On Wed, Jun 10, 2015 at 10:20 AM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
> On Wed, Jun 10, 2015 at 9:31 AM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>> On 06/09, Andy Lutomirski wrote:
>>>
>>> On Tue, Jun 9, 2015 at 5:49 PM, Tycho Andersen
>>> >
>>> > @@ -556,6 +556,15 @@ static int ptrace_setoptions(struct task_struct *child, unsigned long data)
>>> > if (data & ~(unsigned long)PTRACE_O_MASK)
>>> > return -EINVAL;
>>> >
>>> > + if (unlikely(data & PTRACE_O_SUSPEND_SECCOMP)) {
>>
>> Well, we should do this if
>>
>> (data & O_SUSPEND) && !(flags & O_SUSPEND)
>>
>> or at least if
>>
>> (data ^ flags) & O_SUSPEND
>>
>>
>>> > + if (!config_enabled(CONFIG_CHECKPOINT_RESTORE) ||
>>> > + !config_enabled(CONFIG_SECCOMP))
>>> > + return -EINVAL;
>>> > +
>>> > + if (!capable(CAP_SYS_ADMIN))
>>> > + return -EPERM;
>>>
>>> I tend to think that we should also require that current not be using
>>> seccomp. Otherwise, in principle, there's a seccomp bypass for
>>> privileged-but-seccomped programs.
>>
>> Andy, I simply can't understand why do we need any security check at all.
>>
>> OK, yes, in theory we can have a seccomped CAP_SYS_ADMIN process, seccomp
>> doesn't filter ptrace, you hack that process and force it to attach to
>> another CAP_SYS_ADMIN/seccomped process, etc, etc... Looks too paranoid
>> to me.
>
> I've sometimes considered having privileged processes I write fork and
> seccomp their child. Of course, if you're allowing ptrace through
> your seccomp filter, you open a giant can of worms, but I think we
> should take the more paranoid approach to start and relax it later as
> needed. After all, for the intended use of this patch, stuff will
> break regardless of what we do if the ptracer is itself seccomped.
>
> I could be convinced that if the ptracer is outside seccomp then we
> shouldn't need the CAP_SYS_ADMIN check. That would at least make this
> work in a user namespace.
But not if that namespace is running under a manager that has added a
seccomp filter to do things like drop finit_module, as lxc does.
Let's start with CAP_SYS_ADMIN, and when we have an actual use-case,
we can change it then.
>
>>> > @@ -590,6 +590,10 @@ void secure_computing_strict(int this_syscall)
>>> > {
>>> > int mode = current->seccomp.mode;
>>> >
>>> > + if (config_enabled(CONFIG_CHECKPOINT_RESTORE) &&
>>> > + unlikely(current->ptrace & PT_SUSPEND_SECCOMP))
>>> > + return;
>>> > +
>>> > if (mode == 0)
>>> > return;
>>> > else if (mode == SECCOMP_MODE_STRICT)
>>> > @@ -691,6 +695,10 @@ u32 seccomp_phase1(struct seccomp_data *sd)
>>> > int this_syscall = sd ? sd->nr :
>>> > syscall_get_nr(current, task_pt_regs(current));
>>> >
>>> > + if (config_enabled(CONFIG_CHECKPOINT_RESTORE) &&
>>> > + unlikely(current->ptrace & PT_SUSPEND_SECCOMP))
>>> > + return SECCOMP_PHASE1_OK;
>>> > +
>>>
>>> If it's not hard, it might still be nice to try to fold this into
>>> mode. This code is rather hot. If it would be a mess, then don't
>>> worry about it for now.
>>
>> IMO, this would be a mess ;) At least compared to this simple patch.
>>
>> Suppose we add SECCOMP_MODE_SUSPENDED. Not only this adds the problems
>> with detach if the tracer dies.
>>
>> We need to change copy_seccomp(). And it is not clear what should we
>> do if the child is traced too.
>>
>> We need to change prctl_set_seccomp() paths.
>>
>> And even the "tracee->seccomp.mode = SECCOMP_MODE_SUSPENDED" code needs
>> some locking even if the tracee is stopped, we need to avoid the races
>> with SECCOMP_FILTER_FLAG_TSYNC from other threads.
>>
>
> Agreed. Let's hold off until this becomes a problem (if it ever does).
Arg, right, no. I don't want this represented in seccomp.mode. Way too
much would get touched for little benefit.
Thanks! And sorry Tycho as we all disagree about how to disagree with
your patch... :)
--
Kees Cook
Chrome OS Security
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
