Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Wenbin Zeng <wenbin.zeng@gmail.com>
To:     bfields@fieldses.org, davem@davemloft.net, viro@zeniv.linux.org.uk
Cc:     jlayton@kernel.org, trond.myklebust@hammerspace.com,
        anna.schumaker@netapp.com, wenbinzeng@tencent.com,
        dsahern@gmail.com, nicolas.dichtel@6wind.com, willy@infradead.org,
        edumazet@google.com, jakub.kicinski@netronome.com,
        tyhicks@canonical.com, chuck.lever@oracle.com, neilb@suse.com,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        netdev@vger.kernel.org, linux-nfs@vger.kernel.org
Subject: [PATCH v3 0/3] auth_gss: netns refcount leaks when use-gss-proxy==1
Date:   Wed, 12 Jun 2019 20:09:27 +0800
Message-Id: <1560341370-24197-1-git-send-email-wenbinzeng@tencent.com>
In-Reply-To: <1556692945-3996-1-git-send-email-wenbinzeng@tencent.com>
References: <1556692945-3996-1-git-send-email-wenbinzeng@tencent.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

This patch series fixes an auth_gss bug that results in netns refcount
leaks when use-gss-proxy is set to 1.

The problem was found in privileged docker containers with gssproxy service
enabled and /proc/net/rpc/use-gss-proxy set to 1, the corresponding
struct net->count ends up at 2 after container gets killed, the consequence
is that the struct net cannot be freed.

It turns out that write_gssp() called gssp_rpc_create() to create a rpc
client, this increases net->count by 2; rpcsec_gss_exit_net() is supposed
to decrease net->count but it never gets called because its call-path is:
        net->count==0 -> cleanup_net -> ops_exit_list -> rpcsec_gss_exit_net
Before rpcsec_gss_exit_net() gets called, net->count cannot reach 0, this
is a deadlock situation.

To fix the problem, we must break the deadlock, rpcsec_gss_exit_net()
should move out of the put() path and find another chance to get called,
I think nsfs_evict() is a good place to go, when netns inode gets evicted
we call rpcsec_gss_exit_net() to free the rpc client, this requires a new
callback i.e. evict to be added in struct proc_ns_operations, and add
netns_evict() as one of netns_operations as well.

v1->v2:
 * in nsfs_evict(), move ->evict() in front of ->put()
v2->v3:
 * rpcsec_gss_evict_net() directly call gss_svc_shutdown_net() regardless
   if gssp_clnt is null, this is exactly same to what rpcsec_gss_exit_net()
   previously did

Wenbin Zeng (3):
  nsfs: add evict callback into struct proc_ns_operations
  netns: add netns_evict into netns_operations
  auth_gss: fix deadlock that blocks rpcsec_gss_exit_net when
    use-gss-proxy==1

 fs/nsfs.c                      |  2 ++
 include/linux/proc_ns.h        |  1 +
 include/net/net_namespace.h    |  1 +
 net/core/net_namespace.c       | 12 ++++++++++++
 net/sunrpc/auth_gss/auth_gss.c |  4 ++--
 5 files changed, 18 insertions(+), 2 deletions(-)

-- 
1.8.3.1