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 #ifdef HAVE_PTHREAD
  71 #include <pthread.h>
  72 #endif
  73
  74 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
  75 #define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
  76 #endif
  77
  78 #undef ALIGN_DEBUG
  79
  80 #include "ext2_fs.h"
  81 #include "ext2fs.h"
  82 #include "ext2fsP.h"
  83
  84 /*
  85  * For checking structure magic numbers...
  86  */
  87
  88 #define EXT2_CHECK_MAGIC(struct, code) \
  89           if ((struct)->magic != (code)) return (code)
  90
  91 struct unix_cache {
  92         char                    *buf;
  93         unsigned long long      block;
  94         int                     access_time;
  95         unsigned                dirty:1;
  96         unsigned                in_use:1;
  97         unsigned                write_err:1;
  98 };
  99
 100 #define CACHE_SIZE 8
 101 #define WRITE_DIRECT_SIZE 4     /* Must be smaller than CACHE_SIZE */
 102 #define READ_DIRECT_SIZE 4      /* Should be smaller than CACHE_SIZE */
 103
 104 struct unix_private_data {
 105         int     magic;
 106         int     dev;
 107         int     flags;
 108         int     align;
 109         int     access_time;
 110         ext2_loff_t offset;
 111         struct unix_cache cache[CACHE_SIZE];
 112         void    *bounce;
 113         struct struct_io_stats io_stats;
 114 #ifdef HAVE_PTHREAD
 115         pthread_mutex_t cache_mutex;
 116         pthread_mutex_t bounce_mutex;
 117         pthread_mutex_t stats_mutex;
 118 #endif
 119 };
 120
 121 #define IS_ALIGNED(n, align) ((((uintptr_t) n) & \
 122                                ((uintptr_t) ((align)-1))) == 0)
 123
 124 typedef enum lock_kind {
 125         CACHE_MTX, BOUNCE_MTX, STATS_MTX
 126 } kind_t;
 127
 128 #ifdef HAVE_PTHREAD
 129 static inline pthread_mutex_t *get_mutex(struct unix_private_data *data,
 130                                          kind_t kind)
 131 {
 132         if (data->flags & IO_FLAG_THREADS) {
 133                 switch (kind) {
 134                 case CACHE_MTX:
 135                         return &data->cache_mutex;
 136                 case BOUNCE_MTX:
 137                         return &data->bounce_mutex;
 138                 case STATS_MTX:
 139                         return &data->stats_mutex;
 140                 }
 141         }
 142         return NULL;
 143 }
 144 #endif
 145
 146 static inline void mutex_lock(struct unix_private_data *data, kind_t kind)
 147 {
 148 #ifdef HAVE_PTHREAD
 149         pthread_mutex_t *mtx = get_mutex(data,kind);
 150
 151         if (mtx)
 152                 pthread_mutex_lock(mtx);
 153 #endif
 154 }
 155
 156 static inline void mutex_unlock(struct unix_private_data *data, kind_t kind)
 157 {
 158 #ifdef HAVE_PTHREAD
 159         pthread_mutex_t *mtx = get_mutex(data,kind);
 160
 161         if (mtx)
 162                 pthread_mutex_unlock(mtx);
 163 #endif
 164 }
 165
 166 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 167 {
 168         errcode_t       retval = 0;
 169
 170         struct unix_private_data *data;
 171
 172         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 173         data = (struct unix_private_data *) channel->private_data;
 174         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 175
 176         if (stats) {
 177                 mutex_lock(data, STATS_MTX);
 178                 *stats = &data->io_stats;
 179                 mutex_unlock(data, STATS_MTX);
 180         }
 181
 182         return retval;
 183 }
 184
 185 static char *safe_getenv(const char *arg)
 186 {
 187         if ((getuid() != geteuid()) || (getgid() != getegid()))
 188                 return NULL;
 189 #ifdef HAVE_PRCTL
 190         if (prctl(PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 191                 return NULL;
 192 #else
 193 #if (defined(linux) && defined(SYS_prctl))
 194         if (syscall(SYS_prctl, PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
 195                 return NULL;
 196 #endif
 197 #endif
 198
 199 #if defined(HAVE_SECURE_GETENV)
 200         return secure_getenv(arg);
 201 #elif defined(HAVE___SECURE_GETENV)
 202         return __secure_getenv(arg);
 203 #else
 204         return getenv(arg);
 205 #endif
 206 }
 207
 208 /*
 209  * Here are the raw I/O functions
 210  */
 211 static errcode_t raw_read_blk(io_channel channel,
 212                               struct unix_private_data *data,
 213                               unsigned long long block,
 214                               int count, void *bufv)
 215 {
 216         errcode_t       retval;
 217         ssize_t         size;
 218         ext2_loff_t     location;
 219         int             actual = 0;
 220         unsigned char   *buf = bufv;
 221         ssize_t         really_read = 0;
 222         unsigned long long aligned_blk;
 223         int             align_size, offset;
 224
 225         size = (count < 0) ? -count : (ext2_loff_t) count * channel->block_size;
 226         mutex_lock(data, STATS_MTX);
 227         data->io_stats.bytes_read += size;
 228         mutex_unlock(data, STATS_MTX);
 229         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 230
 231         if (data->flags & IO_FLAG_FORCE_BOUNCE)
 232                 goto bounce_read;
 233
 234 #ifdef HAVE_PREAD64
 235         /* Try an aligned pread */
 236         if ((channel->align == 0) ||
 237             (IS_ALIGNED(buf, channel->align) &&
 238              IS_ALIGNED(location, channel->align) &&
 239              IS_ALIGNED(size, channel->align))) {
 240                 actual = pread64(data->dev, buf, size, location);
 241                 if (actual == size)
 242                         return 0;
 243                 actual = 0;
 244         }
 245 #elif HAVE_PREAD
 246         /* Try an aligned pread */
 247         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 248             ((channel->align == 0) ||
 249              (IS_ALIGNED(buf, channel->align) &&
 250               IS_ALIGNED(location, channel->align) &&
 251               IS_ALIGNED(size, channel->align)))) {
 252                 actual = pread(data->dev, buf, size, location);
 253                 if (actual == size)
 254                         return 0;
 255                 actual = 0;
 256         }
 257 #endif /* HAVE_PREAD */
 258
 259         if ((channel->align == 0) ||
 260             (IS_ALIGNED(buf, channel->align) &&
 261              IS_ALIGNED(location, channel->align) &&
 262              IS_ALIGNED(size, channel->align))) {
 263                 mutex_lock(data, BOUNCE_MTX);
 264                 if (ext2fs_llseek(data->dev, location, SEEK_SET) < 0) {
 265                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 266                         goto error_unlock;
 267                 }
 268                 actual = read(data->dev, buf, size);
 269                 if (actual != size) {
 270                 short_read:
 271                         if (actual < 0) {
 272                                 retval = errno;
 273                                 actual = 0;
 274                         } else
 275                                 retval = EXT2_ET_SHORT_READ;
 276                         goto error_unlock;
 277                 }
 278                 goto success_unlock;
 279         }
 280
 281 #ifdef ALIGN_DEBUG
 282         printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
 283                (unsigned long) size);
 284 #endif
 285
 286         /*
 287          * The buffer or size which we're trying to read isn't aligned
 288          * to the O_DIRECT rules, so we need to do this the hard way...
