From mboxrd@z Thu Jan 1 00:00:00 1970 From: Dave Chinner Subject: Re: [PATCH 3/3] writeback: Do not congestion sleep when there are no congested BDIs Date: Fri, 27 Aug 2010 15:13:16 +1000 Message-ID: <20100827051316.GH705@dastard> References: <1282835656-5638-1-git-send-email-mel@csn.ul.ie> <1282835656-5638-4-git-send-email-mel@csn.ul.ie> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-mm@kvack.org, linux-fsdevel@vger.kernel.org, Andrew Morton , Christian Ehrhardt , Johannes Weiner , Wu Fengguang , Jan Kara , linux-kernel@vger.kernel.org To: Mel Gorman Return-path: Received: from bld-mail17.adl2.internode.on.net ([150.101.137.102]:51549 "EHLO mail.internode.on.net" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1753381Ab0H0FNV (ORCPT ); Fri, 27 Aug 2010 01:13:21 -0400 Content-Disposition: inline In-Reply-To: <1282835656-5638-4-git-send-email-mel@csn.ul.ie> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Thu, Aug 26, 2010 at 04:14:16PM +0100, Mel Gorman wrote: > If congestion_wait() is called with no BDIs congested, the caller will > sleep for the full timeout and this is an unnecessary sleep. That, I think, is an invalid assumption. congestion_wait is used in some places as a backoff mechanism that waits for some IO work to be done, with congestion disappearing being a indication that progress has been made and so we can retry sooner than the entire timeout. For example, if _xfs_buf_lookup_pages() fails to allocate page cache pages for a buffer, it will kick the xfsbufd to writeback dirty buffers (so they can be freed) and immediately enter congestion_wait(). If there isn't congestion when we enter congestion_wait(), we still want to give the xfsbufds a chance to clean some pages before we retry the allocation for the new buffer. Removing the congestion_wait() sleep behaviour will effectively _increase_ memory pressure with XFS on fast disk subsystems because it now won't backoff between failed allocation attempts... Perhaps a congestion_wait_iff_congested() variant is needed for the VM? I can certainly see how it benefits the VM from a latency perspective, but it is the opposite behaviour that is expected in other places... Cheers, Dave. -- Dave Chinner david@fromorbit.com