From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756786AbcCaJaT (ORCPT ); Thu, 31 Mar 2016 05:30:19 -0400 Received: from mail-wm0-f42.google.com ([74.125.82.42]:36177 "EHLO mail-wm0-f42.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756289AbcCaJaO (ORCPT ); Thu, 31 Mar 2016 05:30:14 -0400 Date: Thu, 31 Mar 2016 11:30:12 +0200 From: Michal Hocko To: Xishi Qiu Cc: Linux MM , LKML Subject: Re: [RFC] oom, but there is enough memory Message-ID: <20160331093011.GC27831@dhcp22.suse.cz> References: <56FCEAD0.9080806@huawei.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <56FCEAD0.9080806@huawei.com> User-Agent: Mutt/1.5.24 (2015-08-30) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu 31-03-16 17:16:00, Xishi Qiu wrote: > It triggers a lot of ooms, but there is enough memory(many large blocks). > And at last "Kernel panic - not syncing: Out of memory and no killable processes..." > > I find almost the every call trace include "pagefault_out_of_memory" and "gfp_mask=0x0". > If it does oom, why not it triger in mm core path? It seems that somebody in the page fault path has returned with VM_FAULT_OOM without invoking the page allocator and kept returning the same error until there is nothing killable and so the oom killer panics. [...] > <4>[63651.040374s][pid:2912,cpu3,sh]DMA free:550600kB min:5244kB low:27580kB high:28892kB active_anon:343060kB inactive_anon:1224kB active_file:107596kB inactive_file:465156kB unevictable:1040kB isolated(anon):0kB isolated(file):0kB present:2016252kB managed:1720040kB mlocked:1040kB dirty:40kB writeback:0kB mapped:200420kB shmem:1312kB slab_reclaimable:27048kB slab_unreclaimable:73300kB kernel_stack:15248kB pagetables:14484kB unstable:0kB bounce:0kB free_cma:30896kB writeback_tmp:0kB pages_scanned:0 all_unreclaimable? no This is rather weird. DMA zone with 2GB? What kind of architecture is this? -- Michal Hocko SUSE Labs