A 20 Year Old Chipset Workaround Has Been Hurting Modern AMD Linux Systems

AMD engineer K Prateek Nayak recently uncovered that a ~20 year old chipset workaround in the Linux kernel still being applied to modern AMD systems is responsible in some cases for hurting performance on modern Zen hardware. Fortunately, a fix is on the way for limiting that workaround to old systems and in turn helping with performance for modern systems.

Last week was a patch posted for the ACPI processor idle code to avoid an old chipset workaround on modern AMD Zen systems. Since ACPI support was added to the Linux kernel in 2002, there has been a "dummy wait op" to deal with some chipsets where STPCLK# doesn't get asserted in time. The dummy I/O read delays further instruction processing until the CPU is fully stopped. This was a problem with at least some AMD Athlon era systems with a VIA chipset... But not a problem with newer chipsets of roughly the past two decades.

A Linux kernel workaround for the past two decades intended for now-ancient chipsets is still being needlessly applied to modern AMD systems, which in turn can hurt performance in select workloads.

With this workaround still being applied to even modern AMD systems, K Prateek Nayak discovered:

Sampling certain workloads with IBS on AMD Zen3 system shows that a significant amount of time is spent in the dummy op, which incorrectly gets accounted as C-State residency. A large C-State residency value can prime the cpuidle governor to recommend a deeper C-State during the subsequent idle instances, starting a vicious cycle, leading to performance degradation on workloads that rapidly switch between busy and idle phases.

One such workload is tbench where a massive performance degradation can be observed during certain runs.

At least for Tbench, this long-time, unconditional workaround in the Linux kernel has been hurting AMD Ryzen / Threadripper / EPYC performance in select workloads:

This workaround hasn't affected modern Intel systems since those newer Intel platforms use the alternative MWAIT-based intel_idle driver code path instead.

The AMD patch evolved into this patch by Intel Linux engineer Dave Hansen. That patch to limit the "dummy wait" workaround to old systems is already queued into TIP's x86/urgent branch. With it going the route of "x86/urgent" and for fixing a overzealous workaround that isn't needed on modern hardware, it's likely this patch will be submitted this week still for the Linux 6.0 kernel rather than needing to wait until the next (v6.1) merge window.

Update: The fix has been merged for Linux 6.0! I will be having some benchmarks on Phoronix soon.