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;
314 if ((actual + offset) > align_size)
315 actual = align_size - offset;
318 memcpy(buf, (char *)data->bounce + offset, actual);
320 really_read += actual;
327 mutex_unlock(data, BOUNCE_MTX);
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);
340 #define RAW_WRITE_NO_HANDLER 1
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,
349 ext2_loff_t location;
352 const unsigned char *buf = bufv;
353 unsigned long long aligned_blk;
354 int align_size, offset;
357 size = channel->block_size;
362 size = (ext2_loff_t) count * channel->block_size;
364 mutex_lock(data, STATS_MTX);
365 data->io_stats.bytes_written += size;
366 mutex_unlock(data, STATS_MTX);
368 location = ((ext2_loff_t) block * channel->block_size) + data->offset;
370 if (data->flags & IO_FLAG_FORCE_BOUNCE)
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);
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);
394 #endif /* HAVE_PWRITE */
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;
405 actual = write(data->dev, buf, size);
406 mutex_unlock(data, BOUNCE_MTX);
411 if (actual != size) {
413 retval = EXT2_ET_SHORT_WRITE;
420 printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
421 (unsigned long) size);
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...
428 if (channel->align == 0)
430 if ((channel->block_size > channel->align) &&
431 (channel->block_size % channel->align) == 0)
432 align_size = channel->block_size;
434 align_size = channel->align;
435 aligned_blk = location / align_size;
436 offset = location % align_size;
441 mutex_lock(data, BOUNCE_MTX);
442 if (size < align_size || offset) {
443 if (ext2fs_llseek(data->dev, aligned_blk * align_size,
445 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
448 actual = read(data->dev, data->bounce,
450 if (actual != align_size) {
455 memset((char *) data->bounce + actual, 0,
456 align_size - actual);
460 if ((actual + offset) > align_size)
461 actual = align_size - offset;
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;
469 actual_w = write(data->dev, data->bounce, align_size);
470 mutex_unlock(data, BOUNCE_MTX);
475 if (actual_w != align_size)
486 mutex_unlock(data, BOUNCE_MTX);
488 if (((flags & RAW_WRITE_NO_HANDLER) == 0) && channel->write_error)
489 retval = (channel->write_error)(channel, block, count, buf,
490 size, actual, retval);
496 * Here we implement the cache functions
499 /* Allocate the cache buffers */
500 static errcode_t alloc_cache(io_channel channel,
501 struct unix_private_data *data)
504 struct unix_cache *cache;
507 data->access_time = 0;
508 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
510 cache->access_time = 0;
514 ext2fs_free_mem(&cache->buf);
515 retval = io_channel_alloc_buf(channel, 0, &cache->buf);
519 if (channel->align || data->flags & IO_FLAG_FORCE_BOUNCE) {
521 ext2fs_free_mem(&data->bounce);
522 retval = io_channel_alloc_buf(channel, 0, &data->bounce);
527 /* Free the cache buffers */
528 static void free_cache(struct unix_private_data *data)
530 struct unix_cache *cache;
533 data->access_time = 0;
534 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
536 cache->access_time = 0;
540 ext2fs_free_mem(&cache->buf);
543 ext2fs_free_mem(&data->bounce);
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.
552 static struct unix_cache *find_cached_block(struct unix_private_data *data,
553 unsigned long long block,
554 struct unix_cache **eldest)
556 struct unix_cache *cache, *unused_cache, *oldest_cache;
559 unused_cache = oldest_cache = 0;
560 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
561 if (!cache->in_use) {
563 unused_cache = cache;
566 if (cache->block == block) {
567 cache->access_time = ++data->access_time;
571 (cache->access_time < oldest_cache->access_time))
572 oldest_cache = cache;
575 *eldest = (unused_cache) ? unused_cache : oldest_cache;
580 * Reuse a particular cache entry for another block.
