On Wed, Nov 04, 2020 at 05:40:51AM -0600, YiFei Zhu wrote: > On Tue, Nov 3, 2020 at 6:29 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote: > > Yeah, this is very interesting. That there is anything measurably _slower_ > > with the cache is surprising. Though with only 4 runs, I wonder if it's > > still noisy? What happens at 10 runs -- more importantly what is the > > standard deviation? > > I could do that. it just takes such a long time. Each run takes about > 20 minutes so with 10 runs per environment, 3 environments (native + 2 > docker) per boot, and 4 boots (2 bootparam * 2 compile config), it's > 27 hours of compilation. I should probably script it at that point. Yeah, I was facing the same issues. Though perhaps hackbench (with multiple CPUs) would be a better test (and it's much faster): https://lore.kernel.org/lkml/7723ae8d-8333-ba17-6983-a45ec8b11c54@xxxxxxxxxx/ (I usually run this with a CNT of 20 to get quick results.) > > I assume this is from Indirect Branch Prediction Barrier (IBPB) and > > Single Threaded Indirect Branch Prediction (STIBP) (which get enabled > > for threads under seccomp by default). > > > > Try booting with "spectre_v2_user=prctl" > > Hmm, to make sure, boot with just "spectre_v2_user=prctl" on the > command line and test the performance of that? Right, see if that eliminates the 3 minute jump seen for seccomp. -- Kees Cook
