2 * unix_io.c --- This is the Unix (well, really POSIX) implementation
5 * Implements a one-block write-through cache.
7 * Includes support for Windows NT support under Cygwin.
9 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
10 * 2002 by Theodore Ts'o.
13 * This file may be redistributed under the terms of the GNU Library
14 * General Public License, version 2.
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
25 #ifndef _LARGEFILE64_SOURCE
26 #define _LARGEFILE64_SOURCE
45 #include <sys/utsname.h>
48 #include <sys/types.h>
50 #ifdef HAVE_SYS_IOCTL_H
51 #include <sys/ioctl.h>
53 #ifdef HAVE_SYS_MOUNT_H
54 #include <sys/mount.h>
56 #ifdef HAVE_SYS_PRCTL_H
57 #include <sys/prctl.h>
59 #define PR_GET_DUMPABLE 3
64 #if HAVE_SYS_RESOURCE_H
65 #include <sys/resource.h>
67 #if HAVE_LINUX_FALLOC_H
68 #include <linux/falloc.h>
74 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
75 #define BLKROGET _IO(0x12, 94) /* Get read-only status (0 = read_write). */
85 * For checking structure magic numbers...
88 #define EXT2_CHECK_MAGIC(struct, code) \
89 if ((struct)->magic != (code)) return (code)
93 unsigned long long block;
100 #define WRITE_DIRECT_SIZE 4 /* Must be smaller than CACHE_SIZE */
101 #define READ_DIRECT_SIZE 4 /* Should be smaller than CACHE_SIZE */
103 struct unix_private_data {
110 struct unix_cache cache[CACHE_SIZE];
112 struct struct_io_stats io_stats;
114 pthread_mutex_t cache_mutex;
115 pthread_mutex_t bounce_mutex;
116 pthread_mutex_t stats_mutex;
120 #define IS_ALIGNED(n, align) ((((uintptr_t) n) & \
121 ((uintptr_t) ((align)-1))) == 0)
123 typedef enum lock_kind {
124 CACHE_MTX, BOUNCE_MTX, STATS_MTX
128 static inline pthread_mutex_t *get_mutex(struct unix_private_data *data,
131 if (data->flags & IO_FLAG_THREADS) {
134 return &data->cache_mutex;
136 return &data->bounce_mutex;
138 return &data->stats_mutex;
145 static inline void mutex_lock(struct unix_private_data *data, kind_t kind)
148 pthread_mutex_t *mtx = get_mutex(data,kind);
151 pthread_mutex_lock(mtx);
155 static inline void mutex_unlock(struct unix_private_data *data, kind_t kind)
158 pthread_mutex_t *mtx = get_mutex(data,kind);
161 pthread_mutex_unlock(mtx);
165 static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
167 errcode_t retval = 0;
169 struct unix_private_data *data;
171 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
172 data = (struct unix_private_data *) channel->private_data;
173 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
176 mutex_lock(data, STATS_MTX);
177 *stats = &data->io_stats;
178 mutex_unlock(data, STATS_MTX);
184 static char *safe_getenv(const char *arg)
186 if ((getuid() != geteuid()) || (getgid() != getegid()))
189 if (prctl(PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
192 #if (defined(linux) && defined(SYS_prctl))
193 if (syscall(SYS_prctl, PR_GET_DUMPABLE, 0, 0, 0, 0) == 0)
198 #if defined(HAVE_SECURE_GETENV)
199 return secure_getenv(arg);
200 #elif defined(HAVE___SECURE_GETENV)
201 return __secure_getenv(arg);
208 * Here are the raw I/O functions
210 static errcode_t raw_read_blk(io_channel channel,
211 struct unix_private_data *data,
212 unsigned long long block,
213 int count, void *bufv)
217 ext2_loff_t location;
219 unsigned char *buf = bufv;
220 ssize_t really_read = 0;
221 unsigned long long aligned_blk;
222 int align_size, offset;
224 size = (count < 0) ? -count : (ext2_loff_t) count * channel->block_size;
225 mutex_lock(data, STATS_MTX);
226 data->io_stats.bytes_read += size;
227 mutex_unlock(data, STATS_MTX);
228 location = ((ext2_loff_t) block * channel->block_size) + data->offset;
230 if (data->flags & IO_FLAG_FORCE_BOUNCE)
234 /* Try an aligned pread */
235 if ((channel->align == 0) ||
236 (IS_ALIGNED(buf, channel->align) &&
237 IS_ALIGNED(location, channel->align) &&
238 IS_ALIGNED(size, channel->align))) {
239 actual = pread64(data->dev, buf, size, location);
245 /* Try an aligned pread */
246 if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
247 ((channel->align == 0) ||
248 (IS_ALIGNED(buf, channel->align) &&
249 IS_ALIGNED(location, channel->align) &&
250 IS_ALIGNED(size, channel->align)))) {
251 actual = pread(data->dev, buf, size, location);
256 #endif /* HAVE_PREAD */
258 if ((channel->align == 0) ||
259 (IS_ALIGNED(buf, channel->align) &&
260 IS_ALIGNED(location, channel->align) &&
261 IS_ALIGNED(size, channel->align))) {
262 mutex_lock(data, BOUNCE_MTX);
263 if (ext2fs_llseek(data->dev, location, SEEK_SET) < 0) {
264 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
267 actual = read(data->dev, buf, size);
268 if (actual != size) {
274 retval = EXT2_ET_SHORT_READ;
281 printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
282 (unsigned long) size);
286 * The buffer or size which we're trying to read isn't aligned
287 * to the O_DIRECT rules, so we need to do this the hard way...