582 static void reuse_cache(io_channel channel, struct unix_private_data *data,
583 struct unix_cache *cache, unsigned long long block)
585 if (cache->dirty && cache->in_use)
586 raw_write_blk(channel, data, cache->block, 1, cache->buf, 0);
590 cache->block = block;
591 cache->access_time = ++data->access_time;
594 #define FLUSH_INVALIDATE 0x01
595 #define FLUSH_NOLOCK 0x02
598 * Flush all of the blocks in the cache
600 static errcode_t flush_cached_blocks(io_channel channel,
601 struct unix_private_data *data,
604 struct unix_cache *cache;
605 errcode_t retval, retval2;
609 if ((flags & FLUSH_NOLOCK) == 0)
610 mutex_lock(data, CACHE_MTX);
611 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
615 if (flags & FLUSH_INVALIDATE)
621 retval = raw_write_blk(channel, data,
622 cache->block, 1, cache->buf, 0);
628 if ((flags & FLUSH_NOLOCK) == 0)
629 mutex_unlock(data, CACHE_MTX);
632 #endif /* NO_IO_CACHE */
635 #ifndef BLKDISCARDZEROES
636 #define BLKDISCARDZEROES _IO(0x12,124)
640 int ext2fs_open_file(const char *pathname, int flags, mode_t mode)
643 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
644 return open64(pathname, flags, mode);
646 return open64(pathname, flags);
648 return open(pathname, flags, mode);
650 return open(pathname, flags);
654 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf)
656 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
657 return stat64(path, buf);
659 return stat(path, buf);
663 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf)
665 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED)
666 return fstat64(fd, buf);
668 return fstat(fd, buf);
673 static errcode_t unix_open_channel(const char *name, int fd,
674 int flags, io_channel *channel,
677 io_channel io = NULL;
678 struct unix_private_data *data = NULL;
680 ext2fs_struct_stat st;
685 if (safe_getenv("UNIX_IO_FORCE_BOUNCE"))
686 flags |= IO_FLAG_FORCE_BOUNCE;
690 * We need to make sure any previous errors in the block
691 * device are thrown away, sigh.
696 retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
699 memset(io, 0, sizeof(struct struct_io_channel));
700 io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
701 retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
705 io->manager = io_mgr;
706 retval = ext2fs_get_mem(strlen(name)+1, &io->name);
710 strcpy(io->name, name);
711 io->private_data = data;
712 io->block_size = 1024;
718 memset(data, 0, sizeof(struct unix_private_data));
719 data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
720 data->io_stats.num_fields = 2;
724 #if defined(O_DIRECT)
725 if (flags & IO_FLAG_DIRECT_IO)
726 io->align = ext2fs_get_dio_alignment(data->dev);
727 #elif defined(F_NOCACHE)
728 if (flags & IO_FLAG_DIRECT_IO)
733 * If the device is really a block device, then set the
734 * appropriate flag, otherwise we can set DISCARD_ZEROES flag
735 * because we are going to use punch hole instead of discard
736 * and if it succeed, subsequent read from sparse area returns
739 if (ext2fs_fstat(data->dev, &st) == 0) {
740 if (ext2fsP_is_disk_device(st.st_mode))
741 io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE;
743 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
746 #ifdef BLKDISCARDZEROES
749 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 &&
751 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
755 #if defined(__CYGWIN__)
757 * Some operating systems require that the buffers be aligned,
758 * regardless of O_DIRECT
764 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
765 if (io->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
766 int dio_align = ext2fs_get_dio_alignment(fd);
768 if (io->align < dio_align)
769 io->align = dio_align;
773 if ((retval = alloc_cache(io, data)))
777 if (flags & IO_FLAG_RW) {
781 /* Is the block device actually writable? */
782 error = ioctl(data->dev, BLKROGET, &readonly);
783 if (!error && readonly) {
792 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
793 #define RLIM_INFINITY ((unsigned long)(~0UL>>1))
795 #define RLIM_INFINITY (~0UL)
798 * Work around a bug in 2.4.10-2.4.18 kernels where writes to
799 * block devices are wrongly getting hit by the filesize
800 * limit. This workaround isn't perfect, since it won't work
801 * if glibc wasn't built against 2.2 header files. (Sigh.)
