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                         actual = really_read;
 310                         buf -= really_read;
 311                         size += really_read;
 312                         goto short_read;
 313                 }
 314                 if ((actual + offset) > align_size)
 315                         actual = align_size - offset;
 316                 if (actual > size)
 317                         actual = size;
 318                 memcpy(buf, (char *)data->bounce + offset, actual);
 319
 320                 really_read += actual;
 321                 size -= actual;
 322                 buf += actual;
 323                 offset = 0;
 324                 aligned_blk++;
 325         }
 326 success_unlock:
 327         mutex_unlock(data, BOUNCE_MTX);
 328         return 0;
 329
 330 error_unlock:
 331         mutex_unlock(data, BOUNCE_MTX);
 332         if (actual >= 0 && actual < size)
 333                 memset((char *) buf+actual, 0, size-actual);
 334         if (channel->read_error)
 335                 retval = (channel->read_error)(channel, block, count, buf,
 336                                                size, actual, retval);
 337         return retval;
 338 }
 339
 340 #define RAW_WRITE_NO_HANDLER    1
 341
 342 static errcode_t raw_write_blk(io_channel channel,
 343                                struct unix_private_data *data,
 344                                unsigned long long block,
 345                                int count, const void *bufv,
 346                                int flags)
 347 {
 348         ssize_t         size;
 349         ext2_loff_t     location;
 350         int             actual = 0;
 351         errcode_t       retval;
 352         const unsigned char *buf = bufv;
 353         unsigned long long aligned_blk;
 354         int             align_size, offset;
 355
 356         if (count == 1)
 357                 size = channel->block_size;
 358         else {
 359                 if (count < 0)
 360                         size = -count;
 361                 else
 362                         size = (ext2_loff_t) count * channel->block_size;
 363         }
 364         mutex_lock(data, STATS_MTX);
 365         data->io_stats.bytes_written += size;
 366         mutex_unlock(data, STATS_MTX);
 367
 368         location = ((ext2_loff_t) block * channel->block_size) + data->offset;
 369
 370         if (data->flags & IO_FLAG_FORCE_BOUNCE)
 371                 goto bounce_write;
 372
 373 #ifdef HAVE_PWRITE64
 374         /* Try an aligned pwrite */
 375         if ((channel->align == 0) ||
 376             (IS_ALIGNED(buf, channel->align) &&
 377              IS_ALIGNED(location, channel->align) &&
 378              IS_ALIGNED(size, channel->align))) {
 379                 actual = pwrite64(data->dev, buf, size, location);
 380                 if (actual == size)
 381                         return 0;
 382         }
 383 #elif HAVE_PWRITE
 384         /* Try an aligned pwrite */
 385         if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
 386             ((channel->align == 0) ||
 387              (IS_ALIGNED(buf, channel->align) &&
 388               IS_ALIGNED(location, channel->align) &&
 389               IS_ALIGNED(size, channel->align)))) {
 390                 actual = pwrite(data->dev, buf, size, location);
 391                 if (actual == size)
 392                         return 0;
 393         }
 394 #endif /* HAVE_PWRITE */
 395
 396         if ((channel->align == 0) ||
 397             (IS_ALIGNED(buf, channel->align) &&
 398              IS_ALIGNED(location, channel->align) &&
 399              IS_ALIGNED(size, channel->align))) {
 400                 mutex_lock(data, BOUNCE_MTX);
 401                 if (ext2fs_llseek(data->dev, location, SEEK_SET) < 0) {
 402                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 403                         goto error_unlock;
 404                 }
 405                 actual = write(data->dev, buf, size);
 406                 mutex_unlock(data, BOUNCE_MTX);
 407                 if (actual < 0) {
 408                         retval = errno;
 409                         goto error_out;
 410                 }
 411                 if (actual != size) {
 412                 short_write:
 413                         retval = EXT2_ET_SHORT_WRITE;
 414                         goto error_out;
 415                 }
 416                 return 0;
 417         }
 418
 419 #ifdef ALIGN_DEBUG
 420         printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
 421                (unsigned long) size);
 422 #endif
 423         /*
 424          * The buffer or size which we're trying to write isn't aligned
 425          * to the O_DIRECT rules, so we need to do this the hard way...
