From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S933014AbaEMKAh (ORCPT ); Tue, 13 May 2014 06:00:37 -0400 Received: from cantor2.suse.de ([195.135.220.15]:50417 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932952AbaEMKAd (ORCPT ); Tue, 13 May 2014 06:00:33 -0400 Message-ID: <5371ED3F.6070505@suse.cz> Date: Tue, 13 May 2014 12:00:31 +0200 From: Vlastimil Babka User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:24.0) Gecko/20100101 Thunderbird/24.4.0 MIME-Version: 1.0 To: David Rientjes , Andrew Morton CC: Mel Gorman , Rik van Riel , Joonsoo Kim , Greg Thelen , Hugh Dickins , linux-kernel@vger.kernel.org, linux-mm@kvack.org Subject: Re: [patch -mm] mm, thp: avoid excessive compaction latency during fault fix References: In-Reply-To: Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 05/08/2014 07:30 AM, David Rientjes wrote: > mm-thp-avoid-excessive-compaction-latency-during-fault.patch excludes sync > compaction for all high order allocations other than thp. What we really > want to do is suppress sync compaction for thp, but only during the page > fault path. > > Orders greater than PAGE_ALLOC_COSTLY_ORDER aren't necessarily going to > loop again so this is the only way to exhaust our capabilities before > declaring that we can't allocate. > > Reported-by: Hugh Dickins > Signed-off-by: David Rientjes > --- > mm/page_alloc.c | 17 +++++++---------- > 1 file changed, 7 insertions(+), 10 deletions(-) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -2585,16 +2585,13 @@ rebalance: > if (page) > goto got_pg; > > - if (gfp_mask & __GFP_NO_KSWAPD) { > - /* > - * Khugepaged is allowed to try MIGRATE_SYNC_LIGHT, the latency > - * of this allocation isn't critical. Everything else, however, > - * should only be allowed to do MIGRATE_ASYNC to avoid excessive > - * stalls during fault. > - */ > - if ((current->flags & (PF_KTHREAD | PF_KSWAPD)) == PF_KTHREAD) > - migration_mode = MIGRATE_SYNC_LIGHT; > - } > + /* > + * It can become very expensive to allocate transparent hugepages at > + * fault, so use asynchronous memory compaction for THP unless it is > + * khugepaged trying to collapse. > + */ > + if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD)) > + migration_mode = MIGRATE_SYNC_LIGHT; I wonder what about a process doing e.g. mmap() with MAP_POPULATE. It seems to me that it would get only MIGRATE_ASYNC here, right? Since gfp_mask would include __GFP_NO_KSWAPD and it won't have PF_KTHREAD. I think that goes against the idea that with MAP_POPULATE you say you are willing to wait to have everything in place before you actually use the memory. So I guess you are also willing to wait for hugepages in that situation? > > /* > * If compaction is deferred for high-order allocations, it is because >