On Wed, Jan 3, 2024 at 8:52 AM Martin KaFai Lau [off-list ref] wrote:

On 1/2/24 10:11 AM, Stanislav Fomichev wrote:

quoted

On 12/29, Menglong Dong wrote:

quoted

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.

There is bpf_csum_update(), does it work?
A helper call should be acceptable comparing with the csum calculation itself.

Yeah, this helper works in this case! Now we miss the last
piece for the tx path: ip_summed. We need to know if it is
CHECKSUM_PARTIAL to decide if we should update the
csum in the packet. In the tx path, the csum in the L4 is the
pseudo header only if skb->ip_summed is CHECKSUM_PARTIAL.

Maybe we can introduce a lightweight kfunc to get its
value? Such as bpf_skb_csum_mode(). As we need only call
it once, there shouldn't be overhead on it.

Thanks!
Menglong Dong

quoted

quoted

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

+1. Please no new __sk_buff extension and no new conversion in
bpf_convert_ctx_access().

quoted

that :-) Should we add a lightweight kfunc to reset the fields instead?
Or will it still have an unacceptable overhead?