 289          */
 290 bounce_read:
 291         if (channel->align == 0)
 292                 channel->align = 1;
 293         if ((channel->block_size > channel->align) &&
 294             (channel->block_size % channel->align) == 0)
 295                 align_size = channel->block_size;
 296         else
 297                 align_size = channel->align;
 298         aligned_blk = location / align_size;
 299         offset = location % align_size;
 300
 301         mutex_lock(data, BOUNCE_MTX);
 302         if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
 303                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 304                 goto error_unlock;
 305         }
 306         while (size > 0) {
 307                 actual = read(data->dev, data->bounce, align_size);
 308                 if (actual != align_size) {
 309                         mutex_unlock(data, BOUNCE_MTX);
 310                         actual = really_read;
 311                         buf -= really_read;
 312                         size += really_read;
 313                         goto short_read;
 314                 }
 315                 if ((actual + offset) > align_size)
 316                         actual = align_size - offset;
 317                 if (actual > size)
 318                         actual = size;
 319                 memcpy(buf, (char *)data->bounce + offset, actual);
 320
 321                 really_read += actual;
 322                 size -= actual;
 323                 buf += actual;
 324                 offset = 0;
 325                 aligned_blk++;
 326         }
 327 success_unlock:
 328         mutex_unlock(data, BOUNCE_MTX);
 329         return 0;
 330
 331 error_unlock:
 332         mutex_unlock(data, BOUNCE_MTX);
 333         if (actual >= 0 && actual < size)
 334                 memset((char *) buf+actual, 0, size-actual);
 335         if (channel->read_error)
 336                 retval = (channel->read_error)(channel, block, count, buf,
 337                                                size, actual, retval);
 338         return retval;
 339 }
 340
 341 #define RAW_WRITE_NO_HANDLER    1
 342
 343 static errcode_t raw_write_blk(io_channel channel,
 344                                struct unix_private_data *data,
 345                                unsigned long long block,
 346                                int count, const void *bufv,
 347                                int flags)
 348 {
 349         ssize_t         size;
 350         ext2_loff_t     location;
 351         int             actual = 0;
 352         errcode_t       retval;
 353         const unsigned char *buf = bufv;
 354         unsigned long long aligned_blk;
 355         int             align_size, offset;
 356
 357         if (count == 1)
 358                 size = channel->block_size;
 359         else {
 360                 if (count < 0)
 361                         size = -count;
 362                 else
 363                         size = (ext2_loff_t) count * channel->block_size;
 364         }
 365         mutex_lock(data, STATS_MTX);
 366         data->io_stats.bytes_written += size;
 367         mutex_unlock(data, STATS_MTX);
 368
 369         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 370
 371         if (data->flags & IO_FLAG_FORCE_BOUNCE)
 372                 goto bounce_write;
 373
 374 #ifdef HAVE_PWRITE64
 375         /* Try an aligned pwrite */
 376         if ((channel->align == 0) ||
 377             (IS_ALIGNED(buf, channel->align) &&
 378              IS_ALIGNED(location, channel->align) &&
 379              IS_ALIGNED(size, channel->align))) {
 380                 actual = pwrite64(data->dev, buf, size, location);
 381                 if (actual == size)
 382                         return 0;
 383         }
 384 #elif HAVE_PWRITE
 385         /* Try an aligned pwrite */
 386         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 387             ((channel->align == 0) ||
 388              (IS_ALIGNED(buf, channel->align) &&
 389               IS_ALIGNED(location, channel->align) &&
 390               IS_ALIGNED(size, channel->align)))) {
 391                 actual = pwrite(data->dev, buf, size, location);
 392                 if (actual == size)
 393                         return 0;
 394         }
 395 #endif /* HAVE_PWRITE */
 396
 397         if ((channel->align == 0) ||
 398             (IS_ALIGNED(buf, channel->align) &&
 399              IS_ALIGNED(location, channel->align) &&
 400              IS_ALIGNED(size, channel->align))) {
 401                 mutex_lock(data, BOUNCE_MTX);
 402                 if (ext2fs_llseek(data->dev, location, SEEK_SET) < 0) {
 403                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 404                         goto error_unlock;
 405                 }
 406                 actual = write(data->dev, buf, size);
 407                 mutex_unlock(data, BOUNCE_MTX);
 408                 if (actual < 0) {
 409                         retval = errno;
 410                         goto error_out;
 411                 }
 412                 if (actual != size) {
 413                 short_write:
 414                         retval = EXT2_ET_SHORT_WRITE;
 415                         goto error_out;
 416                 }
 417                 return 0;
 418         }
 419
 420 #ifdef ALIGN_DEBUG
 421         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 422                (unsigned long) size);
 423 #endif
 424         /*
 425          * The buffer or size which we're trying to write isn't aligned
 426          * to the O_DIRECT rules, so we need to do this the hard way...
