Re: [PATCH v7 3/6] seccomp: add a way to get a listener fd from ptrace

[Christian Brauner]

On Tue, Oct 09, 2018 at 03:36:04PM +0200, Jann Horn wrote:
> +cc selinux people explicitly, since they probably have opinions on this
> 
> On Tue, Oct 9, 2018 at 3:29 PM Christian Brauner <christian@xxxxxxxxxx> wrote:
> > On Tue, Oct 09, 2018 at 02:39:53PM +0200, Jann Horn wrote:
> > > On Mon, Oct 8, 2018 at 8:18 PM Christian Brauner <christian@xxxxxxxxxx> wrote:
> > > > On Mon, Oct 08, 2018 at 06:42:00PM +0200, Jann Horn wrote:
> > > > > On Mon, Oct 8, 2018 at 6:21 PM Christian Brauner <christian@xxxxxxxxxx> wrote:
> > > > > > On Mon, Oct 08, 2018 at 05:33:22PM +0200, Jann Horn wrote:
> > > > > > > On Mon, Oct 8, 2018 at 5:16 PM Christian Brauner <christian@xxxxxxxxxx> wrote:
> > > > > > > > On Thu, Sep 27, 2018 at 09:11:16AM -0600, Tycho Andersen wrote:
> > > > > > > > > diff --git a/kernel/seccomp.c b/kernel/seccomp.c
> > > > > > > > > index 44a31ac8373a..17685803a2af 100644
> > > > > > > > > --- a/kernel/seccomp.c
> > > > > > > > > +++ b/kernel/seccomp.c
> > > > > > > > > @@ -1777,4 +1777,35 @@ static struct file *init_listener(struct task_struct *task,
> > > > > > > > >
> > > > > > > > >       return ret;
> > > > > > > > >  }
> > > > > > > > > +
> > > > > > > > > +long seccomp_new_listener(struct task_struct *task,
> > > > > > > > > +                       unsigned long filter_off)
> > > > > > > > > +{
> > > > > > > > > +     struct seccomp_filter *filter;
> > > > > > > > > +     struct file *listener;
> > > > > > > > > +     int fd;
> > > > > > > > > +
> > > > > > > > > +     if (!capable(CAP_SYS_ADMIN))
> > > > > > > > > +             return -EACCES;
> > > > > > > >
> > > > > > > > I know this might have been discussed a while back but why exactly do we
> > > > > > > > require CAP_SYS_ADMIN in init_userns and not in the target userns? What
> > > > > > > > if I want to do a setns()fd, CLONE_NEWUSER) to the target process and
> > > > > > > > use ptrace from in there?
> > > > > > >
> > > > > > > See https://lore.kernel.org/lkml/CAG48ez3R+ZJ1vwGkDfGzKX2mz6f=jjJWsO5pCvnH68P+RKO8Ow@xxxxxxxxxxxxxx/
> > > > > > > . Basically, the problem is that this doesn't just give you capability
> > > > > > > over the target task, but also over every other task that has the same
> > > > > > > filter installed; you need some sort of "is the caller capable over
> > > > > > > the filter and anyone who uses it" check.
> > > > > >
> > > > > > Thanks.
> > > > > > But then this new ptrace feature as it stands is imho currently broken.
> > > > > > If you can install a seccomp filter with SECCOMP_RET_USER_NOTIF if you
> > > > > > are ns_cpabable(CAP_SYS_ADMIN) and also get an fd via seccomp() itself
> > > > > > if you are ns_cpabable(CAP_SYS_ADMIN)
> > >
> > > Actually, you don't need CAP_SYS_ADMIN for seccomp() at all as long as
> > > you enable the NNP flag, I think?
> >
> > Yes, if you turn on NNP you don't even need sys_admin.
> >
> > >
> > > > > > then either the new ptrace() api
> > > > > > extension should be fixed to allow for this too or the seccomp() way of
> > > > > > retrieving the pid - which I really think we want - needs to be fixed to
> > > > > > require capable(CAP_SYS_ADMIN) too.
> > > > > > The solution where both require ns_capable(CAP_SYS_ADMIN) is - imho -
> > > > > > the preferred way to solve this.
> > > > > > Everything else will just be confusing.
> > > > >
> > > > > First you say "broken", then you say "confusing". Which one do you mean?
> > > >
> > > > Both. It's broken in so far as it places a seemingly unnecessary
> > > > restriction that could be fixed. You outlined one possible fix yourself
> > > > in the link you provided.
> > >
> > > If by "possible fix" you mean "check whether the seccomp filter is
> > > only attached to a single task": That wouldn't fundamentally change
> > > the situation, it would only add an additional special case.
> > >
