Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756209AbbFPNDe (ORCPT ); Tue, 16 Jun 2015 09:03:34 -0400 Received: from mail-ob0-f175.google.com ([209.85.214.175]:33852 "EHLO mail-ob0-f175.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755984AbbFPNDO (ORCPT ); Tue, 16 Jun 2015 09:03:14 -0400 MIME-Version: 1.0 In-Reply-To: <5580177F.6070303@suse.cz> References: <1433928754-966-1-git-send-email-vbabka@suse.cz> <1433928754-966-7-git-send-email-vbabka@suse.cz> <20150616061013.GF12641@js1304-P5Q-DELUXE> <5580177F.6070303@suse.cz> Date: Tue, 16 Jun 2015 22:03:13 +0900 Message-ID: Subject: Re: [PATCH 6/6] mm, compaction: decouple updating pageblock_skip and cached pfn From: Joonsoo Kim To: Vlastimil Babka Cc: Joonsoo Kim , Andrew Morton , Linux Memory Management List , LKML , Minchan Kim , Mel Gorman , Michal Nazarewicz , Naoya Horiguchi , Christoph Lameter , Rik van Riel , David Rientjes Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3292 Lines: 63 2015-06-16 21:33 GMT+09:00 Vlastimil Babka : > On 06/16/2015 08:10 AM, Joonsoo Kim wrote: >> On Wed, Jun 10, 2015 at 11:32:34AM +0200, Vlastimil Babka wrote: >>> The pageblock_skip bitmap and cached scanner pfn's are two mechanisms in >>> compaction to prevent rescanning pages where isolation has recently failed >>> or they were scanned during the previous compaction attempt. >>> >>> Currently, both kinds of information are updated via update_pageblock_skip(), >>> which is suboptimal for the cached scanner pfn's: >>> >>> - The condition "isolation has failed in the pageblock" checked by >>> update_pageblock_skip() may be valid for the pageblock_skip bitmap, but makes >>> less sense for cached pfn's. There's little point for the next compaction >>> attempt to scan again a pageblock where all pages that could be isolated were >>> already processed. >> >> In async compaction, compaction could be stopped due to cc->contended >> in freepage scanner so sometimes isolated pages were not migrated. Your >> change makes next async compaction skip these pages. This possibly causes >> compaction complete prematurely by async compaction. > > Hm, I see, thanks. That could be fixed when returning the non-migrated pages, > just like we do for the unused freepages and cached free scanner position. Yes, that would work. >> And, rescan previous attempted range could solve some race problem. >> If allocated page waits to set PageLRU in pagevec, compaction will >> pass it. If we try rescan after short time, page will have PageLRU and >> compaction can isolate and migrate it and make high order freepage. This >> requires some rescanning overhead but migration overhead which is more bigger >> than scanning overhead is just a little. If compaction pass it like as >> this change, pages on this area would be allocated for other requestor, and, >> when compaction revisit, there would be more page to migrate. > > The same "race problem" (and many others) can happen when we don't abort and > later restart from cached pfn's, but just continue on to later pageblocks within > single compaction run. Still I would expect that it's statistically higher > chance to succeed in the next pageblock than rescanning pageblock(s) that we > just scanned. If we consider just one compaction attempt, yes, scanning next pageblock is more promising way to succeed. But, if we rescan pageblock that we just scanned and, fortunately, we succeed to migrate and make high order freeepage from that pageblock, more compaction attempt can be successful before both scanner meet and this may result in less overhead in view of overall compaction attempts. >> I basically agree with this change because it is more intuitive. But, >> I'd like to see some improvement result or test this patch myself before merging >> it. > > Sure, please test. I don't expect much difference, the primary motivation was > really that the recorded pfn's by tracepoints looked much saner. Yes. that's what I'd like to say from *intuitive*. :) Thanks. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/