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 : 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 = 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         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 573         if (retval)
 574                 goto cleanup;
 575         memset(io, 0, sizeof(struct struct_io_channel));
 576         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 577         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 578         if (retval)
 579                 goto cleanup;
 580
 581         io->manager = io_mgr;
 582         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 583         if (retval)
 584                 goto cleanup;
 585
 586         strcpy(io->name, name);
 587         io->private_data = data;
 588         io->block_size = 1024;
 589         io->read_error = 0;
 590         io->write_error = 0;
 591         io->refcount = 1;
 592
 593         memset(data, 0, sizeof(struct unix_private_data));
 594         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 595         data->io_stats.num_fields = 2;
 596         data->flags = flags;
 597         data->dev = fd;
 598
 599 #if defined(O_DIRECT)
 600         if (flags & IO_FLAG_DIRECT_IO)
 601                 io->align = ext2fs_get_dio_alignment(data->dev);
 602 #elif defined(F_NOCACHE)
 603         if (flags & IO_FLAG_DIRECT_IO)
 604                 io->align = 4096;
 605 #endif
 606
 607         /*
 608          * If the device is really a block device, then set the
 609          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 610          * because we are going to use punch hole instead of discard
 611          * and if it succeed, subsequent read from sparse area returns
 612          * zero.
 613          */
 614         if (ext2fs_fstat(data->dev, &st) == 0) {
 615                 if (ext2fsP_is_disk_device(st.st_mode))
 616                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 617                 else
 618                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 619         }
 620
 621 #ifdef BLKDISCARDZEROES
 622         {
 623                 int zeroes = 0;
 624                 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 625                     zeroes)
 626                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 627         }
 628 #endif
 629
 630 #if defined(__CYGWIN__)
 631         /*
 632          * Some operating systems require that the buffers be aligned,
 633          * regardless of O_DIRECT
 634          */
 635         if (!io->align)
 636                 io->align = 512;
 637 #endif
 638
 639 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 640         if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 641                 int dio_align = ext2fs_get_dio_alignment(fd);
 642
 643                 if (io->align < dio_align)
 644                         io->align = dio_align;
 645         }
 646 #endif
 647
 648         if ((retval = alloc_cache(io, data)))
 649                 goto cleanup;
 650
 651 #ifdef BLKROGET
 652         if (flags & IO_FLAG_RW) {
 653                 int error;
 654                 int readonly = 0;
 655
 656                 /* Is the block device actually writable? */
 657                 error = ioctl(data->dev, BLKROGET, &readonly);
 658                 if (!error && readonly) {
 659                         retval = EPERM;
 660                         goto cleanup;
 661                 }
 662         }
 663 #endif
 664
 665 #ifdef __linux__
 666 #undef RLIM_INFINITY
 667 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 668 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 669 #else
 670 #define RLIM_INFINITY  (~0UL)
 671 #endif
 672         /*
 673          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 674          * block devices are wrongly getting hit by the filesize
 675          * limit.  This workaround isn't perfect, since it won't work
 676          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 677          *
 678          */
 679         if ((flags & IO_FLAG_RW) &&
 680             (uname(&ut) == 0) &&
 681             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 682              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 683              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 684              (ut.release[5] < '8')) &&
 685             (ext2fs_fstat(data->dev, &st) == 0) &&
 686             (ext2fsP_is_disk_device(st.st_mode))) {
 687                 struct rlimit   rlim;
 688
 689                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 690                 setrlimit(RLIMIT_FSIZE, &rlim);
 691                 getrlimit(RLIMIT_FSIZE, &rlim);
 692                 if (((unsigned long) rlim.rlim_cur) <
 693                     ((unsigned long) rlim.rlim_max)) {
 694                         rlim.rlim_cur = rlim.rlim_max;