804 if ((flags & IO_FLAG_RW) &&
806 ((ut.release[0] == '2') && (ut.release[1] == '.') &&
807 (ut.release[2] == '4') && (ut.release[3] == '.') &&
808 (ut.release[4] == '1') && (ut.release[5] >= '0') &&
809 (ut.release[5] < '8')) &&
810 (ext2fs_fstat(data->dev, &st) == 0) &&
811 (ext2fsP_is_disk_device(st.st_mode))) {
814 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
815 setrlimit(RLIMIT_FSIZE, &rlim);
816 getrlimit(RLIMIT_FSIZE, &rlim);
817 if (((unsigned long) rlim.rlim_cur) <
818 ((unsigned long) rlim.rlim_max)) {
819 rlim.rlim_cur = rlim.rlim_max;
820 setrlimit(RLIMIT_FSIZE, &rlim);
825 if (flags & IO_FLAG_THREADS) {
826 io->flags |= CHANNEL_FLAGS_THREADS;
827 retval = pthread_mutex_init(&data->cache_mutex, NULL);
830 retval = pthread_mutex_init(&data->bounce_mutex, NULL);
832 pthread_mutex_destroy(&data->cache_mutex);
835 retval = pthread_mutex_init(&data->stats_mutex, NULL);
837 pthread_mutex_destroy(&data->cache_mutex);
838 pthread_mutex_destroy(&data->bounce_mutex);
851 ext2fs_free_mem(&data);
855 ext2fs_free_mem(&io->name);
857 ext2fs_free_mem(&io);
862 static errcode_t unixfd_open(const char *str_fd, int flags,
869 #if defined(HAVE_FCNTL)
870 fd_flags = fcntl(fd, F_GETFD);
875 if (fd_flags & O_RDWR)
877 if (fd_flags & O_EXCL)
878 flags |= IO_FLAG_EXCLUSIVE;
879 #if defined(O_DIRECT)
880 if (fd_flags & O_DIRECT)
881 flags |= IO_FLAG_DIRECT_IO;
883 #endif /* HAVE_FCNTL */
885 return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager);
888 static errcode_t unix_open(const char *name, int flags,
895 return EXT2_ET_BAD_DEVICE_NAME;
897 open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
898 if (flags & IO_FLAG_EXCLUSIVE)
899 open_flags |= O_EXCL;
900 #if defined(O_DIRECT)
901 if (flags & IO_FLAG_DIRECT_IO)
902 open_flags |= O_DIRECT;
904 fd = ext2fs_open_file(name, open_flags, 0);
907 #if defined(F_NOCACHE) && !defined(IO_DIRECT)
908 if (flags & IO_FLAG_DIRECT_IO) {
909 if (fcntl(fd, F_NOCACHE, 1) < 0)
913 return unix_open_channel(name, fd, flags, channel, unix_io_manager);
916 static errcode_t unix_close(io_channel channel)
918 struct unix_private_data *data;
919 errcode_t retval = 0;
921 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
922 data = (struct unix_private_data *) channel->private_data;
923 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
925 if (--channel->refcount > 0)
929 retval = flush_cached_blocks(channel, data, 0);
932 if (close(data->dev) < 0)
936 if (data->flags & IO_FLAG_THREADS) {
937 pthread_mutex_destroy(&data->cache_mutex);
938 pthread_mutex_destroy(&data->bounce_mutex);
939 pthread_mutex_destroy(&data->stats_mutex);
943 ext2fs_free_mem(&channel->private_data);
945 ext2fs_free_mem(&channel->name);
946 ext2fs_free_mem(&channel);
950 static errcode_t unix_set_blksize(io_channel channel, int blksize)
952 struct unix_private_data *data;
953 errcode_t retval = 0;
955 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
956 data = (struct unix_private_data *) channel->private_data;
957 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
959 if (channel->block_size != blksize) {
960 mutex_lock(data, CACHE_MTX);
961 mutex_lock(data, BOUNCE_MTX);