 427          */
 428 bounce_write:
 429         if (channel->align == 0)
 430                 channel->align = 1;
 431         if ((channel->block_size > channel->align) &&
 432             (channel->block_size % channel->align) == 0)
 433                 align_size = channel->block_size;
 434         else
 435                 align_size = channel->align;
 436         aligned_blk = location / align_size;
 437         offset = location % align_size;
 438
 439         while (size > 0) {
 440                 int actual_w;
 441
 442                 mutex_lock(data, BOUNCE_MTX);
 443                 if (size < align_size || offset) {
 444                         if (ext2fs_llseek(data->dev, aligned_blk * align_size,
 445                                           SEEK_SET) < 0) {
 446                                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 447                                 goto error_unlock;
 448                         }
 449                         actual = read(data->dev, data->bounce,
 450                                       align_size);
 451                         if (actual != align_size) {
 452                                 if (actual < 0) {
 453                                         retval = errno;
 454                                         goto error_unlock;
 455                                 }
 456                                 memset((char *) data->bounce + actual, 0,
 457                                        align_size - actual);
 458                         }
 459                 }
 460                 actual = size;
 461                 if ((actual + offset) > align_size)
 462                         actual = align_size - offset;
 463                 if (actual > size)
 464                         actual = size;
 465                 memcpy(((char *)data->bounce) + offset, buf, actual);
 466                 if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
 467                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 468                         goto error_unlock;
 469                 }
 470                 actual_w = write(data->dev, data->bounce, align_size);
 471                 mutex_unlock(data, BOUNCE_MTX);
 472                 if (actual_w < 0) {
 473                         retval = errno;
 474                         goto error_out;
 475                 }
 476                 if (actual_w != align_size)
 477                         goto short_write;
 478                 size -= actual;
 479                 buf += actual;
 480                 location += actual;
 481                 aligned_blk++;
 482                 offset = 0;
 483         }
 484         return 0;
 485
 486 error_unlock:
 487         mutex_unlock(data, BOUNCE_MTX);
 488 error_out:
 489         if (((flags & RAW_WRITE_NO_HANDLER) == 0) && channel->write_error)
 490                 retval = (channel->write_error)(channel, block, count, buf,
 491                                                 size, actual, retval);
 492         return retval;
 493 }
 494
 495
 496 /*
 497  * Here we implement the cache functions
 498  */
 499
 500 /* Allocate the cache buffers */
 501 static errcode_t alloc_cache(io_channel channel,
 502                              struct unix_private_data *data)
 503 {
 504         errcode_t               retval;
 505         struct unix_cache       *cache;
 506         int                     i;
 507
 508         data->access_time = 0;
 509         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 510                 cache->block = 0;
 511                 cache->access_time = 0;
 512                 cache->dirty = 0;
 513                 cache->in_use = 0;
 514                 if (cache->buf)
 515                         ext2fs_free_mem(&cache->buf);
 516                 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
 517                 if (retval)
 518                         return retval;
 519         }
 520         if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
 521                 if (data->bounce)
 522                         ext2fs_free_mem(&data->bounce);
 523                 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
 524         }
 525         return retval;
 526 }
 527
 528 /* Free the cache buffers */
 529 static void free_cache(struct unix_private_data *data)
 530 {
 531         struct unix_cache       *cache;
 532         int                     i;
 533
 534         data->access_time = 0;
 535         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 536                 cache->block = 0;
 537                 cache->access_time = 0;
 538                 cache->dirty = 0;
 539                 cache->in_use = 0;
 540                 if (cache->buf)
 541                         ext2fs_free_mem(&cache->buf);
 542         }
 543         if (data->bounce)
 544                 ext2fs_free_mem(&data->bounce);
 545 }
 546
 547 #ifndef NO_IO_CACHE
 548 /*
 549  * Try to find a block in the cache.  If the block is not found, and
 550  * eldest is a non-zero pointer, then fill in eldest with the cache
 551  * entry to that should be reused.
 552  */
 553 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 554                                             unsigned long long block,
 555                                             struct unix_cache **eldest)
 556 {
 557         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 558         int                     i;
 559
 560         unused_cache = oldest_cache = 0;
 561         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 562                 if (!cache->in_use) {
 563                         if (!unused_cache)
 564                                 unused_cache = cache;
 565                         continue;
 566                 }
 567                 if (cache->block == block) {
 568                         cache->access_time = ++data->access_time;
 569                         return cache;
 570                 }
 571                 if (!oldest_cache ||
 572                     (cache->access_time < oldest_cache->access_time))
 573                         oldest_cache = cache;
 574         }
 575         if (eldest)
 576                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 577         return 0;
 578 }
 579
 580 /*
 581  * Reuse a particular cache entry for another block.
