Messages in this thread | | | Subject | Re: [PATCH 0/2] RFC: Precise TSC migration | From | Maxim Levitsky <> | Date | Thu, 03 Dec 2020 13:51:18 +0200 |
| |
On Tue, 2020-12-01 at 12:02 -0300, Marcelo Tosatti wrote: > On Tue, Dec 01, 2020 at 02:48:11PM +0100, Thomas Gleixner wrote: > > On Mon, Nov 30 2020 at 16:16, Marcelo Tosatti wrote: > > > > Besides, Linux guests don't sync the TSC via IA32_TSC write, > > > > but rather use IA32_TSC_ADJUST which currently doesn't participate > > > > in the tsc sync heruistics. > > > > > > Linux should not try to sync the TSC with IA32_TSC_ADJUST. It expects > > > the BIOS to boot with synced TSCs. > > > > That's wishful thinking. > > > > Reality is that BIOS tinkerers fail to get it right. TSC_ADJUST allows > > us to undo the wreckage they create. > > > > Thanks, > > > > tglx > > Have not seen any multicore Dell/HP systems require that. > > Anyway, for QEMU/KVM it should be synced (unless there is a bug > in the sync logic in the first place). >
I agree with that, and that is why I suggested to make the guest avoid TSC syncing when KVM is detected.
I don't mind how to implement this.
It can be either done with new CPUID bit, or always when KVM is detected, (or even when *any* hypervisor is detected)
I also don't mind if we only disable tsc sync logic or set X86_FEATURE_TSC_RELIABLE which will disable it and the clocksource watchdog.
Best regards, Maxim Levitsky
| |