 426          */
 427 bounce_write:
 428         if (channel->align == 0)
 429                 channel->align = 1;
 430         if ((channel->block_size > channel->align) &&
 431             (channel->block_size % channel->align) == 0)
 432                 align_size = channel->block_size;
 433         else
 434                 align_size = channel->align;
 435         aligned_blk = location / align_size;
 436         offset = location % align_size;
 437
 438         while (size > 0) {
 439                 int actual_w;
 440
 441                 mutex_lock(data, BOUNCE_MTX);
 442                 if (size < align_size || offset) {
 443                         if (ext2fs_llseek(data->dev, aligned_blk * align_size,
 444                                           SEEK_SET) < 0) {
 445                                 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 446                                 goto error_unlock;
 447                         }
 448                         actual = read(data->dev, data->bounce,
 449                                       align_size);
 450                         if (actual != align_size) {
 451                                 if (actual < 0) {
 452                                         retval = errno;
 453                                         goto error_unlock;
 454                                 }
 455                                 memset((char *) data->bounce + actual, 0,
 456                                        align_size - actual);
 457                         }
 458                 }
 459                 actual = size;
 460                 if ((actual + offset) > align_size)
 461                         actual = align_size - offset;
 462                 if (actual > size)
 463                         actual = size;
 464                 memcpy(((char *)data->bounce) + offset, buf, actual);
 465                 if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
 466                         retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
 467                         goto error_unlock;
 468                 }
 469                 actual_w = write(data->dev, data->bounce, align_size);
 470                 mutex_unlock(data, BOUNCE_MTX);
 471                 if (actual_w < 0) {
 472                         retval = errno;
 473                         goto error_out;
 474                 }
 475                 if (actual_w != align_size)
 476                         goto short_write;
 477                 size -= actual;
 478                 buf += actual;
 479                 location += actual;
 480                 aligned_blk++;
 481                 offset = 0;
 482         }
 483         return 0;
 484
 485 error_unlock:
 486         mutex_unlock(data, BOUNCE_MTX);
 487 error_out:
 488         if (((flags & RAW_WRITE_NO_HANDLER) == 0) && channel->write_error)
 489                 retval = (channel->write_error)(channel, block, count, buf,
 490                                                 size, actual, retval);
 491         return retval;
 492 }
 493
 494
 495 /*
 496  * Here we implement the cache functions
 497  */
 498
 499 /* Allocate the cache buffers */
 500 static errcode_t alloc_cache(io_channel channel,
 501                              struct unix_private_data *data)
 502 {
 503         errcode_t               retval;
 504         struct unix_cache       *cache;
 505         int                     i;
 506
 507         data->access_time = 0;
 508         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 509                 cache->block = 0;
 510                 cache->access_time = 0;
 511                 cache->dirty = 0;
 512                 cache->in_use = 0;
 513                 if (cache->buf)
 514                         ext2fs_free_mem(&cache->buf);
 515                 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
 516                 if (retval)
 517                         return retval;
 518         }
 519         if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
 520                 if (data->bounce)
 521                         ext2fs_free_mem(&data->bounce);
 522                 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
 523         }
 524         return retval;
 525 }
 526
 527 /* Free the cache buffers */
 528 static void free_cache(struct unix_private_data *data)
 529 {
 530         struct unix_cache       *cache;
 531         int                     i;
 532
 533         data->access_time = 0;
 534         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 535                 cache->block = 0;
 536                 cache->access_time = 0;
 537                 cache->dirty = 0;
 538                 cache->in_use = 0;
 539                 if (cache->buf)
 540                         ext2fs_free_mem(&cache->buf);
 541         }
 542         if (data->bounce)
 543                 ext2fs_free_mem(&data->bounce);
 544 }
 545
 546 #ifndef NO_IO_CACHE
 547 /*
 548  * Try to find a block in the cache.  If the block is not found, and
 549  * eldest is a non-zero pointer, then fill in eldest with the cache
 550  * entry to that should be reused.
 551  */
 552 static struct unix_cache *find_cached_block(struct unix_private_data *data,
 553                                             unsigned long long block,
 554                                             struct unix_cache **eldest)
 555 {
 556         struct unix_cache       *cache, *unused_cache, *oldest_cache;
 557         int                     i;
 558
 559         unused_cache = oldest_cache = 0;
 560         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 561                 if (!cache->in_use) {
 562                         if (!unused_cache)
 563                                 unused_cache = cache;
 564                         continue;
 565                 }
 566                 if (cache->block == block) {
 567                         cache->access_time = ++data->access_time;
 568                         return cache;
 569                 }
 570                 if (!oldest_cache ||
 571                     (cache->access_time < oldest_cache->access_time))
 572                         oldest_cache = cache;
 573         }
 574         if (eldest)
 575                 *eldest = (unused_cache) ? unused_cache : oldest_cache;
 576         return 0;
 577 }
 578
 579 /*
 580  * Reuse a particular cache entry for another block.
