Re: [Qemu-devel] [PATCH for-1.7] seccomp: setting "-sandbox on" by default

[Stefan Hajnoczi <stefanha@gmail.com>]

On Thu, Dec 05, 2013 at 10:12:00AM -0600, Will Drewry wrote:
> On Thu, Dec 5, 2013 at 7:15 AM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> > On Wed, Dec 04, 2013 at 11:21:12AM -0200, Eduardo Otubo wrote:
> >> On 12/04/2013 07:39 AM, Stefan Hajnoczi wrote:
> >> >On Fri, Nov 22, 2013 at 11:00:24AM -0500, Paul Moore wrote:
> >> >>>Developers will only be happy with seccomp if it's easy and rewarding to
> >> >>>support/debug.
> >> >>
> >> >>Agreed.
> >> >>
> >> >>As a developer, how do you feel about the audit/syslog based approach I
> >> >>mentioned earlier?
> >> >
> >> >I used the commands you posted (I think that's what you mean).  They
> >> >produce useful output.
> >> >
> >> >The problem is that without an error message on stderr or from the
> >> >shell, no one will think "QEMU process dead and hung == check seccomp"
> >> >immediately.  It's frustrating to deal with a "silent" failure.
> >>
> >> The process dies with a SIGKILL, and sig handling in Qemu is hard to
> >> implement due to dozen of external linked libraries that has their
> >> own signal masks and conflicts with seccomp. I've already tried this
> >> approach in the past (you can find in the list by searching for
> >> debug mode)
> >
> > I now realize we may be talking past each other.  Dying with
> > SIGKILL/SIGSYS is perfectly reasonable and I would be happy with that
> > :-).
> >
> > But I think there's a bug in seccomp: a multi-threaded process can be
> > left in a zombie state.  In my case the primary thread was killed by
> > seccomp but another thread was deadlocked on a futex.
> >
> > The result is the process isn't quite dead yet.  The shell will not reap
> > it and we're stuck with a zombie.
> >
> > I can reproduce it reliably when I run "qemu-system-x86_64 -sandbox on"
> > on Fedora 20 (qemu-system-x86-1.6.1-2).
> >
> > Should seccomp use do_group_exit() for SIGKILL?
> 
> Is the problem that the SECCOMP_RET_KILL didn't take down the thread
> group (which would be a departure from how seccomp(mode=1) worked) and
> causes the deadlock somehow, or is it that the other thread is
> deadlocked?

The former.

When the first thread is killed by seccomp, the second thread in the
process is left waiting on a futex forever.  Therefore the process never
exits after the seccomp violation occurs.

Directing the signal at a thread makes perfect sense for
SECCOMP_RET_TRAP since the thread can handle the signal and recover.
But for SECCOMP_RET_KILL it's probably more useful to kill the entire
process rather than just a single thread.

> Regardless, adding a SECCOMP_RET_TGKILL probably isn't a bad idea :)

Yes.  Do you have time for that or would you like me to send a patch?

Stefan