 582  */
 583 static errcode_t reuse_cache(io_channel channel,
 584                 struct unix_private_data *data, struct unix_cache *cache,
 585                 unsigned long long block)
 586 {
 587         if (cache->dirty && cache->in_use) {
 588                 errcode_t retval;
 589
 590                 retval = raw_write_blk(channel, data, cache->block, 1,
 591                                        cache->buf, RAW_WRITE_NO_HANDLER);
 592                 if (retval) {
 593                         cache->write_err = 1;
 594                         return retval;
 595                 }
 596         }
 597
 598         cache->in_use = 1;
 599         cache->dirty = 0;
 600         cache->write_err = 0;
 601         cache->block = block;
 602         cache->access_time = ++data->access_time;
 603         return 0;
 604 }
 605
 606 #define FLUSH_INVALIDATE        0x01
 607 #define FLUSH_NOLOCK            0x02
 608
 609 /*
 610  * Flush all of the blocks in the cache
 611  */
 612 static errcode_t flush_cached_blocks(io_channel channel,
 613                                      struct unix_private_data *data,
 614                                      int flags)
 615 {
 616         struct unix_cache       *cache;
 617         errcode_t               retval, retval2;
 618         int                     i;
 619
 620         retval2 = 0;
 621         if ((flags & FLUSH_NOLOCK) == 0)
 622                 mutex_lock(data, CACHE_MTX);
 623         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 624                 if (!cache->in_use)
 625                         continue;
 626
 627                 if (flags & FLUSH_INVALIDATE)
 628                         cache->in_use = 0;
 629
 630                 if (!cache->dirty)
 631                         continue;
 632
 633                 retval = raw_write_blk(channel, data,
 634                                        cache->block, 1, cache->buf, 0);
 635                 if (retval)
 636                         retval2 = retval;
 637                 else
 638                         cache->dirty = 0;
 639         }
 640         if ((flags & FLUSH_NOLOCK) == 0)
 641                 mutex_unlock(data, CACHE_MTX);
 642         return retval2;
 643 }
 644 #endif /* NO_IO_CACHE */
 645
 646 #ifdef __linux__
 647 #ifndef BLKDISCARDZEROES
 648 #define BLKDISCARDZEROES _IO(0x12,124)
 649 #endif
 650 #endif
 651
 652 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
 653 {
 654         if (mode)
 655 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 656                 return open64(pathname, flags, mode);
 657         else
 658                 return open64(pathname, flags);
 659 #else
 660                 return open(pathname, flags, mode);
 661         else
 662                 return open(pathname, flags);
 663 #endif
 664 }
 665
 666 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
 667 {
 668 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 669         return stat64(path, buf);
 670 #else
 671         return stat(path, buf);
 672 #endif
 673 }
 674
 675 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
 676 {
 677 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 678         return fstat64(fd, buf);
 679 #else
 680         return fstat(fd, buf);
 681 #endif
 682 }
 683
 684
 685 static errcode_t unix_open_channel(const char *name, int fd,
 686                                    int flags, io_channel *channel,
 687                                    io_manager io_mgr)
 688 {
 689         io_channel      io = NULL;
 690         struct unix_private_data *data = NULL;
 691         errcode_t       retval;
 692         ext2fs_struct_stat st;
 693 #ifdef __linux__
 694         struct          utsname ut;
 695 #endif
 696
 697         if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
 698                 flags |= IO_FLAG_FORCE_BOUNCE;
 699
 700 #ifdef __linux__
 701         /*
 702          * We need to make sure any previous errors in the block
 703          * device are thrown away, sigh.
 704          */
 705         (void) fsync(fd);
 706 #endif
 707
 708         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 709         if (retval)
 710                 goto cleanup;
 711         memset(io, 0, sizeof(struct struct_io_channel));
 712         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 713         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 714         if (retval)
 715                 goto cleanup;
 716
 717         io->manager = io_mgr;
 718         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 719         if (retval)
 720                 goto cleanup;
 721
 722         strcpy(io->name, name);
 723         io->private_data = data;
 724         io->block_size = 1024;
 725         io->read_error = 0;
 726         io->write_error = 0;
 727         io->refcount = 1;
 728         io->flags = 0;
 729
 730         memset(data, 0, sizeof(struct unix_private_data));
 731         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 732         data->io_stats.num_fields = 2;
 733         data->flags = flags;
 734         data->dev = fd;
 735
 736 #if defined(O_DIRECT)
 737         if (flags & IO_FLAG_DIRECT_IO)
 738                 io->align = ext2fs_get_dio_alignment(data->dev);
 739 #elif defined(F_NOCACHE)
 740         if (flags & IO_FLAG_DIRECT_IO)
 741                 io->align = 4096;
 742 #endif
 743
 744         /*
 745          * If the device is really a block device, then set the
 746          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 747          * because we are going to use punch hole instead of discard
 748          * and if it succeed, subsequent read from sparse area returns
 749          * zero.