 581  */
 582 static errcode_t reuse_cache(io_channel channel,
 583                 struct unix_private_data *data, struct unix_cache *cache,
 584                 unsigned long long block)
 585 {
 586         if (cache->dirty && cache->in_use) {
 587                 errcode_t retval;
 588
 589                 retval = raw_write_blk(channel, data, cache->block, 1,
 590                                        cache->buf, RAW_WRITE_NO_HANDLER);
 591                 if (retval) {
 592                         cache->write_err = 1;
 593                         return retval;
 594                 }
 595         }
 596
 597         cache->in_use = 1;
 598         cache->dirty = 0;
 599         cache->write_err = 0;
 600         cache->block = block;
 601         cache->access_time = ++data->access_time;
 602         return 0;
 603 }
 604
 605 #define FLUSH_INVALIDATE        0x01
 606 #define FLUSH_NOLOCK            0x02
 607
 608 /*
 609  * Flush all of the blocks in the cache
 610  */
 611 static errcode_t flush_cached_blocks(io_channel channel,
 612                                      struct unix_private_data *data,
 613                                      int flags)
 614 {
 615         struct unix_cache       *cache;
 616         errcode_t               retval, retval2 = 0;
 617         int                     i;
 618         int                     errors_found = 0;
 619
 620         if ((flags & FLUSH_NOLOCK) == 0)
 621                 mutex_lock(data, CACHE_MTX);
 622         for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 623                 if (!cache->in_use || !cache->dirty)
 624                         continue;
 625                 retval = raw_write_blk(channel, data,
 626                                        cache->block, 1, cache->buf,
 627                                        RAW_WRITE_NO_HANDLER);
 628                 if (retval) {
 629                         cache->write_err = 1;
 630                         errors_found = 1;
 631                         retval2 = retval;
 632                 } else {
 633                         cache->dirty = 0;
 634                         cache->write_err = 0;
 635                         if (flags & FLUSH_INVALIDATE)
 636                                 cache->in_use = 0;
 637                 }
 638         }
 639         if ((flags & FLUSH_NOLOCK) == 0)
 640                 mutex_unlock(data, CACHE_MTX);
 641 retry:
 642         while (errors_found) {
 643                 if ((flags & FLUSH_NOLOCK) == 0)
 644                         mutex_lock(data, CACHE_MTX);
 645                 errors_found = 0;
 646                 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
 647                         if (!cache->in_use || !cache->write_err)
 648                                 continue;
 649                         errors_found = 1;
 650                         if (cache->write_err && channel->write_error) {
 651                                 char *err_buf = NULL;
 652                                 unsigned long long err_block = cache->block;
 653
 654                                 cache->dirty = 0;
 655                                 cache->in_use = 0;
 656                                 cache->write_err = 0;
 657                                 if (io_channel_alloc_buf(channel, 0,
 658                                                          &err_buf))
 659                                         err_buf = NULL;
 660                                 else
 661                                         memcpy(err_buf, cache->buf,
 662                                                channel->block_size);
 663                                 mutex_unlock(data, CACHE_MTX);
 664                                 (channel->write_error)(channel, err_block,
 665                                         1, err_buf, channel->block_size, -1,
 666                                         retval2);
 667                                 if (err_buf)
 668                                         ext2fs_free_mem(&err_buf);
 669                                 goto retry;
 670                         } else
 671                                 cache->write_err = 0;
 672                 }
 673                 if ((flags & FLUSH_NOLOCK) == 0)
 674                         mutex_unlock(data, CACHE_MTX);
 675         }
 676         return retval2;
 677 }
 678 #endif /* NO_IO_CACHE */
 679
 680 #ifdef __linux__
 681 #ifndef BLKDISCARDZEROES
 682 #define BLKDISCARDZEROES _IO(0x12,124)
 683 #endif
 684 #endif
 685
 686 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
 687 {
 688         if (mode)
 689 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 690                 return open64(pathname, flags, mode);
 691         else
 692                 return open64(pathname, flags);
 693 #else
 694                 return open(pathname, flags, mode);
 695         else
 696                 return open(pathname, flags);
 697 #endif
 698 }
 699
 700 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
 701 {
 702 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 703         return stat64(path, buf);
 704 #else
 705         return stat(path, buf);
 706 #endif
 707 }
 708
 709 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
 710 {
 711 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
 712         return fstat64(fd, buf);
 713 #else
 714         return fstat(fd, buf);
 715 #endif
 716 }
 717
 718
 719 static errcode_t unix_open_channel(const char *name, int fd,
 720                                    int flags, io_channel *channel,
 721                                    io_manager io_mgr)
 722 {
 723         io_channel      io = NULL;
 724         struct unix_private_data *data = NULL;
 725         errcode_t       retval;
 726         ext2fs_struct_stat st;
 727 #ifdef __linux__
 728         struct          utsname ut;
 729 #endif
 730
 731         if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
 732                 flags |= IO_FLAG_FORCE_BOUNCE;
 733
 734 #ifdef __linux__
 735         /*
 736          * We need to make sure any previous errors in the block
 737          * device are thrown away, sigh.
 738          */
 739         (void) fsync(fd);
 740 #endif
 741
 742         retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
 743         if (retval)
 744                 goto cleanup;
 745         memset(io, 0, sizeof(struct struct_io_channel));
 746         io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
 747         retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
 748         if (retval)
 749                 goto cleanup;
 750
 751         io->manager = io_mgr;
 752         retval = ext2fs_get_mem(strlen(name)+1, &io->name);
 753         if (retval)
 754                 goto cleanup;
 755
 756         strcpy(io->name, name);
 757         io->private_data = data;
 758         io->block_size = 1024;
 759         io->read_error = 0;
 760         io->write_error = 0;
 761         io->refcount = 1;
 762         io->flags = 0;
 763
 764         if (safe_getenv("UNIX_IO_NOZEROOUT"))
 765                 io->flags |= CHANNEL_FLAGS_NOZEROOUT;
 766
 767         memset(data, 0, sizeof(struct unix_private_data));
 768         data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
 769         data->io_stats.num_fields = 2;
 770         data->flags = flags;
 771         data->dev = fd;
 772
 773 #if defined(O_DIRECT)
 774         if (flags & IO_FLAG_DIRECT_IO)
 775                 io->align = ext2fs_get_dio_alignment(data->dev);
 776 #elif defined(F_NOCACHE)
 777         if (flags & IO_FLAG_DIRECT_IO)
 778                 io->align = 4096;
 779 #endif
 780
 781         /*
 782          * If the device is really a block device, then set the
 783          * appropriate flag, otherwise we can set DISCARD_ZEROES flag
 784          * because we are going to use punch hole instead of discard
 785          * and if it succeed, subsequent read from sparse area returns
 786          * zero.
