Btrfs: take i_mutex before generic_write_checks
authorChris Mason <chris.mason@oracle.com>
Thu, 1 Oct 2009 16:29:10 +0000 (12:29 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 1 Oct 2009 16:29:10 +0000 (12:29 -0400)
btrfs_file_write was incorrectly calling generic_write_checks without
taking i_mutex.  This lead to problems with racing around i_size when
doing O_APPEND writes.

The fix here is to move i_mutex higher.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/file.c

index 1be96ba..f155179 100644 (file)
@@ -920,26 +920,35 @@ static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
        start_pos = pos;
 
        vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
+
+       /* do the reserve before the mutex lock in case we have to do some
+        * flushing.  We wouldn't deadlock, but this is more polite.
+        */
+       err = btrfs_reserve_metadata_for_delalloc(root, inode, 1);
+       if (err)
+               goto out_nolock;
+
+       mutex_lock(&inode->i_mutex);
+
        current->backing_dev_info = inode->i_mapping->backing_dev_info;
        err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
        if (err)
-               goto out_nolock;
+               goto out;
+
        if (count == 0)
-               goto out_nolock;
+               goto out;
 
        err = file_remove_suid(file);
        if (err)
-               goto out_nolock;
-
-       err = btrfs_reserve_metadata_for_delalloc(root, inode, 1);
-       if (err)
-               goto out_nolock;
+               goto out;
 
        file_update_time(file);
 
        pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
 
-       mutex_lock(&inode->i_mutex);
+       /* generic_write_checks can change our pos */
+       start_pos = pos;
+
        BTRFS_I(inode)->sequence++;
        first_index = pos >> PAGE_CACHE_SHIFT;
        last_index = (pos + count) >> PAGE_CACHE_SHIFT;