From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757204Ab2BBTaF (ORCPT ); Thu, 2 Feb 2012 14:30:05 -0500 Received: from mx1.redhat.com ([209.132.183.28]:35888 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757038Ab2BBTaE (ORCPT ); Thu, 2 Feb 2012 14:30:04 -0500 Date: Thu, 2 Feb 2012 14:29:58 -0500 From: Vivek Goyal To: Tejun Heo Cc: axboe@kernel.dk, ctalbott@google.com, rni@google.com, linux-kernel@vger.kernel.org Subject: Re: [PATCHSET] blkcg: unify blkgs for different policies Message-ID: <20120202192958.GA1723@redhat.com> References: <1328131156-13290-1-git-send-email-tj@kernel.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1328131156-13290-1-git-send-email-tj@kernel.org> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Wed, Feb 01, 2012 at 01:19:05PM -0800, Tejun Heo wrote: [..] > > * use unified stats updated under queue lock and drop percpu stats > which should fix locking / context bug across percpu allocation. Hi Tejun, Does that mean that stat updation will happen under queue lock even if there are no throttling rules? That will introduce extra queue lock on fast path those who have throttling compiled in but are not using (common case for distributions). IMHO, we should keep the lockless per cpu stats and do the allocation in worker thread. I was going through my messages and noticed that for a workload queue lock contention had come down by 11% and in a separate testing I could gain 4-5% with PCIe based flash drive for random reads while I saturated the cpus. Thanks Vivek