Re: [PATCH 4/5] lguest: use KVM hypercalls

[ebiederm-aS9lmoZGLiVWk0Htik3J/w@public.gmane.org (Eric W. Biederman)]

Herbert Xu <herbert-lOAM2aK0SrRLBo1qDEOMRrpzq4S04n8Q@public.gmane.org> writes:

> On Wed, Apr 15, 2009 at 07:18:44AM -0700, Eric W. Biederman wrote:
>>
>> So holding the reference only blocks us indefinitely in
>> netdev_wait_allrefs, blocking the network namespace exit, and holding
>> net_mutex indefinitely.
>
> OK that's a killer because process A in my previous scenario can
> kill the device itself and cause a dead-lock.
>
> So how about this? We replace the dev destructor with our own that
> doesn't immediately call free_netdev.  We only call free_netdev once
> all tun fd's attached to the device have been closed.
>
> This can be achieved by simply adding a counter to tun_struct.
> We'd also change the async detach path to set a marker instead
> of detaching.  That marker can then be checked in places like
> tun_get.

That sounds like it would work, and allow us to have big tun_struct.
Which is sounds simpler overall.

I still have the feeling that putting the socket in tun_file instead
of in tun_struct might be conceptually cleaner, but one big blob that
is allocated and destroyed together is certainly easier and a lot
less racy to deal with.

Eric