Re: performance bug in virtio net xdp

[Jesper Dangaard Brouer]

On Wed, 6 May 2020 04:08:27 -0400
"Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:

> So for mergeable bufs, we use ewma machinery to guess the correct buffer
> size. If we don't guess correctly, XDP has to do aggressive copies.
> 
> Problem is, xdp paths do not update the ewma at all, except
> sometimes with XDP_PASS. So whatever we happen to have
> before we attach XDP, will mostly stay around.
> 
> The fix is probably to update ewma unconditionally.

I personally find the code hard to follow, and (I admit) that it took
me some time to understand this code path (so I might still be wrong).

In patch[1] I tried to explain (my understanding):

  In receive_mergeable() the frame size is more dynamic. There are two
  basic cases: (1) buffer size is based on a exponentially weighted
  moving average (see DECLARE_EWMA) of packet length. Or (2) in case
  virtnet_get_headroom() have any headroom then buffer size is
  PAGE_SIZE. The ctx pointer is this time used for encoding two values;
  the buffer len "truesize" and headroom. In case (1) if the rx buffer
  size is underestimated, the packet will have been split over more
  buffers (num_buf info in virtio_net_hdr_mrg_rxbuf placed in top of
  buffer area). If that happens the XDP path does a xdp_linearize_page
  operation.


The EWMA code is not used when headroom is defined, which e.g. gets
enabled when running XDP.


[1] https://lore.kernel.org/netdev/158824572816.2172139.1358700000273697123.stgit@firesoul/
-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer