Re: [Xen-devel] [PATCH v2 13/48] xen/sched: add is_running indicator to struct sched_unit

[Juergen Gross <jgross@xxxxxxxx>]

On 11.09.19 17:06, Jan Beulich wrote:
> On 11.09.2019 15:44, Juergen Gross wrote:
> > On 04.09.19 17:06, Jan Beulich wrote:
> > > On 09.08.2019 16:57, Juergen Gross wrote:
> > > > --- a/xen/common/schedule.c
> > > > +++ b/xen/common/schedule.c
> > > > @@ -407,6 +407,8 @@ int sched_init_vcpu(struct vcpu *v, unsigned int processor)
> > > >        {
> > > >            get_sched_res(v->processor)->curr = unit;
> > > >            v->is_running = 1;
> > > > +        unit->is_running = 1;
> > > > +        unit->state_entry_time = NOW();
> > > >        }
> > > >        else
> > > >        {
> > > Are idle vCPU-s also going to get grouped into units (I'm sorry,
> > > I don't recall)? If so, just like further down I'd be putting
> > > under question the multiple writing of the field. I'd kind of
> > > expect the unit and all vCPU-s within it to get an identical
> > > state entry time stored.
> > When an idle vcpu is being created its cpu is in no cpupool yet
> > (a free cpu). Those cpus are always subject to cpu scheduling
> > (granularity 1). Only when cpus are put into a cpupool the
> > granularity is adjusted accordingly and the idle-vcpus are
> > possibly grouped in units (see patch 45).
> >
> > Regarding the state entry time: different vcpus of a unit can be
> > in different states (blocked/running, etc.), so their state entry
> > times will generally differ. A unit's state entry time will
> > reflect its scheduling events (being scheduled/descheduled).
> But this doesn't (to me) clarify whether the multiple writing
> (once per vCPU) is actually correct. Obviously whether this is
> merely cosmetic depends on what the consumers of this data are.
There is no multiple writing. This is a preparatory patch for being
able to remove the vcpu dependencies from the single schedulers.

Later when multiple vcpus per unit are introduced there will be a
clear distinction between vcpu switching (context switches) and
sched_unit switching (scheduling events). vcpu data will be adjusted
at context_switch time, while unit data will be adjusted only when
scheduling units.

> > > Also both here and further down I'm puzzled to see that the
> > > unit's field doesn't get set at the same place in code where
> > > the vCPU's field gets set.
> > Right. The states of a vcpu and the unit it is belonging to are
> > coupled, but not identical.
> And this is not (going to be) a problem?
With a proper distinction as explained above, no.


Juergen


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel