounce, channel->block_size);
 365                 if (actual < 0) {
 366                         retval = errno;
 367                         goto error_out;
 368                 }
 369                 if (actual != channel->block_size)
 370                         goto short_write;
 371                 size -= actual;
 372                 buf += actual;
 373                 location += actual;
 374         }
 375         return 0;
 376
 377 error_out:
 378         if (channel->write_error)
 379                 retval = (channel->write_error)(channel, block, count, buf,
 380                                                 size, actual, retval);
 381         return retval;
 382 }
 383
 384
 385 /*
 386  * Here we implement the cache functions
 387  */
 388
 389 /* Allocate the cache buffers */
 390 static errcode_t alloc_cache(io_channel channel,
 391                              struct unix_private_data *data)
 392 {
 393         errcode_t               retval;
 394         struct unix_cache       *cache;
 395         int                     i;
 396
 397         data->access_time = 0;
 398         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 399                 cache->block = 0;
 400                 cache->access_time = 0;
 401                 cache->dirty = 0;
 402                 cache->in_use = 0;
 403                 if (cache->buf)
 404                         ext2fs_free_mem(&cache->buf);
 405                 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
 406                 if (retval)
 407                         return retval;
 408         }
 409         if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
 410                 if (data->bounce)
 411                         ext2fs_free_mem(&data->bounce);
 412                 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
 413         }
 414         return retval;
 415 }
 416
 417 /* Free the cache buffers */
 418 static void free_cache(struct unix_private_data *data)
 419 {
 420         struct unix_cache       *cache;
 421         int                     i;
 422
 423         data->access_time = 0;
 424         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 425                 cache->block = 0;
 426                 cache->access_time = 0;
 427                 cache->dirty = 0;
 428                 cache->in_use = 0;
 429                 if (cache->buf)
 430                         ext2fs_free_mem(&cache->buf);
 431         }
 432         if (data->bounce)
 433                 ext2fs_free_mem(&data->bounce);
 434 }
 435
 436 #ifndef NO_IO_CACHE
 437 /*
 438  * Try to find a block in the cache.  If the block is not found, and
 439  * eldest is a non-zero pointer, then fill in eldest with the cache
 440  * entry to that should be reused.
 441  */
 442 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 443                                             unsigned long long block,
 444                                             struct unix_cache **eldest)
 445 {
 446         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 447         int                     i;
 448
 449         unused_cache = oldest_cache = 0;
 450         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 451                 if (!cache->in_use) {
 452                         if (!unused_cache)
 453                                 unused_cache = cache;
 454                         continue;
 455                 }
 456                 if (cache->block == block) {
 457                         cache->access_time = ++data->access_time;
 458                         return cache;
 459                 }
 460                 if (!oldest_cache ||
 461                     (cache->access_time < oldest_cache->access_time))
 462                         oldest_cache = cache;
 463         }
 464         if (eldest)
 465                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 466         return 0;
 467 }
 468
 469 /*
 470  * Reuse a particular cache entry for another block.
