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:   ebiederm@xmission.com (Eric W. Biederman)
To:     Christian Brauner <christian.brauner@canonical.com>
Cc:     davem@davemloft.net, gregkh@linuxfoundation.org,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        avagin@virtuozzo.com, ktkhai@virtuozzo.com, serge@hallyn.com
References: <20180404194857.29375-1-christian.brauner@ubuntu.com>
        <20180404203048.GA21118@gmail.com>
Date:   Wed, 04 Apr 2018 17:38:02 -0500
In-Reply-To: <20180404203048.GA21118@gmail.com> (Christian Brauner's message
        of "Wed, 4 Apr 2018 22:30:49 +0200")
Message-ID: <871sfuha2d.fsf@xmission.com>
User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/25.1 (gnu/linux)
MIME-Version: 1.0
Content-Type: text/plain
Subject: Re: [PATCH net] netns: filter uevents correctly
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Christian Brauner <christian.brauner@canonical.com> writes:

> On Wed, Apr 04, 2018 at 09:48:57PM +0200, Christian Brauner wrote:
>> commit 07e98962fa77 ("kobject: Send hotplug events in all network namespaces")
>> 
>> enabled sending hotplug events into all network namespaces back in 2010.
>> Over time the set of uevents that get sent into all network namespaces has
>> shrunk. We have now reached the point where hotplug events for all devices
>> that carry a namespace tag are filtered according to that namespace.
>> 
>> Specifically, they are filtered whenever the namespace tag of the kobject
>> does not match the namespace tag of the netlink socket. One example are
>> network devices. Uevents for network devices only show up in the network
>> namespaces these devices are moved to or created in.
>> 
>> However, any uevent for a kobject that does not have a namespace tag
>> associated with it will not be filtered and we will *try* to broadcast it
>> into all network namespaces.
>> 
>> The original patchset was written in 2010 before user namespaces were a
>> thing. With the introduction of user namespaces sending out uevents became
>> partially isolated as they were filtered by user namespaces:
>> 
>> net/netlink/af_netlink.c:do_one_broadcast()
>> 
>> if (!net_eq(sock_net(sk), p->net)) {
>>         if (!(nlk->flags & NETLINK_F_LISTEN_ALL_NSID))
>>                 return;
>> 
>>         if (!peernet_has_id(sock_net(sk), p->net))
>>                 return;
>> 
>>         if (!file_ns_capable(sk->sk_socket->file, p->net->user_ns,
>>                              CAP_NET_BROADCAST))
>>         j       return;
>> }
>> 
>> The file_ns_capable() check will check whether the caller had
>> CAP_NET_BROADCAST at the time of opening the netlink socket in the user
>> namespace of interest. This check is fine in general but seems insufficient
>> to me when paired with uevents. The reason is that devices always belong to
>> the initial user namespace so uevents for kobjects that do not carry a
>> namespace tag should never be sent into another user namespace. This has
>> been the intention all along. But there's one case where this breaks,
>> namely if a new user namespace is created by root on the host and an
>> identity mapping is established between root on the host and root in the
>> new user namespace. Here's a reproducer:
>> 
>>  sudo unshare -U --map-root
>>  udevadm monitor -k
>>  # Now change to initial user namespace and e.g. do
>>  modprobe kvm
>>  # or
>>  rmmod kvm
>> 
>> will allow the non-initial user namespace to retrieve all uevents from the
>> host. This seems very anecdotal given that in the general case user
>> namespaces do not see any uevents and also can't really do anything useful
>> with them.
>> 
>> Additionally, it is now possible to send uevents from userspace. As such we
>> can let a sufficiently privileged (CAP_SYS_ADMIN in the owning user
>> namespace of the network namespace of the netlink socket) userspace process
>> make a decision what uevents should be sent.
>> 
>> This makes me think that we should simply ensure that uevents for kobjects
>> that do not carry a namespace tag are *always* filtered by user namespace
>> in kobj_bcast_filter(). Specifically:
>> - If the owning user namespace of the uevent socket is not init_user_ns the
>>   event will always be filtered.
>> - If the network namespace the uevent socket belongs to was created in the
>>   initial user namespace but was opened from a non-initial user namespace
>>   the event will be filtered as well.
>> Put another way, uevents for kobjects not carrying a namespace tag are now
>> always only sent to the initial user namespace. The regression potential
>> for this is near to non-existent since user namespaces can't really do
>> anything with interesting devices.
>> 
>> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
>
> That was supposed to be [PATCH net] not [PATCH net-next] which is
> obviously closed. Sorry about that.

This does not appear to be a fix.
This looks like feature work.
The motivation appears to be that looks wrong let's change it.

So let's please leave this for when net-next opens again so we can
have time to fully consider a change in semantics.

Thank you,
Eric