290 if (channel->align == 0)
292 if ((channel->block_size > channel->align) &&
293 (channel->block_size % channel->align) == 0)
294 align_size = channel->block_size;
296 align_size = channel->align;
297 aligned_blk = location / align_size;
298 offset = location % align_size;
300 mutex_lock(data, BOUNCE_MTX);
301 if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
302 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
306 actual = read(data->dev, data->bounce, align_size);
307 if (actual != align_size) {
308 mutex_unlock(data, BOUNCE_MTX);
309 actual = really_read;
315 if (actual > align_size)
318 memcpy(buf, data->bounce + offset, actual);
320 really_read += actual;
327 mutex_unlock(data, BOUNCE_MTX);
331 mutex_unlock(data, BOUNCE_MTX);
333 if (actual >= 0 && actual < size)
334 memset((char *) buf+actual, 0, size-actual);
335 if (channel->read_error)
336 retval = (channel->read_error)(channel, block, count, buf,
337 size, actual, retval);
341 static errcode_t raw_write_blk(io_channel channel,
342 struct unix_private_data *data,
343 unsigned long long block,
344 int count, const void *bufv)
347 ext2_loff_t location;
350 const unsigned char *buf = bufv;
351 unsigned long long aligned_blk;
352 int align_size, offset;
355 size = channel->block_size;
360 size = (ext2_loff_t) count * channel->block_size;
362 mutex_lock(data, STATS_MTX);
363 data->io_stats.bytes_written += size;
364 mutex_unlock(data, STATS_MTX);
366 location = ((ext2_loff_t) block * channel->block_size) + data->offset;
368 if (data->flags & IO_FLAG_FORCE_BOUNCE)
372 /* Try an aligned pwrite */
373 if ((channel->align == 0) ||
374 (IS_ALIGNED(buf, channel->align) &&
375 IS_ALIGNED(location, channel->align) &&
376 IS_ALIGNED(size, channel->align))) {
377 actual = pwrite64(data->dev, buf, size, location);
382 /* Try an aligned pwrite */
383 if ((sizeof(off_t) >= sizeof(ext2_loff_t)) &&
384 ((channel->align == 0) ||
385 (IS_ALIGNED(buf, channel->align) &&
386 IS_ALIGNED(location, channel->align) &&
387 IS_ALIGNED(size, channel->align)))) {
388 actual = pwrite(data->dev, buf, size, location);
392 #endif /* HAVE_PWRITE */
394 if ((channel->align == 0) ||
395 (IS_ALIGNED(buf, channel->align) &&
396 IS_ALIGNED(location, channel->align) &&
397 IS_ALIGNED(size, channel->align))) {
398 mutex_lock(data, BOUNCE_MTX);
399 if (ext2fs_llseek(data->dev, location, SEEK_SET) < 0) {
400 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
403 actual = write(data->dev, buf, size);
404 mutex_unlock(data, BOUNCE_MTX);
409 if (actual != size) {
411 retval = EXT2_ET_SHORT_WRITE;
418 printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
419 (unsigned long) size);
422 * The buffer or size which we're trying to write isn't aligned
423 * to the O_DIRECT rules, so we need to do this the hard way...