 471  */
 472 static void reuse_cache(io_channel channel, struct unix_private_data *data,
 473                  struct unix_cache *cache, unsigned long long block)
 474 {
 475         if (cache->dirty && cache->in_use)
 476                 raw_write_blk(channel, data, cache->block, 1, cache->buf);
 477
 478         cache->in_use = 1;
 479         cache->dirty = 0;
 480         cache->block = block;
 481         cache->access_time = ++data->access_time;
 482 }
 483
 484 /*
 485  * Flush all of the blocks in the cache
 486  */
 487 static errcode_t flush_cached_blocks(io_channel channel,
 488                                      struct unix_private_data *data,
 489                                      int invalidate)
 490
 491 {
 492         struct unix_cache       *cache;
 493         errcode_t               retval, retval2;
 494         int                     i;
 495
 496         retval2 = 0;
 497         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 498                 if (!cache->in_use)
 499                         continue;
 500
 501                 if (invalidate)
 502                         cache->in_use = 0;
 503
 504                 if (!cache->dirty)
 505                         continue;
 506
 507                 retval = raw_write_blk(channel, data,
 508                                        cache->block, 1, cache->buf);
 509                 if (retval)
 510                         retval2 = retval;
 511                 else
 512                         cache->dirty = 0;
 513         }
 514         return retval2;
 515 }
 516 #endif /* NO_IO_CACHE */
 517
 518 #ifdef __linux__
 519 #ifndef BLKDISCARDZEROES
 520 #define BLKDISCARDZEROES _IO(0x12,124)
 521 #endif
 522 #endif
 523
 524 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
 525 {
 526         if (mode)
 527 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 528                 return open64(pathname, flags, mode);
 529         else
 530                 return open64(pathname, flags);
 531 #else
 532                 return open(pathname, flags, mode);
 533         else
 534                 return open(pathname, flags);
 535 #endif
 536 }
 537
 538 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
 539 {
 540 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 541         return stat64(path, buf);
 542 #else
 543         return stat(path, buf);
 544 #endif
 545 }
 546
 547 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
 548 {
 549 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 550         return fstat64(fd, buf);
 551 #else
 552         return fstat(fd, buf);
 553 #endif
 554 }
 555
 556
 557 static errcode_t unix_open_channel(const char *name, int fd,
 558                                    int flags, io_channel *channel,
 559                                    io_manager io_mgr)
 560 {
 561         io_channel      io = NULL;
 562         struct unix_private_data *data = NULL;
 563         errcode_t       retval;
 564         ext2fs_struct_stat st;
 565 #ifdef __linux__
 566         struct          utsname ut;
 567 #endif
 568
 569         if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
 570                 flags |= IO_FLAG_FORCE_BOUNCE;
 571
 572 #ifdef __linux__
 573         /*
 574          * We need to make sure any previous errors in the block
 575          * device are thrown away, sigh.
 576          */
 577         (void) fsync(fd);
 578 #endif
 579
 580         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 581         if (retval)
 582                 goto cleanup;
 583         memset(io, 0, sizeof(struct struct_io_channel));
 584         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 585         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 586         if (retval)
 587                 goto cleanup;
 588
 589         io->manager = io_mgr;
 590         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 591         if (retval)
 592                 goto cleanup;
 593
 594         strcpy(io->name, name);
 595         io->private_data = data;
 596         io->block_size = 1024;
 597         io->read_error = 0;
 598         io->write_error = 0;
 599         io->refcount = 1;
 600
 601         memset(data, 0, sizeof(struct unix_private_data));
 602         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 603         data->io_stats.num_fields = 2;
 604         data->flags = flags;
 605         data->dev = fd;
 606
 607 #if defined(O_DIRECT)
 608         if (flags & IO_FLAG_DIRECT_IO)
 609                 io->align = ext2fs_get_dio_alignment(data->dev);
 610 #elif defined(F_NOCACHE)
 611         if (flags & IO_FLAG_DIRECT_IO)
 612                 io->align = 4096;
 613 #endif
 614
 615         /*
 616          * If the device is really a block device, then set the
 617          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 618          * because we are going to use punch hole instead of discard
 619          * and if it succeed, subsequent read from sparse area returns
 620          * zero.
