Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <cover.1604335819.git.gladkov.alexey@gmail.com> <2718f7b13189dfd159414efb68e3533552593140.1604335819.git.gladkov.alexey@gmail.com>
In-Reply-To: <2718f7b13189dfd159414efb68e3533552593140.1604335819.git.gladkov.alexey@gmail.com>
From:   Jann Horn <jannh@google.com>
Date:   Mon, 2 Nov 2020 18:10:06 +0100
Message-ID: <CAG48ez0zGoB4Pr_+nLKaycCgEUtUrAvLJ89JG1ZbcbjKChMcng@mail.gmail.com>
Subject: Re: [RFC PATCH v1 4/4] Allow to change the user namespace in which
 user rlimits are counted
To:     Alexey Gladkov <gladkov.alexey@gmail.com>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        Linux Containers <containers@lists.linux-foundation.org>,
        Kernel Hardening <kernel-hardening@lists.openwall.com>,
        Alexey Gladkov <legion@kernel.org>,
        "Eric W . Biederman" <ebiederm@xmission.com>,
        Kees Cook <keescook@chromium.org>,
        Christian Brauner <christian@brauner.io>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Nov 2, 2020 at 5:52 PM Alexey Gladkov <gladkov.alexey@gmail.com> wrote:
> Add a new prctl to change the user namespace in which the process
> counter is located. A pointer to the user namespace is in cred struct
> to be inherited by all child processes.
[...]
> +       case PR_SET_RLIMIT_USER_NAMESPACE:
> +               if (!capable(CAP_SYS_RESOURCE))
> +                       return -EPERM;
> +
> +               switch (arg2) {
> +               case PR_RLIMIT_BIND_GLOBAL_USERNS:
> +                       error = set_rlimit_ns(&init_user_ns);
> +                       break;
> +               case PR_RLIMIT_BIND_CURRENT_USERNS:
> +                       error = set_rlimit_ns(current_user_ns());
> +                       break;
> +               default:
> +                       error = -EINVAL;
> +               }
> +               break;

I don't see how this can work. capable() requires that
current_user_ns()==&init_user_ns, so you can't use this API to bind
rlimits to any other user namespace.

Fundamentally, if it requires CAP_SYS_RESOURCE, this probably can't be
done as an API that a process uses to change its own rlimit scope. In
that case I would implement this as part of clone3() instead of
prctl(). (Then init_user_ns can set it if the caller has
CAP_SYS_RESOURCE. If you want to have support for doing the same thing
with nested namespaces, you'd also need a flag that the first-level
clone3() can set on the namespace to say "further rlimit splitting
should be allowed".)

Or alternatively, we could say that CAP_SYS_RESOURCE doesn't matter,
and instead you're allowed to move the rlimit scope if your current
hard rlimit is INFINITY. That might make more sense? Maybe?