 695                         setrlimit(RLIMIT_FSIZE, &rlim);
 696                 }
 697         }
 698 #endif
 699         *channel = io;
 700         return 0;
 701
 702 cleanup:
 703         if (data) {
 704                 if (data->dev >= 0)
 705                         close(data->dev);
 706                 free_cache(data);
 707                 ext2fs_free_mem(&data);
 708         }
 709         if (io) {
 710                 if (io->name) {
 711                         ext2fs_free_mem(&io->name);
 712                 }
 713                 ext2fs_free_mem(&io);
 714         }
 715         return retval;
 716 }
 717
 718 static errcode_t unixfd_open(const char *str_fd, int flags,
 719                              io_channel *channel)
 720 {
 721         int fd;
 722         int fd_flags;
 723
 724         fd = atoi(str_fd);
 725 #if defined(HAVE_FCNTL)
 726         fd_flags = fcntl(fd, F_GETFD);
 727         if (fd_flags == -1)
 728                 return -EBADF;
 729
 730         flags = 0;
 731         if (fd_flags & O_RDWR)
 732                 flags |= IO_FLAG_RW;
 733         if (fd_flags & O_EXCL)
 734                 flags |= IO_FLAG_EXCLUSIVE;
 735 #if defined(O_DIRECT)
 736         if (fd_flags & O_DIRECT)
 737                 flags |= IO_FLAG_DIRECT_IO;
 738 #endif
 739 #endif  /* HAVE_FCNTL */
 740
 741         return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
 742 }
 743
 744 static errcode_t unix_open(const char *name, int flags,
 745                            io_channel *channel)
 746 {
 747         int fd = -1;
 748         int open_flags;
 749
 750         if (name == 0)
 751                 return EXT2_ET_BAD_DEVICE_NAME;
 752
 753         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 754         if (flags & IO_FLAG_EXCLUSIVE)
 755                 open_flags |= O_EXCL;
 756 #if defined(O_DIRECT)
 757         if (flags & IO_FLAG_DIRECT_IO)
 758                 open_flags |= O_DIRECT;
 759 #endif
 760         fd = ext2fs_open_file(name, open_flags, 0);
 761         if (fd < 0)
 762                 return errno;
 763 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
 764         if (flags & IO_FLAG_DIRECT_IO) {
 765                 if (fcntl(fd, F_NOCACHE, 1) < 0)
 766                         return errno;
 767         }
 768 #endif
 769         return unix_open_channel(name, fd, flags, channel, unix_io_manager);
 770 }
 771
 772 static errcode_t unix_close(io_channel channel)
 773 {
 774         struct unix_private_data *data;
 775         errcode_t       retval = 0;
 776
 777         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 778         data = (struct unix_private_data *) channel->private_data;
 779         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 780
 781         if (--channel->refcount > 0)
 782                 return 0;
 783
 784 #ifndef NO_IO_CACHE
 785         retval = flush_cached_blocks(channel, data, 0);
 786 #endif
 787
 788         if (close(data->dev) < 0)
 789                 retval = errno;
 790         free_cache(data);
 791
 792         ext2fs_free_mem(&channel->private_data);
 793         if (channel->name)
 794                 ext2fs_free_mem(&channel->name);
 795         ext2fs_free_mem(&channel);
 796         return retval;
 797 }
 798
 799 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 800 {
 801         struct unix_private_data *data;
 802         errcode_t               retval;
 803
 804         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 805         data = (struct unix_private_data *) channel->private_data;
 806         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 807
 808         if (channel->block_size != blksize) {
 809 #ifndef NO_IO_CACHE
 810                 if ((retval = flush_cached_blocks(channel, data, 0)))
 811                         return retval;
 812 #endif
 813
 814                 channel->block_size = blksize;
 815                 free_cache(data);
 816                 if ((retval = alloc_cache(channel, data)))
 817                         return retval;
 818         }
 819         return 0;
 820 }
 821
 822 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 823                                int count, void *buf)
 824 {
 825         struct unix_private_data *data;
 826         struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
 827         errcode_t       retval;
 828         char            *cp;
 829         int             i, j;
 830
 831         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 832         data = (struct unix_private_data *) channel->private_data;
 833         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 834
 835 #ifdef NO_IO_CACHE
 836         return raw_read_blk(channel, data, block, count, buf);
 837 #else
 838         /*
 839          * If we're doing an odd-sized read or a very large read,
 840          * flush out the cache and then do a direct read.
