Re: [PATCH 6/5] KVM: x86: fix periodic lapic timer with hrtimers

[Radim KrÄmÃÅ]

2016-10-26 14:02+0800, Wanpeng Li:
> 2016-10-25 19:43 GMT+08:00 Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>:
>> 2016-10-25 07:39+0800, Wanpeng Li:
>>> 2016-10-24 23:27 GMT+08:00 Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>:
>>>> 2016-10-24 17:09+0200, Paolo Bonzini:
>>>>> On 24/10/2016 17:03, Radim KrÄmÃÅ wrote:
>>>> [...]
>>>>>
>>>>> Reviewed-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>>>>>
>>>>> Go ahead, squash it into 5/5 and commit to kvm/queue. :)
>>>>
>>>> Did that, thanks.
>>>>
>>>> Wanpeng, the code is now under your name so please check it and/or
>>>> complain.
>>>
>>> This patch 6/5 incurred regressions.
>>>
>>> - The latency of the periodic mode which is emulated by VMX preemption
>>> is almost the same as periodic mode which is emulated by hrtimer.
>>
>> Hm, what numbers are you getting?
> 
> The two fixes look good to me. However, the codes which you remove in
> kvm_lapic_switch_to_hv_timer() results in different numbers.

Which of those two results is closer to the expected duration of the
period?

> w/o remove    hlt average latency = 2398462
> w/ remove      hlt average latency = 2403845

Some increase is expected when removing the code, because
kvm_lapic_switch_to_hv_timer() decreased the period by mistake:
it called

  now = get_time()

first and then did

  remaining = target - get_time()  // = hrtimer_get_remaining()

but some time has passed in between calls of get_time(), let's call the
time that passed in between as "delta", so when the function later set
the new target,

  new_target = now + remaining  // = now + target - (now + delta)

the new_target was "delta" earlier.

5k cycles is a huge difference, though ...
You tested the original kvm_lapic_switch_to_hv_timer(), with fixed
advance_periodic_target_expiration()?

>> When I ran the test with the original series, then it actually had worse
> 
> Did you test this by running my kvm-unit-tests/apic_timer_latency.flat?

Yes, I used numbers from Intel(R) Xeon(R) CPU E5-2620 v3 @ 2.40GHz,
which had TSC calibrated to 2397.223 MHz, so the expected "average
latency" with with the default 0x100000 ns period was

  0x100000 * 2.397223 - 0x100000 = 1465094.5044479999

The expected value is pretty close to what I actually measured:

>> [...]
>> If I run the test with [6/5], it gets sane numbers:
>>
>>   hlt average latency   = 1465107
>>   pause average latency = 1465093