[Virtio-fs] Question on ACLs support in virtiofs

[Miklos Szeredi]

On Mon, Feb 15, 2021 at 9:52 PM Vivek Goyal <vgoyal at redhat.com> wrote:
>
> On Fri, Feb 12, 2021 at 10:30:13AM +0000, Luis Henriques wrote:
> > Hi!
> >
> > I've recently executed the generic fstests on virtiofs and decided to have
> > a closer look at generic/099 failure.  In a nutshell, here's the sequence
> > of commands that reproduce that failure:
> >
> > # umask 0
> > # mkdir acldir
> > # chacl -b "u::rwx,g::rwx,o::rwx" "u::r-x,g::r--,o::---" acldir
> > # touch acldir/file1
> > # umask 722
> > # touch acldir/file2
> > # ls -l acldir
> > total 0
> > -r--r----- 1 root root 0 Feb 12 10:04 file1
> > ----r----- 1 root root 0 Feb 12 10:05 file2
> >
> > The failure is that setting umask to 722 shouldn't affect the new file2
> > because acldir has a default ACL (from umask(2): "... if the parent
> > directory has a default ACL (see acl(5)), the umask is ignored...").
> >
> > So... I tried to have look at the code, and initially I thought that the
> > problem was in (kernel) function fuse_create_open(), where we have this:
> >
> >       if (!fm->fc->dont_mask)
> >               mode &= ~current_umask();
> >
> > but then I went down the rabbit hole, into the user-space code, and
> > couldn't reach a conclusion.  Maybe the issue is that there's in fact no
> > support for this POSIX ACLs in virtiofs/FUSE?  Any ideas?
>
> Hi,
>
> [ CC Miklos and linux-fsdevel ]
>
> I debugged into this a little. There are many knobs and it is little
> confusing that what are right set of fixes.
>
> So what's happening in this case is that fc->dont_mask is not set. That
> means fuse client is modifying mode using umask. First time you
> touch file, umask is 0, so there is no modification. But next time,
> you set umask to 722, and fuse modifies mode before sending file
> create request to server. virtiofs server is already running with
> umask 0, so it does not touch the mode.
>
> So that means, that in case of default acl, fuse client should not
> be modifying mode using umask. But question is when should fuse
> skip applying umask.
>
> I see that fuse always sets SB_POSIXACL. That means VFS is not
> going to apply umask and all the umask handling is with-in fuse.
>
> sb->s_flags |= SB_POSIXACL;
>
> Currently fuse sets fc->dont_mask in two conditions.
>
> - If the caller mounted with flag MS_POSIXACL, then fc->dont_mask is set.
> - If fuse server opted in for option FUSE_DONT_MASK, then fc->dont_mask
>   is set.
>
> I see that for virtiofs, both the conditions are not true out of the
> box. In fact looks like ACL support is not fully enabled, because
> I don't see fuse server opting in for FUSE_POSIX_ACL.
>
> I suspect that we probably should provide an option in virtiofsd to
> enable/disable acl support.

Sounds good.

> Setting FUSE_DONT_MASK is tricky. If we leave it to fuse, that means
> fuse will have to query acl to figure out if default acl is set or
> not on parent dir. And that data could be stale and there could be
> races w.r.t setting acls from other client.
>
> If we do set FUSE_DONT_MASK, that means in file creation path virtiofsd
> server will have to switch its umask to one provided in request. Given
> its a per process property, we will have to have some locks to make
> sure other create requests are not progressing in parallel. And that
> hope host does the right thing. That is apply umask if parent dir does
> not have default acl otherwise apply umask (as set by virtiofsd process).
>
> Miklos, does above sound reasonable. You might have more thoughts on
> how to handle this best in fuse/virtiofs.

fv_queue_worker() does unshare(CLONE_FS) for the fchdir() call in
xattr ops, which means that umask is now a per-thread propery in
virtiofsd.

So setting umask before create ops sounds like a good solution.

Thanks,
Miklos

>
> Vivek
>