Re: [EXT.MSG] Re: [PATCH bpf-next v3 0/4] bpf: Check recursion per task for preemptible programs

From: JP Kobryn (CrowdStrike)

Date: Fri Nov 20 2026 - 22:01:54 EST


On 11/19/26 3:22 AM, Mike Bazov wrote:
> Thanks for testing, JP.
>
>> fexit (marked) 61 135
>
> Marked programs take the same per-CPU counter as before, so I expect no
> change there. How many runs are behind these numbers, and how much do
> they vary from run to run?

That was 3 runs taken over a 30 min sapling window of a specific workload.
We changed from a timed window to the full duration of the workload and on
5 new runs, we got mixed results. In one run, we got the inverse where the
miss count was higher before the patch. But this is only in terms of marked
fexit. I think this is noise from system services. We're in the process of
building up a new harness for this issue so I expect by next rev we have
more consistency.