Re: [RFC PATCH v1 1/2] sched: unified sched_powersavings sysfs tunable

From: Jens Axboe <axboe@kernel.dk>
To: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>,
	Vincent Guittot <vincent.guittot@linaro.org>,
	Indan Zupancic <indan@nul.nu>,
	Youquan Song <youquan.song@intel.com>,
	Ingo Molnar <mingo@elte.hu>,
	Arjan van de Ven <arjan@linux.intel.com>,
	Suresh Siddha <suresh.b.siddha@intel.com>,
	Linux Kernel <linux-kernel@vger.kernel.org>
Subject: Re: [RFC PATCH v1 1/2] sched: unified sched_powersavings sysfs tunable
Date: Thu, 26 Jan 2012 12:26:36 +0100	[thread overview]
Message-ID: <4F21386C.5000809@kernel.dk> (raw)
In-Reply-To: <1327576091.2446.87.camel@twins>

On 01/26/2012 12:08 PM, Peter Zijlstra wrote:
> On Thu, 2012-01-26 at 11:42 +0100, Jens Axboe wrote:
>> On 01/25/2012 03:53 PM, Peter Zijlstra wrote:
> 
>>> Jens, what is this thing trying to do?
>>
>> The intent of the code
>> is to return the first CPU in the "group" that the passed in core/thread
>> belongs to. This is used to decide whether to perform a completion
>> locally, or to send it off to a different "group".
> 
> Would you perhaps have meant to identify some shared cache domain?
> 
> In the scheduler core code we have (for CONFIG_SMP):
> 
>   static int ttwu_share_cache(int this_cpu, int that_cpu);
> 
> which returns true if this and that share a cache and false otherwise.
> Would that suffice or do you need a slightly different form? That is, we
> should provide you with some API and avoid you having to poke around
> with CONFIG_SCHED* and topology bits methinks.

Yeah, I think that would suit my purpose nicely, in fact. What level of
cache sharing is being used here? The block code wanted a per-socket
type operation, but since it's a heuristic, perhaps the above is even
better (or equivelant, perhaps).

-- 
Jens Axboe