[RFC PATCH] sched: select_idle_sibling macro optimize

[RFC PATCH] sched: select_idle_sibling macro optimize

[Alex Shi]

If the sd domain just has one group, then we must be caught the
	i == target later, and then goes to deeper level domain.
So just skip this domain checking to save some instructions.

Signed-off-by: Alex Shi <alex.shi@linaro.org>
---
 kernel/sched/fair.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index c7395d9..3265fbc 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4196,6 +4196,11 @@ static int select_idle_sibling(struct task_struct *p, int target)
 	sd = rcu_dereference(per_cpu(sd_llc, target));
 	for_each_lower_domain(sd) {
 		sg = sd->groups;
+
+		/* skip single group domain */
+		if (sg == sg->next)
+			continue;
+
 		do {
 			if (!cpumask_intersects(sched_group_cpus(sg),
 						tsk_cpus_allowed(p)))
-- 
1.8.1.2

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Alex Shi]

Add Mike Galbraith.

Any one like to give some comments?

On 01/15/2014 10:23 PM, Alex Shi wrote:
> If the sd domain just has one group, then we must be caught the
> 	i == target later, and then goes to deeper level domain.
> So just skip this domain checking to save some instructions.
> 
> Signed-off-by: Alex Shi <alex.shi@linaro.org>
> ---
>  kernel/sched/fair.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index c7395d9..3265fbc 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -4196,6 +4196,11 @@ static int select_idle_sibling(struct task_struct *p, int target)
>  	sd = rcu_dereference(per_cpu(sd_llc, target));
>  	for_each_lower_domain(sd) {
>  		sg = sd->groups;
> +
> +		/* skip single group domain */
> +		if (sg == sg->next)
> +			continue;
> +
>  		do {
>  			if (!cpumask_intersects(sched_group_cpus(sg),
>  						tsk_cpus_allowed(p)))
> 


-- 
Thanks
    Alex

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Mike Galbraith]

On Thu, 2014-01-16 at 21:13 +0800, Alex Shi wrote: 
> Add Mike Galbraith.
> 
> Any one like to give some comments?
> 
> On 01/15/2014 10:23 PM, Alex Shi wrote:
> > If the sd domain just has one group, then we must be caught the
> > 	i == target later, and then goes to deeper level domain.
> > So just skip this domain checking to save some instructions.
> > 
> > Signed-off-by: Alex Shi <alex.shi@linaro.org>
> > ---
> >  kernel/sched/fair.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index c7395d9..3265fbc 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -4196,6 +4196,11 @@ static int select_idle_sibling(struct task_struct *p, int target)
> >  	sd = rcu_dereference(per_cpu(sd_llc, target));
> >  	for_each_lower_domain(sd) {
> >  		sg = sd->groups;
> > +
> > +		/* skip single group domain */
> > +		if (sg == sg->next)
> > +			continue;

When is that gonna happen?

-Mike

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Alex Shi]

On 01/16/2014 09:52 PM, Mike Galbraith wrote:
> On Thu, 2014-01-16 at 21:13 +0800, Alex Shi wrote: 
>> Add Mike Galbraith.
>>
>> Any one like to give some comments?
>>
>> On 01/15/2014 10:23 PM, Alex Shi wrote:
>>> If the sd domain just has one group, then we must be caught the
>>> 	i == target later, and then goes to deeper level domain.
>>> So just skip this domain checking to save some instructions.
>>>
>>> Signed-off-by: Alex Shi <alex.shi@linaro.org>
>>> ---
>>>  kernel/sched/fair.c | 5 +++++
>>>  1 file changed, 5 insertions(+)
>>>
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index c7395d9..3265fbc 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -4196,6 +4196,11 @@ static int select_idle_sibling(struct task_struct *p, int target)
>>>  	sd = rcu_dereference(per_cpu(sd_llc, target));
>>>  	for_each_lower_domain(sd) {
>>>  		sg = sd->groups;
>>> +
>>> +		/* skip single group domain */
>>> +		if (sg == sg->next)
>>> +			continue;
> 
> When is that gonna happen?

I had seen this in a Intel platform, you may have both CPU domain and MC
domain layer, because the domain flag is different, then they can not be
merged. and then the CPU domain just has one group.
> 
> -Mike
> 
> 


-- 
Thanks
    Alex

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Michael wang]

On 01/17/2014 10:14 AM, Alex Shi wrote:
[snip]
>>
>> When is that gonna happen?
> 
> I had seen this in a Intel platform, you may have both CPU domain and MC
> domain layer, because the domain flag is different, then they can not be
> merged. and then the CPU domain just has one group.

CPU and MC are different domains, isn't it?

General should like:

CPU	cpu0				cpu1
MC	core0		core1		core0		core1
SMT	cpu 0	cpu 1	cpu 2	cpu 3	cpu 4	cpu 5	cpu 6	cpu 7

So for cpu0:

CPU	sg0:cpu0,1,2,3	sg1:cpu4,5,6,7
MC	sg0:cpu0,1	sg1:cpu2,3
SMT	sg0:cpu0	sg1:cpu1

If one domain only have one group, that's sounds really a weird topology...

Regards,
Michael Wang

>>
>> -Mike
>>
>>
> 
> 

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Mike Galbraith]

On Fri, 2014-01-17 at 10:14 +0800, Alex Shi wrote: 
> On 01/16/2014 09:52 PM, Mike Galbraith wrote:
> > On Thu, 2014-01-16 at 21:13 +0800, Alex Shi wrote: 
> >> Add Mike Galbraith.
> >>
> >> Any one like to give some comments?
> >>
> >> On 01/15/2014 10:23 PM, Alex Shi wrote:
> >>> If the sd domain just has one group, then we must be caught the
> >>> 	i == target later, and then goes to deeper level domain.
> >>> So just skip this domain checking to save some instructions.
> >>>
> >>> Signed-off-by: Alex Shi <alex.shi@linaro.org>
> >>> ---
> >>>  kernel/sched/fair.c | 5 +++++
> >>>  1 file changed, 5 insertions(+)
> >>>
> >>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>> index c7395d9..3265fbc 100644
> >>> --- a/kernel/sched/fair.c
> >>> +++ b/kernel/sched/fair.c
> >>> @@ -4196,6 +4196,11 @@ static int select_idle_sibling(struct task_struct *p, int target)
> >>>  	sd = rcu_dereference(per_cpu(sd_llc, target));
> >>>  	for_each_lower_domain(sd) {
> >>>  		sg = sd->groups;
> >>> +
> >>> +		/* skip single group domain */
> >>> +		if (sg == sg->next)
> >>> +			continue;
> > 
> > When is that gonna happen?
> 
> I had seen this in a Intel platform, you may have both CPU domain and MC
> domain layer, because the domain flag is different, then they can not be
> merged. and then the CPU domain just has one group.

But sd starts at MC.

-Mike

Re: [RFC PATCH] sched: select_idle_sibling macro optimize

[Alex Shi]

On 01/17/2014 11:36 AM, Mike Galbraith wrote:
>>>>> +
>>>>> > >>> +		/* skip single group domain */
>>>>> > >>> +		if (sg == sg->next)
>>>>> > >>> +			continue;
>>> > > 
>>> > > When is that gonna happen?
>> > 
>> > I had seen this in a Intel platform, you may have both CPU domain and MC
>> > domain layer, because the domain flag is different, then they can not be
>> > merged. and then the CPU domain just has one group.
> But sd starts at MC.

That's true. Sorry for noise.

-- 
Thanks
    Alex