426 if (channel->align == 0)
428 if ((channel->block_size > channel->align) &&
429 (channel->block_size % channel->align) == 0)
430 align_size = channel->block_size;
432 align_size = channel->align;
433 aligned_blk = location / align_size;
434 offset = location % align_size;
439 mutex_lock(data, BOUNCE_MTX);
440 if (size < align_size || offset) {
441 if (ext2fs_llseek(data->dev, aligned_blk * align_size,
443 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
446 actual = read(data->dev, data->bounce,
448 if (actual != align_size) {
453 memset((char *) data->bounce + actual, 0,
454 align_size - actual);
458 if (actual > align_size)
461 memcpy(((char *)data->bounce) + offset, buf, actual);
462 if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
463 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
466 actual_w = write(data->dev, data->bounce, align_size);
467 mutex_unlock(data, BOUNCE_MTX);
472 if (actual_w != align_size)
483 mutex_unlock(data, BOUNCE_MTX);
485 if (channel->write_error)
486 retval = (channel->write_error)(channel, block, count, buf,
487 size, actual, retval);
493 * Here we implement the cache functions
496 /* Allocate the cache buffers */
497 static errcode_t alloc_cache(io_channel channel,
498 struct unix_private_data *data)
501 struct unix_cache *cache;
504 data->access_time = 0;
505 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
507 cache->access_time = 0;
511 ext2fs_free_mem(&cache->buf);
512 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
516 if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
518 ext2fs_free_mem(&data->bounce);
519 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
524 /* Free the cache buffers */
525 static void free_cache(struct unix_private_data *data)
527 struct unix_cache *cache;
530 data->access_time = 0;
531 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
533 cache->access_time = 0;
537 ext2fs_free_mem(&cache->buf);
540 ext2fs_free_mem(&data->bounce);
545 * Try to find a block in the cache. If the block is not found, and
546 * eldest is a non-zero pointer, then fill in eldest with the cache
547 * entry to that should be reused.
549 static struct unix_cache *find_cached_block(struct unix_private_data *data,
550 unsigned long long block,
551 struct unix_cache **eldest)
553 struct unix_cache *cache, *unused_cache, *oldest_cache;
556 unused_cache = oldest_cache = 0;
557 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
558 if (!cache->in_use) {
560 unused_cache = cache;
563 if (cache->block == block) {
564 cache->access_time = ++data->access_time;
568 (cache->access_time < oldest_cache->access_time))
569 oldest_cache = cache;
572 *eldest = (unused_cache) ? unused_cache : oldest_cache;
577 * Reuse a particular cache entry for another block.
579 static void reuse_cache(io_channel channel, struct unix_private_data *data,
580 struct unix_cache *cache, unsigned long long block)
582 if (cache->dirty && cache->in_use)
583 raw_write_blk(channel, data, cache->block, 1, cache->buf);
587 cache->block = block;
588 cache->access_time = ++data->access_time;
591 #define FLUSH_INVALIDATE 0x01
592 #define FLUSH_NOLOCK 0x02
595 * Flush all of the blocks in the cache
597 static errcode_t flush_cached_blocks(io_channel channel,
598 struct unix_private_data *data,
601 struct unix_cache *cache;
602 errcode_t retval, retval2;
606 if ((flags & FLUSH_NOLOCK) == 0)
607 mutex_lock(data, CACHE_MTX);
608 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
612 if (flags & FLUSH_INVALIDATE)
618 retval = raw_write_blk(channel, data,
619 cache->block, 1, cache->buf);
625 if ((flags & FLUSH_NOLOCK) == 0)
626 mutex_unlock(data, CACHE_MTX);
629 #endif /* NO_IO_CACHE */
632 #ifndef BLKDISCARDZEROES
633 #define BLKDISCARDZEROES _IO(0x12,124)
637 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
640 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
641 return open64(pathname, flags, mode);
643 return open64(pathname, flags);
645 return open(pathname, flags, mode);
647 return open(pathname, flags);
651 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
653 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
654 return stat64(path, buf);
656 return stat(path, buf);
660 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
662 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
663 return fstat64(fd, buf);
665 return fstat(fd, buf);
670 static errcode_t unix_open_channel(const char *name, int fd,
671 int flags, io_channel *channel,
674 io_channel io = NULL;
675 struct unix_private_data *data = NULL;
677 ext2fs_struct_stat st;
682 if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
683 flags |= IO_FLAG_FORCE_BOUNCE;
687 * We need to make sure any previous errors in the block
688 * device are thrown away, sigh.
