Re: [PATCH bpf-next v2 3/4] bpf: Add support for writing to nf_conn:mark

[Daniel Xu]

Hi Toke,

On Thu, Aug 18, 2022 at 09:52:08PM +0200, Toke Høiland-Jørgensen wrote:
> Daniel Xu <dxu@xxxxxxxxx> writes:
> 
> > Support direct writes to nf_conn:mark from TC and XDP prog types. This
> > is useful when applications want to store per-connection metadata. This
> > is also particularly useful for applications that run both bpf and
> > iptables/nftables because the latter can trivially access this
> > metadata.
> 
> Looking closer at the nf_conn definition, the mark field (and possibly
> secmark) seems to be the only field that is likely to be feasible to
> support direct writes to, as everything else either requires special
> handling (like status and timeout), or they are composite field that
> will require helpers anyway to use correctly.
> 
> Which means we're in the process of creating an API where users have to
> call helpers to fill in all fields *except* this one field that happens
> to be directly writable. That seems like a really confusing and
> inconsistent API, so IMO it strengthens the case for just making a
> helper for this field as well, even though it adds a bit of overhead
> (and then solving the overhead issue in a more generic way such as by
> supporting clever inlining).
> 
> -Toke

I don't particularly have a strong opinion here. But to play devil's
advocate:

* It may be confusing now, but over time I expect to see more direct
  write support via BTF, especially b/c there is support for unstable
  helpers now. So perhaps in the future it will seem more sensible.

* The unstable helpers do not have external documentation. Nor should
  they in my opinion as their unstableness + stale docs may lead to
  undesirable outcomes. So users of the unstable API already have to
  splunk through kernel code and/or selftests to figure out how to wield
  the APIs. All this to say there may not be an argument for
  discoverability.

* Direct writes are slightly more ergnomic than using a helper.

Thanks,
Daniel