ext4: fix race in xattr block allocation path

Ceph users reported that when using Ceph on ext4, the filesystem would often become corrupted, containing inodes with incorrect i_blocks counters.

I managed to reproduce this with a very hacked-up "streamtest" binary from the Ceph tree.

Ceph is doing a lot of xattr writes, to out-of-inode blocks. There is also another thread which does sync_file_range and close, of the same files. The problem appears to happen due to this race:

sync/flush thread xattr-set thread
----------------- ----------------

do_writepages                   ext4_xattr_set
ext4_da_writepages              ext4_xattr_set_handle
mpage_da_map_blocks             ext4_xattr_block_set

set DELALLOC_RESERVE
ext4_new_meta_blocks
ext4_mb_new_blocks
if (!i_delalloc_reserved_flag)
vfs_dq_alloc_block
ext4_get_blocks

	down_write(i_data_sem)
        set i_delalloc_reserved_flag
	...
	up_write(i_data_sem)
                                        if (i_delalloc_reserved_flag)
                                                vfs_dq_alloc_block_nofail

In other words, the sync/flush thread pops in and sets i_delalloc_reserved_flag on the inode, which makes the xattr thread think that it's in a delalloc path in ext4_new_meta_blocks(), and add the block for a second time, after already having added it once in the !i_delalloc_reserved_flag case in ext4_mb_new_blocks

I think this can be as simple as taking i_data_sem before we go down the new metablock path, so that we won't be testing i_delalloc_reserved_flag in a race with xt4_get_blocks setting it.

---

This seems like slight abuse of i_data_sem though, since we are not modifying i_data[] in the inode on the xattr path. But nothing else protects readers of i_delalloc_reserved_flag in other threads. Thoughts?

p.s. this almost feels like it needs quite a rework, I'm trying to remember why we set a "we are in a delalloc path" in an inode flag rather than as something passed through the callchain arguments, but I'm fuzzy on these twisty paths.

index e4d0fca..dcebdbd 100644
@@ -819,8 +819,14 @@ inserted:
 			if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 				goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
+			/*
+			 * take i_data_sem because we will test
+			 * i_delalloc_reserved_flag in ext4_mb_new_blocks */
+			 */
+			down_read((&EXT4_I(inode)->i_data_sem));
 			block = ext4_new_meta_blocks(handle, inode,
 						  goal, NULL, &error);
+			up_read((&EXT4_I(inode)->i_data_sem));
 			if (error)
 				goto cleanup;
 
 

Thanks, applied for now.

Agreed, we really should be depending on the EXT4_GET_BLOCK_DELALLOC_RESERVE and EXT4_MB_DELALLOC_RESERVED flags. I can't remember what caused us to plumb in the i_state flag, but that's clearly a hack which is biting us hard.

I'll take this patch for now since it fixes a real problem, but we should figure out a better way of fixing this, hopefully that results in the i_state flag going away.

- Ted