693 retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
696 memset(io, 0, sizeof(struct struct_io_channel));
697 io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
698 retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
702 io->manager = io_mgr;
703 retval = ext2fs_get_mem(strlen(name)+1, &io->name);
707 strcpy(io->name, name);
708 io->private_data = data;
709 io->block_size = 1024;
715 memset(data, 0, sizeof(struct unix_private_data));
716 data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
717 data->io_stats.num_fields = 2;
721 #if defined(O_DIRECT)
722 if (flags & IO_FLAG_DIRECT_IO)
723 io->align = ext2fs_get_dio_alignment(data->dev);
724 #elif defined(F_NOCACHE)
725 if (flags & IO_FLAG_DIRECT_IO)
730 * If the device is really a block device, then set the
731 * appropriate flag, otherwise we can set DISCARD_ZEROES flag
732 * because we are going to use punch hole instead of discard
733 * and if it succeed, subsequent read from sparse area returns
736 if (ext2fs_fstat(data->dev, &st) == 0) {
737 if (ext2fsP_is_disk_device(st.st_mode))
738 io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
740 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
743 #ifdef BLKDISCARDZEROES
746 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
748 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
752 #if defined(__CYGWIN__)
754 * Some operating systems require that the buffers be aligned,
755 * regardless of O_DIRECT
761 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
762 if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
763 int dio_align = ext2fs_get_dio_alignment(fd);
765 if (io->align < dio_align)
766 io->align = dio_align;
770 if ((retval = alloc_cache(io, data)))
774 if (flags & IO_FLAG_RW) {
778 /* Is the block device actually writable? */
779 error = ioctl(data->dev, BLKROGET, &readonly);
780 if (!error && readonly) {
789 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
790 #define RLIM_INFINITY ((unsigned long)(~0UL>>1))
792 #define RLIM_INFINITY (~0UL)
795 * Work around a bug in 2.4.10-2.4.18 kernels where writes to
796 * block devices are wrongly getting hit by the filesize
797 * limit. This workaround isn't perfect, since it won't work
798 * if glibc wasn't built against 2.2 header files. (Sigh.)
801 if ((flags & IO_FLAG_RW) &&
803 ((ut.release[0] == '2') && (ut.release[1] == '.') &&
804 (ut.release[2] == '4') && (ut.release[3] == '.') &&
805 (ut.release[4] == '1') && (ut.release[5] >= '0') &&
806 (ut.release[5] < '8')) &&
807 (ext2fs_fstat(data->dev, &st) == 0) &&
808 (ext2fsP_is_disk_device(st.st_mode))) {
811 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
812 setrlimit(RLIMIT_FSIZE, &rlim);
813 getrlimit(RLIMIT_FSIZE, &rlim);
814 if (((unsigned long) rlim.rlim_cur) <
815 ((unsigned long) rlim.rlim_max)) {
816 rlim.rlim_cur = rlim.rlim_max;
817 setrlimit(RLIMIT_FSIZE, &rlim);
822 if (flags & IO_FLAG_THREADS) {
823 io->flags |= CHANNEL_FLAGS_THREADS;
824 retval = pthread_mutex_init(&data->cache_mutex, NULL);
827 retval = pthread_mutex_init(&data->bounce_mutex, NULL);
829 pthread_mutex_destroy(&data->cache_mutex);
832 retval = pthread_mutex_init(&data->stats_mutex, NULL);
834 pthread_mutex_destroy(&data->cache_mutex);
835 pthread_mutex_destroy(&data->bounce_mutex);
848 ext2fs_free_mem(&data);
852 ext2fs_free_mem(&io->name);
854 ext2fs_free_mem(&io);
859 static errcode_t unixfd_open(const char *str_fd, int flags,
866 #if defined(HAVE_FCNTL)
867 fd_flags = fcntl(fd, F_GETFD);
872 if (fd_flags & O_RDWR)
874 if (fd_flags & O_EXCL)
875 flags |= IO_FLAG_EXCLUSIVE;
876 #if defined(O_DIRECT)
877 if (fd_flags & O_DIRECT)
878 flags |= IO_FLAG_DIRECT_IO;
880 #endif /* HAVE_FCNTL */
882 return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
885 static errcode_t unix_open(const char *name, int flags,
892 return EXT2_ET_BAD_DEVICE_NAME;
894 open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
895 if (flags & IO_FLAG_EXCLUSIVE)
896 open_flags |= O_EXCL;
897 #if defined(O_DIRECT)
898 if (flags & IO_FLAG_DIRECT_IO)
899 open_flags |= O_DIRECT;
901 fd = ext2fs_open_file(name, open_flags, 0);
904 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
