Btrfs: fix race between mmap writes and compression
[linux-2.6.git] / fs / btrfs / inode.c
index 0df0d1f..9e51325 100644 (file)
@@ -349,6 +349,7 @@ static noinline int compress_file_range(struct inode *inode,
        int i;
        int will_compress;
        int compress_type = root->fs_info->compress_type;
+       int redirty = 0;
 
        /* if this is a small write inside eof, kick off a defrag */
        if ((end - start + 1) < 16 * 1024 &&
@@ -411,6 +412,17 @@ again:
                if (BTRFS_I(inode)->force_compress)
                        compress_type = BTRFS_I(inode)->force_compress;
 
+               /*
+                * we need to call clear_page_dirty_for_io on each
+                * page in the range.  Otherwise applications with the file
+                * mmap'd can wander in and change the page contents while
+                * we are compressing them.
+                *
+                * If the compression fails for any reason, we set the pages
+                * dirty again later on.
+                */
+               extent_range_clear_dirty_for_io(inode, start, end);
+               redirty = 1;
                ret = btrfs_compress_pages(compress_type,
                                           inode->i_mapping, start,
                                           total_compressed, pages,
@@ -552,6 +564,8 @@ cleanup_and_bail_uncompressed:
                        __set_page_dirty_nobuffers(locked_page);
                        /* unlocked later on in the async handlers */
                }
+               if (redirty)
+                       extent_range_redirty_for_io(inode, start, end);
                add_async_extent(async_cow, start, end - start + 1,
                                 0, NULL, 0, BTRFS_COMPRESS_NONE);
                *num_added += 1;