On Sun, 13 Sep 2026 06:31:51 +0300 Anton Danilov wrote: > If such a test is wanted at all, I would rather come back to it > separately, once the conversion has landed, and in a smaller form: > > - perf record -e skb:kfree_skb and perf script, the way > openvswitch.sh already does it, instead of driving a trace instance > by hand;
IIRC, FWIW the perf invocations are quite painful for the CI. They require a lot of DRAM and IO with debug kernels for some reason. > - only the cases that are not evident from reading the code: a peer > that reboots and restarts its sequence number, two ends configured > with different options, plus a control case where nothing is > dropped; > - every case tied to the function the drop has to come from, so that > the same reason reported by unrelated code cannot make it pass. > > Is that worth having, or do you consider drop reason coverage out of > scope for selftests/net? I personally have never seen a production issue related to this stuff at Meta. So I don't see a point adding selftests. Maybe someone else had seen regressions and can guide on what shape the solution should take.