 787          */
 788         if (ext2fs_fstat(data->dev, &st) == 0) {
 789                 if (ext2fsP_is_disk_device(st.st_mode)) {
 790 #ifdef BLKDISCARDZEROES
 791                         int zeroes = 0;
 792
 793                         if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
 794                             zeroes)
 795                                 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 796 #endif
 797                         io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
 798                 } else {
 799                         io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
 800                 }
 801         }
 802
 803 #if defined(__CYGWIN__)
 804         /*
 805          * Some operating systems require that the buffers be aligned,
 806          * regardless of O_DIRECT
 807          */
 808         if (!io->align)
 809                 io->align = 512;
 810 #endif
 811
 812 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 813         if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
 814                 int dio_align = ext2fs_get_dio_alignment(fd);
 815
 816                 if (io->align < dio_align)
 817                         io->align = dio_align;
 818         }
 819 #endif
 820
 821         if ((retval = alloc_cache(io, data)))
 822                 goto cleanup;
 823
 824 #ifdef BLKROGET
 825         if (flags & IO_FLAG_RW) {
 826                 int error;
 827                 int readonly = 0;
 828
 829                 /* Is the block device actually writable? */
 830                 error = ioctl(data->dev, BLKROGET, &readonly);
 831                 if (!error && readonly) {
 832                         retval = EPERM;
 833                         goto cleanup;
 834                 }
 835         }
 836 #endif
 837
 838 #ifdef __linux__
 839 #undef RLIM_INFINITY
 840 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
 841 #define RLIM_INFINITY   ((unsigned long)(~0UL>>1))
 842 #else
 843 #define RLIM_INFINITY  (~0UL)
 844 #endif
 845         /*
 846          * Work around a bug in 2.4.10-2.4.18 kernels where writes to
 847          * block devices are wrongly getting hit by the filesize
 848          * limit.  This workaround isn't perfect, since it won't work
 849          * if glibc wasn't built against 2.2 header files.  (Sigh.)
 850          *
 851          */
 852         if ((flags & IO_FLAG_RW) &&
 853             (uname(&ut) == 0) &&
 854             ((ut.release[0] == '2') && (ut.release[1] == '.') &&
 855              (ut.release[2] == '4') && (ut.release[3] == '.') &&
 856              (ut.release[4] == '1') && (ut.release[5] >= '0') &&
 857              (ut.release[5] < '8')) &&
 858             (ext2fs_fstat(data->dev, &st) == 0) &&
 859             (ext2fsP_is_disk_device(st.st_mode))) {
 860                 struct rlimit   rlim;
 861
 862                 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
 863                 setrlimit(RLIMIT_FSIZE, &rlim);
 864                 getrlimit(RLIMIT_FSIZE, &rlim);
 865                 if (((unsigned long) rlim.rlim_cur) <
 866                     ((unsigned long) rlim.rlim_max)) {
 867                         rlim.rlim_cur = rlim.rlim_max;
 868                         setrlimit(RLIMIT_FSIZE, &rlim);
 869                 }
 870         }
 871 #endif
 872 #ifdef HAVE_PTHREAD
 873         if (flags & IO_FLAG_THREADS) {
 874                 io->flags |= CHANNEL_FLAGS_THREADS;
 875                 retval = pthread_mutex_init(&data->cache_mutex, NULL);
 876                 if (retval)
 877                         goto cleanup;
 878                 retval = pthread_mutex_init(&data->bounce_mutex, NULL);
 879                 if (retval) {
 880                         pthread_mutex_destroy(&data->cache_mutex);
 881                         goto cleanup;
 882                 }
 883                 retval = pthread_mutex_init(&data->stats_mutex, NULL);
 884                 if (retval) {
 885                         pthread_mutex_destroy(&data->cache_mutex);
 886                         pthread_mutex_destroy(&data->bounce_mutex);
 887                         goto cleanup;
 888                 }
 889         }
 890 #endif
 891         *channel = io;
 892         return 0;
 893
 894 cleanup:
 895         if (data) {
 896                 if (data->dev >= 0)
 897                         close(data->dev);
 898                 free_cache(data);
 899                 ext2fs_free_mem(&data);
 900         }
 901         if (io) {
 902                 if (io->name) {
 903                         ext2fs_free_mem(&io->name);
 904                 }
 905                 ext2fs_free_mem(&io);
 906         }
 907         return retval;
 908 }
 909
 910 static errcode_t unixfd_open(const char *str_fd, int flags,
 911                              io_channel *channel)
 912 {
