On Fri, Aug 14, 2020 at 05:32:32PM +0200, Sedat Dilek wrote: > commit 94c7eb54c4b8e81618ec79f414fe1ca5767f9720 > "random32: add a tracepoint for prandom_u32()" > > ...I gave Willy's patches a try and used the Linux Test Project (LTP) > for testing.
Just FWIW today I could run several relevant tests with a 40 Gbps NIC at high connection rates and under SYN flood to stress SYN cookies. I couldn't post earlier due to a net outage but will post the results here. In short, what I'm seeing is very good. The only thing is that the noise collection as-is with the 4 longs takes a bit too much CPU (0.2% measured) but if keeping only one word we're back to tausworthe performance, while using siphash all along. The noise generation function is so small that we're wasting cycles calling it and renaming registers. I'll run one more test by inlining it and exporting the noise. So provided quite some cleanup now, I really think we're about to reach a solution which will satisfy everyone. More on this after I extract the results. Willy