 621          */
 622         if (ext2fs_fstat(data->dev, &st) == 0) {
 623                 if (ext2fsP_is_disk_device(st.st_mode))
 624                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 625                 else
 626                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 627         }
 628
 629 #ifdef BLKDISCARDZEROES
 630         {
 631                 int zeroes = 0;
 632                 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 633                     zeroes)
 634                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 635         }
 636 #endif
 637
 638 #if defined(__CYGWIN__)
 639         /*
 640          * Some operating systems require that the buffers be aligned,
 641          * regardless of O_DIRECT
 642          */
 643         if (!io->align)
 644                 io->align = 512;
 645 #endif
 646
 647 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 648         if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 649                 int dio_align = ext2fs_get_dio_alignment(fd);
 650
 651                 if (io->align < dio_align)
 652                         io->align = dio_align;
 653         }
 654 #endif
 655
 656         if ((retval = alloc_cache(io, data)))
 657                 goto cleanup;
 658
 659 #ifdef BLKROGET
 660         if (flags & IO_FLAG_RW) {
 661                 int error;
 662                 int readonly = 0;
 663
 664                 /* Is the block device actually writable? */
 665                 error = ioctl(data->dev, BLKROGET, &readonly);
 666                 if (!error && readonly) {
 667                         retval = EPERM;
 668                         goto cleanup;
 669                 }
 670         }
 671 #endif
 672
 673 #ifdef __linux__
 674 #undef RLIM_INFINITY
 675 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 676 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 677 #else
 678 #define RLIM_INFINITY  (~0UL)
 679 #endif
 680         /*
 681          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 682          * block devices are wrongly getting hit by the filesize
 683          * limit.  This workaround isn't perfect, since it won't work
 684          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 685          *
 686          */
 687         if ((flags & IO_FLAG_RW) &&
 688             (uname(&ut) == 0) &&
 689             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 690              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 691              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 692              (ut.release[5] < '8')) &&
 693             (ext2fs_fstat(data->dev, &st) == 0) &&
 694             (ext2fsP_is_disk_device(st.st_mode))) {
 695                 struct rlimit   rlim;
 696
 697                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 698                 setrlimit(RLIMIT_FSIZE, &rlim);
 699                 getrlimit(RLIMIT_FSIZE, &rlim);
 700                 if (((unsigned long) rlim.rlim_cur) <
 701                     ((unsigned long) rlim.rlim_max)) {
 702                         rlim.rlim_cur = rlim.rlim_max;
 703                         setrlimit(RLIMIT_FSIZE, &rlim);
 704                 }
 705         }
 706 #endif
 707         *channel = io;
 708         return 0;
 709
 710 cleanup:
 711         if (data) {
 712                 if (data->dev >= 0)
 713                         close(data->dev);
 714                 free_cache(data);
 715                 ext2fs_free_mem(&data);
 716         }
 717         if (io) {
 718                 if (io->name) {
 719                         ext2fs_free_mem(&io->name);
 720                 }
 721                 ext2fs_free_mem(&io);
 722         }
 723         return retval;
 724 }
 725
 726 static errcode_t unixfd_open(const char *str_fd, int flags,
 727                              io_channel *channel)
 728 {
 729         int fd;
 730         int fd_flags;
 731
 732         fd = atoi(str_fd);
 733 #if defined(HAVE_FCNTL)
 734         fd_flags = fcntl(fd, F_GETFD);
 735         if (fd_flags == -1)
 736                 return -EBADF;
 737
 738         flags = 0;
 739         if (fd_flags & O_RDWR)
 740                 flags |= IO_FLAG_RW;
 741         if (fd_flags & O_EXCL)
 742                 flags |= IO_FLAG_EXCLUSIVE;
 743 #if defined(O_DIRECT)
 744         if (fd_flags & O_DIRECT)
 745                 flags |= IO_FLAG_DIRECT_IO;
 746 #endif
 747 #endif  /* HAVE_FCNTL */
 748
 749         return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
 750 }
 751
 752 static errcode_t unix_open(const char *name, int flags,
 753                            io_channel *channel)
 754 {
 755         int fd = -1;
 756         int open_flags;
 757
 758         if (name == 0)
 759                 return EXT2_ET_BAD_DEVICE_NAME;
 760
 761         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 762         if (flags & IO_FLAG_EXCLUSIVE)
 763                 open_flags |= O_EXCL;
 764 #if defined(O_DIRECT)
 765         if (flags & IO_FLAG_DIRECT_IO)
 766                 open_flags |= O_DIRECT;
 767 #endif
 768         fd = ext2fs_open_file(name, open_flags, 0);
 769         if (fd < 0)
 770                 return errno;
 771 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
 772         if (flags & IO_FLAG_DIRECT_IO) {
 773                 if (fcntl(fd, F_NOCACHE, 1) < 0)
 774                         return errno;
 775         }
 776 #endif
 777         return unix_open_channel(name, fd, flags, channel, unix_io_manager);
 778 }
 779
 780 static errcode_t unix_close(io_channel channel)
 781 {
 782         struct unix_private_data *data;
 783         errcode_t       retval = 0;
 784
 785         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 786         data = (struct unix_private_data *) channel->private_data;
 787         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 788
 789         if (--channel->refcount > 0)
 790                 return 0;
 791
 792 #ifndef NO_IO_CACHE
 793         retval = flush_cached_blocks(channel, data, 0);
 794 #endif
 795
 796         if (close(data->dev) < 0)
 797                 retval = errno;
 798         free_cache(data);
 799
 800         ext2fs_free_mem(&channel->private_data);
 801         if (channel->name)
 802                 ext2fs_free_mem(&channel->name);
 803         ext2fs_free_mem(&channel);
 804         return retval;
 805 }
 806
 807 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 808 {
 809         struct unix_private_data *data;
 810         errcode_t               retval;
 811
 812         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 813         data = (struct unix_private_data *) channel->private_data;
 814         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 815
 816         if (channel->block_size != blksize) {
 817 #ifndef NO_IO_CACHE
 818                 if ((retval = flush_cached_blocks(channel, data, 0)))
 819                         return retval;
 820 #endif
 821
 822                 channel->block_size = blksize;
 823                 free_cache(data);
 824                 if ((retval = alloc_cache(channel, data)))
 825                         return retval;
 826         }
 827         return 0;
 828 }
 829
 830 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 831                                int count, void *buf)
 832 {
 833         struct unix_private_data *data;
 834         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 835         errcode_t       retval;
 836         char            *cp;
 837         int             i, j;
 838
 839         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 840         data = (struct unix_private_data *) channel->private_data;
 841         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 842
 843 #ifdef NO_IO_CACHE
 844         return raw_read_blk(channel, data, block, count, buf);
 845 #else
 846         /*
 847          * If we're doing an odd-sized read or a very large read,
 848          * flush out the cache and then do a direct read.
 849          */
 850         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 851                 if ((retval = flush_cached_blocks(channel, data, 0)))
 852                         return retval;
 853                 return raw_read_blk(channel, data, block, count, buf);
 854         }
 855
 856         cp = buf;
 857         while (count > 0) {
 858                 /* If it's in the cache, use it! */
 859                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 860 #ifdef DEBUG
 861                         printf("Using cached block %lu\n", block);
 862 #endif
 863                         memcpy(cp, cache->buf, channel->block_size);
 864                         count--;
 865                         block++;
 866                         cp += channel->block_size;
 867                         continue;
 868                 }
 869                 if (count == 1) {
 870                         /*
 871                          * Special case where we read directly into the
 872                          * cache buffer; important in the O_DIRECT case
 873                          */
 874                         cache = reuse[0];
 875                         reuse_cache(channel, data, cache, block);
 876                         if ((retval = raw_read_blk(channel, data, block, 1,
 877                                                    cache->buf))) {
 878                                 cache->in_use = 0;
 879                                 return retval;
 880                         }
 881                         memcpy(cp, cache->buf, channel->block_size);
 882                         return 0;
 883                 }
 884
 885                 /*
 886                  * Find the number of uncached blocks so we can do a
 887                  * single read request
 888                  */
 889                 for (i=1; i < count; i++)
 890                         if (find_cached_block(data, block+i, &reuse[i]))
 891                                 break;
 892 #ifdef DEBUG
 893                 printf("Reading %d blocks starting at %lu\n", i, block);
 894 #endif
 895                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 896                         return retval;
 897
