Whamcloud - gitweb
libext2fs: fix regression in ext2fs_new_block2() for uninit_bg file systems
[tools/e2fsprogs.git] / lib / ext2fs / unix_io.c
index ae498da..dddaf6f 100644 (file)
@@ -4,19 +4,22 @@
  *
  * Implements a one-block write-through cache.
  *
- * Includes support for Windows NT support under Cygwin. 
+ * Includes support for Windows NT support under Cygwin.
  *
  * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
  *     2002 by Theodore Ts'o.
  *
  * %Begin-Header%
- * This file may be redistributed under the terms of the GNU Public
- * License.
+ * This file may be redistributed under the terms of the GNU Library
+ * General Public License, version 2.
  * %End-Header%
  */
 
 #define _LARGEFILE_SOURCE
 #define _LARGEFILE64_SOURCE
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
 
 #include <stdio.h>
 #include <string.h>
 #ifdef __linux__
 #include <sys/utsname.h>
 #endif
+#ifdef HAVE_SYS_IOCTL_H
+#include <sys/ioctl.h>
+#endif
+#ifdef HAVE_SYS_MOUNT_H
+#include <sys/mount.h>
+#endif
 #if HAVE_SYS_STAT_H
 #include <sys/stat.h>
 #endif
 #include <sys/resource.h>
 #endif
 
+#if defined(__linux__) && defined(_IO) && !defined(BLKROGET)
+#define BLKROGET   _IO(0x12, 94) /* Get read-only status (0 = read_write).  */
+#endif
+
+#if defined(__linux__) && defined(_IO) && !defined(BLKSSZGET)
+#define BLKSSZGET  _IO(0x12,104)/* get block device sector size */
+#endif
+
+#undef ALIGN_DEBUG
+
 #include "ext2_fs.h"
 #include "ext2fs.h"
 
@@ -55,8 +74,8 @@ struct unix_cache {
        char            *buf;
        unsigned long   block;
        int             access_time;
-       int             dirty:1;
-       int             in_use:1;
+       unsigned        dirty:1;
+       unsigned        in_use:1;
 };
 
 #define CACHE_SIZE 8
@@ -67,10 +86,17 @@ struct unix_private_data {
        int     magic;
        int     dev;
        int     flags;
+       int     align;
        int     access_time;
+       ext2_loff_t offset;
        struct unix_cache cache[CACHE_SIZE];
+       void    *bounce;
+       struct struct_io_stats io_stats;
 };
 
+#define IS_ALIGNED(n, align) ((((unsigned long) n) & \
+                              ((unsigned long) ((align)-1))) == 0)
+
 static errcode_t unix_open(const char *name, int flags, io_channel *channel);
 static errcode_t unix_close(io_channel channel);
 static errcode_t unix_set_blksize(io_channel channel, int blksize);
@@ -81,9 +107,18 @@ static errcode_t unix_write_blk(io_channel channel, unsigned long block,
 static errcode_t unix_flush(io_channel channel);
 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
                                int size, const void *data);
-
+static errcode_t unix_set_option(io_channel channel, const char *option,
+                                const char *arg);
+static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
+;
 static void reuse_cache(io_channel channel, struct unix_private_data *data,
-                struct unix_cache *cache, unsigned long block);
+                struct unix_cache *cache, unsigned long long block);
+static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
+                              int count, void *data);
+static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
+                               int count, const void *data);
+static errcode_t unix_discard(io_channel channel, unsigned long long block,
+                             unsigned long long count);
 
 static struct struct_io_manager struct_unix_manager = {
        EXT2_ET_MAGIC_IO_MANAGER,
@@ -94,112 +129,101 @@ static struct struct_io_manager struct_unix_manager = {
        unix_read_blk,
        unix_write_blk,
        unix_flush,
-#ifdef __CYGWIN__
-       0
-#else
-       unix_write_byte
-#endif
+       unix_write_byte,
+       unix_set_option,
+       unix_get_stats,
+       unix_read_blk64,
+       unix_write_blk64,
+       unix_discard,
 };
 
 io_manager unix_io_manager = &struct_unix_manager;
 