905 if (flags & IO_FLAG_DIRECT_IO) {
906 if (fcntl(fd, F_NOCACHE, 1) < 0)
910 return unix_open_channel(name, fd, flags, channel, unix_io_manager);
913 static errcode_t unix_close(io_channel channel)
915 struct unix_private_data *data;
916 errcode_t retval = 0;
918 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
919 data = (struct unix_private_data *) channel->private_data;
920 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
922 if (--channel->refcount > 0)
926 retval = flush_cached_blocks(channel, data, 0);
929 if (close(data->dev) < 0)
933 if (data->flags & IO_FLAG_THREADS) {
934 pthread_mutex_destroy(&data->cache_mutex);
935 pthread_mutex_destroy(&data->bounce_mutex);
936 pthread_mutex_destroy(&data->stats_mutex);
940 ext2fs_free_mem(&channel->private_data);
942 ext2fs_free_mem(&channel->name);
943 ext2fs_free_mem(&channel);
947 static errcode_t unix_set_blksize(io_channel channel, int blksize)
949 struct unix_private_data *data;
950 errcode_t retval = 0;
952 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
953 data = (struct unix_private_data *) channel->private_data;
954 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
956 if (channel->block_size != blksize) {
957 mutex_lock(data, CACHE_MTX);
958 mutex_lock(data, BOUNCE_MTX);
960 if ((retval = flush_cached_blocks(channel, data, FLUSH_NOLOCK)))
964 channel->block_size = blksize;
966 retval = alloc_cache(channel, data);
967 mutex_unlock(data, BOUNCE_MTX);
968 mutex_unlock(data, CACHE_MTX);
973 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
974 int count, void *buf)
976 struct unix_private_data *data;
977 struct unix_cache *cache, *reuse[READ_DIRECT_SIZE];
978 errcode_t retval = 0;
982 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
983 data = (struct unix_private_data *) channel->private_data;
984 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
987 return raw_read_blk(channel, data, block, count, buf);
989 if (data->flags & IO_FLAG_NOCACHE)
990 return raw_read_blk(channel, data, block, count, buf);
992 * If we're doing an odd-sized read or a very large read,
993 * flush out the cache and then do a direct read.
995 if (count < 0 || count > WRITE_DIRECT_SIZE) {
996 if ((retval = flush_cached_blocks(channel, data, 0)))
998 return raw_read_blk(channel, data, block, count, buf);
1002 mutex_lock(data, CACHE_MTX);
1004 /* If it's in the cache, use it! */
1005 if ((cache = find_cached_block(data, block, &reuse[0]))) {
1007 printf("Using cached block %lu\n", block);
1009 memcpy(cp, cache->buf, channel->block_size);
1012 cp += channel->block_size;
1017 * Special case where we read directly into the
1018 * cache buffer; important in the O_DIRECT case
1021 reuse_cache(channel, data, cache, block);
1022 if ((retval = raw_read_blk(channel, data, block, 1,
1027 memcpy(cp, cache->buf, channel->block_size);
1033 * Find the number of uncached blocks so we can do a
1034 * single read request
1036 for (i=1; i < count; i++)
1037 if (find_cached_block(data, block+i, &reuse[i]))
1040 printf("Reading %d blocks starting at %lu\n", i, block);
1042 if ((retval = raw_read_blk(channel, data, block, i, cp)))
1045 /* Save the results in the cache */
1046 for (j=0; j < i; j++) {
1049 reuse_cache(channel, data, cache, block++);
1050 memcpy(cache->buf, cp, channel->block_size);
1051 cp += channel->block_size;
1054 mutex_unlock(data, CACHE_MTX);
1056 #endif /* NO_IO_CACHE */
1059 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
1060 int count, void *buf)
1062 return unix_read_blk64(channel, block, count, buf);
1065 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
1066 int count, const void *buf)
1068 struct unix_private_data *data;
1069 struct unix_cache *cache, *reuse;
1070 errcode_t retval = 0;
1074 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1075 data = (struct unix_private_data *) channel->private_data;
1076 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1079 return raw_write_blk(channel, data, block, count, buf);
1081 if (data->flags & IO_FLAG_NOCACHE)
1082 return raw_write_blk(channel, data, block, count, buf);
1084 * If we're doing an odd-sized write or a very large write,
1085 * flush out the cache completely and then do a direct write.