 750          */
 751         if (ext2fs_fstat(data->dev, &st) == 0) {
 752                 if (ext2fsP_is_disk_device(st.st_mode))
 753                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 754                 else
 755                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 756         }
 757
 758 #ifdef BLKDISCARDZEROES
 759         {
 760                 int zeroes = 0;
 761                 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 762                     zeroes)
 763                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 764         }
 765 #endif
 766
 767 #if defined(__CYGWIN__)
 768         /*
 769          * Some operating systems require that the buffers be aligned,
 770          * regardless of O_DIRECT
 771          */
 772         if (!io->align)
 773                 io->align = 512;
 774 #endif
 775
 776 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 777         if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 778                 int dio_align = ext2fs_get_dio_alignment(fd);
 779
 780                 if (io->align < dio_align)
 781                         io->align = dio_align;
 782         }
 783 #endif
 784
 785         if ((retval = alloc_cache(io, data)))
 786                 goto cleanup;
 787
 788 #ifdef BLKROGET
 789         if (flags & IO_FLAG_RW) {
 790                 int error;
 791                 int readonly = 0;
 792
 793                 /* Is the block device actually writable? */
 794                 error = ioctl(data->dev, BLKROGET, &readonly);
 795                 if (!error && readonly) {
 796                         retval = EPERM;
 797                         goto cleanup;
 798                 }
 799         }
 800 #endif
 801
 802 #ifdef __linux__
 803 #undef RLIM_INFINITY
 804 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 805 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 806 #else
 807 #define RLIM_INFINITY  (~0UL)
 808 #endif
 809         /*
 810          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 811          * block devices are wrongly getting hit by the filesize
 812          * limit.  This workaround isn't perfect, since it won't work
 813          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 814          *
 815          */
 816         if ((flags & IO_FLAG_RW) &&
 817             (uname(&ut) == 0) &&
 818             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 819              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 820              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 821              (ut.release[5] < '8')) &&
 822             (ext2fs_fstat(data->dev, &st) == 0) &&
 823             (ext2fsP_is_disk_device(st.st_mode))) {
 824                 struct rlimit   rlim;
 825
 826                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 827                 setrlimit(RLIMIT_FSIZE, &rlim);
 828                 getrlimit(RLIMIT_FSIZE, &rlim);
 829                 if (((unsigned long) rlim.rlim_cur) <
 830                     ((unsigned long) rlim.rlim_max)) {
 831                         rlim.rlim_cur = rlim.rlim_max;
 832                         setrlimit(RLIMIT_FSIZE, &rlim);
 833                 }
 834         }
 835 #endif
 836 #ifdef HAVE_PTHREAD
 837         if (flags & IO_FLAG_THREADS) {
 838                 io->flags |= CHANNEL_FLAGS_THREADS;
 839                 retval = pthread_mutex_init(&data->cache_mutex, NULL);
 840                 if (retval)
 841                         goto cleanup;
 842                 retval = pthread_mutex_init(&data->bounce_mutex, NULL);
 843                 if (retval) {
 844                         pthread_mutex_destroy(&data->cache_mutex);
 845                         goto cleanup;
 846                 }
 847                 retval = pthread_mutex_init(&data->stats_mutex, NULL);
 848                 if (retval) {
 849                         pthread_mutex_destroy(&data->cache_mutex);
 850                         pthread_mutex_destroy(&data->bounce_mutex);
 851                         goto cleanup;
 852                 }
 853         }
 854 #endif
 855         *channel = io;
 856         return 0;
 857
 858 cleanup:
 859         if (data) {
 860                 if (data->dev >= 0)
 861                         close(data->dev);
 862                 free_cache(data);
 863                 ext2fs_free_mem(&data);
 864         }
 865         if (io) {
 866                 if (io->name) {
 867                         ext2fs_free_mem(&io->name);
 868                 }
 869                 ext2fs_free_mem(&io);
 870         }
 871         return retval;
 872 }
 873
 874 static errcode_t unixfd_open(const char *str_fd, int flags,
 875                              io_channel *channel)
 876 {
 877         int fd;
 878         int fd_flags;
 879
 880         fd = atoi(str_fd);
 881 #if defined(HAVE_FCNTL)
 882         fd_flags = fcntl(fd, F_GETFD);
 883         if (fd_flags == -1)
 884                 return EBADF;
 885
 886         flags = 0;
 887         if (fd_flags & O_RDWR)
 888                 flags |= IO_FLAG_RW;
 889         if (fd_flags & O_EXCL)
 890                 flags |= IO_FLAG_EXCLUSIVE;
 891 #if defined(O_DIRECT)
 892         if (fd_flags & O_DIRECT)
 893                 flags |= IO_FLAG_DIRECT_IO;
 894 #endif
 895 #endif  /* HAVE_FCNTL */
 896
 897         return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
 898 }
 899
 900 static errcode_t unix_open(const char *name, int flags,
 901                            io_channel *channel)
 902 {
 903         int fd = -1;
 904         int open_flags;
 905
 906         if (name == 0)
 907                 return EXT2_ET_BAD_DEVICE_NAME;
 908
 909         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 910         if (flags & IO_FLAG_EXCLUSIVE)
 911                 open_flags |= O_EXCL;
 912 #if defined(O_DIRECT)
 913         if (flags & IO_FLAG_DIRECT_IO)
 914                 open_flags |= O_DIRECT;
 915 #endif
 916         fd = ext2fs_open_file(name, open_flags, 0);
 917         if (fd < 0)
 918                 return errno;
 919 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
 920         if (flags & IO_FLAG_DIRECT_IO) {
 921                 if (fcntl(fd, F_NOCACHE, 1) < 0)
 922                         return errno;
 923         }
 924 #endif
 925         return unix_open_channel(name, fd, flags, channel, unix_io_manager);
 926 }
 927
 928 static errcode_t unix_close(io_channel channel)
 929 {
 930         struct unix_private_data *data;
 931         errcode_t       retval = 0;
 932
 933         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 934         data = (struct unix_private_data *) channel->private_data;
 935         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 936
 937         if (--channel->refcount > 0)
 938                 return 0;
 939
 940 #ifndef NO_IO_CACHE
 941         retval = flush_cached_blocks(channel, data, 0);
 942 #endif
 943
 944         if (close(data->dev) < 0)
 945                 retval = errno;
 946         free_cache(data);
 947 #ifdef HAVE_PTHREAD
 948         if (data->flags & IO_FLAG_THREADS) {
 949                 pthread_mutex_destroy(&data->cache_mutex);
 950                 pthread_mutex_destroy(&data->bounce_mutex);
 951                 pthread_mutex_destroy(&data->stats_mutex);
 952         }
 953 #endif
 954
 955         ext2fs_free_mem(&channel->private_data);
 956         if (channel->name)
 957                 ext2fs_free_mem(&channel->name);
 958         ext2fs_free_mem(&channel);
 959         return retval;
 960 }
 961
 962 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 963 {
 964         struct unix_private_data *data;
 965         errcode_t               retval = 0;
 966
 967         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 968         data = (struct unix_private_data *) channel->private_data;
 969         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 970
 971         if (channel->block_size != blksize) {
 972                 mutex_lock(data, CACHE_MTX);
 973                 mutex_lock(data, BOUNCE_MTX);
 974 #ifndef NO_IO_CACHE
 975                 if ((retval = flush_cached_blocks(channel, data, FLUSH_NOLOCK))){
 976                         mutex_unlock(data, BOUNCE_MTX);
 977                         mutex_unlock(data, CACHE_MTX);
 978                         return retval;
 979                 }
 980 #endif
 981
 982                 channel->block_size = blksize;
 983                 free_cache(data);
 984                 retval = alloc_cache(channel, data);
 985                 mutex_unlock(data, BOUNCE_MTX);
 986                 mutex_unlock(data, CACHE_MTX);
 987         }
 988         return retval;
 989 }
 990
 991 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
 992                                int count, void *buf)
 993 {
 994         struct unix_private_data *data;
 995         struct unix_cache *cache;
 996         errcode_t       retval;
 997         char            *cp;
 998         int             i, j;
 999
1000         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1001         data = (struct unix_private_data *) channel->private_data;
1002         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1003
1004 #ifdef NO_IO_CACHE
1005         return raw_read_blk(channel, data, block, count, buf);
1006 #else
1007         if (data->flags & IO_FLAG_NOCACHE)
1008                 return raw_read_blk(channel, data, block, count, buf);
1009         /*
1010          * If we're doing an odd-sized read or a very large read,
1011          * flush out the cache and then do a direct read.
