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

[Stefan Hajnoczi <stefanha@gmail.com>]

On Wed, Oct 30, 2013 at 11:04:39AM +0100, Stefan Hajnoczi wrote:
> On Wed, Oct 23, 2013 at 12:42:34PM -0200, Eduardo Otubo wrote:
> > On 10/22/2013 11:00 AM, Anthony Liguori wrote:
> > >On Tue, Oct 22, 2013 at 12:21 PM, Eduardo Otubo
> > ><otubo@linux.vnet.ibm.com> wrote:
> > >>Inverting the way sandbox handles arguments, making possible to have no
> > >>argument and still have '-sandbox on' enabled.
> > >>
> > >>Signed-off-by: Eduardo Otubo <otubo@linux.vnet.ibm.com>
> > >>---
> > >>
> > >>The option '-sandbox on' is now used by default by virt-test[0] -- it has been
> > >>merged into the 'next' branch and will be available in the next release,
> > >>meaning we have a back support for regression tests if anything breaks because
> > >>of some missing system call not listed in the whitelist.
> > >>
> > >>This being said, I think it makes sense to have this option set to 'on' by
> > >>default in the next Qemu version. It's been a while since no missing syscall is
> > >>reported and at this point the whitelist seems to be pretty mature.
> > >>
> > >>[0] - https://github.com/autotest/virt-test/commit/50e1f7d47a94f4c770880cd8ec0f18365dcba714
> > >
> > >This breaks hot_add of a network device that uses a script= argument, correct?
> > >
> > >If so, this cannot be made default.
> > 
> > Anthony, I believe you're talking about the blacklist feature. This
> > is the old whitelist that is already upstream and it does not block
> > any network device to be hot plugged.
> 
> The following fails to start here (the shell hangs and ps shows QEMU is
> a <defunct> process):
> 
> qemu-system-x86_64 -sandbox on -enable-kvm -m 1024 -cpu host \
>                    -drive if=virtio,cache=none,file=test.img
> 
> It is using the GTK UI.

IMO this seccomp approach is doomed since QEMU does not practice
privilege separation.  QEMU is monolithic so it's really hard to create
a meaningful sets of system calls.  To avoid breaking stuff you need to
be too liberal, defeating the purpose of seccomp.

For each QEMU command-line there may be a different set of syscalls that
should be allowed/forbidden.

The existing approach clearly doesn't support the full range of options
that users specify on the command-line.  So I guess the options are:

1. Don't make it the default since it breaks stuff but use it for very
specific scenarios (e.g. libvirt use cases that have been well tested).

2. Provide a kind of syscall set for various QEMU options and apply the
union of them at launch.  This still seems fragile but in theory it
could work.

Stefan