ext4: fix ZERO_RANGE bug hidden by flag aliasing
We accidently aliased EXT4_EX_NOCACHE and EXT4_GET_CONVERT_UNWRITTEN
falgs, which apparently was hiding a bug that was unmasked when this
flag aliasing issue was addressed (see the subsequent commit). The
reproduction case was:
fsx -N 10000 -l 500000 -r 4096 -t 4096 -w 4096 -Z -R -W /vdb/junk
... which would cause fsx to report corruption in the data file.
The fix we have is a bit of an overkill, but I'd much rather be
conservative for now, and we can optimize ZERO_RANGE_FL handling
later. The fact that we need to zap the extent_status cache for the
inode is unfortunate, but correctness is far more important than
performance.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: Namjae Jeon <namjae.jeon@samsung.com>
diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
index d009373..bf205f7 100644
--- a/fs/ext4/extents.c
+++ b/fs/ext4/extents.c
@@ -4802,7 +4802,8 @@
max_blocks -= lblk;
flags = EXT4_GET_BLOCKS_CREATE_UNWRIT_EXT |
- EXT4_GET_BLOCKS_CONVERT_UNWRITTEN;
+ EXT4_GET_BLOCKS_CONVERT_UNWRITTEN |
+ EXT4_EX_NOCACHE;
if (mode & FALLOC_FL_KEEP_SIZE)
flags |= EXT4_GET_BLOCKS_KEEP_SIZE;
@@ -4840,17 +4841,23 @@
ext4_inode_block_unlocked_dio(inode);
inode_dio_wait(inode);
- /*
- * Remove entire range from the extent status tree.
- */
- ret = ext4_es_remove_extent(inode, lblk, max_blocks);
- if (ret)
- goto out_dio;
-
ret = ext4_alloc_file_blocks(file, lblk, max_blocks, new_size,
flags, mode);
if (ret)
goto out_dio;
+ /*
+ * Remove entire range from the extent status tree.
+ *
+ * ext4_es_remove_extent(inode, lblk, max_blocks) is
+ * NOT sufficient. I'm not sure why this is the case,
+ * but let's be conservative and remove the extent
+ * status tree for the entire inode. There should be
+ * no outstanding delalloc extents thanks to the
+ * filemap_write_and_wait_range() call above.
+ */
+ ret = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+ if (ret)
+ goto out_dio;
}
if (!partial_begin && !partial_end)
goto out_dio;