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

[Radim KrÄmÃÅ]

2016-10-25 12:48+0000, Wanpeng Li:
>> å 2016å10æ24æïäå11:03ïRadim KrÄmÃÅ <rkrcmar@xxxxxxxxxx> åéï
>> 
>> I have only compile-tested it, but it should optimize the switch and
>> also fix two bugs.  The first one is major.
>> This needs that the deadline clearing in [5/5] is fixed.
>> ---8<---
>> We must start the hrimer even if the expiration is already in the past,
>> otherwise the periodic timer would not rearm the hrtimer.
>> 
>> And computing next expiration of a period timer does not require current
>> time.  The period should be constant, so it is more precise to add the
>> period to the last expiration time.  This fixes a time difference
>> between hrtimer expiration and apic->lapic_timer.target_expiration.
>> 
>> Signed-off-by: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
>> ---
>> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
>> @@ -1481,19 +1482,6 @@ void kvm_lapic_switch_to_hv_timer(struct kvm_vcpu *vcpu)
>> 
>>    WARN_ON(apic->lapic_timer.hv_timer_in_use);
>> 
>> -    if (apic_lvtt_period(apic)) {
>> -        ktime_t remaining, now;
>> -        u64 tscl = rdtsc();
>> -
>> -        now = apic->lapic_timer.timer.base->get_time();
>> -        remaining = hrtimer_get_remaining(&apic->lapic_timer.timer);
>> -        if (ktime_to_ns(remaining) < 0)
>> -            remaining = ktime_set(0, 0);
>> -
>> -        apic->lapic_timer.tscdeadline = kvm_read_l1_tsc(apic->vcpu, tscl) +
>> -            nsec_to_cycles(apic->vcpu, ktime_to_ns(remaining));
>> -        apic->lapic_timer.target_expiration = ktime_add_ns(now, ktime_to_ns(remaining));
>> -    }
> 
> Why you remove this? 

We computed those values in advance_periodic_target_expiration() and can
re-use them here.

hrtimer expiration == apic->lapic_timer.target_expiration.  (Otherwise
we're doing something wrong.) If that holds then the code does

  target_expiration = now + (target_expiration - now)

Which can be optimized to 

  target_expiration = target_expiration

and to nothing.  The same principle holds for
apic->lapic_timer.tscdeadline as well.

In other words: It doesn't matter if the timer switches between hrtimer
and VMX deadline -- the target expiration is still the same.

This hunk only added imprecision, because kvm_read_l1_tsc() and
ktime_to_ns() were not using the same time for computation.