 841          */
 842         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 843                 if ((retval = flush_cached_blocks(channel, data, 0)))
 844                         return retval;
 845                 return raw_read_blk(channel, data, block, count, buf);
 846         }
 847
 848         cp = buf;
 849         while (count > 0) {
 850                 /* If it's in the cache, use it! */
 851                 if ((cache = find_cached_block(data, block, &reuse[0]))) {
 852 #ifdef DEBUG
 853                         printf("Using cached block %lu\n", block);
 854 #endif
 855                         memcpy(cp, cache->buf, channel->block_size);
 856                         count--;
 857                         block++;
 858                         cp += channel->block_size;
 859                         continue;
 860                 }
 861                 if (count == 1) {
 862                         /*
 863                          * Special case where we read directly into the
 864                          * cache buffer; important in the O_DIRECT case
 865                          */
 866                         cache = reuse[0];
 867                         reuse_cache(channel, data, cache, block);
 868                         if ((retval = raw_read_blk(channel, data, block, 1,
 869                                                    cache->buf))) {
 870                                 cache->in_use = 0;
 871                                 return retval;
 872                         }
 873                         memcpy(cp, cache->buf, channel->block_size);
 874                         return 0;
 875                 }
 876
 877                 /*
 878                  * Find the number of uncached blocks so we can do a
 879                  * single read request
 880                  */
 881                 for (i=1; i < count; i++)
 882                         if (find_cached_block(data, block+i, &reuse[i]))
 883                                 break;
 884 #ifdef DEBUG
 885                 printf("Reading %d blocks starting at %lu\n", i, block);
 886 #endif
 887                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
 888                         return retval;
 889
 890                 /* Save the results in the cache */
 891                 for (j=0; j < i; j++) {
 892                         count--;
 893                         cache = reuse[j];
 894                         reuse_cache(channel, data, cache, block++);
 895                         memcpy(cache->buf, cp, channel->block_size);
 896                         cp += channel->block_size;
 897                 }
 898         }
 899         return 0;
 900 #endif /* NO_IO_CACHE */
 901 }
 902
 903 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
 904                                int count, void *buf)
 905 {
 906         return unix_read_blk64(channel, block, count, buf);
 907 }
 908
 909 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
 910                                 int count, const void *buf)
 911 {
 912         struct unix_private_data *data;
 913         struct unix_cache *cache, *reuse;
 914         errcode_t       retval = 0;
 915         const char      *cp;
 916         int             writethrough;
 917
 918         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 919         data = (struct unix_private_data *) channel->private_data;
 920         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 921
 922 #ifdef NO_IO_CACHE
 923         return raw_write_blk(channel, data, block, count, buf);
 924 #else
 925         /*
 926          * If we're doing an odd-sized write or a very large write,
 927          * flush out the cache completely and then do a direct write.
 928          */
 929         if (count < 0 || count > WRITE_DIRECT_SIZE) {
 930                 if ((retval = flush_cached_blocks(channel, data, 1)))
 931                         return retval;
 932                 return raw_write_blk(channel, data, block, count, buf);
 933         }
 934
 935         /*
 936          * For a moderate-sized multi-block write, first force a write
 937          * if we're in write-through cache mode, and then fill the
 938          * cache with the blocks.
 939          */
 940         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
 941         if (writethrough)
 942                 retval = raw_write_blk(channel, data, block, count, buf);
 943
 944         cp = buf;
 945         while (count > 0) {
 946                 cache = find_cached_block(data, block, &reuse);
 947                 if (!cache) {
 948                         cache = reuse;
 949                         reuse_cache(channel, data, cache, block);
 950                 }
 951                 if (cache->buf != cp)
 952                         memcpy(cache->buf, cp, channel->block_size);
 953                 cache->dirty = !writethrough;
 954                 count--;
 955                 block++;
 956                 cp += channel->block_size;
 957         }
 958         return retval;
 959 #endif /* NO_IO_CACHE */
