Re: [PATCH 4/8] vfs: Fold casefolding into vfs

[Eric Biggers <ebiggers@kernel.org>]

On Tue, Dec 03, 2019 at 04:21:02PM -0500, Gabriel Krisman Bertazi wrote:
> Eric Biggers <ebiggers@kernel.org> writes:
> 
> > On Tue, Dec 03, 2019 at 02:42:10PM -0500, Gabriel Krisman Bertazi wrote:
> >> Gao Xiang <gaoxiang25@huawei.com> writes:
> 
> >> I think Daniel's approach of moving this into VFS is the simplest way to
> >> actually solve the issue, instead of extending and duplicating a lot of
> >> functionality into filesystem hooks to support the possible mixes of
> >> case-insensitive, overlayfs and fscrypt.
> >> 
> >
> > I think we can actually get everything we want using dentry_operations only,
> > since the filesystem can set ->d_op during ->lookup() (like what is done for
> > encrypted filenames now) rather than at dentry allocation time.  And fs/crypto/
> > can export fscrypt_d_revalidate() rather than setting ->d_op itself.
> 
> Problem is, differently from fscrypt, case-insensitive uses the d_hash()
> hook and for a lookup, we actually use
> dentry->d_parent->d_ops->d_hash().  Which works well, until you are flipping the
> casefold flag.  Then the dentry already exists and you need to modify
> the d_ops on the fly, which I couldn't find precedent anywhere.  I tried
> invalidating the dentry whenever we flip the flag, but then if it has
> negative dentries as children,I wasn't able to reliably invalidate it,
> and that's when I reached the limit of my knowledge in VFS.  In
> particular, in every attempt I made to implement it like this, I was
> able to race and do a case-insensitive lookup on a directory that was
> just made case sensitive.
> 
> I'm not saying there isn't a way.  But it is a bit harder than this
> proposal. I tried it already and still didn't manage to make it work.
> Maybe someone who better understands vfs.

Yes you're right, I forgot that for ->d_hash() and ->d_compare() it's actually
the parent's directory dentry_operations that are used.

> 
> > It's definitely ugly to have to handle the 3 cases of encrypt, casefold, and
> > encrypt+casefold separately -- and this will need to be duplicated for each
> > filesystem.  But we do have to weigh that against adding additional complexity
> > and overhead to the VFS for everyone.  If we do go with the VFS changes, please
> > try to make them as simple and unobtrusive as possible.
> 
> Well, it is just not case-insensitive+fscrypt. Also overlayfs
> there. Probably more.  So we have much more cases.  I understand the VFS
> changes need to be very well thought, but when I worked on this it
> started to look a more correct solution than using the hooks.

Well the point of my proof-of-concept patch having separate ext4_ci_dentry_ops,
ext4_encrypted_dentry_ops, and ext4_encrypted_ci_dentry_ops is supposed to be
for overlayfs support -- since overlayfs requires that some operations are not
present.  If we didn't need overlayfs support, we could just use a single
ext4_dentry_ops for all dentries instead.

I think we could still support fscrypt, casefold, fscrypt+casefold, and
fscrypt+overlayfs with dentry_operations only.  It's casefold+overlayfs that's
the biggest problem, due to the possibility of the casefold flag being set on a
directory later as you pointed out.

- Eric