1087 if (count < 0 || count > WRITE_DIRECT_SIZE) {
1088 if ((retval = flush_cached_blocks(channel, data,
1091 return raw_write_blk(channel, data, block, count, buf);
1095 * For a moderate-sized multi-block write, first force a write
1096 * if we're in write-through cache mode, and then fill the
1097 * cache with the blocks.
1099 writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
1101 retval = raw_write_blk(channel, data, block, count, buf);
1104 mutex_lock(data, CACHE_MTX);
1106 cache = find_cached_block(data, block, &reuse);
1109 reuse_cache(channel, data, cache, block);
1111 if (cache->buf != cp)
1112 memcpy(cache->buf, cp, channel->block_size);
1113 cache->dirty = !writethrough;
1116 cp += channel->block_size;
1118 mutex_unlock(data, CACHE_MTX);
1120 #endif /* NO_IO_CACHE */
1123 static errcode_t unix_cache_readahead(io_channel channel,
1124 unsigned long long block,
1125 unsigned long long count)
1127 #ifdef POSIX_FADV_WILLNEED
1128 struct unix_private_data *data;
1130 data = (struct unix_private_data *)channel->private_data;
1131 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1132 return posix_fadvise(data->dev,
1133 (ext2_loff_t)block * channel->block_size + data->offset,
1134 (ext2_loff_t)count * channel->block_size,
1135 POSIX_FADV_WILLNEED);
1137 return EXT2_ET_OP_NOT_SUPPORTED;
1141 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
1142 int count, const void *buf)
1144 return unix_write_blk64(channel, block, count, buf);
1147 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
1148 int size, const void *buf)
1150 struct unix_private_data *data;
1151 errcode_t retval = 0;
1154 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1155 data = (struct unix_private_data *) channel->private_data;
1156 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1158 if (channel->align != 0) {
1160 printf("unix_write_byte: O_DIRECT fallback\n");
1162 return EXT2_ET_UNIMPLEMENTED;
1167 * Flush out the cache completely
1169 if ((retval = flush_cached_blocks(channel, data, FLUSH_INVALIDATE)))
1173 if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1176 actual = write(data->dev, buf, size);
1180 return EXT2_ET_SHORT_WRITE;
1186 * Flush data buffers to disk and invalidate cache if needed
1188 static errcode_t _unix_flush(io_channel channel, int invalidate)
1190 struct unix_private_data *data;
1191 errcode_t retval = 0;
1193 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1194 data = (struct unix_private_data *) channel->private_data;
1195 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1198 retval = flush_cached_blocks(channel, data, invalidate);
1201 if (!retval && fsync(data->dev) != 0)
1208 * Flush data buffers to disk.
1210 static errcode_t unix_flush(io_channel channel)
1212 return _unix_flush(channel, 0);
1216 * Flush data buffers to disk and invalidate cache.
1218 static errcode_t unix_flush_cleanup(io_channel channel)
1220 return _unix_flush(channel, 1);
1223 static errcode_t unix_set_option(io_channel channel, const char *option,
1226 struct unix_private_data *data;
1227 unsigned long long tmp;
1231 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1232 data = (struct unix_private_data *) channel->private_data;
1233 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1235 if (!strcmp(option, "offset")) {
1237 return EXT2_ET_INVALID_ARGUMENT;
1239 tmp = strtoull(arg, &end, 0);
1241 return EXT2_ET_INVALID_ARGUMENT;
1243 if (data->offset < 0)
1244 return EXT2_ET_INVALID_ARGUMENT;
1247 if (!strcmp(option, "cache")) {
1249 return EXT2_ET_INVALID_ARGUMENT;
1250 if (!strcmp(arg, "on")) {
1251 data->flags &= ~IO_FLAG_NOCACHE;
1254 if (!strcmp(arg, "off")) {
1255 retval = flush_cached_blocks(channel, data, 0);
1256 data->flags |= IO_FLAG_NOCACHE;
1259 return EXT2_ET_INVALID_ARGUMENT;
1261 return EXT2_ET_INVALID_ARGUMENT;
1264 #if defined(__linux__) && !defined(BLKDISCARD)
1265 #define BLKDISCARD _IO(0x12,119)
1268 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1269 unsigned long long count)
1271 struct unix_private_data *data;
1274 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1275 data = (struct unix_private_data *) channel->private_data;
1276 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1278 if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1282 range[0] = (__u64)(block) * channel->block_size + data->offset;
1283 range[1] = (__u64)(count) * channel->block_size;
1285 ret = ioctl(data->dev, BLKDISCARD, &range);
1290 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1292 * If we are not on block device, try to use punch hole
1293 * to reclaim free space.
