Re: [PATCH] ptrace: being capable wrt a process requires mapped uids/gids

[Jann Horn]

On Sat, Dec 26, 2015 at 02:23:45PM -0600, Serge E. Hallyn wrote:
> On Sat, Dec 26, 2015 at 02:10:38AM +0100, Jann Horn wrote:
> > On Sat, Dec 12, 2015 at 09:12:41PM +0100, Jann Horn wrote:
> > > With this change, the entering process can first enter the
> > > namespace and then safely inspect the namespace's
> > > properties, e.g. through /proc/self/{uid_map,gid_map},
> > > assuming that the namespace owner doesn't have access to
> > > uid 0.
> > 
> > Actually, I think I missed something there. Well, at least it
> > should not directly lead to a container escape.
> > 
> > 
> > > -static int ptrace_has_cap(struct user_namespace *ns, unsigned int mode)
> > > +static bool ptrace_has_cap(const struct cred *tcred, unsigned int mode)
> > >  {
> > > +	struct user_namespace *tns = tcred->user_ns;
> > > +	struct user_namespace *curns = current_cred()->user_ns;
> > > +
> > > +	/* When a root-owned process enters a user namespace created by a
> > > +	 * malicious user, the user shouldn't be able to execute code under
> > > +	 * uid 0 by attaching to the root-owned process via ptrace.
> > > +	 * Therefore, similar to the capable_wrt_inode_uidgid() check,
> > > +	 * verify that all the uids and gids of the target process are
> > > +	 * mapped into the current namespace.
> > > +	 * No fsuid/fsgid check because __ptrace_may_access doesn't do it
> > > +	 * either.
> > > +	 */
> > > +	if (!kuid_has_mapping(curns, tcred->euid) ||
> > > +			!kuid_has_mapping(curns, tcred->suid) ||
> > > +			!kuid_has_mapping(curns, tcred->uid)  ||
> > > +			!kgid_has_mapping(curns, tcred->egid) ||
> > > +			!kgid_has_mapping(curns, tcred->sgid) ||
> > > +			!kgid_has_mapping(curns, tcred->gid))
> > > +		return false;
> > > +
> > >  	if (mode & PTRACE_MODE_NOAUDIT)
> > > -		return has_ns_capability_noaudit(current, ns, CAP_SYS_PTRACE);
> > > +		return has_ns_capability_noaudit(current, tns, CAP_SYS_PTRACE);
> > >  	else
> > > -		return has_ns_capability(current, ns, CAP_SYS_PTRACE);
> > > +		return has_ns_capability(current, tns, CAP_SYS_PTRACE);
> > >  }
> > 
> > If the namespace owner can run code in the init namespace, the kuids are
> > mapped into curns but he is still capable wrt the target namespace.
> > 
> > I think a proper fix should first determine the highest parent of
> > tcred->user_ns in which the caller still has privs, then do the
> > kxid_has_mapping() checks in there.
> 
> Hi,
> 
> I don't quite follow what you are concerned about.  Based on the new
> patch you sent, I assume it's not the case where the tcred's kuid is
> actually mapped into the container.  So is it the case where I
> unshare a userns which unshares a userns, then setns from the grandparent
> into the child?  And if so, the concern is that if the setns()ing task's
> kuid is mappable all along into the grandhild, then container root should
> be able to ptrace it?

Consider the following scenario:

init_user_ns has a child namespace (I'll call it child_ns).
child_ns is owned by an attacker (child_ns->owner == attacker_kuid).
The attacking process has current_cred()->euid == attacker_kuid and lives
in init_user_ns (which means it's capable in child_ns).

The victim process (with euid==0) enters the namespace, then the attacking
process tries to attach to it. ptrace_has_cap(tcred, mode) would, with my
old patch, still be true because current is capable in tcred->user_ns and
all kuids are mapped into init_user_ns.

The new patch uses the following rule for cap checks:

    The caller is capable relative to a target if a user namespace exists
    for which all of the following conditions are true:

     - The caller has the requested capability inside the namespace.
     - The target is inside the namespace (either directly or in a child).
     - The target's kuids and kgids are mapped into the namespace.

The first rule is enforced by the has_ns_capability(..., tns, ...) or
has_ns_capability_noaudit(..., tns, ...) call at the bottom.

The second rule is implicitly true because tns initially is the target's
user namespace and then moves up through ->parent.

The third rule is enforced by the while loop.


This prevents the attack I described, but e.g. still allows someone who
is capable in init_user_ns to ptrace anything, no matter in what weird
namespace the target is - if a task was ptrace-able for a process before
it called clone(CLONE_NEWUSER) / unshare(CLONE_NEWUSER) /
setns(, CLONE_NEWUSER), it should still be ptrace-able afterwards.
(Unless access was permitted based on the introspection rule.)

Attachment: signature.asc
Description: Digital signature