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