SECCOMP_IOCTL_NOTIF_ADDFD race condition

[Kees Cook]

On Tue, Dec 01, 2020 at 08:13:34AM -0500, Tycho Andersen wrote:
> On Tue, Dec 01, 2020 at 01:08:25PM +0000, Sargun Dhillon wrote:
> > On Tue, Dec 01, 2020 at 07:41:05AM -0500, Tycho Andersen wrote:
> > > On Mon, Nov 30, 2020 at 06:20:09PM -0500, Tycho Andersen wrote:
> > > > Idea 1 sounds best to me, but maybe that's because it's the way I
> > > > originally did the fd support that never landed :)
> > > > 
> > > > But here's an Idea 4: we add a way to remotely close an fd (I don't
> > > > see that the current infra can do this, but perhaps I didn't look hard
> > > > enough), and then when you get ENOENT you have to close the fd. Of
> > > > course, this can't be via seccomp, so maybe it's even more racy.
> > > 
> > > Or better yet: what if the kernel closed everything it had added via
> > > ADDFD if it didn't get a valid response from the supervisor? Then
> > > everyone gets this bug fixed for free.
> > > 
> > > Tycho
> > > _______________________________________________
> > > Containers mailing list
> > > Containers at lists.linux-foundation.org
> > > https://lists.linuxfoundation.org/mailman/listinfo/containers
> > 
> > This doesn't solve the problem universally because of the (Go) preemption 
> > problem. Unless we can guarantee that the supervisor can always handle the 
> > request in fewer than 10ms, or if it implements resumption behaviour. I know 
> > that resumption behaviour is a requirement no matter what, but the easier we can 
> > make it to implement resumption, the better chance we are giving users to get 
> > this right.
> 
> Doesn't automatic cleanup of fds make things easier? I'm not sure I
> understand the argument.

I doubt Al would ever allow the "cleanup" approach: his observation was
that the instant a file has been added to the fdtable, it's not possible
to "unwind" that ever, since it could be cloned away, etc, etc.

> I agree it doesn't fix the problem of uncooperative userspace.

IIUC, I see two issues:

- a slow monitor might cause a child to loop forever retrying the same
  interrupted syscall.

- a syscall-interrupted process may have had an fd added that it has no
  idea about.

The former problem seems like a userspace issue. :P But, to help, yeah, is
signal blocking best? Either explicit (at filter apply time) or implicit
(all user_notif-triggering syscalls get all signals blocks automatically)?

For the latter problem, I think we need to get back to Tycho's original
method: add fd and finish syscall in a single action. I can't see any
other way to get around the need for atomicity...

-- 
Kees Cook