Re: [CRIU] Introspecting userns relationships to other namespaces?

From: Andrew Vagin
Date: Sat Jul 09 2016 - 03:41:14 EST

On Fri, Jul 08, 2016 at 10:05:18PM -0500, Eric W. Biederman wrote:
> James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> writes:
> > On Fri, 2016-07-08 at 18:52 -0500, Eric W. Biederman wrote:
> >> James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> writes:
> >>
> >> > On July 8, 2016 1:38:19 PM PDT, Andrew Vagin <avagin@xxxxxxxxxxxxx>
> >> > wrote:
> >>
> >> > > What do you think about the idea to mount nsfs and be able to
> >> > > look up any alive namespace by inum:
> >> >
> >> > I think I like it. It will give us a way to enter any extant
> >> > namespace. It will work for Eric's fs namespaces as well. Perhaps
> >> > a /process/ns/<inum> Directory?
> >
> > As you understood, I meant /proc/ns/<inum> (damn mobile phone
> > completions).
> >
> >> *Shivers*
> >>
> >> That makes it very easy to bypass any existing controls that exist
> >> for getting at namespaces. It is true that everything of that kind
> >> is directory based but still.
> >>
> >> Plus I think it would serve as information leak to information
> >> outside of the container.
> >>
> >> An operation to get a user namespace file descriptor from some kernel
> >> object sounds reasonably sane.
> >>
> >> A great big list of things sounds about as scary as it can get. This
> >> is not the time to be making it easier to escape from containers.
> >
> > To be honest, I think this argument is rubbish. If we're afraid of
> > giving out a list of all the namespaces, it means we're afraid there's
> > some security bug and we're trying to obscure it by making the list
> > hard to get. All we've done is allayed fears about the bug but the
> > hackers still know the portals to get through.
> >
> > If such a bug exists, it will be possible to exploit it by simply
> > reconstructing the information from the individual process directories,
> > so obscurity doesn't protect us and all it does is give us a false
> > sense of security. If such a bug doesn't exist, then all the security
> > mechanisms currently in place (like no re-entry to prior namespace)
> > should protect us and we can give out the list.
> >
> > Let's deal with the world as we'd like it to be (no obscure namespace
> > bugs) and accept the consequences and the responsibility for fixing
> > them if we turn out to be slightly incorrect. We'll end up in a far
> > better place than security by obscurity would land us.
> No. That is not the fear. The permission checks on /proc/self/ns/xxx
> are different than if the namespace is bind mounted somewhere.
> That was done deliberately and with a reasonable amount of forethought.
> You are asking to throw those permission checks out. The answer is no.
> Furthermore there is a much clearer reason not to go with a list of all
> namespaces. A list of all namespaces breaks CRIU. As you have described
> it the list will change depending upon which machine you restore a
> checkpoint on. I honestly don't know what kind of havoc that will cause
> but it is certainly something we won't be able to checkpoint no matter
> how hard we try.

It's right. I hadn't thought about this.

> A global list of namespaces especially of the kind that you can open
> and get a handle to the namespace is just not appropriate.
> I know inode numbers comes darn close to names but they aren't really
> names and if it comes to it we can figure out how to preserve an
> applications view of it all across a checkpoint/restart. So far it
> hasn't proven necessary to preserve any inode numbers across
> checkpoint/restart but again it is theoretically possible if it becomes
> necessary.
> Throwing away checkpoint/restart support for the sake of
> checkpoint/restart is a no-go.
> Containers fundamentally imply you don't have global visibility,
> and that is a good thing.

All these thoughts about security make me thinking that kcmp is what we
should use here. It's maybe something like this:

kcmp(pid1, pid2, KCMP_NS_USERNS, fd1, fd2)

- to check if userns of the fd1 namepsace is equal to the fd2 userns

kcmp(pid1, pid2, KCMP_NS_PARENT, fd1, fd2)

- to check if a parent namespace of the fd1 pidns is equal to fd pidns.

fd1 and fd2 is file descriptors to namespace files.

So if we want to build a hierarchy, we need to collect all namespaces
and then enumerate them to check dependencies with help of kcmp.

> Eric