Re: [PATCH 1/3] libfs: reject casefolded names that fold to ".", "..", or empty

From: Sarthak Kukreti

Date: Mon Nov 16 2026 - 21:45:11 EST


On Mon, Nov 16, 2026 at 02:34:26PM -0500, Gabriel Krisman Bertazi wrote:
> Ugh. I agree this is a problem, but I'm also trying to understand first
> the real impact beyond filesystem complaints during lookup, if any:
>
> 1. Looking up ".\u202e" folds to "." and matches the on-disk "." entry
> (de->inode == dir->i_ino) in generic_ci_match(), triggering
> EXT4_ERROR_INODE ("linked to parent dir") in ext4_lookup().
> 2. Looking up "..\u202e" folds to ".." and matches the on-disk ".."
> entry, returning -ELOOP via d_splice_alias() (or triggering
> EXT4_ERROR_INODE on the root directory).
>
> These two look like a "do stupid things, get stupid prizes" with no
> further repercussions beyond the error message. Sure, we can fix them
> to throw EINVAL, but are there bigger repercussions I'm missing?

In addition to what Eric mentioned re: side-effects due to
errors=remount-ro and errors=panic, encountering EXT4_ERROR_INODE marks
the superblock with an error flag which can trigger a full fsck pass on
the next boot.

> 3. Filenames consisting solely of ignorable characters fold to "" (0
> bytes), causing distinct ignorable-only filenames to hash and compare
> equal and creating entries that e2fsck flags as corruption.
>
> This seems concerning. Do we actually writes multiple files to the disk?
> shouldn't they all be folded into the same empty string and rejected or
> at least written as a single file? I'm not sure how it'd be writing
> multiple files to the disk, that seems like a bigger problem.

Sorry, that should have been specifically for f2fs: ext4 will only write
the first ignorable-only file to disk and subsequent files fail with
-EEXIST. On f2fs, f2fs_init_casefolded_name() treats `len <= 0` as a
fallback to raw bytewise hashing/matching, which allows multiple
ignorable-only files to be written to disk (which fsck.f2fs then flags
with hash mismatches).

> > Fixes: c843843e714c ("fs: Add standard casefolding support")
>
> This bug was just made generic by c843843e714c, but the issue was
> introduced by b886ee3e778e in ext4 and the equivalent f2fs support.

Ack, I will update the Fixes tags in v2.

> > + if (name_is_dot_dotdot(str, len) ||
> > + name_is_dot_dotdot(name->name, name->len))
> > + return 1;
>
> This is checking for strictly encoded '.' and '..'. But I'm not sure
> that happens? Do we really reach the dcache for '.' and '..' instead of
> the name of the parent or cwd?
>
> > + um = dentry->d_sb->s_encoding;
> > + cf_len = utf8_casefold(um, name, buf, sizeof(buf));
> > + if (cf_len == 0 || (cf_len > 0 && name_is_dot_dotdot(buf, cf_len)))
> > + return 1;
>
> I think this is a bit too late in the lookup to detect these
> code-points. I'd much rather we fail the lookup outright instead of
> attempting to treat them as different spellings of . and .. . But isn't
> this dead code because we are rejecting those strings in d_hash through
> your hunk below? Is there a way we get here without the -EINVAL in
> d_hash first?

Ah right - let me get rid of this for v2.

> But the way I'd suggest be handled is by rejecting the ignorable code
> points lookup (all, but variant selectors for the reasons above), early
> in the lookup path for case-insensitive directories so they never even
> get to the filesystem. Wdyt?

As Eric pointed out, since variation selectors are also
Default_Ignorable_Code_Point and stripped to 0 bytes by utf8_casefold(),
allowing variation selectors means names like ".\ufe00", "..\ufe00", and
"\ufe00" still fold to ".", "..", and "" and hit the exact same bug.

Thanks,
Sarthak