[PATCH] nsproxy: attach to namespaces via pidfds

[Jann Horn]

On Mon, Apr 27, 2020 at 8:15 PM Christian Brauner
<christian.brauner at ubuntu.com> wrote:
> On Mon, Apr 27, 2020 at 07:28:56PM +0200, Jann Horn wrote:
> > On Mon, Apr 27, 2020 at 4:47 PM Christian Brauner
> > <christian.brauner at ubuntu.com> wrote:
[...]
> > > That means
> > > setns(nsfd, CLONE_NEWNET) equals setns(pidfd, CLONE_NEWNET). However,
> > > when a pidfd is passed, multiple namespace flags can be specified in the
> > > second setns() argument and setns() will attach the caller to all the
> > > specified namespaces all at once or to none of them. If 0 is specified
> > > together with a pidfd then setns() will interpret it the same way 0 is
> > > interpreted together with a nsfd argument, i.e. attach to any/all
> > > namespaces.
> > [...]
> > > Apart from significiantly reducing the number of syscalls from double
> > > digit to single digit which is a decent reason post-spectre/meltdown
> > > this also allows to switch to a set of namespaces atomically, i.e.
> > > either attaching to all the specified namespaces succeeds or we fail.
> >
> > Apart from the issues I've pointed out below, I think it's worth
> > calling out explicitly that with the current design, the switch will
> > not, in fact, be fully atomic - the process will temporarily be in
> > intermediate stages where the switches to some namespaces have
> > completed while the switches to other namespaces are still pending;
> > and while there will be less of these intermediate stages than before,
> > it also means that they will be less explicit to userspace.
>
> Right, that can be fixed by switching to the unshare model of getting a
> new set of credentials and committing it after the nsproxy has been
> installed? Then there shouldn't be an intermediate state anymore or
> rather an intermediate stage where we can still fail somehow.

It still wouldn't be atomic (in the sense of parallelism, not in the
sense of intermediate error handling) though; for example, if task B
does setns(<pidfd_of_task_a>, 0) and task C concurrently does
setns(<pidfd_of_task_b>, 0), then task C may end up with the new mount
namespace of task B but the old user namespace, or something like
that. If C is more privileged than B, that may cause C to have more
privileges through its configuration of namespaces than B does (e.g.
by running in the &init_user_ns but with a mount namespace owned by an
unprivileged user), which C may not expect. Same thing for racing
between unshare() and setns().

[...]
> > > +               put_user_ns(user_ns);
> > > +       }
> > > +#else
> > > +       if (flags & CLONE_NEWUSER)
> > > +               ret = -EINVAL;
> > > +#endif
> > > +
> > > +       if (!ret && wants_ns(flags, CLONE_NEWNS))
> > > +               ret = __ns_install(nsproxy, mnt_ns_to_common(nsp->mnt_ns));
> >
> > And this one might be even worse, because the mount namespace change
> > itself is only stored in the nsproxy at this point, but the cwd and
> > root paths have already been overwritten on the task's fs_struct.
> >
> > To actually make sys_set_ns() atomic, I think you'd need some
> > moderately complicated prep work, splitting the ->install handlers up
> > into prep work and a commit phase that can't fail.
>
> Wouldn't it be sufficient to move to an unshare like model, i.e.
> creating a new set of creds, and passing the new user_ns to
> create_new_namespaces() as well as having a temporary new_fs struct?
> That should get rid of all intermediate stages.

Ah, good point, I didn't realize that that already exists for unshare().