Date: Fri, 16 Jan 2009 22:59:46 +0100
From: Ingo Molnar <mingo@elte.hu>
To: Rusty Russell <rusty@rustcorp.com.au>
Cc: Herbert Xu <herbert@gondor.apana.org.au>, akpm@linux-foundation.org,
       tj@kernel.org, hpa@zytor.com, brgerst@gmail.com, ebiederm@xmission.com,
       cl@linux-foundation.org, travis@sgi.com, linux-kernel@vger.kernel.org,
       steiner@sgi.com, hugh@veritas.com
Subject: Re: [PATCH] percpu: add optimized generic percpu accessors
Message-ID: <20090116215946.GA20653@elte.hu>
References: <20090116001200.GA9137@gondor.apana.org.au> <200901170804.18622.rusty@rustcorp.com.au>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <200901170804.18622.rusty@rustcorp.com.au>
User-Agent: Mutt/1.5.18 (2008-05-17)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1811
Lines: 52


* Rusty Russell <rusty@rustcorp.com.au> wrote:

> On Friday 16 January 2009 10:42:00 Herbert Xu wrote:
> > Ingo Molnar <mingo@elte.hu> wrote:
> > >
> > >> Of course.  But do any architectures actually _need_ that for a single 
> > >> read?
> > > 
> > > not for a read i guess - but for the other ops like add/and/or/xor.
> > 
> > One of the things I'd like to see happen with this work is for
> > us to have a cheap per-cpu atomic counter that we can use for
> > SNMP stats.
> > 
> > If we can make the inc/add variants into a single instruction,
> > then it won't need to disable preemption or interrupts.
> > 
> > So if you could design the API such that we have a variant of
> > add/inc that automatically disables/enables preemption then we
> > can optimise that away on x86.
> 
> Yep, already on it.  It's called local_t; that's what it was originally 
> designed for.
> 
> Unfortunately, to use it efficiently, we need large per-cpu areas.

Do you mean constructs like:

        local_inc(&__get_cpu_var(var));

?

If yes then i think you are missing the point here.

Yes, local_t can be useful when something is in an object and we know only 
a local IRQ context can update it and we dont want to disable irqs or use 
heavy atomics.

But percpu_read()/write()/add()/sub() ops are about optimizing _percpu_ 
variables. local_t alone does not solve that problem - because to use 
local_t as a percpu variable you have to get to the address of that 
variable - and that alone is not well optimized.

Or do you propose some new API that would allow that?

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