On Tue, Jun 14, 2022 at 8:31 AM Tatsuyuki Ishi <ishitatsuyuki@xxxxxxxxx> wrote: > > Hi Andrii, > Thanks for looking into this. > > > I'd like to get to this in next few days, but meanwhile have you tried > > to benchmark what are the implications of stricter memory ordering > > changes on libbpf side? Do you have example changes you were thinking > > about for libbpf side? I can try benchmarking it on my side as well. > > I don't have a benchmark yet. I'll try to prepare a benchmark when I > have time to do so. > > The proposed change to libbpf is simply to replace the two > smp_store_release with smp_store_mb. I just realized that the Linux > kernel memory model doesn't have direct mappings to seq_cst loads and > stores though, so this will lead to a redundant barrier on AArch64 > etc. Hey Tatsuyuki, Just to follow up on this. We do have a bunch of benchmarks in selftests/bpf/bench, so I did run after replacing smp_store_release() in libbpf code with atomic_store(SEQ_CST). Generally it didn't show significant performance differences, except for "Single-producer, back-to-back mode (rb-libbpf case)". You can run these benchmarks yourself from selftest/bpf with just benchs/run_bench_ringbufs.sh. But before we make any changes, can you please share a reproducer for this issue? And which architecture (x86-64? arm64?) did you manage to reproduce this issue on? > > Regards, > Tatsuyuki
