Leaking path or inconsistency LSM checking observed in fs/net

[Stephen Smalley]

On 09/25/2018 07:36 PM, TongZhang wrote:
> ocfs2 is using sock_create instead of sock_create_kern in kernel v4.18.5.
> 
> fs/ocfs2/cluster/tcp.c: 1636
> https://elixir.bootlin.com/linux/v4.18.5/source/fs/ocfs2/cluster/tcp.c#L1636
>> ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
> 
> fs/ocfs2/cluster/tcp.c: 2035
> https://elixir.bootlin.com/linux/v4.18.5/source/fs/ocfs2/cluster/tcp.c#L2035
>> ret = sock_create(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);

Yes, I think that's the real bug here.  Unless the socket is in fact 
exposed for direct use by userspace, it ought to be using 
sock_create_kern() or similar.  I would suggest that you verify that the 
socket is never returned to userspace, and then submit a patch switching 
the code to using sock_create_kern().

> 
> 
>> On Sep 25, 2018, at 2:44 PM, Stephen Smalley <sds at tycho.nsa.gov> wrote:
>>
>> On 09/25/2018 01:27 PM, Tong Zhang wrote:
>>> Kernel Version: 4.18.5
>>> Problem Description:
>>> We found several leaking path or inconsistency LSM design issue in fs/net.
>>> Currently we can only observe sock creation from kernel and all bind/listen/connect are not sent to LSM.
>>> So, we think that those net/socket related stuff should all go through LSM check and being audited
>>> even it is not a user thread or process.
>>> Here’s an example where we have a check:
>>> in fs/ocfs2/cluster/tcp.c:2035 o2net_open_listening_sock() a sock is created using sock_create(),
>>> where a LSM check security_socket_create is called(net/socket.c:1242)
>>> And where we don’t have a check
>>> fs/ocfs2/cluster/tcp.c:2052 bind
>>> fs/ocfs2/cluster/tcp.c:2059 listen
>>> fs/dlm/lowcomms.c:1264 bind
>>> fs/dlm/lowcomms.c:1278 listen
>>> fs/dlm/lowcomms.c:1354 listen
>>> several places that use kernel_bind/kernel_listen/kernel_connect
>>> net/socket.c:3231 kernel_bind
>>> net/socket.c:3237 kernel_listen
>>> net/socket.c:3286 kernel_connect
>>
>> That's intentional.  LSM isn't trying to mediate kernel-internal operations, and we do not want to apply permission checks against the credentials of the current userspace process for such operations.  ocfs2 should likely be using sock_create_kern.
>>
>