On 12/29, Menglong Dong wrote:

For now, we have to call some helpers when we need to update the csum,
such as bpf_l4_csum_replace, bpf_l3_csum_replace, etc. These helpers are
not inlined, which causes poor performance.

In fact, we can define our own csum update functions in BPF program
instead of bpf_l3_csum_replace, which is totally inlined and efficient.
However, we can't do this for bpf_l4_csum_replace for now, as we can't
update skb->csum, which can cause skb->csum invalid in the rx path with
CHECKSUM_COMPLETE mode.

What's more, we can't use the direct data access and have to use
skb_store_bytes() with the BPF_F_RECOMPUTE_CSUM flag in some case, such
as modifing the vni in the vxlan header and the underlay udp header has
no checksum.

In the first patch, we make skb->csum readable and writable, and we make
skb->ip_summed readable. For now, for tc only. With these 2 fields, we
don't need to call bpf helpers for csum update any more.

In the second patch, we add some testcases for the read/write testing for
skb->csum and skb->ip_summed.

If this series is acceptable, we can define the inlined functions for csum
update in libbpf in the next step.

One downside of exposing those as __sk_buff fields is that all this
skb internal csum stuff now becomes a UAPI. And I'm not sure we want
that :-) Should we add a lightweight kfunc to reset the fields instead?
Or will it still have an unacceptable overhead?