lkml.org 
[lkml]   [2018]   [Oct]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v7 3/6] seccomp: add a way to get a listener fd from ptrace
On Wed, Oct 10, 2018 at 05:33:43PM +0200, Jann Horn wrote:
> On Wed, Oct 10, 2018 at 5:32 PM Paul Moore <paul@paul-moore.com> wrote:
> > On Tue, Oct 9, 2018 at 9:36 AM Jann Horn <jannh@google.com> wrote:
> > > +cc selinux people explicitly, since they probably have opinions on this
> >
> > I just spent about twenty minutes working my way through this thread,
> > and digging through the containers archive trying to get a good
> > understanding of what you guys are trying to do, and I'm not quite
> > sure I understand it all. However, from what I have seen, this
> > approach looks very ptrace-y to me (I imagine to others as well based
> > on the comments) and because of this I think ensuring the usual ptrace
> > access controls are evaluated, including the ptrace LSM hooks, is the
> > right thing to do.
>
> Basically the problem is that this new ptrace() API does something
> that doesn't just influence the target task, but also every other task
> that has the same seccomp filter. So the classic ptrace check doesn't
> work here.

Just to throw this into the mix: then maybe ptrace() isn't the right
interface and we should just go with the native seccomp() approach for
now.

>
> > If I've missed something, or I'm thinking about this wrong, please
> > educate me; just a heads-up that I'm largely offline for most of this
> > week so responses on my end are going to be delayed much more than
> > usual.
> >
> > > On Tue, Oct 9, 2018 at 3:29 PM Christian Brauner <christian@brauner.io> 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@brauner.io> 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@brauner.io> 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@brauner.io> 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@mail.gmail.com/
> > > > > > > > > . 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".
> > >
> > > > > 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@mail.gmail.com/
> >
> >
> >
> > --
> > paul moore
> > www.paul-moore.com

\
 
 \ /
  Last update: 2018-10-10 17:40    [W:0.200 / U:2.908 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site