Re: [RFC 2/3] KVM: arm64: pmu: Fix chained SW_INCR counters

[Andrew Murray]

On Fri, Dec 06, 2019 at 03:35:03PM +0000, Marc Zyngier wrote:
> On 2019-12-06 15:21, Andrew Murray wrote:
> > On Thu, Dec 05, 2019 at 02:52:26PM +0000, Marc Zyngier wrote:
> > > On 2019-12-05 14:06, Auger Eric wrote:
> > > > Hi Marc,


> > > > >
> > > > > I think the whole function is a bit of a mess, and could be
> > > better
> > > > > structured to treat 64bit counters as a first class citizen.
> > > > >
> > > > > I'm suggesting something along those lines, which tries to
> > > > > streamline things a bit and keep the flow uniform between the
> > > > > two word sizes. IMHO, it helps reasonning about it and gives
> > > > > scope to the ARMv8.5 full 64bit counters... It is of course
> > > > > completely untested.
> > > >
> > > > Looks OK to me as well. One remark though, don't we need to test
> > > if the
> > > > n+1th reg is enabled before incrementing it?
> > 
> > Indeed - we don't want to indicate an overflow on a disabled counter.
> > 
> > 
> > > 
> > > Hmmm. I'm not sure. I think we should make sure that we don't flag
> > > a counter as being chained if the odd counter is disabled, rather
> > > than checking it here. As long as the odd counter is not chained
> > > *and* enabled, we shouldn't touch it.
> > 
> > Does this mean that we don't care if the low counter is enabled or not
> > when deciding if the pair is chained?
> > 
> > I would find the code easier to follow if we had an explicit 'is the
> > high counter enabled here' check (at the point of deciding where to
> > put the overflow).
> 
> Sure. But the point is that we're spreading that kind of checks all over
> the map, and that we don't have a way to even reason about the state of
> a 64bit counter. Doesn't it strike you as being mildly broken?
> 

Yup! To the point where I can't trust the function names and have to look
at what the code does...


> > > @@ -645,7 +647,8 @@ static void kvm_pmu_update_pmc_chained(struct
> > > kvm_vcpu
> > > *vcpu, u64 select_idx)
> > >  	struct kvm_pmu *pmu = &vcpu->arch.pmu;
> > >  	struct kvm_pmc *pmc = &pmu->pmc[select_idx];
> > > 
> > > -	if (kvm_pmu_idx_has_chain_evtype(vcpu, pmc->idx)) {
> > > +	if (kvm_pmu_idx_has_chain_evtype(vcpu, pmc->idx) &&
> > > +	    kvm_pmu_counter_is_enabled(vcpu, pmc->idx)) {
> > 
> > I.e. here we don't care what the state of enablement is for the low
> > counter.
> > 
> > Also at present, this may break the following use-case
> > 
> >  - User creates and uses a pair of chained counters
> >  - User disables odd/high counter
> >  - User reads values of both counters
> >  - User rewrites CHAIN event to odd/high counter OR user re-enables
> > just the even/low counter
> >  - User reads value of both counters <- this may now different to the
> > last read
> 
> Hey, I didn't say it was perfect ;-). But for sure we can't let the
> PMU bitrot more than it already has, and I'm not sure this is heading
> the right way.

I think we're aligned here. To me this code is becoming very fragile, difficult
for me to make sense of and is stretching the abstractions we've made. This is
why it is difficult to enhance it without breaking something. It's why I felt it
was safer to add 'an extra check' in the SWINCR than to risk touching something
that I didn't have the confidence I could be sure was correct. 


> 
> I'm certainly going to push back on new PMU features until we can properly
> reason about 64bit counters as a top-level entity (as opposed to a bunch
> of discrete counters).

Thanks,

Andrew Murray

> 
> Thanks,
> 
>         M.
> -- 
> Jazz is not dead. It just smells funny...