-/*
- * Here are the raw I/O functions
- */
-#ifndef __CYGWIN__
-static errcode_t raw_read_blk(io_channel channel,
-                             struct unix_private_data *data,
-                             unsigned long block,
-                             int count, void *buf)
+static errcode_t unix_get_stats(io_channel channel, io_stats *stats)
 {
-       errcode_t       retval;
-       size_t          size;
-       ext2_loff_t     location;
-       int             actual = 0;
+       errcode_t       retval = 0;
+
+       struct unix_private_data *data;
+
+       EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
+       data = (struct unix_private_data *) channel->private_data;
+       EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
+
+       if (stats)
+               *stats = &data->io_stats;
 
-       size = (count < 0) ? -count : count * channel->block_size;
-       location = (ext2_loff_t) block * channel->block_size;
-       if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
-               retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
-               goto error_out;
-       }
-       actual = read(data->dev, buf, size);
-       if (actual != size) {
-               if (actual < 0)
-                       actual = 0;
-               retval = EXT2_ET_SHORT_READ;
-               goto error_out;
-       }
-       return 0;
-       
-error_out:
-       memset((char *) buf+actual, 0, size-actual);
-       if (channel->read_error)
-               retval = (channel->read_error)(channel, block, count, buf,
-                                              size, actual, retval);
        return retval;
 }
-#else /* __CYGWIN__ */
+
 /*
- * Windows block devices only allow sector alignment IO in offset and size
+ * Here are the raw I/O functions
  */
 static errcode_t raw_read_blk(io_channel channel,
                              struct unix_private_data *data,
-                             unsigned long block,
+                             unsigned long long block,
                              int count, void *buf)
 {
        errcode_t       retval;
-       size_t          size, alignsize, fragment;
+       ssize_t         size;
        ext2_loff_t     location;
-       int             total = 0, actual;
-#define BLOCKALIGN 512
-       char            sector[BLOCKALIGN];
+       int             actual = 0;
 
        size = (count < 0) ? -count : count * channel->block_size;
-       location = (ext2_loff_t) block * channel->block_size;
-#ifdef DEBUG
-       printf("count=%d, size=%d, block=%d, blk_size=%d, location=%lx\n",
-                       count, size, block, channel->block_size, location);
-#endif
+       data->io_stats.bytes_read += size;
+       location = ((ext2_loff_t) block * channel->block_size) + data->offset;
        if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
                retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
                goto error_out;
        }
-       fragment = size % BLOCKALIGN;
-       alignsize = size - fragment;
-       if (alignsize) {
-               actual = read(data->dev, buf, alignsize);
-               if (actual != alignsize)
-                       goto short_read;
+       if ((data->align == 0) ||
+           ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
+               actual = read(data->dev, buf, size);
+               if (actual != size) {
+               short_read:
+                       if (actual < 0)
+                               actual = 0;
+                       retval = EXT2_ET_SHORT_READ;
+                       goto error_out;
+               }
+               return 0;
        }
-       if (fragment) {
-               actual = read(data->dev, sector, BLOCKALIGN);
-               if (actual != BLOCKALIGN)
+
+#ifdef ALIGN_DEBUG
+       printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf,
+              (unsigned long) size);
+#endif
+
+       /*
+        * The buffer or size which we're trying to read isn't aligned
+        * to the O_DIRECT rules, so we need to do this the hard way...
+        */
+       while (size > 0) {
+               actual = read(data->dev, data->bounce, channel->block_size);
+               if (actual != channel->block_size)
                        goto short_read;
-               memcpy(buf+alignsize, sector, fragment);
+               actual = size;
+               if (size > channel->block_size)
+                       actual = channel->block_size;
+               memcpy(buf, data->bounce, actual);
+               size -= actual;
+               buf += actual;
        }
        return 0;
 