1012          */
1013         if (count < 0 || count > WRITE_DIRECT_SIZE) {
1014                 if ((retval = flush_cached_blocks(channel, data, 0)))
1015                         return retval;
1016                 return raw_read_blk(channel, data, block, count, buf);
1017         }
1018
1019         cp = buf;
1020         mutex_lock(data, CACHE_MTX);
1021         while (count > 0) {
1022                 /* If it's in the cache, use it! */
1023                 if ((cache = find_cached_block(data, block, NULL))) {
1024 #ifdef DEBUG
1025                         printf("Using cached block %lu\n", block);
1026 #endif
1027                         memcpy(cp, cache->buf, channel->block_size);
1028                         count--;
1029                         block++;
1030                         cp += channel->block_size;
1031                         continue;
1032                 }
1033
1034                 /*
1035                  * Find the number of uncached blocks so we can do a
1036                  * single read request
1037                  */
1038                 for (i=1; i < count; i++)
1039                         if (find_cached_block(data, block+i, NULL))
1040                                 break;
1041 #ifdef DEBUG
1042                 printf("Reading %d blocks starting at %lu\n", i, block);
1043 #endif
1044                 mutex_unlock(data, CACHE_MTX);
1045                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
1046                         return retval;
1047                 mutex_lock(data, CACHE_MTX);
1048
1049                 /* Save the results in the cache */
1050                 for (j=0; j < i; j++) {
1051                         if (!find_cached_block(data, block, &cache)) {
1052                                 retval = reuse_cache(channel, data,
1053                                                      cache, block);
1054                                 if (retval)
1055                                         goto call_write_handler;
1056                                 memcpy(cache->buf, cp, channel->block_size);
1057                         }
1058                         count--;
1059                         block++;
1060                         cp += channel->block_size;
1061                 }
1062         }
1063         mutex_unlock(data, CACHE_MTX);
1064         return 0;
1065
1066 call_write_handler:
1067         if (cache->write_err && channel->write_error) {
1068                 char *err_buf = NULL;
1069                 unsigned long long err_block = cache->block;
1070
1071                 cache->dirty = 0;
1072                 cache->in_use = 0;
1073                 cache->write_err = 0;
1074                 if (io_channel_alloc_buf(channel, 0, &err_buf))
1075                         err_buf = NULL;
1076                 else
1077                         memcpy(err_buf, cache->buf, channel->block_size);
1078                 mutex_unlock(data, CACHE_MTX);
1079                 (channel->write_error)(channel, err_block, 1, err_buf,
1080                                        channel->block_size, -1,
1081                                        retval);
1082                 if (err_buf)
1083                         ext2fs_free_mem(&err_buf);
1084         } else
1085                 mutex_unlock(data, CACHE_MTX);
1086         return retval;
1087 #endif /* NO_IO_CACHE */
1088 }
1089
1090 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
1091                                int count, void *buf)
1092 {
1093         return unix_read_blk64(channel, block, count, buf);
1094 }
1095
1096 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
1097                                 int count, const void *buf)
1098 {
1099         struct unix_private_data *data;
1100         struct unix_cache *cache, *reuse;
1101         errcode_t       retval = 0;
1102         const char      *cp;
1103         int             writethrough;
1104
1105         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1106         data = (struct unix_private_data *) channel->private_data;
1107         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1108
1109 #ifdef NO_IO_CACHE
1110         return raw_write_blk(channel, data, block, count, buf, 0);
1111 #else
1112         if (data->flags & IO_FLAG_NOCACHE)
1113                 return raw_write_blk(channel, data, block, count, buf, 0);
1114         /*
1115          * If we're doing an odd-sized write or a very large write,
1116          * flush out the cache completely and then do a direct write.
1117          */
1118         if (count < 0 || count > WRITE_DIRECT_SIZE) {
1119                 if ((retval = flush_cached_blocks(channel, data,
1120                                                   FLUSH_INVALIDATE)))
1121                         return retval;
1122                 return raw_write_blk(channel, data, block, count, buf, 0);
1123         }
1124
1125         /*
1126          * For a moderate-sized multi-block write, first force a write
1127          * if we're in write-through cache mode, and then fill the
1128          * cache with the blocks.
