Re: [PATCH] KVM: stats: add stats to detect if vcpu is currently halted

[Sean Christopherson <seanjc@google.com>]

On Fri, Aug 20, 2021, Jing Zhang wrote:
> Hi Sean,
> 
> On Thu, Aug 19, 2021 at 3:37 PM Sean Christopherson <seanjc@google.com> wrote:
> >
> > On Wed, Aug 18, 2021, Cannon Matthews wrote:
> > > Since a guest has explictly asked for a vcpu to HLT, this is "useful work on
> > > behalf of the guest" even though the thread is "blocked" from running.
> > >
> > > This allows answering questions like, are we spending too much time waiting
> > > on mutexes, or long running kernel routines rather than running the vcpu in
> > > guest mode, or did the guest explictly tell us to not doing anything.
> > >
> > > So I would suggest keeping the "halt" part of the counters' name, and remove
> > > the "blocked" part rather than the other way around. We explicitly do not
> > > want to include non-halt blockages in this.
> >
> > But this patch does include non-halt blockages, which is why I brought up the
> > technically-wrong naming.  Specifically, x86 reaches this path for any !RUNNABLE
> > vCPU state, e.g. if the vCPU is in WFS.  Non-x86 usage appears to mostly call
> > this for halt-like behavior, but PPC looks like it has at least one path that's
> > not halt-like.
> >
> > I doubt anyone actually cares if the stat is a misnomer in some cases, but at the
> > same time I think there's opportunity for clean up here.  E.g. halt polling if a
> > vCPU is in WFS or UNINITIALIZED is a waste of cycles.  Ditto for the calls to
> > kvm_arch_vcpu_blocking() and kvm_arch_vcpu_unblocking() when halt polling is
> > successful, e.g. arm64 puts and reloads the vgic, which I assume is a complete
> > waste of cycles if the vCPU doesn't actually block.  And kvm_arch_vcpu_block_finish()
> > can be dropped by moving the one line of code into s390, which can add its own
> > wrapper if necessary.
> >
> > So with a bit of massaging and a slight change in tracing behavior, I believe we
> > can isolate the actual wait/halt and avoid "halted" being technically-wrong, and
> > fix some inefficiencies at the same time.
> >
> > Jing, can you do a v2 of this patch and send it to me off-list?  With luck, my
> > idea will work and I can fold your patch in, and if not we can always post v2
> > standalone in a few weeks.

Circling back to this with fresh eyes, limiting the state to "halted" would be
wrong.  I still stand by my assertion that non-halt states such as WFS should not
go through halt polling, but the intent of the proposed stat is to differentiate
between not running a vCPU because of a guest action and not running a vCPU because
the host is not scheduling its task.

E.g. on x86, if a vCPU is put into WFS for an extended time, the vCPU will not be
run because of a guest action, not because of any host activity.  But again, WFS
has very different meaning than "halt", which was the basis for my original
objection to the "halt_block" terminology.

One option would be to invert the stat, e.g. vcpu->stat.runnable, but that has the
downside of needed to be set somewhere outside of kvm_vcpu_block/halt(), and it
would likely be difficult to come up with a name that isn't confusing, e.g. the
vCPU would show up as "runnable" when mp_state!=RUNNABLE and it's not actively
blocking.

Back to bikeshedding the "halted" name, what about "blocking" or "waiting"?  I.e.
switch from past tense to present tense to convey that the _vCPU_ is "actively"
blocking/waiting, as opposed to the vCPU being blocked by some host condition.