From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf0-f197.google.com (mail-pf0-f197.google.com [209.85.192.197]) by kanga.kvack.org (Postfix) with ESMTP id 08A486B0033 for ; Thu, 14 Dec 2017 02:40:58 -0500 (EST) Received: by mail-pf0-f197.google.com with SMTP id r88so3913618pfi.23 for ; Wed, 13 Dec 2017 23:40:57 -0800 (PST) Received: from mx2.suse.de (mx2.suse.de. [195.135.220.15]) by mx.google.com with ESMTPS id s15si2502139pgf.602.2017.12.13.23.40.56 for (version=TLS1 cipher=AES128-SHA bits=128/128); Wed, 13 Dec 2017 23:40:56 -0800 (PST) Date: Thu, 14 Dec 2017 08:40:53 +0100 From: Michal Hocko Subject: Re: [RFC PATCH 3/5] mm, hugetlb: do not rely on overcommit limit during migration Message-ID: <20171214074053.GC16951@dhcp22.suse.cz> References: <20171204140117.7191-1-mhocko@kernel.org> <20171204140117.7191-4-mhocko@kernel.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Sender: owner-linux-mm@kvack.org List-ID: To: Mike Kravetz Cc: linux-mm@kvack.org, Naoya Horiguchi , Andrew Morton , LKML On Wed 13-12-17 15:35:33, Mike Kravetz wrote: > On 12/04/2017 06:01 AM, Michal Hocko wrote: [...] > > Before migration > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/free_hugepages:0 > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages:1 > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/surplus_hugepages:0 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/free_hugepages:0 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/nr_hugepages:0 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/surplus_hugepages:0 > > > > After > > > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/free_hugepages:0 > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages:0 > > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/surplus_hugepages:0 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/free_hugepages:0 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/nr_hugepages:1 > > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/surplus_hugepages:0 > > > > with the previous implementation, both nodes would have nr_hugepages:1 > > until the page is freed. > > With the previous implementation, the migration would have failed unless > nr_overcommit_hugepages was explicitly set. Correct? yes [...] > In the previous version of this patch, I asked about handling of 'free' huge > pages. I did a little digging and IIUC, we do not attempt migration of > free huge pages. The routine isolate_huge_page() has this check: > > if (!page_huge_active(page) || !get_page_unless_zero(page)) { > ret = false; > goto unlock; > } > > I believe one of your motivations for this effort was memory offlining. > So, this implies that a memory area can not be offlined if it contains > a free (not in use) huge page? do_migrate_range will ignore this free huge page and then we will free it up in dissolve_free_huge_pages > Just FYI and may be something we want to address later. Maybe yes. The free pool might be reserved which would make dissolve_free_huge_pages to fail. Maybe we can be more clever and allocate a new huge page in that case. > My other issues were addressed. > > Reviewed-by: Mike Kravetz Thanks! -- Michal Hocko SUSE Labs -- 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