domain setup in 2.6.9rc1-bk6

domain setup in 2.6.9rc1-bk6

[Andi Kleen]

Someone deleted arch/x86_64/kernel/domain.c in current BK 
and moved the SMT domain setup to generic code.

Problem is that I need a architecture specific hook to 
turn off SMT idle for AMD dual core CPUs, which otherwise
look like SMT siblings. I suppose I will do more tunings
in this area in the future anyways, so having it in the 
architecture would be better.

It seems a bit counterintuitive that we have such a configurable
scheduler, and then there is only a "one size fits all" way
to configure it.

Can part of this code please be moved back to the architecture? 

Thanks,

-Andi

Re: domain setup in 2.6.9rc1-bk6

[Anton Blanchard]

> Problem is that I need a architecture specific hook to 
> turn off SMT idle for AMD dual core CPUs, which otherwise
> look like SMT siblings. I suppose I will do more tunings
> in this area in the future anyways, so having it in the 
> architecture would be better.

From a ppc64 perspective I liked the merging of the domains code.

We could have the ability to override generic domain setup per
architecture but leave all the default stuff in there.

Anton

Re: domain setup in 2.6.9rc1-bk6

[Andi Kleen]

On Tue, Aug 31, 2004 at 08:09:24PM +1000, Anton Blanchard wrote:
> 
> > Problem is that I need a architecture specific hook to 
> > turn off SMT idle for AMD dual core CPUs, which otherwise
> > look like SMT siblings. I suppose I will do more tunings
> > in this area in the future anyways, so having it in the 
> > architecture would be better.
> 
> From a ppc64 perspective I liked the merging of the domains code.
> 
> We could have the ability to override generic domain setup per
> architecture but leave all the default stuff in there.

Currently all I want is to disable SD_SHARE_POWER when one flag 
is set. A simple way to do this would be appreciated.

Later overriding would be fine. A simple way to do this would
be to declare it weak.

-Andi

Re: domain setup in 2.6.9rc1-bk6

[Nick Piggin]

Andi Kleen wrote:
> On Tue, Aug 31, 2004 at 08:09:24PM +1000, Anton Blanchard wrote:
> 
>>>Problem is that I need a architecture specific hook to 
>>>turn off SMT idle for AMD dual core CPUs, which otherwise
>>>look like SMT siblings. I suppose I will do more tunings
>>>in this area in the future anyways, so having it in the 
>>>architecture would be better.
>>
>>From a ppc64 perspective I liked the merging of the domains code.
>>
>>We could have the ability to override generic domain setup per
>>architecture but leave all the default stuff in there.
> 
> 
> Currently all I want is to disable SD_SHARE_POWER when one flag 
> is set. A simple way to do this would be appreciated.
> 

You can define ARCH_HAS_SCHED_TUNE, and define your own
SD_SIBLING_INIT, SD_CPU_INIT, SD_NODE_INIT. Also, you just
need to define a cpu_sibling_map in order to use CONFIG_SCHED_SMT.

> Later overriding would be fine. A simple way to do this would
> be to declare it weak.
> 

Ingo preferred to do it this way. I can't say I blame him because
the setup code is not exactly something trivial that you want
duplicated everywhere (although it is now much improved).

It would be simple enough to just re-export it to the arch code,
however I guess I agree with tackling that when the need arises.

I think it may not be *too* difficult to even come up with a
decent API to handle more arbitrary setups now. Definitely let
me know when you run into a limitation, and I'll be interested in
fixing it.

Thanks
Nick

Re: domain setup in 2.6.9rc1-bk6

[Andi Kleen]

On Wed, 01 Sep 2004 10:16:16 +1000
Nick Piggin <nickpiggin@yahoo.com.au> wrote:

> > Currently all I want is to disable SD_SHARE_POWER when one flag 
> > is set. A simple way to do this would be appreciated.
> > 
> 
> You can define ARCH_HAS_SCHED_TUNE, and define your own
> SD_SIBLING_INIT, SD_CPU_INIT, SD_NODE_INIT. Also, you just

Hmm, it seems to be evaluated at run time. That may actually
work. 

Basically i just need a function call to determine the flags
for SD_SIBLING_INIT.


-Andi