-short_read:
-       if (actual>0)
-               total += actual;
-       retval = EXT2_ET_SHORT_READ;
-
 error_out:
-       memset((char *) buf+total, 0, size-actual);
+       memset((char *) buf+actual, 0, size-actual);
        if (channel->read_error)
                retval = (channel->read_error)(channel, block, count, buf,
                                               size, actual, retval);
        return retval;
 }
-#endif
 
 static errcode_t raw_write_blk(io_channel channel,
                               struct unix_private_data *data,
-                              unsigned long block,
+                              unsigned long long block,
                               int count, const void *buf)
 {
-       size_t          size;
+       ssize_t         size;
        ext2_loff_t     location;
        int             actual = 0;
        errcode_t       retval;
@@ -212,20 +236,54 @@ static errcode_t raw_write_blk(io_channel channel,
                else
                        size = count * channel->block_size;
        }
+       data->io_stats.bytes_written += size;
 
-       location = (ext2_loff_t) block * channel->block_size;
+       location = ((ext2_loff_t) block * channel->block_size) + data->offset;
        if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) {
                retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
                goto error_out;
        }
-       
-       actual = write(data->dev, buf, size);
-       if (actual != size) {
-               retval = EXT2_ET_SHORT_WRITE;
-               goto error_out;
+
+       if ((data->align == 0) ||
+           ((IS_ALIGNED(buf, data->align)) && IS_ALIGNED(size, data->align))) {
+               actual = write(data->dev, buf, size);
+               if (actual != size) {
+               short_write:
+                       retval = EXT2_ET_SHORT_WRITE;
+                       goto error_out;
+               }
+               return 0;
+       }
+
+#ifdef ALIGN_DEBUG
+       printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf,
+              (unsigned long) size);
+#endif
+       /*
+        * The buffer or size which we're trying to write isn't aligned
+        * to the O_DIRECT rules, so we need to do this the hard way...
+        */
+       while (size > 0) {
+               if (size < channel->block_size) {
+                       actual = read(data->dev, data->bounce,
+                                     channel->block_size);
+                       if (actual != channel->block_size) {
+                               retval = EXT2_ET_SHORT_READ;
+                               goto error_out;
+                       }
+               }
+               actual = size;
+               if (size > channel->block_size)
+                       actual = channel->block_size;
+               memcpy(data->bounce, buf, actual);
+               actual = write(data->dev, data->bounce, channel->block_size);
+               if (actual != channel->block_size)
+                       goto short_write;
+               size -= actual;
+               buf += actual;
        }
        return 0;
-       
+
 error_out:
        if (channel->write_error)
                retval = (channel->write_error)(channel, block, count, buf,
@@ -245,27 +303,35 @@ static errcode_t alloc_cache(io_channel channel,
        errcode_t               retval;
        struct unix_cache       *cache;
        int                     i;
-       
+
        data->access_time = 0;
        for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
                cache->block = 0;
                cache->access_time = 0;
                cache->dirty = 0;
                cache->in_use = 0;
-               if ((retval = ext2fs_get_mem(channel->block_size,
-                                            (void **) &cache->buf)))
+               if (cache->buf)
+                       ext2fs_free_mem(&cache->buf);
+               retval = ext2fs_get_memalign(channel->block_size,
+                                            data->align, &cache->buf);
+               if (retval)
                        return retval;
        }
-       return 0;
+       if (data->align) {
+               if (data->bounce)
+                       ext2fs_free_mem(&data->bounce);
+               retval = ext2fs_get_memalign(channel->block_size, data->align,
+                                            &data->bounce);
+       }
+       return retval;
 }
 
 /* Free the cache buffers */
-static void free_cache(io_channel channel, 
-                      struct unix_private_data *data)
+static void free_cache(struct unix_private_data *data)
 {
        struct unix_cache       *cache;
        int                     i;
-       
+
        data->access_time = 0;
        for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
                cache->block = 0;
@@ -273,24 +339,25 @@ static void free_cache(io_channel channel,
                cache->dirty = 0;
                cache->in_use = 0;
                if (cache->buf)
-                       ext2fs_free_mem((void **) &cache->buf);
-               cache->buf = 0;
+                       ext2fs_free_mem(&cache->buf);
        }
+       if (data->bounce)
+               ext2fs_free_mem(&data->bounce);
 }
 