963 if ((retval = flush_cached_blocks(channel, data, FLUSH_NOLOCK))){
964 mutex_unlock(data, BOUNCE_MTX);
965 mutex_unlock(data, CACHE_MTX);
970 channel->block_size = blksize;
972 retval = alloc_cache(channel, data);
973 mutex_unlock(data, BOUNCE_MTX);
974 mutex_unlock(data, CACHE_MTX);
979 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
980 int count, void *buf)
982 struct unix_private_data *data;
983 struct unix_cache *cache;
988 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
989 data = (struct unix_private_data *) channel->private_data;
990 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
993 return raw_read_blk(channel, data, block, count, buf);
995 if (data->flags & IO_FLAG_NOCACHE)
996 return raw_read_blk(channel, data, block, count, buf);
998 * If we're doing an odd-sized read or a very large read,
999 * flush out the cache and then do a direct read.
1001 if (count < 0 || count > WRITE_DIRECT_SIZE) {
1002 if ((retval = flush_cached_blocks(channel, data, 0)))
1004 return raw_read_blk(channel, data, block, count, buf);
1008 mutex_lock(data, CACHE_MTX);
1010 /* If it's in the cache, use it! */
1011 if ((cache = find_cached_block(data, block, NULL))) {
1013 printf("Using cached block %lu\n", block);
1015 memcpy(cp, cache->buf, channel->block_size);
1018 cp += channel->block_size;
1023 * Find the number of uncached blocks so we can do a
1024 * single read request
1026 for (i=1; i < count; i++)
1027 if (find_cached_block(data, block+i, NULL))
1030 printf("Reading %d blocks starting at %lu\n", i, block);
1032 mutex_unlock(data, CACHE_MTX);
1033 if ((retval = raw_read_blk(channel, data, block, i, cp)))
1035 mutex_lock(data, CACHE_MTX);
1037 /* Save the results in the cache */
1038 for (j=0; j < i; j++) {
1039 if (!find_cached_block(data, block, &cache)) {
1040 reuse_cache(channel, data, cache, block);
1041 memcpy(cache->buf, cp, channel->block_size);
1045 cp += channel->block_size;
1048 mutex_unlock(data, CACHE_MTX);
1050 #endif /* NO_IO_CACHE */
1053 static errcode_t unix_read_blk(io_channel channel, unsigned long block,
1054 int count, void *buf)
1056 return unix_read_blk64(channel, block, count, buf);
1059 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
1060 int count, const void *buf)
1062 struct unix_private_data *data;
1063 struct unix_cache *cache, *reuse;
1064 errcode_t retval = 0;
1068 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1069 data = (struct unix_private_data *) channel->private_data;
1070 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1073 return raw_write_blk(channel, data, block, count, buf, 0);
1075 if (data->flags & IO_FLAG_NOCACHE)
1076 return raw_write_blk(channel, data, block, count, buf, 0);
1078 * If we're doing an odd-sized write or a very large write,
1079 * flush out the cache completely and then do a direct write.
1081 if (count < 0 || count > WRITE_DIRECT_SIZE) {
1082 if ((retval = flush_cached_blocks(channel, data,
1085 return raw_write_blk(channel, data, block, count, buf, 0);
1089 * For a moderate-sized multi-block write, first force a write
1090 * if we're in write-through cache mode, and then fill the
1091 * cache with the blocks.
