[PATCH] Forbid invocation of kexec_load() outside initial PID namespace

Serge E. Hallyn serge at hallyn.com
Mon Aug 6 19:24:46 UTC 2012

Quoting Eric W. Biederman (ebiederm at xmission.com):
> "Serge E. Hallyn" <serge at hallyn.com> writes:
> > Quoting Daniel P. Berrange (berrange at redhat.com):
> >> From: "Daniel P. Berrange" <berrange at redhat.com>
> >> 
> >> The following commit
> >> 
> >>     commit cf3f89214ef6a33fad60856bc5ffd7bb2fc4709b
> >>     Author: Daniel Lezcano <daniel.lezcano at free.fr>
> >>     Date:   Wed Mar 28 14:42:51 2012 -0700
> >> 
> >>     pidns: add reboot_pid_ns() to handle the reboot syscall
> >> 
> >> introduced custom handling of the reboot() syscall when invoked
> >> from a non-initial PID namespace. The intent was that a process
> >> in a container can be allowed to keep CAP_SYS_BOOT and execute
> >> reboot() to shutdown/reboot just their private container, rather
> >> than the host.
> >> 
> >> Unfortunately the kexec_load() syscall also relies on the
> >> CAP_SYS_BOOT capability. So by allowing a container to keep
> >> this capability to safely invoke reboot(), they mistakenly
> >> also gain the ability to use kexec_load(). The solution is
> >> to make kexec_load() return -EPERM if invoked from a PID
> >> namespace that is not the initial namespace
> >> 
> >> Signed-off-by: Daniel P. Berrange <berrange at redhat.com>
> >> Cc: Serge Hallyn <serge.hallyn at canonical.com>
> >
> > Acked-by: Serge Hallyn <serge.hallyn at canonical.com>
> >
> > (Please see my previous email explaining why I believe the pidns
> > is an appropriate check)
> Serge as to your objects.
> If we define kexec_load in terms of the pid namespace then something
> makes sense, but the error should be EINVAL, or something of that sort.

Makes sense.

> That is what we did with reboot.  We defined reboot in terms of the pid
> namespace.
> Not defining kexec_load in terms of the pid namespace and then returning
> EPERM because having we happen to have CAP_SYS_BOOT for other reasons is
> semantically horrible.
> At the end of the day the effect is the same, but I think it matters a
> great deal in how we think about things.
> We have CAP_SYS_BOOT in the initial user namespace.  We do have
> permission to make the system call.
> So I continue to see this patch the way it is current constructed as
> broken.
> Nacked-by: "Eric W. Biederman" <ebiederm at xmission.com>

I do also prefer splitting the capability.  Michael Kerrisk, do you
have any good suggestions for better names than CAP_RESTART (for
killing or restarting /sbin/init) and CAP_BOOT (for kexec and/or
hardware resets)?  Maybe CAP_RESTART_USER and CAP_RESTART_HW?
(CAP_SYS_BOOT being an alias for both for backward compatibility)

More information about the Containers mailing list