1129          */
1130         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
1131         if (writethrough)
1132                 retval = raw_write_blk(channel, data, block, count, buf, 0);
1133
1134         cp = buf;
1135         mutex_lock(data, CACHE_MTX);
1136         while (count > 0) {
1137                 cache = find_cached_block(data, block, &reuse);
1138                 if (!cache) {
1139                         errcode_t err;
1140
1141                         cache = reuse;
1142                         err = reuse_cache(channel, data, cache, block);
1143                         if (err)
1144                                 goto call_write_handler;
1145                 }
1146                 if (cache->buf != cp)
1147                         memcpy(cache->buf, cp, channel->block_size);
1148                 cache->dirty = !writethrough;
1149                 count--;
1150                 block++;
1151                 cp += channel->block_size;
1152         }
1153         mutex_unlock(data, CACHE_MTX);
1154         return retval;
1155
1156 call_write_handler:
1157         if (cache->write_err && channel->write_error) {
1158                 char *err_buf = NULL;
1159                 unsigned long long err_block = cache->block;
1160
1161                 cache->dirty = 0;
1162                 cache->in_use = 0;
1163                 cache->write_err = 0;
1164                 if (io_channel_alloc_buf(channel, 0, &err_buf))
1165                         err_buf = NULL;
1166                 else
1167                         memcpy(err_buf, cache->buf, channel->block_size);
1168                 mutex_unlock(data, CACHE_MTX);
1169                 (channel->write_error)(channel, err_block, 1, err_buf,
1170                                        channel->block_size, -1,
1171                                        retval);
1172                 if (err_buf)
1173                         ext2fs_free_mem(&err_buf);
1174         } else
1175                 mutex_unlock(data, CACHE_MTX);
1176         return retval;
1177 #endif /* NO_IO_CACHE */
1178 }
1179
1180 static errcode_t unix_cache_readahead(io_channel channel,
1181                                       unsigned long long block,
1182                                       unsigned long long count)
1183 {
1184 #ifdef POSIX_FADV_WILLNEED
1185         struct unix_private_data *data;
1186
1187         data = (struct unix_private_data *)channel->private_data;
1188         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1189         return posix_fadvise(data->dev,
1190                              (ext2_loff_t)block * channel->block_size + data->offset,
1191                              (ext2_loff_t)count * channel->block_size,
1192                              POSIX_FADV_WILLNEED);
1193 #else
1194         return EXT2_ET_OP_NOT_SUPPORTED;
1195 #endif
1196 }
1197
1198 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
1199                                 int count, const void *buf)
1200 {
1201         return unix_write_blk64(channel, block, count, buf);
1202 }
1203
1204 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
1205                                  int size, const void *buf)
1206 {
1207         struct unix_private_data *data;
1208         errcode_t       retval = 0;
1209         ssize_t         actual;
1210
1211         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1212         data = (struct unix_private_data *) channel->private_data;
1213         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1214
1215         if (channel->align != 0) {
1216 #ifdef ALIGN_DEBUG
1217                 printf("unix_write_byte: O_DIRECT fallback\n");
1218 #endif
1219                 return EXT2_ET_UNIMPLEMENTED;
1220         }
1221
1222 #ifndef NO_IO_CACHE
1223         /*
1224          * Flush out the cache completely
1225          */
1226         if ((retval = flush_cached_blocks(channel, data, FLUSH_INVALIDATE)))
1227                 return retval;
1228 #endif
1229
1230         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1231                 return errno;
1232
1233         actual = write(data->dev, buf, size);
1234         if (actual < 0)
1235                 return errno;
1236         if (actual != size)
1237                 return EXT2_ET_SHORT_WRITE;
1238
1239         return 0;
1240 }
1241
1242 /*
1243  * Flush data buffers to disk.
1244  */
1245 static errcode_t unix_flush(io_channel channel)
1246 {
1247         struct unix_private_data *data;
1248         errcode_t retval = 0;
1249
1250         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1251         data = (struct unix_private_data *) channel->private_data;
1252         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1253
1254 #ifndef NO_IO_CACHE
1255         retval = flush_cached_blocks(channel, data, 0);
1256 #endif
1257 #ifdef HAVE_FSYNC
1258         if (!retval && fsync(data->dev) != 0)
1259                 return errno;
1260 #endif
1261         return retval;
1262 }
1263
1264 static errcode_t unix_set_option(io_channel channel, const char *option,
1265                                  const char *arg)
1266 {
1267         struct unix_private_data *data;
1268         unsigned long long tmp;
1269         errcode_t retval;
1270         char *end;
1271
1272         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1273         data = (struct unix_private_data *) channel->private_data;
1274         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1275
1276         if (!strcmp(option, "offset")) {
1277                 if (!arg)
1278                         return EXT2_ET_INVALID_ARGUMENT;
1279
1280                 tmp = strtoull(arg, &end, 0);
1281                 if (*end)
1282                         return EXT2_ET_INVALID_ARGUMENT;
1283                 data->offset = tmp;
1284                 if (data->offset < 0)
1285                         return EXT2_ET_INVALID_ARGUMENT;
1286                 return 0;
1287         }
1288         if (!strcmp(option, "cache")) {
1289                 if (!arg)
1290                         return EXT2_ET_INVALID_ARGUMENT;
1291                 if (!strcmp(arg, "on")) {
1292                         data->flags &= ~IO_FLAG_NOCACHE;
1293                         return 0;
1294                 }
1295                 if (!strcmp(arg, "off")) {
1296                         retval = flush_cached_blocks(channel, data, 0);
1297                         data->flags |= IO_FLAG_NOCACHE;
1298                         return retval;
1299                 }
1300                 return EXT2_ET_INVALID_ARGUMENT;
1301         }
1302         return EXT2_ET_INVALID_ARGUMENT;
1303 }
1304
1305 #if defined(__linux__) && !defined(BLKDISCARD)
1306 #define BLKDISCARD              _IO(0x12,119)
1307 #endif
1308
1309 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1310                               unsigned long long count)
1311 {
1312         struct unix_private_data *data;
1313         int             ret;
1314
1315         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1316         data = (struct unix_private_data *) channel->private_data;
1317         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1318
1319         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1320 #ifdef BLKDISCARD
1321                 __u64 range[2];
1322
1323                 range[0] = (__u64)(block) * channel->block_size + data->offset;
1324                 range[1] = (__u64)(count) * channel->block_size;
1325
1326                 ret = ioctl(data->dev, BLKDISCARD, &range);
1327 #else
1328                 goto unimplemented;
1329 #endif
1330         } else {
1331 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1332                 /*
1333                  * If we are not on block device, try to use punch hole
1334                  * to reclaim free space.