 898                 /* Save the results in the cache */
 899                 for (j=0; j < i; j++) {
 900                         count--;
 901                         cache = reuse[j];
 902                         reuse_cache(channel, data, cache, block++);
 903                         memcpy(cache->buf, cp, channel->block_size);
 904                         cp += channel->block_size;
 905                 }
 906         }
 907         return 0;
 908 #endif /* NO_IO_CACHE */
 909 }
 910
 911 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 912                                int count, void *buf)
 913 {
 914         return unix_read_blk64(channel, block, count, buf);
 915 }
 916
 917 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 918                                 int count, const void *buf)
 919 {
 920         struct unix_private_data *data;
 921         struct unix_cache *cache, *reuse;
 922         errcode_t       retval = 0;
 923         const char      *cp;
 924         int             writethrough;
 925
 926         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 927         data = (struct unix_private_data *) channel->private_data;
 928         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 929
 930 #ifdef NO_IO_CACHE
 931         return raw_write_blk(channel, data, block, count, buf);
 932 #else
 933         /*
 934          * If we're doing an odd-sized write or a very large write,
 935          * flush out the cache completely and then do a direct write.
 936          */
 937         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 938                 if ((retval = flush_cached_blocks(channel, data, 1)))
 939                         return retval;
 940                 return raw_write_blk(channel, data, block, count, buf);
 941         }
 942
 943         /*
 944          * For a moderate-sized multi-block write, first force a write
 945          * if we're in write-through cache mode, and then fill the
 946          * cache with the blocks.
 947          */
 948         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 949         if (writethrough)
 950                 retval = raw_write_blk(channel, data, block, count, buf);
 951
 952         cp = buf;
 953         while (count > 0) {
 954                 cache = find_cached_block(data, block, &reuse);
 955                 if (!cache) {
 956                         cache = reuse;
 957                         reuse_cache(channel, data, cache, block);
 958                 }
 959                 if (cache->buf != cp)
 960                         memcpy(cache->buf, cp, channel->block_size);
 961                 cache->dirty = !writethrough;
 962                 count--;
 963                 block++;
 964                 cp += channel->block_size;
 965         }
 966         return retval;
 967 #endif /* NO_IO_CACHE */
 968 }
 969
 970 static errcode_t unix_cache_readahead(io_channel channel,
 971                                       unsigned long long block,
 972                                       unsigned long long count)
 973 {
 974 #ifdef POSIX_FADV_WILLNEED
 975         struct unix_private_data *data;
 976
 977         data = (struct unix_private_data *)channel->private_data;
 978         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 979         return posix_fadvise(data->dev,
 980                              (ext2_loff_t)block * channel->block_size + data->offset,
 981                              (ext2_loff_t)count * channel->block_size,
 982                              POSIX_FADV_WILLNEED);
 983 #else
 984         return EXT2_ET_OP_NOT_SUPPORTED;
 985 #endif
 986 }
 987
 988 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 989                                 int count, const void *buf)
 990 {
 991         return unix_write_blk64(channel, block, count, buf);
 992 }
 993
 994 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 995                                  int size, const void *buf)
 996 {
 997         struct unix_private_data *data;
 998         errcode_t       retval = 0;
 999         ssize_t         actual;
1000
1001         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1002         data = (struct unix_private_data *) channel->private_data;
1003         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1004
1005         if (channel->align != 0) {
1006 #ifdef ALIGN_DEBUG
1007                 printf("unix_write_byte: O_DIRECT fallback\n");
1008 #endif
1009                 return EXT2_ET_UNIMPLEMENTED;
1010         }
1011
1012 #ifndef NO_IO_CACHE
1013         /*
1014          * Flush out the cache completely
1015          */
1016         if ((retval = flush_cached_blocks(channel, data, 1)))
1017                 return retval;
1018 #endif
1019
1020         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1021                 return errno;
1022
1023         actual = write(data->dev, buf, size);
1024         if (actual < 0)
1025                 return errno;
1026         if (actual != size)
1027                 return EXT2_ET_SHORT_WRITE;
1028
1029         return 0;
1030 }
1031
1032 /*
1033  * Flush data buffers to disk.