 960 }
 961
 962 static errcode_t unix_cache_readahead(io_channel channel,
 963                                       unsigned long long block,
 964                                       unsigned long long count)
 965 {
 966 #ifdef POSIX_FADV_WILLNEED
 967         struct unix_private_data *data;
 968
 969         data = (struct unix_private_data *)channel->private_data;
 970         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 971         return posix_fadvise(data->dev,
 972                              (ext2_loff_t)block * channel->block_size + data->offset,
 973                              (ext2_loff_t)count * channel->block_size,
 974                              POSIX_FADV_WILLNEED);
 975 #else
 976         return EXT2_ET_OP_NOT_SUPPORTED;
 977 #endif
 978 }
 979
 980 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 981                                 int count, const void *buf)
 982 {
 983         return unix_write_blk64(channel, block, count, buf);
 984 }
 985
 986 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 987                                  int size, const void *buf)
 988 {
 989         struct unix_private_data *data;
 990         errcode_t       retval = 0;
 991         ssize_t         actual;
 992
 993         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 994         data = (struct unix_private_data *) channel->private_data;
 995         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 996
 997         if (channel->align != 0) {
 998 #ifdef ALIGN_DEBUG
 999                 printf("unix_write_byte: O_DIRECT fallback\n");
1000 #endif
1001                 return EXT2_ET_UNIMPLEMENTED;
1002         }
1003
1004 #ifndef NO_IO_CACHE
1005         /*
1006          * Flush out the cache completely
1007          */
1008         if ((retval = flush_cached_blocks(channel, data, 1)))
1009                 return retval;
1010 #endif
1011
1012         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1013                 return errno;
1014
1015         actual = write(data->dev, buf, size);
1016         if (actual < 0)
1017                 return errno;
1018         if (actual != size)
1019                 return EXT2_ET_SHORT_WRITE;
1020
1021         return 0;
1022 }
1023
1024 /*
1025  * Flush data buffers to disk.
1026  */
1027 static errcode_t unix_flush(io_channel channel)
1028 {
1029         struct unix_private_data *data;
1030         errcode_t retval = 0;
1031
1032         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1033         data = (struct unix_private_data *) channel->private_data;
1034         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1035
1036 #ifndef NO_IO_CACHE
1037         retval = flush_cached_blocks(channel, data, 0);
1038 #endif
1039 #ifdef HAVE_FSYNC
1040         if (!retval && fsync(data->dev) != 0)
1041                 return errno;
1042 #endif
1043         return retval;
1044 }
1045
1046 static errcode_t unix_set_option(io_channel channel, const char *option,
1047                                  const char *arg)
1048 {
1049         struct unix_private_data *data;
1050         unsigned long long tmp;
1051         char *end;
1052
1053         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1054         data = (struct unix_private_data *) channel->private_data;
1055         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1056
1057         if (!strcmp(option, "offset")) {
1058                 if (!arg)
1059                         return EXT2_ET_INVALID_ARGUMENT;
1060
1061                 tmp = strtoull(arg, &end, 0);
1062                 if (*end)
1063                         return EXT2_ET_INVALID_ARGUMENT;
1064                 data->offset = tmp;
1065                 if (data->offset < 0)
1066                         return EXT2_ET_INVALID_ARGUMENT;
1067                 return 0;
1068         }
1069         return EXT2_ET_INVALID_ARGUMENT;
1070 }
1071
1072 #if defined(__linux__) && !defined(BLKDISCARD)
1073 #define BLKDISCARD              _IO(0x12,119)
1074 #endif
1075
1076 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1077                               unsigned long long count)
1078 {
1079         struct unix_private_data *data;
1080         int             ret;
1081
1082         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1083         data = (struct unix_private_data *) channel->private_data;
1084         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1085
1086         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1087 #ifdef BLKDISCARD
1088                 __u64 range[2];
1089
1090                 range[0] = (__u64)(block) * channel->block_size + data->offset;
1091                 range[1] = (__u64)(count) * channel->block_size;
1092
1093                 ret = ioctl(data->dev, BLKDISCARD, &range);
1094 #else
1095                 goto unimplemented;
1096 #endif
1097         } else {
1098 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1099                 /*
1100                  * If we are not on block device, try to use punch hole
1101                  * to reclaim free space.