 913         int fd;
 914         int fd_flags;
 915
 916         fd = atoi(str_fd);
 917 #if defined(HAVE_FCNTL)
 918         fd_flags = fcntl(fd, F_GETFD);
 919         if (fd_flags == -1)
 920                 return EBADF;
 921
 922         flags = 0;
 923         if (fd_flags & O_RDWR)
 924                 flags |= IO_FLAG_RW;
 925         if (fd_flags & O_EXCL)
 926                 flags |= IO_FLAG_EXCLUSIVE;
 927 #if defined(O_DIRECT)
 928         if (fd_flags & O_DIRECT)
 929                 flags |= IO_FLAG_DIRECT_IO;
 930 #endif
 931 #endif  /* HAVE_FCNTL */
 932
 933         return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
 934 }
 935
 936 static errcode_t unix_open(const char *name, int flags,
 937                            io_channel *channel)
 938 {
 939         int fd = -1;
 940         int open_flags;
 941
 942         if (name == 0)
 943                 return EXT2_ET_BAD_DEVICE_NAME;
 944
 945         open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
 946         if (flags & IO_FLAG_EXCLUSIVE)
 947                 open_flags |= O_EXCL;
 948 #if defined(O_DIRECT)
 949         if (flags & IO_FLAG_DIRECT_IO)
 950                 open_flags |= O_DIRECT;
 951 #endif
 952         fd = ext2fs_open_file(name, open_flags, 0);
 953         if (fd < 0)
 954                 return errno;
 955 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
 956         if (flags & IO_FLAG_DIRECT_IO) {
 957                 if (fcntl(fd, F_NOCACHE, 1) < 0)
 958                         return errno;
 959         }
 960 #endif
 961         return unix_open_channel(name, fd, flags, channel, unix_io_manager);
 962 }
 963
 964 static errcode_t unix_close(io_channel channel)
 965 {
 966         struct unix_private_data *data;
 967         errcode_t       retval = 0;
 968
 969         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
 970         data = (struct unix_private_data *) channel->private_data;
 971         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 972
 973         if (--channel->refcount > 0)
 974                 return 0;
 975
 976 #ifndef NO_IO_CACHE
 977         retval = flush_cached_blocks(channel, data, 0);
 978 #endif
 979
 980         if (close(data->dev) < 0)
 981                 retval = errno;
 982         free_cache(data);
 983 #ifdef HAVE_PTHREAD
 984         if (data->flags & IO_FLAG_THREADS) {
 985                 pthread_mutex_destroy(&data->cache_mutex);
 986                 pthread_mutex_destroy(&data->bounce_mutex);
 987                 pthread_mutex_destroy(&data->stats_mutex);
 988         }
 989 #endif
 990
 991         ext2fs_free_mem(&channel->private_data);
 992         if (channel->name)
 993                 ext2fs_free_mem(&channel->name);
 994         ext2fs_free_mem(&channel);
 995         return retval;
 996 }
 997
 998 static errcode_t unix_set_blksize(io_channel channel, int blksize)
 999 {
1000         struct unix_private_data *data;
1001         errcode_t               retval = 0;
1002
1003         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1004         data = (struct unix_private_data *) channel->private_data;
1005         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1006
1007         if (channel->block_size != blksize) {
1008                 mutex_lock(data, CACHE_MTX);
1009                 mutex_lock(data, BOUNCE_MTX);
1010 #ifndef NO_IO_CACHE
1011                 if ((retval = flush_cached_blocks(channel, data, FLUSH_NOLOCK))){
1012                         mutex_unlock(data, BOUNCE_MTX);
1013                         mutex_unlock(data, CACHE_MTX);
1014                         return retval;
1015                 }
1016 #endif
1017
1018                 channel->block_size = blksize;
1019                 free_cache(data);
1020                 retval = alloc_cache(channel, data);
1021                 mutex_unlock(data, BOUNCE_MTX);
1022                 mutex_unlock(data, CACHE_MTX);
1023         }
1024         return retval;
1025 }
1026
1027 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
1028                                int count, void *buf)
1029 {
1030         struct unix_private_data *data;
1031         struct unix_cache *cache;
1032         errcode_t       retval;
1033         char            *cp;
1034         int             i, j;
1035
1036         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1037         data = (struct unix_private_data *) channel->private_data;
1038         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1039
1040 #ifdef NO_IO_CACHE
1041         return raw_read_blk(channel, data, block, count, buf);
1042 #else
1043         if (data->flags & IO_FLAG_NOCACHE)
1044                 return raw_read_blk(channel, data, block, count, buf);
1045         /*
1046          * If we're doing an odd-sized read or a very large read,
1047          * flush out the cache and then do a direct read.
