From: Lorenz Bauer lmb@isovalent.com Date: Wed, 21 Jun 2023 09:01:15 +0100
On Tue, Jun 20, 2023 at 7:31 PM Kuniyuki Iwashima kuniyu@amazon.com wrote:
Good point. This is based on an assumption that all SO_REUSEPORT sockets have the same score, which is wrong for two corner cases if reuseport_has_conns() == true :
SO_INCOMING_CPU is set -> selected sk might have +1 score
BPF prog returns ESTABLISHED and/or SO_INCOMING_CPU sk -> selected sk will have more than 8
Using the old score could trigger more lookups depending on the order that sockets are created.
So the result will still be correct, but it's less performant? Happy to fix a perf regression, but if the result is incorrect this might need a separate fix?
Right, the result is always correct.
If BPF prog selects a different socket per lookup, there is no consistency, but it _is_ corret.
I did some more digging. I think this was introduced by commit efc6b6f6c311 ("udp: Improve load balancing for SO_REUSEPORT.") which unfortunately ran into a merge conflict. That resulted in Dave Miller moving the bug around in commit a57066b1a019 ("Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net"). Can you take a look and let me know if you think that is correct?
Yes, I should have updated the score too in efc6b6f6c311 to save unneeded lookups. The conflict itself was resolved properly.