Re: [PATCH v5 13/14] x86/tsc: Mark Secure TSC as reliable clocksource

[Nikunj A. Dadhania]

On 11/2/2023 6:08 PM, Kirill A. Shutemov wrote:
> On Thu, Nov 02, 2023 at 05:46:26PM +0530, Nikunj A. Dadhania wrote:
>> On 11/2/2023 5:37 PM, Nikunj A. Dadhania wrote:
>>> On 11/2/2023 4:03 PM, Kirill A. Shutemov wrote:
>>>> On Thu, Nov 02, 2023 at 11:23:34AM +0530, Nikunj A. Dadhania wrote:
>>>>> On 10/30/2023 10:48 PM, Dave Hansen wrote:
>>>>>> On 10/29/23 23:36, Nikunj A Dadhania wrote:
>>>>>> ...
>>>>>>> diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
>>>>>>> index 15f97c0abc9d..b0a8546d3703 100644
>>>>>>> --- a/arch/x86/kernel/tsc.c
>>>>>>> +++ b/arch/x86/kernel/tsc.c
>>>>>>> @@ -1241,7 +1241,7 @@ static void __init check_system_tsc_reliable(void)
>>>>>>>  			tsc_clocksource_reliable = 1;
>>>>>>>  	}
>>>>>>>  #endif
>>>>>>> -	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
>>>>>>> +	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE) || cc_platform_has(CC_ATTR_GUEST_SECURE_TSC))
>>>>>>>  		tsc_clocksource_reliable = 1;
>>>>>>
>>>>>> Why can't you just set X86_FEATURE_TSC_RELIABLE?
>>>>>
>>>>> Last time when I tried, I had removed my kvmclock changes and I had set
>>>>> the X86_FEATURE_TSC_RELIABLE similar to Kirill's patch[1], this did not
>>>>> select the SecureTSC.
>>>>>
>>>>> Let me try setting X86_FEATURE_TSC_RELIABLE and retaining my patch for
>>>>> skipping kvmclock.
>>>>
>>>> kvmclock lowers its rating if TSC is good enough:
>>>>
>>>> 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
>>>> 	    boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
>>>> 	    !check_tsc_unstable())
>>>> 		kvm_clock.rating = 299;
>>>>
>>>> Does your TSC meet the requirements?
>>>
>>> I have set TscInvariant (bit 8) in CPUID_8000_0007_edx and TSC is set as reliable.
>>>
>>> With this I see kvm_clock rating being lowered, but kvm-clock is still being picked as clock-source.
>>
>> Ah.. at later point TSC is picked up, is this expected ?
>>
>> [    2.564052] clocksource: Switched to clocksource kvm-clock
>> [    2.678136] clocksource: Switched to clocksource tsc
> 
> On bare metal I see switch from tsc-early to tsc. tsc-early rating is
> equal to kvmclock rating after it gets lowered.

For SNP guest with secure tsc enabled, kvm-clock and tsc-early both are at 299.
Initially, kvm-clock is selected as clocksource and when tsc with 300 rating is enqueued, 
clocksource then switches to tsc.

[    0.004231] clocksource: clocksource_enqueue: name kvm-clock rating 299
[...]
[    2.046319] clocksource: clocksource_enqueue: name tsc-early rating 299
[...]
[    3.399179] clocksource: Switched to clocksource kvm-clock
[...]
[    3.513652] clocksource: clocksource_enqueue: name tsc rating 300
[    3.517314] clocksource: Switched to clocksource tsc
 
> Maybe kvmclock rating has to be even lower after detecting sane TSC?

If I set kvmclock rating to 298, I do see exact behavior as you have seen on the bare-metal.

[    0.004520] clocksource: clocksource_enqueue: name kvm-clock rating 298
[...]
[    1.827422] clocksource: clocksource_enqueue: name tsc-early rating 299
[...]
[    3.485059] clocksource: Switched to clocksource tsc-early
[...]
[    3.623625] clocksource: clocksource_enqueue: name tsc rating 300
[    3.628954] clocksource: Switched to clocksource tsc

Regards
Nikunj