]> bbs.cooldavid.org Git - net-next-2.6.git/commitdiff
ext4: fix EOFBLOCKS_FL handling
authorTheodore Ts'o <tytso@mit.edu>
Thu, 28 Oct 2010 01:23:12 +0000 (21:23 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Thu, 28 Oct 2010 01:23:12 +0000 (21:23 -0400)
It turns out we have several problems with how EOFBLOCKS_FL is
handled.  First of all, there was a fencepost error where we were not
clearing the EOFBLOCKS_FL when fill in the last uninitialized block,
but rather when we allocate the next block _after_ the uninitalized
block.  Secondly we were not testing to see if we needed to clear the
EOFBLOCKS_FL when writing to the file O_DIRECT or when were converting
an uninitialized block (which is the most common case).

Google-Bug-Id: 2928259

Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
fs/ext4/extents.c

index 06328d3e5717fd368ce7872fe0aa3db879cf6817..82027841022009346fbf638d9fe658c79e64adb8 100644 (file)
@@ -3180,6 +3180,57 @@ static void unmap_underlying_metadata_blocks(struct block_device *bdev,
                 unmap_underlying_metadata(bdev, block + i);
 }
 
+/*
+ * Handle EOFBLOCKS_FL flag, clearing it if necessary
+ */
+static int check_eofblocks_fl(handle_t *handle, struct inode *inode,
+                             struct ext4_map_blocks *map,
+                             struct ext4_ext_path *path,
+                             unsigned int len)
+{
+       int i, depth;
+       struct ext4_extent_header *eh;
+       struct ext4_extent *ex, *last_ex;
+
+       if (!ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
+               return 0;
+
+       depth = ext_depth(inode);
+       eh = path[depth].p_hdr;
+       ex = path[depth].p_ext;
+
+       if (unlikely(!eh->eh_entries)) {
+               EXT4_ERROR_INODE(inode, "eh->eh_entries == 0 and "
+                                "EOFBLOCKS_FL set");
+               return -EIO;
+       }
+       last_ex = EXT_LAST_EXTENT(eh);
+       /*
+        * We should clear the EOFBLOCKS_FL flag if we are writing the
+        * last block in the last extent in the file.  We test this by
+        * first checking to see if the caller to
+        * ext4_ext_get_blocks() was interested in the last block (or
+        * a block beyond the last block) in the current extent.  If
+        * this turns out to be false, we can bail out from this
+        * function immediately.
+        */
+       if (map->m_lblk + len < le32_to_cpu(last_ex->ee_block) +
+           ext4_ext_get_actual_len(last_ex))
+               return 0;
+       /*
+        * If the caller does appear to be planning to write at or
+        * beyond the end of the current extent, we then test to see
+        * if the current extent is the last extent in the file, by
+        * checking to make sure it was reached via the rightmost node
+        * at each level of the tree.
+        */
+       for (i = depth-1; i >= 0; i--)
+               if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
+                       return 0;
+       ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
+       return ext4_mark_inode_dirty(handle, inode);
+}
+
 static int
 ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
                        struct ext4_map_blocks *map,
@@ -3217,8 +3268,12 @@ ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
        if ((flags & EXT4_GET_BLOCKS_CONVERT)) {
                ret = ext4_convert_unwritten_extents_endio(handle, inode,
                                                        path);
-               if (ret >= 0)
+               if (ret >= 0) {
                        ext4_update_inode_fsync_trans(handle, inode, 1);
+                       err = check_eofblocks_fl(handle, inode, map, path,
+                                                map->m_len);
+               } else
+                       err = ret;
                goto out2;
        }
        /* buffered IO case */
@@ -3244,8 +3299,13 @@ ext4_ext_handle_uninitialized_extents(handle_t *handle, struct inode *inode,
 
        /* buffered write, writepage time, convert*/
        ret = ext4_ext_convert_to_initialized(handle, inode, map, path);
-       if (ret >= 0)
+       if (ret >= 0) {
                ext4_update_inode_fsync_trans(handle, inode, 1);
+               err = check_eofblocks_fl(handle, inode, map, path, map->m_len);
+               if (err < 0)
+                       goto out2;
+       }
+
 out:
        if (ret <= 0) {
                err = ret;
@@ -3292,6 +3352,7 @@ out2:
        }
        return err ? err : allocated;
 }
+
 /*
  * Block allocation/map/preallocation routine for extents based files
  *
@@ -3315,9 +3376,9 @@ int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
 {
        struct ext4_ext_path *path = NULL;
        struct ext4_extent_header *eh;
-       struct ext4_extent newex, *ex, *last_ex;
+       struct ext4_extent newex, *ex;
        ext4_fsblk_t newblock;
-       int i, err = 0, depth, ret, cache_type;
+       int err = 0, depth, ret, cache_type;
        unsigned int allocated = 0;
        struct ext4_allocation_request ar;
        ext4_io_end_t *io = EXT4_I(inode)->cur_aio_dio;
@@ -3497,31 +3558,10 @@ int ext4_ext_map_blocks(handle_t *handle, struct inode *inode,
                        map->m_flags |= EXT4_MAP_UNINIT;
        }
 
-       if (unlikely(ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))) {
-               if (unlikely(!eh->eh_entries)) {
-                       EXT4_ERROR_INODE(inode,
-                                        "eh->eh_entries == 0 and "
-                                        "EOFBLOCKS_FL set");
-                       err = -EIO;
-                       goto out2;
-               }
-               last_ex = EXT_LAST_EXTENT(eh);
-               /*
-                * If the current leaf block was reached by looking at
-                * the last index block all the way down the tree, and
-                * we are extending the inode beyond the last extent
-                * in the current leaf block, then clear the
-                * EOFBLOCKS_FL flag.
-                */
-               for (i = depth-1; i >= 0; i--) {
-                       if (path[i].p_idx != EXT_LAST_INDEX(path[i].p_hdr))
-                               break;
-               }
-               if ((i < 0) &&
-                   (map->m_lblk + ar.len > le32_to_cpu(last_ex->ee_block) +
-                    ext4_ext_get_actual_len(last_ex)))
-                       ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
-       }
+       err = check_eofblocks_fl(handle, inode, map, path, ar.len);
+       if (err)
+               goto out2;
+
        err = ext4_ext_insert_extent(handle, inode, path, &newex, flags);
        if (err) {
                /* free data blocks we just allocated */