Re: [PATCH bpf-next] bpf, tcx: Skip RCU synchronization when the entry was newly created

[Qi Wu <shijing.wuqi@linux.alibaba.com>]

On Sat, 14 Nov 2026 19:06:55 +0000, Alexei Starovoitov wrote:
> Lets do the same in bpf_mprog. A new bundle starts with
> get_completed_synchronize_rcu(). Then there is no need for the
> 'created' check. netkit needs no special case either.

Thanks. Recording the poll state in bpf_mprog_commit() and waiting in
bpf_mprog_peer() before the peer is overwritten makes sense, and it
covers the clsact-first and netkit cases the created check misses.
I'll implement it this way, test it and send v2.

> synchronize_net() is expedited under RTNL.
> Should tcx_entry_sync() and netkit_entry_sync() use it?

For individual detach, using synchronize_net() looks reasonable.

For bulk uninstall, I will also investigate batching the updates and
waiting once, as the synchronize_rcu_expedited() documentation recommends.
There are already synchronize_net() calls around the teardown loop, but
reusing one requires checking the notifier and ndo_uninit() ordering
before concluding that the per-device waits can be removed.

I will keep v2 focused on the bpf_mprog polling change and evaluate the
teardown changes separately.

> Way too long. That's 100+ lines of commit log for a two line change.
> Keep the problem, the idea and one table of numbers.

Will do.