Re: [PATCH net-next V2 3/3] tun: add eBPF based queue selection method

[Jason Wang]

On 2017å11æ02æ 03:12, Alexei Starovoitov wrote:
> On Wed, Nov 01, 2017 at 03:59:48PM +0200, Michael S. Tsirkin wrote:
> > On Wed, Nov 01, 2017 at 09:02:03PM +0800, Jason Wang wrote:
> > > On 2017å11æ01æ 00:45, Michael S. Tsirkin wrote:
> > > > > +static void __tun_set_steering_ebpf(struct tun_struct *tun,
> > > > > +				    struct bpf_prog *new)
> > > > > +{
> > > > > +	struct bpf_prog *old;
> > > > > +
> > > > > +	old = rtnl_dereference(tun->steering_prog);
> > > > > +	rcu_assign_pointer(tun->steering_prog, new);
> > > > > +
> > > > > +	if (old) {
> > > > > +		synchronize_net();
> > > > > +		bpf_prog_destroy(old);
> > > > > +	}
> > > > > +}
> > > > > +
> > > > Is this really called under rtnl?
> > > Yes it is __tun_chr_ioctl() will call rtnl_lock().
> > Is the call from tun_free_netdev under rtnl too?
> >
> > > > If no then rtnl_dereference
> > > > is wrong. If yes I'm not sure you can call synchronize_net
> > > > under rtnl.
> > > Are you worrying about the long wait? Looking at synchronize_net(), it does:
> > >
> > > void synchronize_net(void)
> > > {
> > >  ÂÂÂ might_sleep();
> > >  ÂÂÂ if (rtnl_is_locked())
> > >  ÂÂÂ ÂÂÂ synchronize_rcu_expedited();
> > >  ÂÂÂ else
> > >  ÂÂÂ ÂÂÂ synchronize_rcu();
> > > }
> > > EXPORT_SYMBOL(synchronize_net);
> > >
> > > Thanks
> > Not the wait - expedited is not a good thing to allow unpriveledged
> > userspace to do, it interrupts all VMs running on the same box.
> >
> > We could use a callback though the docs warn userspace can use that
> > to cause a DOS and needs to be limited.
> the whole __tun_set_steering_ebpf() looks odd to me.
> There is tun_attach_filter/tun_detach_filter pattern
> that works for classic BPF. Why for eBPF this strange
> synchronize_net() is there?

I'm not sure I get the question. eBPF here is used to do queue 
selection, so we could not reuse socket filter (tun_detach_filter use 
call_rcu()). cBPF could be used here, but I'm not quite sure it's worth 
to support it. And I agree we should use call_rcu() here.

Hope this answer your question.

Thanks