1048          */
1049         if (count < 0 || count > WRITE_DIRECT_SIZE) {
1050                 if ((retval = flush_cached_blocks(channel, data, 0)))
1051                         return retval;
1052                 return raw_read_blk(channel, data, block, count, buf);
1053         }
1054
1055         cp = buf;
1056         mutex_lock(data, CACHE_MTX);
1057         while (count > 0) {
1058                 /* If it's in the cache, use it! */
1059                 if ((cache = find_cached_block(data, block, NULL))) {
1060 #ifdef DEBUG
1061                         printf("Using cached block %lu\n", block);
1062 #endif
1063                         memcpy(cp, cache->buf, channel->block_size);
1064                         count--;
1065                         block++;
1066                         cp += channel->block_size;
1067                         continue;
1068                 }
1069
1070                 /*
1071                  * Find the number of uncached blocks so we can do a
1072                  * single read request
1073                  */
1074                 for (i=1; i < count; i++)
1075                         if (find_cached_block(data, block+i, NULL))
1076                                 break;
1077 #ifdef DEBUG
1078                 printf("Reading %d blocks starting at %lu\n", i, block);
1079 #endif
1080                 mutex_unlock(data, CACHE_MTX);
1081                 if ((retval = raw_read_blk(channel, data, block, i, cp)))
1082                         return retval;
1083                 mutex_lock(data, CACHE_MTX);
1084
1085                 /* Save the results in the cache */
1086                 for (j=0; j < i; j++) {
1087                         if (!find_cached_block(data, block, &cache)) {
1088                                 retval = reuse_cache(channel, data,
1089                                                      cache, block);
1090                                 if (retval)
1091                                         goto call_write_handler;
1092                                 memcpy(cache->buf, cp, channel->block_size);
1093                         }
1094                         count--;
1095                         block++;
1096                         cp += channel->block_size;
1097                 }
1098         }
1099         mutex_unlock(data, CACHE_MTX);
1100         return 0;
1101
1102 call_write_handler:
1103         if (cache->write_err && channel->write_error) {
1104                 char *err_buf = NULL;
1105                 unsigned long long err_block = cache->block;
1106
1107                 cache->dirty = 0;
1108                 cache->in_use = 0;
1109                 cache->write_err = 0;
1110                 if (io_channel_alloc_buf(channel, 0, &err_buf))
1111                         err_buf = NULL;
1112                 else
1113                         memcpy(err_buf, cache->buf, channel->block_size);
1114                 mutex_unlock(data, CACHE_MTX);
1115                 (channel->write_error)(channel, err_block, 1, err_buf,
1116                                        channel->block_size, -1,
1117                                        retval);
1118                 if (err_buf)
1119                         ext2fs_free_mem(&err_buf);
1120         } else
1121                 mutex_unlock(data, CACHE_MTX);
1122         return retval;
1123 #endif /* NO_IO_CACHE */
1124 }
1125
1126 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
1127                                int count, void *buf)
1128 {
1129         return unix_read_blk64(channel, block, count, buf);
1130 }
1131
1132 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
1133                                 int count, const void *buf)
1134 {
1135         struct unix_private_data *data;
1136         struct unix_cache *cache, *reuse;
1137         errcode_t       retval = 0;
1138         const char      *cp;
1139         int             writethrough;
1140
1141         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1142         data = (struct unix_private_data *) channel->private_data;
1143         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1144
1145 #ifdef NO_IO_CACHE
1146         return raw_write_blk(channel, data, block, count, buf, 0);
1147 #else
1148         if (data->flags & IO_FLAG_NOCACHE)
1149                 return raw_write_blk(channel, data, block, count, buf, 0);
1150         /*
1151          * If we're doing an odd-sized write or a very large write,
1152          * flush out the cache completely and then do a direct write.
1153          */
1154         if (count < 0 || count > WRITE_DIRECT_SIZE) {
1155                 if ((retval = flush_cached_blocks(channel, data,
1156                                                   FLUSH_INVALIDATE)))
1157                         return retval;
1158                 return raw_write_blk(channel, data, block, count, buf, 0);
1159         }
1160
1161         /*
1162          * For a moderate-sized multi-block write, first force a write
1163          * if we're in write-through cache mode, and then fill the
1164          * cache with the blocks.
1165          */
1166         writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
1167         if (writethrough)
1168                 retval = raw_write_blk(channel, data, block, count, buf, 0);
1169
1170         cp = buf;
1171         mutex_lock(data, CACHE_MTX);
1172         while (count > 0) {
1173                 cache = find_cached_block(data, block, &reuse);
1174                 if (!cache) {
1175                         errcode_t err;
1176
1177                         cache = reuse;
1178                         err = reuse_cache(channel, data, cache, block);
1179                         if (err)
1180                                 goto call_write_handler;
1181                 }
1182                 if (cache->buf != cp)
1183                         memcpy(cache->buf, cp, channel->block_size);
1184                 cache->dirty = !writethrough;
1185                 count--;
1186                 block++;
1187                 cp += channel->block_size;
1188         }
1189         mutex_unlock(data, CACHE_MTX);
1190         return retval;
1191
1192 call_write_handler:
1193         if (cache->write_err && channel->write_error) {
1194                 char *err_buf = NULL;
1195                 unsigned long long err_block = cache->block;
1196
1197                 cache->dirty = 0;
1198                 cache->in_use = 0;
1199                 cache->write_err = 0;
1200                 if (io_channel_alloc_buf(channel, 0, &err_buf))
1201                         err_buf = NULL;
1202                 else
1203                         memcpy(err_buf, cache->buf, channel->block_size);
1204                 mutex_unlock(data, CACHE_MTX);
1205                 (channel->write_error)(channel, err_block, 1, err_buf,
1206                                        channel->block_size, -1,
1207                                        retval);
1208                 if (err_buf)
1209                         ext2fs_free_mem(&err_buf);
1210         } else
1211                 mutex_unlock(data, CACHE_MTX);
1212         return retval;
1213 #endif /* NO_IO_CACHE */
1214 }
1215
1216 static errcode_t unix_cache_readahead(io_channel channel,
1217                                       unsigned long long block,
1218                                       unsigned long long count)
1219 {
1220 #ifdef POSIX_FADV_WILLNEED
1221         struct unix_private_data *data;
1222
1223         data = (struct unix_private_data *)channel->private_data;
1224         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1225         return posix_fadvise(data->dev,
1226                              (ext2_loff_t)block * channel->block_size + data->offset,
1227                              (ext2_loff_t)count * channel->block_size,
1228                              POSIX_FADV_WILLNEED);
1229 #else
1230         return EXT2_ET_OP_NOT_SUPPORTED;
1231 #endif
1232 }
1233
1234 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
1235                                 int count, const void *buf)
1236 {
1237         return unix_write_blk64(channel, block, count, buf);
1238 }
1239
1240 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
1241                                  int size, const void *buf)
1242 {
1243         struct unix_private_data *data;
1244         errcode_t       retval = 0;
1245         ssize_t         actual;
1246
1247         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1248         data = (struct unix_private_data *) channel->private_data;
1249         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1250
1251         if (channel->align != 0) {
1252 #ifdef ALIGN_DEBUG
1253                 printf("unix_write_byte: O_DIRECT fallback\n");
1254 #endif
1255                 return EXT2_ET_UNIMPLEMENTED;
1256         }
1257
1258 #ifndef NO_IO_CACHE
1259         /*
1260          * Flush out the cache completely
1261          */
1262         if ((retval = flush_cached_blocks(channel, data, FLUSH_INVALIDATE)))
1263                 return retval;
1264 #endif
1265
1266         if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1267                 return errno;
1268
1269         actual = write(data->dev, buf, size);
1270         if (actual < 0)
1271                 return errno;
1272         if (actual != size)
1273                 return EXT2_ET_SHORT_WRITE;
1274
1275         return 0;
1276 }
1277
1278 /*
1279  * Flush data buffers to disk.
1280  */
1281 static errcode_t unix_flush(io_channel channel)
1282 {
1283         struct unix_private_data *data;
1284         errcode_t retval = 0;
1285
1286         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1287         data = (struct unix_private_data *) channel->private_data;
1288         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1289
1290 #ifndef NO_IO_CACHE
1291         retval = flush_cached_blocks(channel, data, 0);
1292 #endif
1293 #ifdef HAVE_FSYNC
1294         if (!retval && fsync(data->dev) != 0)
1295                 return errno;
1296 #endif
1297         return retval;
1298 }
1299
1300 static errcode_t unix_set_option(io_channel channel, const char *option,
1301                                  const char *arg)
1302 {
1303         struct unix_private_data *data;
1304         unsigned long long tmp;
1305         errcode_t retval;
1306         char *end;
1307
1308         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1309         data = (struct unix_private_data *) channel->private_data;
1310         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1311
1312         if (!strcmp(option, "offset")) {
1313                 if (!arg)
1314                         return EXT2_ET_INVALID_ARGUMENT;
1315
1316                 tmp = strtoull(arg, &end, 0);
1317                 if (*end)
1318                         return EXT2_ET_INVALID_ARGUMENT;
1319                 data->offset = tmp;
1320                 if (data->offset < 0)
1321                         return EXT2_ET_INVALID_ARGUMENT;
1322                 return 0;
1323         }
1324         if (!strcmp(option, "cache")) {
1325                 if (!arg)
1326                         return EXT2_ET_INVALID_ARGUMENT;
1327                 if (!strcmp(arg, "on")) {
1328                         data->flags &= ~IO_FLAG_NOCACHE;
1329                         return 0;
1330                 }
1331                 if (!strcmp(arg, "off")) {
1332                         retval = flush_cached_blocks(channel, data, 0);
1333                         data->flags |= IO_FLAG_NOCACHE;
1334                         return retval;
1335                 }
1336                 return EXT2_ET_INVALID_ARGUMENT;
1337         }
1338         return EXT2_ET_INVALID_ARGUMENT;
1339 }
1340
1341 #if defined(__linux__) && !defined(BLKDISCARD)
1342 #define BLKDISCARD              _IO(0x12,119)
1343 #endif
1344
1345 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1346                               unsigned long long count)
1347 {
1348         struct unix_private_data *data;
1349         int             ret = EOPNOTSUPP;
1350
1351         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1352         data = (struct unix_private_data *) channel->private_data;
1353         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1354
1355         if (channel->flags & CHANNEL_FLAGS_NODISCARD)
1356                 goto unimplemented;
1357
1358         if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1359 #ifdef BLKDISCARD
1360                 __u64 range[2];
1361
1362                 range[0] = (__u64)(block) * channel->block_size + data->offset;
1363                 range[1] = (__u64)(count) * channel->block_size;
1364
1365                 ret = ioctl(data->dev, BLKDISCARD, &range);
1366 #else
1367                 goto unimplemented;
1368 #endif
1369         } else {
1370 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1371                 /*
1372                  * If we are not on block device, try to use punch hole
1373                  * to reclaim free space.
