On Fri, Oct 18, 2019 at 12:43:49AM +0100, Will Deacon wrote: > Hi Linus, > > Please pull these arm64 fixes for -rc4. The main thing here is a > long-awaited workaround for a CPU erratum on ThunderX2 which we have > developed in conjunction with engineers from Cavium/Marvell. At the moment, > the workaround is unconditionally enabled for affected CPUs at runtime > but we may add a command-line option to disable it in future if performance > numbers show up indicating a significant cost for real workloads.
As the Cavium/Marvell engineer who was involved in this, I will note that I had suggested a patch providing a runtime override[1] while providing safe defaults. Marc's patchset adds a trap to hypervisor in the system call path when KPTI is enabled, and KPTI is generally enabled on stock VM images. So normal users will see some performance regression (e.g I see something in the range of 3-4% on guest kernel compile). As a policy, I don't agree with having errata workarounds that can be left to the discretion of the admin to be forced at compile time. Since most of these workarounds use run-time code patching with alternatives, there is no need to do this at all. But given that this is already merged and cc:ed to stable, I will see if I can come up with enough data to convince Will. JC [1] https://lore.kernel.org/linux-arm-kernel/20191011232031.ga29...@dc5-eodlnx05.marvell.com/T/