1034  */
1035 static errcode_t unix_flush(io_channel channel)
1036 {
1037         struct unix_private_data *data;
1038         errcode_t retval = 0;
1039
1040         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1041         data = (struct unix_private_data *) channel->private_data;
1042         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1043
1044 #ifndef NO_IO_CACHE
1045         retval = flush_cached_blocks(channel, data, 0);
1046 #endif
1047 #ifdef HAVE_FSYNC
1048         if (!retval && fsync(data->dev) != 0)
1049                 return errno;
1050 #endif
1051         return retval;
1052 }
1053
1054 static errcode_t unix_set_option(io_channel channel, const char *option,
1055                                  const char *arg)
1056 {
1057         struct unix_private_data *data;
1058         unsigned long long tmp;
1059         char *end;
1060
1061         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1062         data = (struct unix_private_data *) channel->private_data;
1063         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1064
1065         if (!strcmp(option, "offset")) {
1066                 if (!arg)
1067                         return EXT2_ET_INVALID_ARGUMENT;
1068
1069                 tmp = strtoull(arg, &end, 0);
1070                 if (*end)
1071                         return EXT2_ET_INVALID_ARGUMENT;
1072                 data->offset = tmp;
1073                 if (data->offset < 0)
1074                         return EXT2_ET_INVALID_ARGUMENT;
1075                 return 0;
1076         }
1077         return EXT2_ET_INVALID_ARGUMENT;
1078 }
1079
1080 #if defined(__linux__) && !defined(BLKDISCARD)
1081 #define BLKDISCARD              _IO(0x12,119)
1082 #endif
1083
1084 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1085                               unsigned long long count)
1086 {
1087         struct unix_private_data *data;
1088         int             ret;
1089
1090         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1091         data = (struct unix_private_data *) channel->private_data;
1092         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1093
1094         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1095 #ifdef BLKDISCARD
1096                 __u64 range[2];
1097
1098                 range[0] = (__u64)(block) * channel->block_size + data->offset;
1099                 range[1] = (__u64)(count) * channel->block_size;
1100
1101                 ret = ioctl(data->dev, BLKDISCARD, &range);
1102 #else
1103                 goto unimplemented;
1104 #endif
1105         } else {
1106 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1107                 /*
1108                  * If we are not on block device, try to use punch hole
1109                  * to reclaim free space.
1110                  */
1111                 ret = fallocate(data->dev,
1112                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1113                                 (off_t)(block) * channel->block_size + data->offset,
1114                                 (off_t)(count) * channel->block_size);
1115 #else
1116                 goto unimplemented;
1117 #endif
1118         }
1119         if (ret < 0) {
1120                 if (errno == EOPNOTSUPP)
1121                         goto unimplemented;
1122                 return errno;
1123         }
1124         return 0;
1125 unimplemented:
1126         return EXT2_ET_UNIMPLEMENTED;
1127 }
1128
1129 /*
1130  * If we know about ZERO_RANGE, try that before we try PUNCH_HOLE because
1131  * ZERO_RANGE doesn't unmap preallocated blocks.  We prefer fallocate because
1132  * it always invalidates page cache, and libext2fs requires that reads after
1133  * ZERO_RANGE return zeroes.
