Hi,
19.11.2024 12:05, Miklos Szeredi wrote:
On Mon, 18 Nov 2024 at 19:54, Amir Goldstein <amir73il@xxxxxxxxx> wrote:
Can you analyse what went wrong with the reproducer?
How did we get to a state where lowerstack of parent
has a dentry which is !d_can_lookup?
Theoretically we could still get a an S_ISDIR inode, because
ovl_get_inode() doesn't look at the is_dir value that lookup found.
I.e. lookup thinks it found a non-dir, but iget will create a dir
because of the backing inode's type.
AFAICS this can only happen if i_op->lookup is not set on S_ISDIR for
the backing inode, which shouldn't happen on normal filesystems.
Reproducer seems to use bfs, which *should* be normal, and bfs_iget
certainly doesn't do anything weird in that case, so I still don't
understand what is happening.
During testing, it was discovered that BFS can return a directory inode
without a lookup operation. Adding the following check in bfs_iget:
struct inode *bfs_iget(struct super_block *sb, unsigned long ino)
{
...
brelse(bh);
+ if (S_ISDIR(inode->i_mode) && !inode->i_op->lookup) {
+ printf("Directory inode missing lookup %s:%08lx\n",
inode->i_sb->s_id, ino);
+ goto error;
+ }
+
unlock_new_inode(inode);
return inode;
error:
iget_failed(inode);
return ERR_PTR(-EIO);
}
prevents the error but exposes an invalid inode:
loop0: detected capacity change from 0 to 64
BFS-fs: bfs_iget(): Directory inode missing lookup loop0:00000002
overlayfs: overlapping lowerdir path
Would this be considered a valid workaround, or does BFS require further
fixes?
In any case something like the following should filter out such weirdness:
bool ovl_dentry_weird(struct dentry *dentry)
{
+ if (!d_can_lookup(dentry) && !d_is_file(dentry) &&
!d_is_symlink(dentry))
+ return true;
+
return dentry->d_flags & (DCACHE_NEED_AUTOMOUNT |
I tested this addition successfully.
Additionally, fixes for BFS seem to be discussed reluctantly.
For instance, this patch set [1] has remained unanswered.
Perhaps it would be worth considering discarding invalid inodes at the
overlayfs level?
[1]
https://lore.kernel.org/all/20240822161219.459054-1-kovalev@xxxxxxxxxxxx/
Thanks,
Miklos
--
Thanks,
Vasiliy Kovalev
