Subject: Re: PATCH] cpuidle: A new variant of the menu governor to boost IO
 performance
From: Peter Zijlstra <peterz@infradead.org>
To: Tejun Heo <tj@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>,
       Arjan van de Ven <arjan@infradead.org>, linux-kernel@vger.kernel.org,
       lenb@kernel.org, mingo@elte.hu, yanmin_zhang@linux.intel.com,
       torvalds@linux-foundation.org, jens.axboe@oracle.com,
       David Howells <dhowells@redhat.com>,
       Ivan Kokshaysky <ink@jurassic.park.msu.ru>
In-Reply-To: <4AADF8FE.1060208@kernel.org>
References: <20090911174019.1ed02737@infradead.org>
	 <20090914004301.105196cc.akpm@linux-foundation.org>
	 <4AADF8FE.1060208@kernel.org>
Content-Type: text/plain
Date: Mon, 14 Sep 2009 10:15:28 +0200
Message-Id: <1252916128.5699.41.camel@laptop>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2455
Lines: 61

On Mon, 2009-09-14 at 17:04 +0900, Tejun Heo wrote:
> Andrew Morton wrote:
> > This fails to compile in linux-next:
> > 
> > drivers/cpuidle/governors/menu-tng.o:(.discard+0x0): multiple definition of `__pcpu_unique_menu_devices'
> > drivers/cpuidle/governors/menu.o:(.discard+0x0): first defined here
> > 
> > because we have
> > 
> > static DEFINE_PER_CPU(struct menu_device, menu_devices);
> > 
> > in both menu.c and menu-tng.c.
> > 
> > Despite the `static', the percpu changes in
> > 
> > commit 7c756e6e19e71f0327760d8955f7077118ebb2b1
> > Author:     Tejun Heo <tj@kernel.org>
> > AuthorDate: Wed Jun 24 15:13:50 2009 +0900
> > Commit:     Tejun Heo <tj@kernel.org>
> > CommitDate: Wed Jun 24 15:13:50 2009 +0900
> > 
> >     percpu: implement optional weak percpu definitions
> > 
> > are emitting global symbol derived from `menu_devices' and they clash.
> > 
> > I'll rename menu_devices to fix that up, but we have a problem...
> 
> Unfortunately, this was the only way we could come up with to get
> alpha and s390 working with the new percpu allocator.  On other archs,
> the global definition isn't necessary but config option
> DEBUG_FORCE_WEAK_PER_CPU enables it so that alpha and s390 don't choke
> on generic code later.
> 
> The core of the problem is that the memory model used by gcc on those
> two archs assume that static symbols are reachable with small offset
> but percpu variables, static or not, always end up way away.  Making
> all percpu variables weak and using global symbols to enforce global
> uniqueness works around the problem but with the side effect you're
> seeing.
> 
> If someone has any better ideas, I would happy to remove the annoying
> restriction.

That sounds like a variable __attribute__((no-explicit-relocs)) should
be proposed to the GCC folks or something. Then we can at least have a
hope of lifting this restriction some time in the future.

(Alternatively we build the full alpha kernel with
--mno-explicit-relocs, but that would be sub-optimal as well)

Not sure if s390 has anything similar..

Alternatively we could file it as a bug, since the symbols are in a
custom .section, which is a strong indication we're going to play funny
games anyway.

--
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/