On Mon, Aug 23, 2021 at 05:52:50PM -0400, Hans Montero wrote:

From: Hans Montero <redacted>

This patch set adds a BPF local storage optimization. The first patch adds the
feature, and the second patch extends the bpf selftests so that the feature is
tested.

We are running BPF programs for each egress packet and noticed that
bpf_sk_storage_get incurs a significant amount of cpu time. By inlining the
storage into struct sock and accessing that instead of performing a map lookup,
we expect to reduce overhead for our specific use-case. 

Looks like a hack to me. Please share the perf numbers and setup details.
I think there should be a different way to address performance concerns
without going into such hacks.

This also has a
side-effect of persisting the socket storage, which can be beneficial.

Without explicit opt-in such sharing will cause multiple bpf progs to corrupt
each other data.