+#ifndef NO_IO_CACHE
 /*
  * Try to find a block in the cache.  If the block is not found, and
  * eldest is a non-zero pointer, then fill in eldest with the cache
  * entry to that should be reused.
  */
-static struct unix_cache *find_cached_block(io_channel channel,
-                                           struct unix_private_data *data,
-                                           unsigned long block,
+static struct unix_cache *find_cached_block(struct unix_private_data *data,
+                                           unsigned long long block,
                                            struct unix_cache **eldest)
 {
        struct unix_cache       *cache, *unused_cache, *oldest_cache;
        int                     i;
-       
+
        unused_cache = oldest_cache = 0;
        for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
                if (!cache->in_use) {
@@ -315,7 +382,7 @@ static struct unix_cache *find_cached_block(io_channel channel,
  * Reuse a particular cache entry for another block.
  */
 static void reuse_cache(io_channel channel, struct unix_private_data *data,
-                struct unix_cache *cache, unsigned long block)
+                struct unix_cache *cache, unsigned long long block)
 {
        if (cache->dirty && cache->in_use)
                raw_write_blk(channel, data, cache->block, 1, cache->buf);
@@ -337,18 +404,18 @@ static errcode_t flush_cached_blocks(io_channel channel,
        struct unix_cache       *cache;
        errcode_t               retval, retval2;
        int                     i;
-       
+
        retval2 = 0;
        for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) {
                if (!cache->in_use)
                        continue;
-               
+
                if (invalidate)
                        cache->in_use = 0;
-               
+
                if (!cache->dirty)
                        continue;
-               
+
                retval = raw_write_blk(channel, data,
                                       cache->block, 1, cache->buf);
                if (retval)
@@ -358,13 +425,20 @@ static errcode_t flush_cached_blocks(io_channel channel,
        }
        return retval2;
 }
+#endif /* NO_IO_CACHE */
+
+#ifdef __linux__
+#ifndef BLKDISCARDZEROES
+#define BLKDISCARDZEROES _IO(0x12,124)
+#endif
+#endif
 
 static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 {
        io_channel      io = NULL;
        struct unix_private_data *data = NULL;
        errcode_t       retval;
-       int             open_flags;
+       int             open_flags, zeroes = 0;
        struct stat     st;
 #ifdef __linux__
        struct          utsname ut;
@@ -372,19 +446,17 @@ static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 
        if (name == 0)
                return EXT2_ET_BAD_DEVICE_NAME;
-       retval = ext2fs_get_mem(sizeof(struct struct_io_channel),
-                               (void **) &io);
+       retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io);
        if (retval)
                return retval;
        memset(io, 0, sizeof(struct struct_io_channel));
        io->magic = EXT2_ET_MAGIC_IO_CHANNEL;
-       retval = ext2fs_get_mem(sizeof(struct unix_private_data),
-                               (void **) &data);
+       retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data);
        if (retval)
                goto cleanup;
 
        io->manager = unix_io_manager;
-       retval = ext2fs_get_mem(strlen(name)+1, (void **) &io->name);
+       retval = ext2fs_get_mem(strlen(name)+1, &io->name);
        if (retval)
                goto cleanup;
 
@@ -397,21 +469,67 @@ static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 
        memset(data, 0, sizeof(struct unix_private_data));
        data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL;
+       data->io_stats.num_fields = 2;
 
