Received-SPF: pass (google.com: domain of linux-kernel+bounces-149946-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <20240416144814.173185-1-aleksandr.mikhalitsyn@canonical.com> <32f56a2e-8142-4391-916a-65fe51a57933@ssi.bg>
In-Reply-To: <32f56a2e-8142-4391-916a-65fe51a57933@ssi.bg>
From: Aleksandr Mikhalitsyn <aleksandr.mikhalitsyn@canonical.com>
Date: Thu, 18 Apr 2024 13:05:19 +0200
Message-ID: <CAEivzxfDLSrHP2H1ou8rccGLOxk5tycZH1+VKt3X8S0QcXNxcA@mail.gmail.com>
Subject: Re: [PATCH net-next] ipvs: allow some sysctls in non-init user namespaces
To: Julian Anastasov <ja@ssi.bg>
Cc: horms@verge.net.au, netdev@vger.kernel.org, lvs-devel@vger.kernel.org, 
	netfilter-devel@vger.kernel.org, linux-kernel@vger.kernel.org, 
	=?UTF-8?Q?St=C3=A9phane_Graber?= <stgraber@stgraber.org>, 
	Christian Brauner <brauner@kernel.org>, Pablo Neira Ayuso <pablo@netfilter.org>, 
	Jozsef Kadlecsik <kadlec@netfilter.org>, Florian Westphal <fw@strlen.de>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Wed, Apr 17, 2024 at 3:02=E2=80=AFPM Julian Anastasov <ja@ssi.bg> wrote:
>
>
>         Hello,

Dear Julian,

>
> On Tue, 16 Apr 2024, Alexander Mikhalitsyn wrote:
>
> > Let's make all IPVS sysctls visible and RO even when
> > network namespace is owned by non-initial user namespace.
> >
> > Let's make a few sysctls to be writable:
> > - conntrack
> > - conn_reuse_mode
> > - expire_nodest_conn
> > - expire_quiescent_template
> >
> > I'm trying to be conservative with this to prevent
> > introducing any security issues in there. Maybe,
> > we can allow more sysctls to be writable, but let's
> > do this on-demand and when we see real use-case.
> >
> > This list of sysctls was chosen because I can't
> > see any security risks allowing them and also
> > Kubernetes uses [2] these specific sysctls.
> >
> > This patch is motivated by user request in the LXC
> > project [1].
> >
> > [1] https://github.com/lxc/lxc/issues/4278
> > [2] https://github.com/kubernetes/kubernetes/blob/b722d017a34b300a2284b=
890448e5a605f21d01e/pkg/proxy/ipvs/proxier.go#L103
> >
> > Cc: St=C3=A9phane Graber <stgraber@stgraber.org>
> > Cc: Christian Brauner <brauner@kernel.org>
> > Cc: Julian Anastasov <ja@ssi.bg>
> > Cc: Simon Horman <horms@verge.net.au>
> > Cc: Pablo Neira Ayuso <pablo@netfilter.org>
> > Cc: Jozsef Kadlecsik <kadlec@netfilter.org>
> > Cc: Florian Westphal <fw@strlen.de>
> > Signed-off-by: Alexander Mikhalitsyn <aleksandr.mikhalitsyn@canonical.c=
om>
> > ---
> >  net/netfilter/ipvs/ip_vs_ctl.c | 18 +++++++++++++++---
> >  1 file changed, 15 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/netfilter/ipvs/ip_vs_ctl.c b/net/netfilter/ipvs/ip_vs_=
ctl.c
> > index 143a341bbc0a..92a818c2f783 100644
> > --- a/net/netfilter/ipvs/ip_vs_ctl.c
> > +++ b/net/netfilter/ipvs/ip_vs_ctl.c
> > @@ -4285,10 +4285,22 @@ static int __net_init ip_vs_control_net_init_sy=
sctl(struct netns_ipvs *ipvs)
>
>         As the list of privileged vars is short I prefer
> to use a bool and to make only some vars read-only:
>
>         bool unpriv =3D false;
>
> >               if (tbl =3D=3D NULL)
> >                       return -ENOMEM;
> >
> > -             /* Don't export sysctls to unprivileged users */
> > +             /* Let's show all sysctls in non-init user namespace-owne=
d
> > +              * net namespaces, but make them read-only.
> > +              *
> > +              * Allow only a few specific sysctls to be writable.
> > +              */
> >               if (net->user_ns !=3D &init_user_ns) {
>
>         Here we should just set: unpriv =3D true;
>
> > -                     tbl[0].procname =3D NULL;
> > -                     ctl_table_size =3D 0;
> > +                     for (idx =3D 0; idx < ARRAY_SIZE(vs_vars); idx++)=
 {
> > +                             if (!tbl[idx].procname)
> > +                                     continue;
> > +
> > +                             if (!((strcmp(tbl[idx].procname, "conntra=
ck") =3D=3D 0) ||
> > +                                   (strcmp(tbl[idx].procname, "conn_re=
use_mode") =3D=3D 0) ||
> > +                                   (strcmp(tbl[idx].procname, "expire_=
nodest_conn") =3D=3D 0) ||
> > +                                   (strcmp(tbl[idx].procname, "expire_=
quiescent_template") =3D=3D 0)))
> > +                                     tbl[idx].mode =3D 0444;
> > +                     }
> >               }
> >       } else
> >               tbl =3D vs_vars;
>
>         And below at every place to use:
>
>         if (unpriv)
>                 tbl[idx].mode =3D 0444;
>
>         for the following 4 privileged sysctl vars:
>
> - sync_qlen_max:
>         - allocates messages in kernel context
>         - this needs better tunning in another patch
>
> - sync_sock_size:
>         - allocates messages in kernel context
>
> - run_estimation:
>         - for now, better init ns to decide if to use est stats
>
> - est_nice:
>         - for now, better init ns to decide the value
>
> - debug_level:
>         - already set to 0444
>
>         I.e. these vars allocate resources (mem, CPU) without
> proper control, so for now we will just copy them from init ns
> without allowing writing. And they are vars that are not tuned
> often. Also we do not know which netns is supposed to be the
> privileged one, some solutions move all devices out of init_net,
> so we can not decide where to use lower limits.

I agree. I have also decided to forbid "est_cpulist" for unprivileged users=
.

>
>         OTOH, "amemthresh" is not privileged but needs single READ_ONCE
> for sysctl_amemthresh in update_defense_level() due to the possible
> div by zero if we allow writing to anyone, eg.:
>
>         int amemthresh =3D max(READ_ONCE(ipvs->sysctl_amemthresh), 0);
>         ...
>         nomem =3D availmem < amemthresh;
>         ... use only amemthresh
>
>         All other vars can be writable.

Have fixed this and sent it as a separate patch! ;-)

Thanks a lot for such a quick review!

Kind regards,
Alex

>
> Regards
>
> --
> Julian Anastasov <ja@ssi.bg>