From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1758420AbXI2Mpp (ORCPT ); Sat, 29 Sep 2007 08:45:45 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1758149AbXI2MpO (ORCPT ); Sat, 29 Sep 2007 08:45:14 -0400 Received: from smtp102.mail.mud.yahoo.com ([209.191.85.212]:32908 "HELO smtp102.mail.mud.yahoo.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1758142AbXI2MpM (ORCPT ); Sat, 29 Sep 2007 08:45:12 -0400 DomainKey-Signature: a=rsa-sha1; q=dns; c=nofws; s=s1024; d=yahoo.com.au; h=Received:X-YMail-OSG:From:To:Subject:Date:User-Agent:Cc:References:In-Reply-To:MIME-Version:Content-Disposition:Content-Type:Content-Transfer-Encoding:Message-Id; b=o5f44/TPHq9a4iPGRHOHF8SXtg3rp4+GqWqV5R7aDe4OKXNUUZURG3RArv5wX429I4H24vx56CzZx3ZmZoIzG8kTMNCtqcUQNp9sndZqRBb5XhNGWGEWygu7NR9ZelcYXirOD/jCBOxZYUcAbRDrGSWq9dmhMfvdzyUh/GkvISM= ; X-YMail-OSG: lgE0V0UVM1liNxBtGvUEeD87qYAPF91ZHW5g4jBmL_CsmYhSW9d6hPvPuw3us1Ck8M9WxkxJUg-- From: Nick Piggin To: Peter Zijlstra Subject: Re: [RFC][PATCH] mm: couple rcu and memory reclaim Date: Sat, 29 Sep 2007 06:13:51 +1000 User-Agent: KMail/1.9.5 Cc: "Paul E. McKenney" , linux-kernel@vger.kernel.org, Rik van Riel References: <20070924104517.7d8126ae@twins> In-Reply-To: <20070924104517.7d8126ae@twins> MIME-Version: 1.0 Content-Disposition: inline Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Message-Id: <200709290613.51790.nickpiggin@yahoo.com.au> Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Monday 24 September 2007 18:45, Peter Zijlstra wrote: > Just an idea I had, it seems like a good idea to wait for RCU callbacks > in reclaim so that we won't get all of memory stuck there. I think it would be much too aggressive (_especially_ with preemptible RCU, I would have thought?). And not just for the case of adding a lot of idle time while other CPUs are busy not being preempted -- also that if you have lots of CPUs reclaiming and asking to synchronize_rcu, then you could have global RCU state getting ping-ponged. > If this location is too aggressive we might stick it next to > disable_swap_token(). I don't see a really good reason to add this kind of wait here. I mean, we have to wait for RCU anyway, so we may as well scan what we can until then? Putting it in a slowpath at some point when we think we're heading for OOM I think would be a bit better.