-       if ((retval = alloc_cache(io, data)))
-               goto cleanup;
-       
        open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY;
+       if (flags & IO_FLAG_EXCLUSIVE)
+               open_flags |= O_EXCL;
+#ifdef O_DIRECT
+       if (flags & IO_FLAG_DIRECT_IO)
+               open_flags |= O_DIRECT;
+#endif
+       data->flags = flags;
+
 #ifdef HAVE_OPEN64
-       data->dev = open64(name, open_flags);
+       data->dev = open64(io->name, open_flags);
 #else
-       data->dev = open(name, open_flags);
+       data->dev = open(io->name, open_flags);
 #endif
        if (data->dev < 0) {
                retval = errno;
                goto cleanup;
        }
 
+#ifdef BLKSSZGET
+       if (flags & IO_FLAG_DIRECT_IO) {
+               if (ioctl(data->dev, BLKSSZGET, &data->align) != 0)
+                       data->align = io->block_size;
+       }
+#endif
+
+#ifdef BLKDISCARDZEROES
+       ioctl(data->dev, BLKDISCARDZEROES, &zeroes);
+       if (zeroes)
+               io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES;
+#endif
+
+#if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
+       /*
+        * Some operating systems require that the buffers be aligned,
+        * regardless of O_DIRECT
+        */
+       data->align = 512;
+#endif
+
+
+       if ((retval = alloc_cache(io, data)))
+               goto cleanup;
+
+#ifdef BLKROGET
+       if (flags & IO_FLAG_RW) {
+               int error;
+               int readonly = 0;
+
+               /* Is the block device actually writable? */
+               error = ioctl(data->dev, BLKROGET, &readonly);
+               if (!error && readonly) {
+                       close(data->dev);
+                       retval = EPERM;
+                       goto cleanup;
+               }
+       }
+#endif
+
 #ifdef __linux__
 #undef RLIM_INFINITY
 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4)))
@@ -424,7 +542,7 @@ static errcode_t unix_open(const char *name, int flags, io_channel *channel)
         * block devices are wrongly getting hit by the filesize
         * limit.  This workaround isn't perfect, since it won't work
         * if glibc wasn't built against 2.2 header files.  (Sigh.)
-        * 
+        *
         */
        if ((flags & IO_FLAG_RW) &&
            (uname(&ut) == 0) &&
@@ -435,7 +553,7 @@ static errcode_t unix_open(const char *name, int flags, io_channel *channel)
            (fstat(data->dev, &st) == 0) &&
            (S_ISBLK(st.st_mode))) {
                struct rlimit   rlim;
-               
+
                rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY;
                setrlimit(RLIMIT_FSIZE, &rlim);
                getrlimit(RLIMIT_FSIZE, &rlim);
@@ -451,11 +569,11 @@ static errcode_t unix_open(const char *name, int flags, io_channel *channel)
 
 cleanup:
        if (data) {
-               free_cache(io, data);
-               ext2fs_free_mem((void **) &data);
+               free_cache(data);
+               ext2fs_free_mem(&data);
        }
        if (io)
-               ext2fs_free_mem((void **) &io);
+               ext2fs_free_mem(&io);
        return retval;
 }
 
@@ -471,16 +589,18 @@ static errcode_t unix_close(io_channel channel)
        if (--channel->refcount > 0)
                return 0;
 
+#ifndef NO_IO_CACHE
        retval = flush_cached_blocks(channel, data, 0);
+#endif
 
        if (close(data->dev) < 0)
                retval = errno;
-       free_cache(channel, data);
+       free_cache(data);
 
-       ext2fs_free_mem((void **) &channel->private_data);
+       ext2fs_free_mem(&channel->private_data);
        if (channel->name)
-               ext2fs_free_mem((void **) &channel->name);
-       ext2fs_free_mem((void **) &channel);
+               ext2fs_free_mem(&channel->name);
+       ext2fs_free_mem(&channel);
        return retval;
 }
 
@@ -494,11 +614,13 @@ static errcode_t unix_set_blksize(io_channel channel, int blksize)
        EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
        if (channel->block_size != blksize) {
+#ifndef NO_IO_CACHE
                if ((retval = flush_cached_blocks(channel, data, 0)))
                        return retval;
-               
+#endif
+
                channel->block_size = blksize;
-               free_cache(channel, data);
+               free_cache(data);
                if ((retval = alloc_cache(channel, data)))
                        return retval;
        }
@@ -506,7 +628,7 @@ static errcode_t unix_set_blksize(io_channel channel, int blksize)
 }
 