1093 writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
1095 retval = raw_write_blk(channel, data, block, count, buf, 0);
1098 mutex_lock(data, CACHE_MTX);
1100 cache = find_cached_block(data, block, &reuse);
1103 reuse_cache(channel, data, cache, block);
1105 if (cache->buf != cp)
1106 memcpy(cache->buf, cp, channel->block_size);
1107 cache->dirty = !writethrough;
1110 cp += channel->block_size;
1112 mutex_unlock(data, CACHE_MTX);
1114 #endif /* NO_IO_CACHE */
1117 static errcode_t unix_cache_readahead(io_channel channel,
1118 unsigned long long block,
1119 unsigned long long count)
1121 #ifdef POSIX_FADV_WILLNEED
1122 struct unix_private_data *data;
1124 data = (struct unix_private_data *)channel->private_data;
1125 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1126 return posix_fadvise(data->dev,
1127 (ext2_loff_t)block * channel->block_size + data->offset,
1128 (ext2_loff_t)count * channel->block_size,
1129 POSIX_FADV_WILLNEED);
1131 return EXT2_ET_OP_NOT_SUPPORTED;
1135 static errcode_t unix_write_blk(io_channel channel, unsigned long block,
1136 int count, const void *buf)
1138 return unix_write_blk64(channel, block, count, buf);
1141 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
1142 int size, const void *buf)
1144 struct unix_private_data *data;
1145 errcode_t retval = 0;
1148 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1149 data = (struct unix_private_data *) channel->private_data;
1150 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1152 if (channel->align != 0) {
1154 printf("unix_write_byte: O_DIRECT fallback\n");
1156 return EXT2_ET_UNIMPLEMENTED;
1161 * Flush out the cache completely
1163 if ((retval = flush_cached_blocks(channel, data, FLUSH_INVALIDATE)))
1167 if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
1170 actual = write(data->dev, buf, size);
1174 return EXT2_ET_SHORT_WRITE;
1180 * Flush data buffers to disk.
1182 static errcode_t unix_flush(io_channel channel)
1184 struct unix_private_data *data;
1185 errcode_t retval = 0;
1187 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1188 data = (struct unix_private_data *) channel->private_data;
1189 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1192 retval = flush_cached_blocks(channel, data, 0);
1195 if (!retval && fsync(data->dev) != 0)
1201 static errcode_t unix_set_option(io_channel channel, const char *option,
1204 struct unix_private_data *data;
1205 unsigned long long tmp;
1209 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1210 data = (struct unix_private_data *) channel->private_data;
1211 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1213 if (!strcmp(option, "offset")) {
1215 return EXT2_ET_INVALID_ARGUMENT;
1217 tmp = strtoull(arg, &end, 0);
1219 return EXT2_ET_INVALID_ARGUMENT;
1221 if (data->offset < 0)
1222 return EXT2_ET_INVALID_ARGUMENT;
1225 if (!strcmp(option, "cache")) {
1227 return EXT2_ET_INVALID_ARGUMENT;
1228 if (!strcmp(arg, "on")) {
1229 data->flags &= ~IO_FLAG_NOCACHE;
1232 if (!strcmp(arg, "off")) {
1233 retval = flush_cached_blocks(channel, data, 0);
1234 data->flags |= IO_FLAG_NOCACHE;
1237 return EXT2_ET_INVALID_ARGUMENT;
1239 return EXT2_ET_INVALID_ARGUMENT;
1242 #if defined(__linux__) && !defined(BLKDISCARD)
1243 #define BLKDISCARD _IO(0x12,119)
1246 static errcode_t unix_discard(io_channel channel, unsigned long long block,
1247 unsigned long long count)
1249 struct unix_private_data *data;
1252 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1253 data = (struct unix_private_data *) channel->private_data;
1254 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1256 if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) {
1260 range[0] = (__u64)(block) * channel->block_size + data->offset;
1261 range[1] = (__u64)(count) * channel->block_size;
1263 ret = ioctl(data->dev, BLKDISCARD, &range);
1268 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE)
1270 * If we are not on block device, try to use punch hole
1271 * to reclaim free space.
