Message-ID: <87r4mt4um7.fsf@xmission.com>
Date:	Thu, 13 Dec 2012 11:31:12 -0800
From:	ebiederm@...ssion.com (Eric W. Biederman)
To:	Tom Parkin <tparkin@...alix.com>
Cc:	netdev@...r.kernel.org
Subject: Re: Network namespace bugs in L2TP

Tom Parkin <tparkin@...alix.com> writes:

> On Wed, Dec 12, 2012 at 11:44:36AM -0800, Eric W. Biederman wrote:
>> Tom Parkin <tparkin@...alix.com> writes:
>> >  1. Why do we need to change the namespace of the socket created in
>> >     l2tp_tunnel_sock_create?  So far as I can tell, sock_create
>> >     defaults to the namespace of the calling process.  Is the issue
>> >     here that this code may run from a work queue or similar?
>> 
>> Something similar.  At the very least l2tp_tunnel_create which calls
>> l2tp_tunnel_sock_create gets called from netlink.  The network namespace
>> of a socket is not necessarily the same as the network namespace of the
>> process that uses that socket.
>> 
>> So since current is not necessarily the right network namespace we need
>> push the desired network namespace of the socket down into
>> l2tp_tunnel_sock_create and use that when creating the socket.
>
> Ah, I see.  I hadn't appreciated that a process might swap between
> namespaces.
>
> I think that raises a question in the case of the L2TP tunnel sockets,
> though.  Currently l2tp_tunnel_sock_create uses the namespace of the
> current process for the socket.  The alternative is to pass in the
> desired namespace from l2tp_tunnel_create -- and this makes sense, I
> think.
>
> However, when l2tp_tunnel_create is called from the netlink code, the
> namespace passed is that of the netlink socket.  At the risk of sounding
> silly, what's the benefit of using the netlink socket namespace over the
> process namespace in this case?

Using the netlink socket namespace ensure that if the netlink socket is
passed between processes the semantics of sending messages down the
netlink socket don't change.

There is another thread on netdev discussing another variant of this
right now.  For some cases it is just a waste of resources to have one
copy of a daemon per network namespace.  In which case a controlling
daemon will open one netlink socket per network namespace and send
commands down the appropriate socket for the network namespace the
daemon wishes to control.

>> >  2. You mentioned the need to keep track of sockets allocated within a
>> >     namespace in order to be able to clean them up when the namespace
>> >     is deleted.  Should we be keeping a list of sockets we create and
>> >     then destroying them in the namespace pernet_ops exit function?
>> 
>> I think the issue that I was referring to and certainly the issue I am
>> thinking about is the issue where normal sockets hold a reference to a
>> network namespace and keep the network namespace alive.  Today l2tp uses
>> sock_create when creating a socket, and as such I think it pins it
>> current network namespace.  So I believe we can effectively have a
>> reference counting loop with l2tp sockets pinning the network namespace
>> and the network namespace keeping the l2tp device alive which keeps the
>> l2tp socket alive.
>
> OK, so presumably the way this would usually work is that a process
> creates sockets, and when the process exits those sockets go away.
> When all the processes in the namespace have exited, the namespace
> can close because there are no sockets holding it open.  Is that
> right?
>
> If that's correct, then I suppose the issue with the L2TP tunnel socket
> for an unmanaged tunnel is that it isn't owned by a process, per-se.
> So there's no obvious way to get rid of it, apart from sending a
> netlink message to tell the kernel to tear it down.
>
> But that doesn't seem too unreasonable.  A user would have to take
> explicit action to create an L2TP tunnel socket, and it might seem
> reasonable for that socket to keep the namespace alive until the user
> explicitly tears it down again.

Sending a netlink message to tear down the socket is not unreasonable.

Having a reference counting loop such that it is possible to close all
other sockets and all other references to a network namespace and not
have the network namespace go away because the L2TP tunnel socket holds
a reference to the unreachable and unuusable network namespace is
unreasonable.

We handle this with arp and icmp control sockets by not creating a
reference count.  And having a pernet cleanup routing clean up those
sockets.  Assuming I am right about the reference counting loop being
possible this is something to look at.

>> I don't remeber the specifics of l2tp as it creates some sockets, and
>> has other sockets passed in, and as such has rules that are not at all
>> normal.
>
> Ack.  Sockets are created in the kernel code for "unmanaged" tunnels,
> which don't run the control protocol over the top -- they're just for
> data encapsulation/de-encapsulation.  "Managed" tunnels have a
> userspace process looking after all the L2TP configuration and
> control/keepalive protocol, and in this case the daemon handles the
> creation of the tunnel socket.

Eric

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists