Re: [RFC PATCH 3/4] kmod - add call_usermodehelper_ns() helper

[Benjamin Coddington]

n Tue, 25 Nov 2014, Eric W. Biederman wrote:
Hi,

> Ian Kent <ikent@xxxxxxxxxx> writes:
>
> > On Tue, 2014-11-25 at 17:19 -0600, Eric W. Biederman wrote:
> >> Ian Kent <ikent@xxxxxxxxxx> writes:
> >>
> >> > On Tue, 2014-11-25 at 16:23 -0600, Eric W. Biederman wrote:
> >> >> Oleg Nesterov <oleg@xxxxxxxxxx> writes:
> >> >>
> >> >> > On 11/25, Oleg Nesterov wrote:
> >> >> >>
> >> >> >> Let me first apologize, I didn't actually read this series yet.
> >> >> >>
> >> >> >> But I have to admit that so far I do not like this approach...
> >> >> >> probably I am biased.
> >> >> >
> >> >> > Yes.
> >> >> >
> >> >> > And I have another concern... this is mostly a feeling, I can be
> >> >> > easily wrong but:
> >> >> >
> >> >> >> On 11/25, Ian Kent wrote:
> >> >> >> >
> >> >> >> > +static int umh_set_ns(struct subprocess_info *info, struct cred *new)
> >> >> >> > +{
> >> >> >> > +	struct nsproxy *ns = info->data;
> >> >> >> > +
> >> >> >> > +	mntns_setfs(ns->mnt_ns);
> >> >> >>
> >> >> >> Firstly, it is not clear to me if we should use the caller's ->mnt_ns.
> >> >> >> Let me remind about the coredump. The dumping task can cloned with
> >> >> >> CLONE_NEWNS or it cam do unshare(NEWNS)... but OK, I do not understand
> >> >> >> this enough.
> >> >> >
> >> >> > And otoh. If we actually want to use the caller's mnt_ns/namespaces we
> >> >> > could simply fork/reparent a child which will do execve ?
> >> >>
> >> >> That would certainly be a better approach, and roughly equivalent to
> >> >> what exists here.  That would even ensure we remain in the proper
> >> >> cgroups, and lsm context.
> >> >>
> >> >> The practical problem with the approach presented here is that I can
> >> >> hijack any user mode helper I wish, and make it run in any executable I
> >> >> wish as the global root user.
> >> >>
> >> >> Ian if we were to merge this I believe you would win the award for
> >> >> easiest path to a root shell.
> >> >
> >> > LOL, OK, so there's a problem with this.
> >> >
> >> > But, how should a user mode helper execute within a namespace (or more
> >> > specifically within a container)?
> >> >
> >> > Suppose a user mode helper program scans through the pid list and
> >> > somehow picks the correct process pid and then does an
> >> > open()/setns()/execve().
> >> >
> >> > Does that then satisfy the requirements?
> >> > What needs to be done to safely do that in kernel?
> >> >
> >> > The other approach I've considered is doing a full open()/setns() in
> >> > kernel (since the caller already knows its pid) but it sounds like
> >> > that's not right either.
> >>
> >> The approach we agreed upon with the core dump helper was to provide
> >> enough information that userspace could figure out what was the
> >> appropriate policy and call nsenter/setns.
> >
> > Your recommending I have a look at the core dump helper, that's fine,
> > I'll do that.
>
> I am just describing it because it came up.  Core dumps are a much
> easier case than nfs.
>
> Frankly if we can figure out how to run the user mode helpers from the
> kernel with an appropriate context and not involve userspace I think
> that will be better for everyone, as it involves fewer moving parts at
> the end of the day.
>
> >> The only sane approach I can think of in the context of nfs is to fork
> >> a kernel thread at mount time that has all of the appropriate context
> >> because it was captured from the privileged mounting process, and use
> >> that kernel as the equivalent of kthreadd.
> >>
> >> There may be some intermediate ground where we capture things or we use
> >> the init process of the pid namespace (captured at mount time) as our
> >> template/reference process.
> >>
> >> If we are going to set this stuff up in the kernel we need a reference
> >> process that we can create children of because what is possible with
> >> respect to containers keeps changing, and it is extremely error prone to
> >> figure out what all othe crazy little bits are, and to update everything
> >> every time someone tweaks the kernel's capabilities.  We have kthreadd
> >> because it was too error prone to scrub a userspace thread of all of the
> >> userspace bits and make it the equivalent of what kthreadd is today.
> >>
> >> Of course it is also rather nice to have something to hang everything
> >> else on.
> >>
> >> In summary we need a reference struct task that is all setup properly
> >> so that we can create an appropriate kernel thread.
> >
> > I'm having trouble understanding what your getting at here but I'm not
> > that sharp so bear with me.
> >
> > When call_usermodehelper() is called it's called from a process that is
> > within the context within which the execution is required.
>
> No it is not.  That is precisely why we have call_usermodehelper instead
> of just forking and exec'ing something.  The context which triggers the
> call can be completely different from where you want to run.
>
> > So what information do we not have available for setup?
> >
> > Are you saying that the problem is that when the user mode helper run
> > thread is invoked we don't have the information available that was
> > present when call_usermodehelper() was called and that's where the
> > challenge lies?
>
> That is part of it.
>
> However in the context of nfs the correct context is determined at mount
> time, and so when call_usermodehelper() is invoked the only link to that
> correct context that we have is the nfs super block.
>
> In a pathological case a userspace application can create a new user
> namespace, and a new mount namespace and completely rearrange the
> mounts, and deliberately trigger an nfs user mode helper.  If we were to
> use that applications context it could control which userspace
> application the kernel invoked.

Then it seems it would never be safe to try to execve /sbin/request-key
within the calling process' mnt_ns, so if we want to do some work
within that namespace the transition to it needs to happen after the
upcall, after /sbin/request-key calls the appropriate helper.

And instead of transitioning into this context before execve, instead
pass it along to allow the userspace helper to do setns()..


> So deeply and fundamentally for the case of nfs you need to capture the
> context at nfs mount time.  The easiest way would be to fork a kernel
> thread when nfs is mounted, that captures the entire context.
>
> We might be able to do something a little more sophisticated and a
> little less resource intensive, like pick on the init process that
> exists when nfs is mounted.

The only namespace that needs to be preserved from mount time would be
the mount namespace which can be pulled from the mount itself and passed
along in the key request.

Or if we tracked the mount namespace when creating a user
namespace, we could climb out of the mount namespaces created within
user namespaces by following them out.  Does that sound sane, or
completely nuts?

> Those are the general parameters.

It does seem very expensive to keep a thread around for every mount; I'm
still trying to find a way around it..

Ben

>
> Eric
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/