> > > > And it's confusing in so far as there is a way
> > > > via seccomp() to get the fd without said requirement.
> > >
> > > I don't find it confusing at all. seccomp() and ptrace() are very
> >
> > Fine, then that's a matter of opinion. I find it counterintuitive that
> > you can get an fd without privileges via one interface but not via
> > another.
> >
> > > different situations: When you use seccomp(), infrastructure is
> >
> > Sure. Note, that this is _one_ of the reasons why I want to make sure we
> > keep the native seccomp() only based way of getting an fd without
> > forcing userspace to switching to a differnet kernel api.
> >
> > > already in place for ensuring that your filter is only applied to
> > > processes over which you are capable, and propagation is limited by
> > > inheritance from your task down. When you use ptrace(), you need a
> > > pretty different sort of access check that checks whether you're
> > > privileged over ancestors, siblings and so on of the target task.
> >
> > So, don't get me wrong I'm not arguing against the ptrace() interface in
> > general. If this is something that people find useful, fine. But, I
> > would like to have a simple single-syscall pure-seccomp() based way of
> > getting an fd, i.e. what we have in patch 1 of this series.
> 
> Yeah, I also prefer the seccomp() one.
> 
> > > But thinking about it more, I think that CAP_SYS_ADMIN over the saved
> > > current->mm->user_ns of the task that installed the filter (stored as
> > > a "struct user_namespace *" in the filter) should be acceptable.
> >
> > Hm... Why not CAP_SYS_PTRACE?
> 
> Because LSMs like SELinux add extra checks that apply even if you have
> CAP_SYS_PTRACE, and this would subvert those. The only capability I
> know of that lets you bypass LSM checks by design (if no LSM blocks
> the capability itself) is CAP_SYS_ADMIN.
> 
> > One more thing. Citing from [1]
> >
> > > I think there's a security problem here. Imagine the following scenario:
> > >
> > > 1. task A (uid==0) sets up a seccomp filter that uses SECCOMP_RET_USER_NOTIF
> > > 2. task A forks off a child B
> > > 3. task B uses setuid(1) to drop its privileges
> > > 4. task B becomes dumpable again, either via prctl(PR_SET_DUMPABLE, 1)
> > > or via execve()
> > > 5. task C (the attacker, uid==1) attaches to task B via ptrace
> > > 6. task C uses PTRACE_SECCOMP_NEW_LISTENER on task B
> >
> > Sorry, to be late to the party but would this really pass
> > __ptrace_may_access() in ptrace_attach()? It doesn't seem obvious to me
> > that it would... Doesn't look like it would get past:
> >
> >         tcred = __task_cred(task);
> >         if (uid_eq(caller_uid, tcred->euid) &&
> >             uid_eq(caller_uid, tcred->suid) &&
> >             uid_eq(caller_uid, tcred->uid)  &&
> >             gid_eq(caller_gid, tcred->egid) &&
> >             gid_eq(caller_gid, tcred->sgid) &&
> >             gid_eq(caller_gid, tcred->gid))
> >                 goto ok;
> >         if (ptrace_has_cap(tcred->user_ns, mode))
> >                 goto ok;
> >         rcu_read_unlock();
> >         return -EPERM;
> > ok:
> >         rcu_read_unlock();
> >         mm = task->mm;
> >         if (mm &&
> >             ((get_dumpable(mm) != SUID_DUMP_USER) &&
> >              !ptrace_has_cap(mm->user_ns, mode)))
> >             return -EPERM;
> 
> Which specific check would prevent task C from attaching to task B? If
> the UIDs match, the first "goto ok" executes; and you're dumpable, so
> you don't trigger the second "return -EPERM".

You'd also need CAP_SYS_PTRACE in the mm->user_ns which you shouldn't
have if you did a setuid to an unpriv user. (But I always find that code
confusing.)

> 
> > > 7. because the seccomp filter is shared by task A and task B, task C
> > > is now able to influence syscall results for syscalls performed by
> > > task A
> >
> > [1]: https://lore.kernel.org/lkml/CAG48ez3R+ZJ1vwGkDfGzKX2mz6f=jjJWsO5pCvnH68P+RKO8Ow@xxxxxxxxxxxxxx/