1273 ret = fallocate(data->dev,
1274 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1275 (off_t)(block) * channel->block_size + data->offset,
1276 (off_t)(count) * channel->block_size);
1282 if (errno == EOPNOTSUPP)
1288 return EXT2_ET_UNIMPLEMENTED;
1292 * If we know about ZERO_RANGE, try that before we try PUNCH_HOLE because
1293 * ZERO_RANGE doesn't unmap preallocated blocks. We prefer fallocate because
1294 * it always invalidates page cache, and libext2fs requires that reads after
1295 * ZERO_RANGE return zeroes.
1297 static int __unix_zeroout(int fd, off_t offset, off_t len)
1301 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_ZERO_RANGE)
1302 ret = fallocate(fd, FALLOC_FL_ZERO_RANGE, offset, len);
1306 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE)
1307 ret = fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1316 /* parameters might not be used if OS doesn't support zeroout */
1317 #if __GNUC_PREREQ (4, 6)
1318 #pragma GCC diagnostic push
1319 #pragma GCC diagnostic ignored "-Wunused-parameter"
1321 static errcode_t unix_zeroout(io_channel channel, unsigned long long block,
1322 unsigned long long count)
1324 struct unix_private_data *data;
1327 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
1328 data = (struct unix_private_data *) channel->private_data;
1329 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
1331 if (safe_getenv("UNIX_IO_NOZEROOUT"))
1334 if (!(channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE)) {
1335 /* Regular file, try to use truncate/punch/zero. */
1336 struct stat statbuf;
1341 * If we're trying to zero a range past the end of the file,
1342 * extend the file size, then truncate everything.
1344 ret = fstat(data->dev, &statbuf);
1347 if ((unsigned long long) statbuf.st_size <
1348 (block + count) * channel->block_size + data->offset) {
1349 ret = ftruncate(data->dev,
1350 (block + count) * channel->block_size + data->offset);
1356 ret = __unix_zeroout(data->dev,
1357 (off_t)(block) * channel->block_size + data->offset,
1358 (off_t)(count) * channel->block_size);
1361 if (errno == EOPNOTSUPP)
1367 return EXT2_ET_UNIMPLEMENTED;
1369 #if __GNUC_PREREQ (4, 6)
1370 #pragma GCC diagnostic pop
1373 static struct struct_io_manager struct_unix_manager = {
1374 .magic = EXT2_ET_MAGIC_IO_MANAGER,
1375 .name = "Unix I/O Manager",
1377 .close = unix_close,
1378 .set_blksize = unix_set_blksize,
1379 .read_blk = unix_read_blk,
1380 .write_blk = unix_write_blk,
1381 .flush = unix_flush,
1382 .write_byte = unix_write_byte,
1383 .set_option = unix_set_option,
1384 .get_stats = unix_get_stats,
1385 .read_blk64 = unix_read_blk64,
1386 .write_blk64 = unix_write_blk64,
1387 .discard = unix_discard,
1388 .cache_readahead = unix_cache_readahead,
1389 .zeroout = unix_zeroout,
1392 io_manager unix_io_manager = &struct_unix_manager;
1394 static struct struct_io_manager struct_unixfd_manager = {
1395 .magic = EXT2_ET_MAGIC_IO_MANAGER,
1396 .name = "Unix fd I/O Manager",
1397 .open = unixfd_open,
1398 .close = unix_close,
1399 .set_blksize = unix_set_blksize,
1400 .read_blk = unix_read_blk,
1401 .write_blk = unix_write_blk,
1402 .flush = unix_flush,
1403 .write_byte = unix_write_byte,
1404 .set_option = unix_set_option,
1405 .get_stats = unix_get_stats,
1406 .read_blk64 = unix_read_blk64,
1407 .write_blk64 = unix_write_blk64,
1408 .discard = unix_discard,
1409 .cache_readahead = unix_cache_readahead,
1410 .zeroout = unix_zeroout,
1413 io_manager unixfd_io_manager = &struct_unixfd_manager;