[PATCH v2 4/4] samples: Add example of using PTRACE_GETFD in conjunction with user trap

[Tycho Andersen]

On Tue, Dec 10, 2019 at 08:07:45AM -0800, Sargun Dhillon wrote:
> On Tue, Dec 10, 2019 at 3:10 AM Christian Brauner
> <christian.brauner at ubuntu.com> wrote:
> >
> > [I'm expanding the Cc to a few Firefox and glibc people since we've been
> >  been talking about replacing SECCOMP_RET_TRAP with
> >  SECCOMP_RET_USER_NOTIF for a bit now because the useage of
> >  SECCOMP_RET_TRAP in the broker blocks desirable core glibc changes.
> >  Even if just for their lurking pleasure. :)]
> >
> > On Mon, Dec 09, 2019 at 09:46:35PM +0100, Oleg Nesterov wrote:
> > > On 12/09, Christian Brauner wrote
> > >
> > > I agree, and I won't really argue...
> > >
> > > but the changelog in 2/4 says
> > >
> > >       The requirement that the tracer has attached to the tracee prior to the
> > >       capture of the file descriptor may be lifted at a later point.
> > >
> > > so may be we should do this right now?
> >
> > I think so, yes. This doesn't strike me as premature optimization but
> > rather as a core design questions.
> >
> > >
> > > plus this part
> > >
> > >       @@ -1265,7 +1295,8 @@ SYSCALL_DEFINE4(ptrace, long, request, long, pid, unsigned long, addr,
> > >               }
> > >
> > >               ret = ptrace_check_attach(child, request == PTRACE_KILL ||
> > >       -                                 request == PTRACE_INTERRUPT);
> > >       +                                 request == PTRACE_INTERRUPT ||
> > >       +                                 request == PTRACE_GETFD);
> > >
> > > actually means "we do not need ptrace, but we do not know where else we
> > > can add this fd_install(get_task_file()).
> >
> > Right, I totally get your point and I'm not a fan of this being in
> > ptrace() either.
> >
> > The way I see is is that the main use-case for this feature is the
> > seccomp notifier and I can see this being useful. So the right place to
> > plumb this into might just be seccomp and specifically on to of the
> > notifier.
> > If we don't care about getting and setting fds at random points of
> > execution it might make sense to add new options to the notify ioctl():
> >
> > #define SECCOMP_IOCTL_NOTIF_GET_FD      SECCOMP_IOWR(3, <sensible struct>)
> > #define SECCOMP_IOCTL_NOTIF_SET_FD      SECCOMP_IOWR(4, <sensible struct>)
> >
> > which would let you get and set fds while the supervisee is blocked.
> >
> > Christian
> Doesn't SECCOMP_IOCTL_NOTIF_GET_FD have some ambiguity to it?
> Specifically, because
> multiple processes can have the same notifier attached to them?

The id member corresponds to a particular syscall from a particular
pid, which makes it unique.

> If we
> choose to go down the
> route of introducing an ioctl (which I'm not at all opposed to), I
> would rather do it on pidfd. We
> can then plumb seccomp notifier to send pidfd instead of raw pid. In
> the mean time, folks
> can just open up /proc/${PID}, and do the check cookie dance.

This might be more generally useful, the problem is synchronization, I
guess.

Tycho