[vfio-users] Users of X99-boards, can you do me a small favor?

[Hristo Iliev]

Hi Martin,

Am 04.10.2016 10:09, schrieb Martin Schrodt:
> Hi Hristo,
> 
>> No need to sleep/wake - my X99-based system starts with TSC disabled:
>> 
>> $ dmesg | grep TSC
>> [    0.000000] tsc: Fast TSC calibration using PIT
>> [    0.077986] TSC deadline timer enabled
>> [    0.203383] TSC synchronization [CPU#0 -> CPU#1]:
>> [    0.203384] Measured 974558547804462 cycles TSC warp between CPUs,
>> turning off TSC clock.
>> [    0.203388] tsc: Marking TSC unstable due to check_tsc_sync_source
>> failed
>> 
>> Consequently, tsc is not among the clock sources listed in
>> available_clocksource. KVM is not happy about that:
>> 
>> [16739.200656] kvm: SMP vm created on host with unstable TSC; guest 
>> TSC
>> will not be reliable
> 
> Ok, so an X99-board that behaves like this even on a fresh start.
> Interesting.
> 
>> But I haven't observed any instabilities of the Windows 10 guest, 
>> which
>> happily runs with 4 virtual CPUs (2 virtual hyperthreaded CPUs) bound 
>> to
>> two cores of my i7-5820K.
> 
> This really makes me think there's something else involved in this
> behaviour. Maybe the CPU configuration (I use "Skylake-Client") exposes
> TSC to the guest, so if you put that on, it'll use it?
> 
> Can you check what kind of virtual CPU you use?
> 

I use 'passthrough', otherwise the default 'Haswell-noTSX' that 
virt-manager picks results in NVIDIA GeForce Experience complaining 
about unrecognised CPU type, so the TSC gets pretty much exposed to the 
guest.

Just found this paste with a boot log from another system with the same 
motherboard, though with an older BIOS version and a Xeon E5-1620 v3 
CPU:

https://pastelink.net/fjm

It shows the same problem with unsynchronised TSCs. Could also be 
something specific to the LGA 2011v3 processors or to the EFI BIOS. I'll 
take a look in the BIOS options when time permits.

> Cheers,
> Martin

Cheers,
Hristo