Re: [RFC PATCH 3/3] idle: store the idle state index in the struct rq

[Lorenzo Pieralisi]

On Thu, Jan 30, 2014 at 09:02:15PM +0000, Nicolas Pitre wrote:
> On Thu, 30 Jan 2014, Lorenzo Pieralisi wrote:
> 
> > On Thu, Jan 30, 2014 at 05:25:27PM +0000, Daniel Lezcano wrote:
> > > On 01/30/2014 05:35 PM, Peter Zijlstra wrote:
> > > > On Thu, Jan 30, 2014 at 05:27:54PM +0100, Daniel Lezcano wrote:
> > > >> IIRC, Alex Shi sent a patchset to improve the choosing of the idlest cpu and
> > > >> the exit_latency was needed.
> > > >
> > > > Right. However if we have a 'natural' order in the state array the index
> > > > itself might often be sufficient to find the least idle state, in this
> > > > specific case the absolute exit latency doesn't matter, all we want is
> > > > the lowest one.
> > > 
> > > Indeed. It could be simple as that. I feel we may need more informations 
> > > in the future but comparing the indexes could be a nice simple and 
> > > efficient solution.
> > 
> > As long as we take into account that some states might require multiple
> > CPUs to be idle in order to be entered, fine by me. But we should
> > certainly avoid waking up a CPU in a cluster that is in eg C2 (all CPUs in
> > C2, so cluster in C2) when there are CPUs in C3 in other clusters with
> > some CPUs running in those clusters, because there C3 means "CPU in C3, not
> > cluster in C3". Overall what I am saying is that what you are doing
> > makes perfect sense but we have to take the above into account.
> > 
> > Some states have CPU and cluster (or we can call it package) components,
> > and that's true on ARM and other architectures too, to the best of my
> > knowledge.
> 
> The notion of cluster or package maps pretty naturally onto scheduling 
> domains.  And the search for an idle CPU to wake up should avoid a 
> scheduling domain with a load of zero (which is obviously a prerequisite 
> for a power save mode to be applied to the cluster level) if there exist 
> idle CPUs in another domain already which load is not zero (all other 
> considerations being equal).  Hence your concern would be addressed 
> without any particular issue even if the individual CPU idle state index 
> is not exactly in sync with reality because of other hardware related 
> constraints.

Yes, just wanted to mention that relying solely on the C-state index
is not enough and can lead to surprises, as long as we take that into
account, it is ok. Highest index does not mean idlest CPU in power
consumption terms, because of cluster/package shared components.

It is probably worth noticing that parameters like eg exit_latency
have a meaning that necessarily depends on other CPUs in a cluster/package
for a given C-state, same logic as to my reasoning above applies.

> The other solution consists in making the index dynamic.  That means 
> letting backend idle drivers change it i.e. when the last man in a 
> cluster goes idle it could update the index for all the other CPUs in 
> the cluster.  There is no locking needed as the scheduler is only 
> consuming this info, and the scheduler getting it wrong on rare 
> occasions is not a big deal either.  But that looks pretty ugly as at 
> least 2 levels of abstractions would be breached in this case.

Yes, that's ugly, and that's the reason why tracing cluster C-states
require external tools (or HW probing) like the one Daniel wrote to
analize the time spent in cluster states.

Overall, it is not a concern, it is something we should take into account,
that's why I mentioned that.

Thanks,
Lorenzo

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/