On Wed, Aug 23, 2023 at 1:59 AM Jesper Dangaard Brouer [off-list ref] wrote:

Loading an XDP bpf-prog on veth device driver results in a significant
performance degradation (for normal unrelated traffic) due to
veth_convert_skb_to_xdp_buff() in most cases fully reallocates an SKB and copy
data over, even when XDP prog does nothing (e.g. XDP_PASS).

This patchset reduce the cases that cause reallocation.
After patchset UDP and AF_XDP sending avoids reallocations.

This approach is a lot more elegant than registering two XDP memory
models and fiddling with the skb and XDP buffer. The tests conducted
in our environment show similar figures in terms of performance
improvements. For example, using pktgen (skb data buffer allocated by
kmalloc) with the following setup: pktgen -> veth1 -> veth0 (XDP_TX)
-> veth1 (XDP_PASS) gives an improvement of around 23%. Thanks!


Thanks,
Liang

Future work will investigate TCP.

---

Jesper Dangaard Brouer (4):
      veth: use same bpf_xdp_adjust_head check as generic-XDP
      veth: use generic-XDP functions when dealing with SKBs
      veth: lift skb_head_is_locked restriction for SKB based XDP
      veth: when XDP is loaded increase needed_headroom


 drivers/net/veth.c | 86 +++++++++++++++++++---------------------------
 net/core/dev.c     |  1 +
 net/core/filter.c  |  1 +
 3 files changed, 38 insertions(+), 50 deletions(-)

--
Jesper