Re: [PATCHv14 bpf-next 1/6] bpf: run devmap xdp_prog on flush instead of bulk enqueue

On Mon, 18 Jan 2021 18:07:17 +0800
Hangbin Liu [off-list ref] wrote:

On Sun, Jan 17, 2021 at 02:57:02PM -0800, John Fastabend wrote:
[...]

quoted

It looks like we could embed xdp_buff in xdp_frame and then keep the metadata
at the end.

Because you are working performance here wdyt? <- @Jesper as well.  

Leave this question to Jesper.

The struct xdp_buff is larger than struct xdp_frame.  The size of
xdp_frame matters. It is a reserved areas in top of the frame.
An XDP BPF-program cannot access this area (and limit headroom grow).
This is why this code works, as afterwards xdp_frame is still valid.
Looking at the code xdp_update_frame_from_buff() we do seem to update
more fields than actually needed.

quoted

quoted

 
-	sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
+	if (unlikely(bq->xdp_prog)) {  

Whats the rational for making above unlikely()? Seems for users its not
unlikely. Can you measure a performance increase/decrease here? I think
its probably fine to just let compiler/prefetcher do its thing here. Or
I'm not reading this right, but seems users of bq->xdp_prog would disagree
on unlikely case?

Either way a comment might be nice to give us some insight in 6 months
why we decided this is unlikely.  

I agree that there is no need to use unlikely() here.

I added the unlikely() to preserve the baseline performance when not
having the 2nd prog loaded.  But I'm fine with removing that.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer