Re: [PATCH v6 4/5] seccomp: add support for passing fds via USER_NOTIF

From: Tycho Andersen
Date: Thu Sep 20 2018 - 19:42:52 EST

On Wed, Sep 19, 2018 at 12:58:20PM -0700, Andy Lutomirski wrote:
> On Wed, Sep 19, 2018 at 7:38 AM, Tycho Andersen <tycho@xxxxxxxx> wrote:
> > On Wed, Sep 19, 2018 at 07:19:56AM -0700, Andy Lutomirski wrote:
> >>
> >>
> >> > On Sep 19, 2018, at 2:55 AM, Tycho Andersen <tycho@xxxxxxxx> wrote:
> >> >
> >> >> On Wed, Sep 12, 2018 at 04:52:38PM -0700, Andy Lutomirski wrote:
> >> >>> On Thu, Sep 6, 2018 at 8:28 AM, Tycho Andersen <tycho@xxxxxxxx> wrote:
> >> >>> The idea here is that the userspace handler should be able to pass an fd
> >> >>> back to the trapped task, for example so it can be returned from socket().
> >> >>>
> >> >>> I've proposed one API here, but I'm open to other options. In particular,
> >> >>> this only lets you return an fd from a syscall, which may not be enough in
> >> >>> all cases. For example, if an fd is written to an output parameter instead
> >> >>> of returned, the current API can't handle this. Another case is that
> >> >>> netlink takes as input fds sometimes (IFLA_NET_NS_FD, e.g.). If netlink
> >> >>> ever decides to install an fd and output it, we wouldn't be able to handle
> >> >>> this either.
> >> >>
> >> >> An alternative could be to have an API (an ioctl on the listener,
> >> >> perhaps) that just copies an fd into the tracee. There would be the
> >> >> obvious set of options: do we replace an existing fd or allocate a new
> >> >> one, and is it CLOEXEC. Then the tracer could add an fd and then
> >> >> return it just like it's a regular number.
> >> >>
> >> >> I feel like this would be more flexible and conceptually simpler, but
> >> >> maybe a little slower for the common cases. What do you think?
> >> >
> >> > I'm just implementing this now, and there's one question: when do we
> >> > actually do the fd install? Should we do it when the user calls
> >> > SECCOMP_NOTIF_PUT_FD, or when the actual response is sent? It feels
> >> > like we should do it when the response is sent, instead of doing it
> >> > right when SECCOMP_NOTIF_PUT_FD is called, since if there's a
> >> > subsequent signal and the tracer decides to discard the response,
> >> > we'll have to implement some delete mechanism to delete the fd, but it
> >> > would have already been visible to the process, etc. So I'll go
> >> > forward with this unless there are strong objections, but I thought
> >> > I'd point it out just to avoid another round trip.
> >> >
> >> >
> >>
> >> Can you do that non-racily? That is, you need to commit to an fd *number* right away, but what if another thread uses the number before you actually install the fd?
> >
> > I was thinking we could just do an __alloc_fd() and then do the
> > fd_install() when the response is sent or clean up the case that the
> > listener or task dies. I haven't actually tried to run the code yet,
> > so it's possible the locking won't work :)
> I would be very surprised if the locking works. How can you run a
> thread in a process when another thread has allocated but not
> installed an fd and is blocked for an arbitrarily long time?

I think the trick is that there's no actual locking required (except
for a brief locking of task->files). I've run the patch below and it
seems to work. But perhaps that's abusing __alloc_fd a little too
hard, I don't really know.

> >
> >> Do we really allow non-âkillâ signals to interrupt the whole process? It might be the case that we donât really need to clean up from signals if thereâs a guarantee that the thread dies.
> >
> > Yes, we do, because of this:
> >
> I'm still not sure I see the problem. Suppose I'm implementing a user
> notifier for a nasty syscall like recvmsg(). If I'm the tracer, by
> the time I decide to install an fd, I've committed to returning
> something other than -EINTR, even if a non-fatal signal is sent before
> I finish. No rollback should be necessary.

I don't understand why this is true. Surely you could stop a handler
on receipt of a new signal, and have it do something else entirely?

> In the (unlikely?) event that some tracer needs to be able to rollback
> an fd installation to return -EINTR, a SECCOMP_NOTIF_CLOSE_FD
> operation should be good enough, I think. Or maybe PUT_FD can put -1
> to delete an fd.

Yes, I think even with something like what I did below we'd need some
sort of REMOVE_FD option, because otherwise there's no way to change
your mind and send -EINTR without the fd you just PUT_FD'd.