1335                  */
1336                 ret = fallocate(data->dev,
1337                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1338                                 (off_t)(block) * channel->block_size + data->offset,
1339                                 (off_t)(count) * channel->block_size);
1340 #else
1341                 goto unimplemented;
1342 #endif
1343         }
1344         if (ret < 0) {
1345                 if (errno == EOPNOTSUPP)
1346                         goto unimplemented;
1347                 return errno;
1348         }
1349         return 0;
1350 unimplemented:
1351         return EXT2_ET_UNIMPLEMENTED;
1352 }
1353
1354 /*
1355  * If we know about ZERO_RANGE, try that before we try PUNCH_HOLE because
1356  * ZERO_RANGE doesn't unmap preallocated blocks.  We prefer fallocate because
1357  * it always invalidates page cache, and libext2fs requires that reads after
1358  * ZERO_RANGE return zeroes.
1359  */
1360 static int __unix_zeroout(int fd, off_t offset, off_t len)
1361 {
1362         int ret = -1;
1363
1364 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_ZERO_RANGE)
1365         ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, offset, len);
1366         if (ret == 0)
1367                 return 0;
1368 #endif
1369 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1370         ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1371                         offset,  len);
1372         if (ret == 0)
1373                 return 0;
1374 #endif
1375         errno = EOPNOTSUPP;
1376         return ret;
1377 }
1378
1379 /* parameters might not be used if OS doesn't support zeroout */
1380 #if __GNUC_PREREQ (4, 6)
1381 #pragma GCC diagnostic push
1382 #pragma GCC diagnostic ignored "-Wunused-parameter"
1383 #endif
1384 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1385                               unsigned long long count)
1386 {
1387         struct unix_private_data *data;
1388         int             ret;
1389
1390         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1391         data = (struct unix_private_data *) channel->private_data;
1392         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1393
1394         if (safe_getenv("UNIX_IO_NOZEROOUT"))
1395                 goto unimplemented;
1396
1397         if (!(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE)) {
1398                 /* Regular file, try to use truncate/punch/zero. */
1399                 struct stat statbuf;
1400
1401                 if (count == 0)
1402                         return 0;
1403                 /*
1404                  * If we're trying to zero a range past the end of the file,
1405                  * extend the file size, then truncate everything.
1406                  */
1407                 ret = fstat(data->dev, &statbuf);
1408                 if (ret)
1409                         goto err;
1410                 if ((unsigned long long) statbuf.st_size <
1411                         (block + count) * channel->block_size + data->offset) {
1412                         ret = ftruncate(data->dev,
1413                                         (block + count) * channel->block_size + data->offset);
1414                         if (ret)
1415                                 goto err;
1416                 }
1417         }
1418
1419         ret = __unix_zeroout(data->dev,
1420                         (off_t)(block) * channel->block_size + data->offset,
1421                         (off_t)(count) * channel->block_size);
1422 err:
1423         if (ret < 0) {
1424                 if (errno == EOPNOTSUPP)
1425                         goto unimplemented;
1426                 return errno;
1427         }
1428         return 0;
1429 unimplemented:
1430         return EXT2_ET_UNIMPLEMENTED;
1431 }
1432 #if __GNUC_PREREQ (4, 6)
1433 #pragma GCC diagnostic pop
1434 #endif
1435
1436 static struct struct_io_manager struct_unix_manager = {
1437         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1438         .name           = "Unix I/O Manager",
1439         .open           = unix_open,
1440         .close          = unix_close,
1441         .set_blksize    = unix_set_blksize,
1442         .read_blk       = unix_read_blk,
1443         .write_blk      = unix_write_blk,
1444         .flush          = unix_flush,
1445         .write_byte     = unix_write_byte,
1446         .set_option     = unix_set_option,
1447         .get_stats      = unix_get_stats,
1448         .read_blk64     = unix_read_blk64,
1449         .write_blk64    = unix_write_blk64,
1450         .discard        = unix_discard,
1451         .cache_readahead        = unix_cache_readahead,
1452         .zeroout        = unix_zeroout,
1453 };
1454
1455 io_manager unix_io_manager = &struct_unix_manager;
1456
1457 static struct struct_io_manager struct_unixfd_manager = {
1458         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1459         .name           = "Unix fd I/O Manager",
1460         .open           = unixfd_open,
1461         .close          = unix_close,
1462         .set_blksize    = unix_set_blksize,
1463         .read_blk       = unix_read_blk,
1464         .write_blk      = unix_write_blk,
1465         .flush          = unix_flush,
1466         .write_byte     = unix_write_byte,
1467         .set_option     = unix_set_option,
1468         .get_stats      = unix_get_stats,
1469         .read_blk64     = unix_read_blk64,
1470         .write_blk64    = unix_write_blk64,
1471         .discard        = unix_discard,
1472         .cache_readahead        = unix_cache_readahead,
1473         .zeroout        = unix_zeroout,
1474 };
1475
1476 io_manager unixfd_io_manager = &struct_unixfd_manager;