Re: [PATCH net-next 0/2] virtio-net: re enable XDP_REDIRECT for mergeable buffer

[Jason Wang]

On 2018å03æ01æ 22:16, Jesper Dangaard Brouer wrote:
> On Thu, 1 Mar 2018 21:15:36 +0800
> Jason Wang <jasowang@xxxxxxxxxx> wrote:
>
> > On 2018å03æ01æ 18:35, Jesper Dangaard Brouer wrote:
> > > On Thu, 1 Mar 2018 17:23:37 +0800
> > > Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > >   
> > > > On 2018å03æ01æ 17:10, Jesper Dangaard Brouer wrote:
> > > > > On Thu,  1 Mar 2018 11:19:03 +0800
> > > > > Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > > >      
> > > > > > This series tries to re-enable XDP_REDIRECT for mergeable buffer which
> > > > > > was removed since commit 7324f5399b06 ("virtio_net: disable
> > > > > > XDP_REDIRECT in receive_mergeable() case"). Main concerns are:
> > > > > >
> > > > > > - not enough tailroom was reserved which breaks cpumap
> > > > > To address this at a more fundamental level, I would suggest that we/you
> > > > > instead extend XDP to know it's buffers "frame" size/end.  (The
> > > > > assumption use to be, xdp_buff->data_hard_start + PAGE_SIZE, but
> > > > > ixgbe+virtio_net broke that assumption).
> > > > >
> > > > > It should actually be fairly easy to implement:
> > > > >     * Simply extend xdp_buff with a "data_hard_end" pointer.
> > > > Right, and then cpumap can warn and drop packets with insufficient
> > > > tailroom.
> > > >
> > > > But it should be a patch on top of this I think.
> > > Hmmm, not really.  If we/you instead fix the issue of XDP doesn't know
> > > the end/size of the frame, then we don't need this mixed XDP
> > > generic/native code path mixing.
> > I know this but I'm still a little bit confused. According to the commit
> > log of 7324f5399b06 ("virtio_net: disable XDP_REDIRECT in
> > receive_mergeable() case"), you said:
> >
> > """
> >   ÂÂÂ The longer explaination is that receive_mergeable() tries to
> >   ÂÂÂ work-around and satisfy these XDP requiresments e.g. by having a
> >   ÂÂÂ function xdp_linearize_page() that allocates and memcpy RX buffers
> >    around (in case packet is scattered across multiple rx buffers). This
> >   ÂÂÂ does currently satisfy XDP_PASS, XDP_DROP and XDP_TX (but only because
> >   ÂÂÂ we have not implemented bpf_xdp_adjust_tail yet).
> > """
> >
> > So I consider the tailroom is a must for the (future) tail adjustment.
> That is true, BUT implementing the "data_hard_end" extension is a
> pre-requisite.  It will also be to catch the issue of too little
> tail-room if/when implementing bpf_xdp_adjust_tail().
>
> It is of-cause a "nice-to-have", to fix this virtio_net driver's
> receive_mergeable() call to have enough tail-room, but I don't see it
> as a solution to the fundamental problem.
>
>
> > > You could re-enable native redirect, and push the responsibility to
> > > cpumap for detecting this too-small frame "missing tailroom" (and avoid
> > > crashing...). (If we really want to support this, cpumap could fallback
> > > to dev_alloc_skb, and handle it gracefully).
> > >   
> > Right but it will be slower than build_skb().
> True, but bad argument in this context, as you are already using a
> similar function call napi_alloc_skb().  And it will be even slower to
> call generic-XDP code path.

Well, there's no generic skb implementation for cpumap redirection so I 
think we're talking about native XDP for cpumap, In this case, we won't 
even use napi_alloc_skb().

Thanks