1134  */
1135 static int __unix_zeroout(int fd, off_t offset, off_t len)
1136 {
1137         int ret = -1;
1138
1139 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_ZERO_RANGE)
1140         ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, offset, len);
1141         if (ret == 0)
1142                 return 0;
1143 #endif
1144 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1145         ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1146                         offset,  len);
1147         if (ret == 0)
1148                 return 0;
1149 #endif
1150         errno = EOPNOTSUPP;
1151         return ret;
1152 }
1153
1154 /* parameters might not be used if OS doesn't support zeroout */
1155 #if __GNUC_PREREQ (4, 6)
1156 #pragma GCC diagnostic push
1157 #pragma GCC diagnostic ignored "-Wunused-parameter"
1158 #endif
1159 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1160                               unsigned long long count)
1161 {
1162         struct unix_private_data *data;
1163         int             ret;
1164
1165         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1166         data = (struct unix_private_data *) channel->private_data;
1167         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1168
1169         if (safe_getenv("UNIX_IO_NOZEROOUT"))
1170                 goto unimplemented;
1171
1172         if (!(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE)) {
1173                 /* Regular file, try to use truncate/punch/zero. */
1174                 struct stat statbuf;
1175
1176                 if (count == 0)
1177                         return 0;
1178                 /*
1179                  * If we're trying to zero a range past the end of the file,
1180                  * extend the file size, then truncate everything.
1181                  */
1182                 ret = fstat(data->dev, &statbuf);
1183                 if (ret)
1184                         goto err;
1185                 if ((unsigned long long) statbuf.st_size <
1186                         (block + count) * channel->block_size + data->offset) {
1187                         ret = ftruncate(data->dev,
1188                                         (block + count) * channel->block_size + data->offset);
1189                         if (ret)
1190                                 goto err;
1191                 }
1192         }
1193
1194         ret = __unix_zeroout(data->dev,
1195                         (off_t)(block) * channel->block_size + data->offset,
1196                         (off_t)(count) * channel->block_size);
1197 err:
1198         if (ret < 0) {
1199                 if (errno == EOPNOTSUPP)
1200                         goto unimplemented;
1201                 return errno;
1202         }
1203         return 0;
1204 unimplemented:
1205         return EXT2_ET_UNIMPLEMENTED;
1206 }
1207 #if __GNUC_PREREQ (4, 6)
1208 #pragma GCC diagnostic pop
1209 #endif
1210
1211 static struct struct_io_manager struct_unix_manager = {
1212         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1213         .name           = "Unix I/O Manager",
1214         .open           = unix_open,
1215         .close          = unix_close,
1216         .set_blksize    = unix_set_blksize,
1217         .read_blk       = unix_read_blk,
1218         .write_blk      = unix_write_blk,
1219         .flush          = unix_flush,
1220         .write_byte     = unix_write_byte,
1221         .set_option     = unix_set_option,
1222         .get_stats      = unix_get_stats,
1223         .read_blk64     = unix_read_blk64,
1224         .write_blk64    = unix_write_blk64,
1225         .discard        = unix_discard,
1226         .cache_readahead        = unix_cache_readahead,
1227         .zeroout        = unix_zeroout,
1228 };
1229
1230 io_manager unix_io_manager = &struct_unix_manager;
1231
1232 static struct struct_io_manager struct_unixfd_manager = {
1233         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1234         .name           = "Unix fd I/O Manager",
1235         .open           = unixfd_open,
1236         .close          = unix_close,
1237         .set_blksize    = unix_set_blksize,
1238         .read_blk       = unix_read_blk,
1239         .write_blk      = unix_write_blk,
1240         .flush          = unix_flush,
1241         .write_byte     = unix_write_byte,
1242         .set_option     = unix_set_option,
1243         .get_stats      = unix_get_stats,
1244         .read_blk64     = unix_read_blk64,
1245         .write_blk64    = unix_write_blk64,
1246         .discard        = unix_discard,
1247         .cache_readahead        = unix_cache_readahead,
1248         .zeroout        = unix_zeroout,
1249 };
1250
1251 io_manager unixfd_io_manager = &struct_unixfd_manager;