Windows 10 Will Banish Spectre Slowdowns With Google's Retpoline Patch (zdnet.com)

Microsoft is including Google's mitigation for the Spectre Variant 2 speculative execution side-channel attack in the next release of Windows 10, currently codenamed 19H1. ZDNet reports: Google developed a software-based mitigation for Spectre Variant 2 called Retpoline that constrains speculative execution behavior sufficiently to mitigate an attack. Google's testing found its fix had a negligible effect on performance. Retpoline was implemented by Linux distributions such as Red Hat and SUSE, as well as by Oracle for Oracle Linux 6 and 7. And now, as MSPoweruser spotted, Microsoft's kernel engineers have confirmed that Retpoline will be part of the next version of Windows 10, 19H1, which is due out next year. Google's Retpoline plus Microsoft's own kernel modifications have reduced the performance impact to "noise level", according to Mehmet Iyigun of Microsoft's Windows and Azure kernel team. "Yes, we have enabled Retpoline by default in our 19H1 flights along with what we call 'import optimization' to further reduce perf impact due to indirect calls in kernel-mode. Combined, these reduce the perf impact of Spectre v2 mitigations to noise-level for most scenarios," wrote Iyigun.

"The bad news is that Microsoft didn't include the Retpoline fix in the latest Windows 10 October 2018 Update Redstone 5, or RS5, release, even though, according to CrowdStrike researcher Alex Ionescu, it could have," reports ZDNet.

Comments

Retpoline does not "banish" slowdowns

[complete loony]

The retpoline hack is a deliberate stack smash, to execute an indirect jump that the CPU will not speculate. Since the CPU cannot speculate it, execution *must* be slower than code from before spectre was discovered. But it does mean you can turn off *really* slow CPU mitigations.

The real trick is avoiding the need for retpoline in the first place. Make sure that indirect jumps have shortcuts for commonly executed branches that aren't affected by Spectre.

BTW, I watched a great talk about spectre, for application developers, by a clang compiler engineer who was involved in the research on spectre.