From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from psmtp.com (na3sys010amx139.postini.com [74.125.245.139]) by kanga.kvack.org (Postfix) with SMTP id 7E4D46B0005 for ; Thu, 24 Jan 2013 10:16:17 -0500 (EST) Date: Thu, 24 Jan 2013 16:16:03 +0100 From: Jan Kara Subject: Re: [PATCH] Negative (setpoint-dirty) in bdi_position_ratio() Message-ID: <20130124151603.GD21818@quack.suse.cz> References: <201301200002.r0K02Atl031280@como.maths.usyd.edu.au> <20130124145707.GB12745@localhost> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20130124145707.GB12745@localhost> Sender: owner-linux-mm@kvack.org List-ID: To: Fengguang Wu Cc: paul.szabo@sydney.edu.au, linux-mm@kvack.org, 695182@bugs.debian.org, linux-kernel@vger.kernel.org, Andrew Morton , Jan Kara On Thu 24-01-13 22:57:07, Wu Fengguang wrote: > Hi Paul, > > > (This patch does not solve the PAE OOM issue.) > > You may try the below debug patch. The only way the writeback patches > should trigger OOM, I think, is for the number of dirty/writeback > pages going out of control. > > Or more simple, you may show us the OOM dmesg which will contain the > number of dirty pages. Or run this in a continuous loop during your > tests, and see how the dirty numbers change before OOM: I think he found the culprit of the problem being min_free_kbytes was not properly reflected in the dirty throttling. But the patch has been already picked up by Andrew so I didn't forward it to you. Paul please correct me if I'm wrong. Honza > > while : > do > grep -E '(Dirty|Writeback)' /proc/meminfo > sleep 1 > done > > Thanks, > Fengguang > > diff --git a/mm/page-writeback.c b/mm/page-writeback.c > index 50f0824..cf1165a 100644 > --- a/mm/page-writeback.c > +++ b/mm/page-writeback.c > @@ -1147,6 +1147,16 @@ pause: > if (task_ratelimit) > break; > > + if (nr_dirty > dirty_thresh + dirty_thresh / 2) { > + if (printk_ratelimit()) > + printk(KERN_WARNING "nr_dirty=%lu dirty_thresh=%lu task_ratelimit=%lu dirty_ratelimit=%lu pos_ratio=%lu\n", > + nr_dirty, > + dirty_thresh, > + task_ratelimit, > + dirty_ratelimit, > + pos_ratio); > + } > + > /* > * In the case of an unresponding NFS server and the NFS dirty > * pages exceeds dirty_thresh, give the other good bdi's a pipe -- Jan Kara SUSE Labs, CR -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org