Hello,
On Mon, 8 Sep 2025, Slavin Liu wrote:
> When exiting a network namespace, in cleanup_net()->ops_undo_list(),
> ip_vs_ftp_ops->exit() is called before ip_vs_core_ops->exit_batch().
> The ip_vs_app ip_vs_ftp and its incarnations will be freed by
> unregister_ip_vs_app().
> However, there could still be connections bound to ip_vs_ftp's incarnation.
> cp->app points to the free'd incarnation, which will be accessed later by
> __ip_vs_cleanup_batch()->ip_vs_conn_net_cleanup()->ip_vs_conn_flush()->ip_vs_conn_del()->
> ip_vs_conn_expire()->ip_vs_unbind_app(), causing a uaf. This vulnarability can
> lead to a local privilege escalation.
>
> Reproduction steps:
> 1. create a ipvs service on (127.0.0.1:21)
> 2. create a ipvs destination on the service, to (127.0.0.1:<any>)
> 3. send a tcp packet to (127.0.0.1:21)
> 4. exit the network namespace
>
> I think the fix should flush all connection to ftp before unregistration.
> The simpler fix is to delete ip_vs_ftp_ops->exit, and defer the unregistration
> of ip_vs_ftp to ip_vs_app_net_cleanup(), which will unregister all ip_vs_app.
> It's after ip_vs_conn_net_cleanup() so there is no uaf issue. This patch
> seems to solve the issue but has't been fully tested yet, and is also not
> graceful.
>
> Signed-off-by: Slavin Liu <slavin...@gmail.com>
> ---
> net/netfilter/ipvs/ip_vs_ftp.c | 13 -------------
> 1 file changed, 13 deletions(-)
>
> diff --git a/net/netfilter/ipvs/ip_vs_ftp.c b/net/netfilter/ipvs/ip_vs_ftp.c
> index d8a284999544..68def1106681 100644
> --- a/net/netfilter/ipvs/ip_vs_ftp.c
> +++ b/net/netfilter/ipvs/ip_vs_ftp.c
> @@ -598,22 +598,9 @@ static int __net_init __ip_vs_ftp_init(struct net *net)
> unregister_ip_vs_app(ipvs, &ip_vs_ftp);
> return ret;
> }
> -/*
> - * netns exit
> - */
> -static void __ip_vs_ftp_exit(struct net *net)
> -{
> - struct netns_ipvs *ipvs = net_ipvs(net);
> -
> - if (!ipvs)
What if we change this 'if' check to:
if (!ipvs || !ipvs->enable)
If netns exits, the cleanup order is:
1. exit handlers for pernet device (&ipvs_core_dev_ops) where
ipvs->enable is set to 0
2. exit handlers for ip_vs_ftp_ops (as last pernet subsys)
By checking for ipvs->enable, we should not call
unregister_ip_vs_app() in __ip_vs_ftp_exit() because
there can be existing conns with valid cp->app
3. exit handlers for pernet subsys (&ipvs_core_ops) where
ip_vs_app_net_cleanup() unregisters all apps for netns.
Here the apps will be freed after all conns are gone
Why we should keep the ftp pernet subsys: because when
there are no conns using the app, the module can be removed and
it must unregister its app from all netns where ipvs->enable will
be 1.
But note that we have a pending patch that changes
the access to ipvs->enable to use READ_ONCE/WRITE_ONCE:
https://archive.linuxvirtualserver.org/html/lvs-devel/2025-09/msg00000.html
> - return;
> -
> - unregister_ip_vs_app(ipvs, &ip_vs_ftp);
> -}
>
> static struct pernet_operations ip_vs_ftp_ops = {
> .init = __ip_vs_ftp_init,
> - .exit = __ip_vs_ftp_exit,
> };
>
> static int __init ip_vs_ftp_init(void)
> --
> 2.34.1
Regards
--
Julian Anastasov <j...@ssi.bg>