 
-static errcode_t unix_read_blk(io_channel channel, unsigned long block,
+static errcode_t unix_read_blk64(io_channel channel, unsigned long long block,
                               int count, void *buf)
 {
        struct unix_private_data *data;
@@ -519,6 +641,9 @@ static errcode_t unix_read_blk(io_channel channel, unsigned long block,
        data = (struct unix_private_data *) channel->private_data;
        EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
+#ifdef NO_IO_CACHE
+       return raw_read_blk(channel, data, block, count, buf);
+#else
        /*
         * If we're doing an odd-sized read or a very large read,
         * flush out the cache and then do a direct read.
@@ -532,10 +657,9 @@ static errcode_t unix_read_blk(io_channel channel, unsigned long block,
        cp = buf;
        while (count > 0) {
                /* If it's in the cache, use it! */
-               if ((cache = find_cached_block(channel, data, block,
-                                              &reuse[0]))) {
+               if ((cache = find_cached_block(data, block, &reuse[0]))) {
 #ifdef DEBUG
-                       printf("Using cached block %d\n", block);
+                       printf("Using cached block %lu\n", block);
 #endif
                        memcpy(cp, cache->buf, channel->block_size);
                        count--;
@@ -543,20 +667,35 @@ static errcode_t unix_read_blk(io_channel channel, unsigned long block,
                        cp += channel->block_size;
                        continue;
                }
+               if (count == 1) {
+                       /*
+                        * Special case where we read directly into the
+                        * cache buffer; important in the O_DIRECT case
+                        */
+                       cache = reuse[0];
+                       reuse_cache(channel, data, cache, block);
+                       if ((retval = raw_read_blk(channel, data, block, 1,
+                                                  cache->buf))) {
+                               cache->in_use = 0;
+                               return retval;
+                       }
+                       memcpy(cp, cache->buf, channel->block_size);
+                       return 0;
+               }
+
                /*
                 * Find the number of uncached blocks so we can do a
                 * single read request
                 */
                for (i=1; i < count; i++)
-                       if (find_cached_block(channel, data, block+i,
-                                             &reuse[i]))
+                       if (find_cached_block(data, block+i, &reuse[i]))
                                break;
 #ifdef DEBUG
-               printf("Reading %d blocks starting at %d\n", i, block);
+               printf("Reading %d blocks starting at %lu\n", i, block);
 #endif
                if ((retval = raw_read_blk(channel, data, block, i, cp)))
                        return retval;
-               
+
                /* Save the results in the cache */
                for (j=0; j < i; j++) {
                        count--;
@@ -567,9 +706,16 @@ static errcode_t unix_read_blk(io_channel channel, unsigned long block,
                }
        }
        return 0;
+#endif /* NO_IO_CACHE */
 }
 
-static errcode_t unix_write_blk(io_channel channel, unsigned long block,
+static errcode_t unix_read_blk(io_channel channel, unsigned long block,
+                              int count, void *buf)
+{
+       return unix_read_blk64(channel, block, count, buf);
+}
+
+static errcode_t unix_write_blk64(io_channel channel, unsigned long long block,
                                int count, const void *buf)
 {
        struct unix_private_data *data;
@@ -582,6 +728,9 @@ static errcode_t unix_write_blk(io_channel channel, unsigned long block,
        data = (struct unix_private_data *) channel->private_data;
        EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
+#ifdef NO_IO_CACHE
+       return raw_write_blk(channel, data, block, count, buf);
+#else
        /*
         * If we're doing an odd-sized write or a very large write,
         * flush out the cache completely and then do a direct write.
@@ -600,10 +749,10 @@ static errcode_t unix_write_blk(io_channel channel, unsigned long block,
        writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH;
        if (writethrough)
                retval = raw_write_blk(channel, data, block, count, buf);
-       
+
        cp = buf;
        while (count > 0) {
-               cache = find_cached_block(channel, data, block, &reuse);
+               cache = find_cached_block(data, block, &reuse);
                if (!cache) {
                        cache = reuse;
                        reuse_cache(channel, data, cache, block);
@@ -615,6 +764,13 @@ static errcode_t unix_write_blk(io_channel channel, unsigned long block,
                cp += channel->block_size;
        }
        return retval;
+#endif /* NO_IO_CACHE */
+}
+
+static errcode_t unix_write_blk(io_channel channel, unsigned long block,
+                               int count, const void *buf)
+{
+       return unix_write_blk64(channel, block, count, buf);
 }
 
