Re: [PATCH v2 3/7] KVM-HV: KVM Steal time implementation

[Glauber Costa <glommer@redhat.com>]

On 06/20/2011 05:56 PM, Marcelo Tosatti wrote:
> On Sun, Jun 19, 2011 at 12:57:53PM +0300, Avi Kivity wrote:
>> On 06/17/2011 01:20 AM, Glauber Costa wrote:
>>> To implement steal time, we need the hypervisor to pass the guest information
>>> about how much time was spent running other processes outside the VM.
>>> This is per-vcpu, and using the kvmclock structure for that is an abuse
>>> we decided not to make.
>>>
>>> In this patchset, I am introducing a new msr, KVM_MSR_STEAL_TIME, that
>>> holds the memory area address containing information about steal time
>>>
>>> This patch contains the hypervisor part for it. I am keeping it separate from
>>> the headers to facilitate backports to people who wants to backport the kernel
>>> part but not the hypervisor, or the other way around.
>>>
>>>
>>>
>>> +#define KVM_STEAL_ALIGNMENT_BITS 5
>>> +#define KVM_STEAL_VALID_BITS ((-1ULL<<   (KVM_STEAL_ALIGNMENT_BITS + 1)))
>>> +#define KVM_STEAL_RESERVED_MASK (((1<<   KVM_STEAL_ALIGNMENT_BITS) - 1 )<<   1)
>>
>> Clumsy, but okay.
>>
>>> +static void record_steal_time(struct kvm_vcpu *vcpu)
>>> +{
>>> +	u64 delta;
>>> +
>>> +	if (vcpu->arch.st.stime&&   vcpu->arch.st.this_time_out) {
>>
>> 0 is a valid value for stime.
>>
>>> +
>>> +		if (unlikely(kvm_read_guest(vcpu->kvm, vcpu->arch.st.stime,
>>> +			&vcpu->arch.st.steal, sizeof(struct kvm_steal_time)))) {
>>> +
>>> +			vcpu->arch.st.stime = 0;
>>> +			return;
>>> +		}
>>> +
>>> +		delta = (get_kernel_ns() - vcpu->arch.st.this_time_out);
>>> +
>>> +		vcpu->arch.st.steal.steal += delta;
>>> +		vcpu->arch.st.steal.version += 2;
>>> +
>>> +		if (unlikely(kvm_write_guest(vcpu->kvm, vcpu->arch.st.stime,
>>> +			&vcpu->arch.st.steal, sizeof(struct kvm_steal_time)))) {
>>> +
>>> +			vcpu->arch.st.stime = 0;
>>> +			return;
>>> +		}
>>> +	}
>>> +
>>> +}
>>> +
>>>
>>> @@ -2158,6 +2206,8 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>>>   			kvm_migrate_timers(vcpu);
>>>   		vcpu->cpu = cpu;
>>>   	}
>>> +
>>> +	record_steal_time(vcpu);
>>>   }
>>
>> This records time spent in userspace in the vcpu thread as steal
>> time.  Is this what we want?  Or just time preempted away?
>
> It also accounts halt time (kvm_vcpu_block) as steal time. Glauber, you
> could instead use the "runnable-state-but-waiting-in-runqueue" field of
> SCHEDSTATS, i forgot the exact name.
>
I thought about it in the past. I let the idea aside because I didn't 
want to introduce a dependency on SCHEDSTATS. But thinking about it 
again now (and after some days of experimentations with it), I think we 
could have both.

use run_delay (the field you were thinking of) when schedstats are 
available, and fallback to an estimate method like the one we're doing 
when it is not.

Objections ?