1295 ret = fallocate(data->dev,
1296 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1297 (off_t)(block) * channel->block_size + data->offset,
1298 (off_t)(count) * channel->block_size);
1304 if (errno == EOPNOTSUPP)
1310 return EXT2_ET_UNIMPLEMENTED;
1314 * If we know about ZERO_RANGE, try that before we try PUNCH_HOLE because
1315 * ZERO_RANGE doesn't unmap preallocated blocks. We prefer fallocate because
1316 * it always invalidates page cache, and libext2fs requires that reads after
1317 * ZERO_RANGE return zeroes.
1319 static int __unix_zeroout(int fd, off_t offset, off_t len)
1323 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_ZERO_RANGE)
1324 ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, offset, len);
1328 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1329 ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1338 /* parameters might not be used if OS doesn't support zeroout */
1339 #if __GNUC_PREREQ (4, 6)
1340 #pragma GCC diagnostic push
1341 #pragma GCC diagnostic ignored "-Wunused-parameter"
1343 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1344 unsigned long long count)
1346 struct unix_private_data *data;
1349 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1350 data = (struct unix_private_data *) channel->private_data;
1351 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1353 if (safe_getenv("UNIX_IO_NOZEROOUT"))
1356 if (!(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE)) {
1357 /* Regular file, try to use truncate/punch/zero. */
1358 struct stat statbuf;
1363 * If we're trying to zero a range past the end of the file,
1364 * extend the file size, then truncate everything.
1366 ret = fstat(data->dev, &statbuf);
1369 if ((unsigned long long) statbuf.st_size <
1370 (block + count) * channel->block_size + data->offset) {
1371 ret = ftruncate(data->dev,
1372 (block + count) * channel->block_size + data->offset);
1378 ret = __unix_zeroout(data->dev,
1379 (off_t)(block) * channel->block_size + data->offset,
1380 (off_t)(count) * channel->block_size);
1383 if (errno == EOPNOTSUPP)
1389 return EXT2_ET_UNIMPLEMENTED;
1391 #if __GNUC_PREREQ (4, 6)
1392 #pragma GCC diagnostic pop
1395 static struct struct_io_manager struct_unix_manager = {
1396 .magic = EXT2_ET_MAGIC_IO_MANAGER,
1397 .name = "Unix I/O Manager",
1399 .close = unix_close,
1400 .set_blksize = unix_set_blksize,
1401 .read_blk = unix_read_blk,
1402 .write_blk = unix_write_blk,
1403 .flush = unix_flush,
1404 .write_byte = unix_write_byte,
1405 .set_option = unix_set_option,
1406 .get_stats = unix_get_stats,
1407 .read_blk64 = unix_read_blk64,
1408 .write_blk64 = unix_write_blk64,
1409 .discard = unix_discard,
1410 .cache_readahead = unix_cache_readahead,
1411 .zeroout = unix_zeroout,
1412 .flush_cleanup = unix_flush_cleanup,
1415 io_manager unix_io_manager = &struct_unix_manager;
1417 static struct struct_io_manager struct_unixfd_manager = {
1418 .magic = EXT2_ET_MAGIC_IO_MANAGER,
1419 .name = "Unix fd I/O Manager",
1420 .open = unixfd_open,
1421 .close = unix_close,
1422 .set_blksize = unix_set_blksize,
1423 .read_blk = unix_read_blk,
1424 .write_blk = unix_write_blk,
1425 .flush = unix_flush,
1426 .write_byte = unix_write_byte,
1427 .set_option = unix_set_option,
1428 .get_stats = unix_get_stats,
1429 .read_blk64 = unix_read_blk64,
1430 .write_blk64 = unix_write_blk64,
1431 .discard = unix_discard,
1432 .cache_readahead = unix_cache_readahead,
1433 .zeroout = unix_zeroout,
1434 .flush_cleanup = unix_flush_cleanup,
1437 io_manager unixfd_io_manager = &struct_unixfd_manager;