 static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
@@ -622,21 +778,30 @@ static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 {
        struct unix_private_data *data;
        errcode_t       retval = 0;
-       size_t          actual;
+       ssize_t         actual;
 
        EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
        data = (struct unix_private_data *) channel->private_data;
        EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
+       if (data->align != 0) {
+#ifdef ALIGN_DEBUG
+               printf("unix_write_byte: O_DIRECT fallback\n");
+#endif
+               return EXT2_ET_UNIMPLEMENTED;
+       }
+
+#ifndef NO_IO_CACHE
        /*
         * Flush out the cache completely
         */
        if ((retval = flush_cached_blocks(channel, data, 1)))
                return retval;
+#endif
 
-       if (lseek(data->dev, offset, SEEK_SET) < 0)
+       if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0)
                return errno;
-       
+
        actual = write(data->dev, buf, size);
        if (actual != size)
                return EXT2_ET_SHORT_WRITE;
@@ -645,19 +810,74 @@ static errcode_t unix_write_byte(io_channel channel, unsigned long offset,
 }
 
 /*
- * Flush data buffers to disk.  
+ * Flush data buffers to disk.
  */
 static errcode_t unix_flush(io_channel channel)
 {
        struct unix_private_data *data;
        errcode_t retval = 0;
-       
+
        EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
        data = (struct unix_private_data *) channel->private_data;
        EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
 
+#ifndef NO_IO_CACHE
        retval = flush_cached_blocks(channel, data, 0);
+#endif
        fsync(data->dev);
        return retval;
 }
 
+static errcode_t unix_set_option(io_channel channel, const char *option,
+                                const char *arg)
+{
+       struct unix_private_data *data;
+       unsigned long long tmp;
+       char *end;
+
+       EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
+       data = (struct unix_private_data *) channel->private_data;
+       EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
+
+       if (!strcmp(option, "offset")) {
+               if (!arg)
+                       return EXT2_ET_INVALID_ARGUMENT;
+
+               tmp = strtoull(arg, &end, 0);
+               if (*end)
+                       return EXT2_ET_INVALID_ARGUMENT;
+               data->offset = tmp;
+               if (data->offset < 0)
+                       return EXT2_ET_INVALID_ARGUMENT;
+               return 0;
+       }
+       return EXT2_ET_INVALID_ARGUMENT;
+}
+
+#if defined(__linux__) && !defined(BLKDISCARD)
+#define BLKDISCARD     _IO(0x12,119)
+#endif
+
+static errcode_t unix_discard(io_channel channel, unsigned long long block,
+                             unsigned long long count)
+{
+#ifdef BLKDISCARD
+       struct unix_private_data *data;
+       __uint64_t      range[2];
+       int             ret;
+
+       EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL);
+       data = (struct unix_private_data *) channel->private_data;
+       EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL);
+
+       range[0] = (__uint64_t)(block) * channel->block_size;
+       range[1] = (__uint64_t)(count) * channel->block_size;
+
+       ret = ioctl(data->dev, BLKDISCARD, &range);
+       if (ret < 0)
+               return errno;
+       return 0;
+#else
+       return EXT2_ET_UNIMPLEMENTED;
+#endif
+}