1374                  */
1375                 ret = fallocate(data->dev,
1376                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1377                                 (off_t)(block) * channel->block_size + data->offset,
1378                                 (off_t)(count) * channel->block_size);
1379 #else
1380                 goto unimplemented;
1381 #endif
1382         }
1383         if (ret < 0) {
1384                 if (errno == EOPNOTSUPP) {
1385                         channel->flags |= CHANNEL_FLAGS_NODISCARD;
1386                         goto unimplemented;
1387                 }
1388                 return errno;
1389         }
1390         return 0;
1391 unimplemented:
1392         return EXT2_ET_UNIMPLEMENTED;
1393 }
1394
1395 /*
1396  * If we know about ZERO_RANGE, try that before we try PUNCH_HOLE because
1397  * ZERO_RANGE doesn't unmap preallocated blocks.  We prefer fallocate because
1398  * it always invalidates page cache, and libext2fs requires that reads after
1399  * ZERO_RANGE return zeroes.
1400  */
1401 static int __unix_zeroout(int fd, off_t offset, off_t len)
1402 {
1403         int ret = -1;
1404
1405 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_ZERO_RANGE)
1406         ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, offset, len);
1407         if (ret == 0)
1408                 return 0;
1409 #endif
1410 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1411         ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1412                         offset,  len);
1413         if (ret == 0)
1414                 return 0;
1415 #endif
1416         errno = EOPNOTSUPP;
1417         return ret;
1418 }
1419
1420 /* parameters might not be used if OS doesn't support zeroout */
1421 #if __GNUC_PREREQ (4, 6)
1422 #pragma GCC diagnostic push
1423 #pragma GCC diagnostic ignored "-Wunused-parameter"
1424 #endif
1425 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1426                               unsigned long long count)
1427 {
1428         struct unix_private_data *data;
1429         int             ret;
1430
1431         EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1432         data = (struct unix_private_data *) channel->private_data;
1433         EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1434
1435         if (!(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE)) {
1436                 /* Regular file, try to use truncate/punch/zero. */
1437                 struct stat statbuf;
1438
1439                 if (count == 0)
1440                         return 0;
1441                 /*
1442                  * If we're trying to zero a range past the end of the file,
1443                  * extend the file size, then truncate everything.
1444                  */
1445                 ret = fstat(data->dev, &statbuf);
1446                 if (ret)
1447                         goto err;
1448                 if ((unsigned long long) statbuf.st_size <
1449                         (block + count) * channel->block_size + data->offset) {
1450                         ret = ftruncate(data->dev,
1451                                         (block + count) * channel->block_size + data->offset);
1452                         if (ret)
1453                                 goto err;
1454                 }
1455         }
1456
1457         if (channel->flags & CHANNEL_FLAGS_NOZEROOUT)
1458                 goto unimplemented;
1459
1460         ret = __unix_zeroout(data->dev,
1461                         (off_t)(block) * channel->block_size + data->offset,
1462                         (off_t)(count) * channel->block_size);
1463 err:
1464         if (ret < 0) {
1465                 if (errno == EOPNOTSUPP) {
1466                         channel->flags |= CHANNEL_FLAGS_NOZEROOUT;
1467                         goto unimplemented;
1468                 }
1469                 return errno;
1470         }
1471         return 0;
1472 unimplemented:
1473         return EXT2_ET_UNIMPLEMENTED;
1474 }
1475 #if __GNUC_PREREQ (4, 6)
1476 #pragma GCC diagnostic pop
1477 #endif
1478
1479 static struct struct_io_manager struct_unix_manager = {
1480         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1481         .name           = "Unix I/O Manager",
1482         .open           = unix_open,
1483         .close          = unix_close,
1484         .set_blksize    = unix_set_blksize,
1485         .read_blk       = unix_read_blk,
1486         .write_blk      = unix_write_blk,
1487         .flush          = unix_flush,
1488         .write_byte     = unix_write_byte,
1489         .set_option     = unix_set_option,
1490         .get_stats      = unix_get_stats,
1491         .read_blk64     = unix_read_blk64,
1492         .write_blk64    = unix_write_blk64,
1493         .discard        = unix_discard,
1494         .cache_readahead        = unix_cache_readahead,
1495         .zeroout        = unix_zeroout,
1496 };
1497
1498 io_manager unix_io_manager = &struct_unix_manager;
1499
1500 static struct struct_io_manager struct_unixfd_manager = {
1501         .magic          = EXT2_ET_MAGIC_IO_MANAGER,
1502         .name           = "Unix fd I/O Manager",
1503         .open           = unixfd_open,
1504         .close          = unix_close,
1505         .set_blksize    = unix_set_blksize,
1506         .read_blk       = unix_read_blk,
1507         .write_blk      = unix_write_blk,
1508         .flush          = unix_flush,
1509         .write_byte     = unix_write_byte,
1510         .set_option     = unix_set_option,
1511         .get_stats      = unix_get_stats,
1512         .read_blk64     = unix_read_blk64,
1513         .write_blk64    = unix_write_blk64,
1514         .discard        = unix_discard,
1515         .cache_readahead        = unix_cache_readahead,
1516         .zeroout        = unix_zeroout,
1517 };
1518
1519 io_manager unixfd_io_manager = &struct_unixfd_manager;