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

[Jason Wang]

On 2017å11æ08æ 14:43, Michael S. Tsirkin wrote:
> On Wed, Nov 08, 2017 at 02:28:53PM +0900, Jason Wang wrote:
> > On 2017å11æ04æ 08:56, Willem de Bruijn wrote:
> > > On Fri, Nov 3, 2017 at 5:56 PM, Willem de Bruijn
> > > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > > On Tue, Oct 31, 2017 at 7:32 PM, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > > > This patch introduces an eBPF based queue selection method based on
> > > > > the flow steering policy ops. Userspace could load an eBPF program
> > > > > through TUNSETSTEERINGEBPF. This gives much more flexibility compare
> > > > > to simple but hard coded policy in kernel.
> > > > >
> > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
> > > > > ---
> > > > > +static int tun_set_steering_ebpf(struct tun_struct *tun, void __user *data)
> > > > > +{
> > > > > +       struct bpf_prog *prog;
> > > > > +       u32 fd;
> > > > > +
> > > > > +       if (copy_from_user(&fd, data, sizeof(fd)))
> > > > > +               return -EFAULT;
> > > > > +
> > > > > +       prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_SOCKET_FILTER);
> > > > If the idea is to allow guests to pass BPF programs down to the host,
> > > > you may want to define a new program type that is more restrictive than
> > > > socket filter.
> > > >
> > > > The external functions allowed for socket filters (sk_filter_func_proto)
> > > > are relatively few (compared to, say, clsact), but may still leak host
> > > > information to a guest. More importantly, guest security considerations
> > > > limits how we can extend socket filters later.
> > > Unless the idea is for the hypervisor to prepared the BPF based on a
> > > limited set of well defined modes that the guest can configure. Then
> > > socket filters are fine, as the BPF is prepared by a regular host process.
> > Yes, I think the idea is to let qemu to build a BPF program now.
> >
> > Passing eBPF program from guest to host is interesting, but an obvious issue
> > is how to deal with the accessing of map.
> >
> > Thanks
> Fundamentally, I suspect the way to solve it is to allow
> the program to specify "should be offloaded to host".
>
> And then it would access the host map rather than the guest map.

This looks a big extension.

> Then add some control path API for guest to poke at the host map.

Actually, as Willem said, we can even forbid using map through a type, 
but this will lose lots of flexibility.

> It's not that there's anything special about the host map -
> it's just separate from the guest - so if we wanted to
> do something that can work on bare-metal we could -
> just do something like a namespace and put all host
> maps there. But I'm not sure it's worth the complexity.
>
> Cc Aaron who wanted to look at this.

Maybe the first step is to let classic BPF to be passed from guest and 
consider eBPF on top.

Thanks