A Linux 6.15 Performance Regression Hits Modern AMD CPUs
To bisect the performance regressions I noticed on the Linux 6.15 Git kernel, I decided to bisect it off the code compilation regressions. It was a solid ~10% regression that would be quite noticeable in CI/CD environments, build farms, etc. Plus it was a clear-cut case and quickly running. Thankfully with dual AMD EPYC 9755 processors providing a combined 256 cores / 512 threads make for quick kernel builds for bisecting the regression.
After spending the Sunday doing speedy kernel builds on the AMD EPYC 9755 server, the problematic commit was found.
This commit introduced during the Linux 6.15 merge window last month to add SRSO_MSR_FIX to KVM as part of the AMD Speculative Return Stack Overflow (SRSO) mitigation was the culprit. But... I was doing all of my benchmarks bare-metal and not even running any virtual machines (VMs) or similar.