1102                  */
1103                 ret = fallocate(data->dev,
1104                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1105                                 (off_t)(block) * channel->block_size + data->offset,
1106                                 (off_t)(count) * channel->block_size);
1107 #else
1108                 goto unimplemented;
1109 #endif
1110         }
1111         if (ret < 0) {
1112                 if (errno == EOPNOTSUPP)
1113                         goto unimplemented;
1114                 return errno;
1115         }
1116         return 0;
1117 unimplemented:
1118         return EXT2_ET_UNIMPLEMENTED;
1119 }
1120
1121 /* parameters might not be used if OS doesn't support zeroout */
1122 #pragma GCC diagnostic push
1123 #pragma GCC diagnostic ignored "-Wunused-parameter"
1124 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1125                               unsigned long long count)
1126 {
1127         struct unix_private_data *data;
1128         int             ret;
1129
1130         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1131         data = (struct unix_private_data *) channel->private_data;
1132         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1133
1134         if (safe_getenv("UNIX_IO_NOZEROOUT"))
1135                 goto unimplemented;
1136
1137         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1138                 /* Not implemented until the BLKZEROOUT mess is fixed */
1139                 goto unimplemented;
1140         } else {
1141                 /* Regular file, try to use truncate/punch/zero. */
1142                 struct stat statbuf;
1143
1144                 if (count == 0)
1145                         return 0;
1146                 /*
1147                  * If we're trying to zero a range past the end of the file,
1148                  * extend the file size, then truncate everything.
1149                  */
1150                 ret = fstat(data->dev, &statbuf);
1151                 if (ret)
1152                         goto err;
1153                 if ((unsigned long long) statbuf.st_size <
1154                         (block + count) * channel->block_size + data->offset) {
1155                         ret = ftruncate(data->dev,
1156                                         (block + count) * channel->block_size + data->offset);
1157                         if (ret)
1158                                 goto err;
1159                 }
1160 #if defined(HAVE_FALLOCATE) && (defined(FALLOC_FL_ZERO_RANGE) || \
1161         (defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)))
1162 #if defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1163                 ret = fallocate(data->dev,
1164                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1165                                 (off_t)(block) * channel->block_size + data->offset,
1166                                 (off_t)(count) * channel->block_size);
1167                 if (ret == 0)
1168                         goto err;
1169 #endif
1170 #ifdef FALLOC_FL_ZERO_RANGE
1171                 ret = fallocate(data->dev,
1172                                 FALLOC_FL_ZERO_RANGE,
1173                                 (off_t)(block) * channel->block_size + data->offset,
1174                                 (off_t)(count) * channel->block_size);
1175 #endif
1176 #else
1177                 goto unimplemented;
1178 #endif /* HAVE_FALLOCATE && (ZERO_RANGE || (PUNCH_HOLE && KEEP_SIZE)) */
1179         }
1180 err:
1181         if (ret < 0) {
1182                 if (errno == EOPNOTSUPP)
1183                         goto unimplemented;
1184                 return errno;
1185         }
1186         return 0;
1187 unimplemented:
1188         return EXT2_ET_UNIMPLEMENTED;
1189 }
1190 #pragma GCC diagnostic pop
1191
1192 static struct struct_io_manager struct_unix_manager = {
1193         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1194         .name           = "Unix I/O Manager",
1195         .open           = unix_open,
1196         .close          = unix_close,
1197         .set_blksize    = unix_set_blksize,
1198         .read_blk       = unix_read_blk,
1199         .write_blk      = unix_write_blk,
1200         .flush          = unix_flush,
1201         .write_byte     = unix_write_byte,
1202         .set_option     = unix_set_option,
1203         .get_stats      = unix_get_stats,
1204         .read_blk64     = unix_read_blk64,
1205         .write_blk64    = unix_write_blk64,
1206         .discard        = unix_discard,
1207         .cache_readahead        = unix_cache_readahead,
1208         .zeroout        = unix_zeroout,
1209 };
1210
1211 io_manager unix_io_manager = &struct_unix_manager;
1212
1213 static struct struct_io_manager struct_unixfd_manager = {
1214         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1215         .name           = "Unix fd I/O Manager",
1216         .open           = unixfd_open,
1217         .close          = unix_close,
1218         .set_blksize    = unix_set_blksize,
1219         .read_blk       = unix_read_blk,
1220         .write_blk      = unix_write_blk,
1221         .flush          = unix_flush,
1222         .write_byte     = unix_write_byte,
1223         .set_option     = unix_set_option,
1224         .get_stats      = unix_get_stats,
1225         .read_blk64     = unix_read_blk64,
1226         .write_blk64    = unix_write_blk64,
1227         .discard        = unix_discard,
1228         .cache_readahead        = unix_cache_readahead,
1229         .zeroout        = unix_zeroout,
1230 };
1231
1232 io_manager unixfd_io_manager = &struct_unixfd_manager;