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

From: David Laight

Date: Tue Nov 17 2026 - 14:41:01 EST


On Tue, 17 Nov 2026 17:59:17 +0000
Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:

> On Tue, Nov 17, 2026 at 09:54:32AM -0500, Theodore Tso wrote:
> > On Tue, Nov 17, 2026 at 08:47:27AM -0500, David Laight wrote:
> > > > Sigh... I really think these characters are *evil*. I wonder if we
> > > > should simply drop those code points unconditionally when they are fed
> > > > to the kernel --- not only folding to themselves, but *not* preserving
> > > > them when a misguided application tries to create a file with one of
> > > > these misbegotten Unicode charactesr.
> > >
> > > Doesn't that assume that filenames are UTF8 rather than just sequences
> > > of bytes with everything except '\0' and '/' being valid.
> > > Casefolding is itself so evil/impossible it shouldn't have been allowed.
> >
> > We would only do this if casefolding is enabled.
>
> Which is to say, on any distro kernel. Besides, it's not just a matter of
> filtering that dreck out on getname() and friends (syscall boundary won't
> do at all due to the things like binfmt_misc opening files, etc.) - you'd
> need to scan what you get upon ->get_link() as well.

I also hope it is a mount option?
So you could cross mountpoints several times (in both directions) to and from
fs with utf8 casefold un/supported while processing a single pathname.

> > For better or for
> > worse, it's better than what it replaced which was this terrible
> > wrapfs hack that was a trivial to trigger deadlocks using fsstress ---
> > which was even more evil.
> >
> > The other way we can deal with this is that if case folding is
> > enabled, we change the VFS to call the case-folding for incoming
> > filename *before* it does its special case checking for '.' and '..'
> > if casefolding is enabled.
>
> That would penalize every pathname resolution, whether that misfeature is
> stepped into or not; hash_name() is fairly hot.
>
> '.' and '..' should be recognized only as literal octet sequences 0x2e
> and 0x2e 0x2e; if some misguided filesystem decides to treat something
> else as equivalent to those, it bloody well ought to recognize and reject
> that something else in its ->lookup() (or ->d_hash(), for that matter).

I remember when "." and ".." weren't actually special.
You could have great 'fun' using the 'link' system call to make them
refer to unexpected places.

David