[PATCH v2 07/39] mount: attach mappings to mounts

[Tycho Andersen]

On Tue, Nov 24, 2020 at 02:40:35PM +0100, Christian Brauner wrote:
> On Tue, Nov 24, 2020 at 08:37:40AM -0500, Tycho Andersen wrote:
> > On Tue, Nov 24, 2020 at 01:30:35PM +0100, Christian Brauner wrote:
> > > On Mon, Nov 23, 2020 at 11:24:28AM -0500, Tycho Andersen wrote:
> > > > On Mon, Nov 23, 2020 at 10:47:19AM -0500, Tycho Andersen wrote:
> > > > > On Sun, Nov 15, 2020 at 11:36:46AM +0100, Christian Brauner wrote:
> > > > > > +static inline struct user_namespace *mnt_user_ns(const struct vfsmount *mnt)
> > > > > > +{
> > > > > > +	return mnt->mnt_user_ns;
> > > > > > +}
> > > > > 
> > > > > I think you might want a READ_ONCE() here. Right now it seems ok, since the
> > > > > mnt_user_ns can't change, but if we ever allow it to change (and I see you have
> > > > > a idmapped_mounts_wip_v2_allow_to_change_idmapping branch on your public tree
> > > > > :D), the pattern of,
> > > > > 
> > > > >         user_ns = mnt_user_ns(path->mnt);
> > > > >         if (mnt_idmapped(path->mnt)) {
> > > > >                 uid = kuid_from_mnt(user_ns, uid);
> > > > >                 gid = kgid_from_mnt(user_ns, gid);
> > > > >         }
> > > > > 
> > > > > could race.
> > > > 
> > > > Actually, isn't a race possible now?
> > > > 
> > > > kuid_from_mnt(mnt_user_ns(path->mnt) /* &init_user_ns */);
> > > > WRITE_ONCE(mnt->mnt.mnt_user_ns, user_ns);
> > > > WRITE_ONCE(m->mnt.mnt_flags, flags);
> > > > kgid_from_mnt(mnt_user_ns(path->mnt) /* the right user ns */);
> > > > 
> > > > So maybe it should be:
> > > > 
> > > >          if (mnt_idmapped(path->mnt)) {
> > > >                  barrier();
> > > >                  user_ns = mnt_user_ns(path->mnt);
> > > >                  uid = kuid_from_mnt(user_ns, uid);
> > > >                  gid = kgid_from_mnt(user_ns, gid);
> > > >          }
> > > > 
> > > > since there's no data dependency between mnt_idmapped() and
> > > > mnt_user_ns()?
> > > 
> > > I think I had something to handle this case in another branch of mine.
> > > The READ_ONCE() you mentioned in another patch I had originally dropped
> > > because I wasn't sure whether it works on pointers but after talking to
> > > Jann and David it seems that it handles pointers fine.
> > > Let me take a look and fix it in the next version. I just finished
> > > porting the test suite to xfstests as Christoph requested and I'm
> > > looking at this now.
> > 
> > Another way would be to just have mnt_idmapped() test
> > mnt_user_ns() != &init_user_ns instead of the flags; then I think you
> > get the data dependency and thus correct ordering for free.
> 
> I indeed dropped mnt_idmapped() which is unnecessary. :)

It still might be a nice helper to prevent people from checking the
flags and forgetting that there's a memory ordering issue, though.

> I think we should still use smp_store_release() in mnt_user_ns() paired
> with smp_load_acquire() in do_idmap_mount() thought.

Sounds reasonable.

Tycho