Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751016AbaKDAf5 (ORCPT ); Mon, 3 Nov 2014 19:35:57 -0500 Received: from lgeamrelo04.lge.com ([156.147.1.127]:53329 "EHLO lgeamrelo04.lge.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750770AbaKDAfx (ORCPT ); Mon, 3 Nov 2014 19:35:53 -0500 X-Original-SENDERIP: 10.177.222.213 X-Original-MAILFROM: iamjoonsoo.kim@lge.com Date: Tue, 4 Nov 2014 09:37:33 +0900 From: Joonsoo Kim To: Vlastimil Babka Cc: Andrew Morton , linux-mm@kvack.org, linux-kernel@vger.kernel.org, Minchan Kim , Mel Gorman , Michal Nazarewicz , Naoya Horiguchi , Christoph Lameter , Rik van Riel , David Rientjes Subject: Re: [PATCH 5/5] mm, compaction: more focused lru and pcplists draining Message-ID: <20141104003733.GB8412@js1304-P5Q-DELUXE> References: <1412696019-21761-1-git-send-email-vbabka@suse.cz> <1412696019-21761-6-git-send-email-vbabka@suse.cz> <20141027074112.GC23379@js1304-P5Q-DELUXE> <545738F1.4010307@suse.cz> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <545738F1.4010307@suse.cz> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, Nov 03, 2014 at 09:12:33AM +0100, Vlastimil Babka wrote: > On 10/27/2014 08:41 AM, Joonsoo Kim wrote: > >On Tue, Oct 07, 2014 at 05:33:39PM +0200, Vlastimil Babka wrote: > >>The goal of memory compaction is to create high-order freepages through page > >>migration. Page migration however puts pages on the per-cpu lru_add cache, > >>which is later flushed to per-cpu pcplists, and only after pcplists are > >>drained the pages can actually merge. This can happen due to the per-cpu > >>caches becoming full through further freeing, or explicitly. > >> > >>During direct compaction, it is useful to do the draining explicitly so that > >>pages merge as soon as possible and compaction can detect success immediately > >>and keep the latency impact at minimum. However the current implementation is > >>far from ideal. Draining is done only in __alloc_pages_direct_compact(), > >>after all zones were already compacted, and the decisions to continue or stop > >>compaction in individual zones was done without the last batch of migrations > >>being merged. It is also missing the draining of lru_add cache before the > >>pcplists. > >> > >>This patch moves the draining for direct compaction into compact_zone(). It > >>adds the missing lru_cache draining and uses the newly introduced single zone > >>pcplists draining to reduce overhead and avoid impact on unrelated zones. > >>Draining is only performed when it can actually lead to merging of a page of > >>desired order (passed by cc->order). This means it is only done when migration > >>occurred in the previously scanned cc->order aligned block(s) and the > >>migration scanner is now pointing to the next cc->order aligned block. > >> > >>The patch has been tested with stress-highalloc benchmark from mmtests. > >>Although overal allocation success rates of the benchmark were not affected, > >>the number of detected compaction successes has doubled. This suggests that > >>allocations were previously successful due to implicit merging caused by > >>background activity, making a later allocation attempt succeed immediately, > >>but not attributing the success to compaction. Since stress-highalloc always > >>tries to allocate almost the whole memory, it cannot show the improvement in > >>its reported success rate metric. However after this patch, compaction should > >>detect success and terminate earlier, reducing the direct compaction latencies > >>in a real scenario. > >> > >>Signed-off-by: Vlastimil Babka > >>Cc: Minchan Kim > >>Cc: Mel Gorman > >>Cc: Joonsoo Kim > >>Cc: Michal Nazarewicz > >>Cc: Naoya Horiguchi > >>Cc: Christoph Lameter > >>Cc: Rik van Riel > >>Cc: David Rientjes > >>--- > >> mm/compaction.c | 41 ++++++++++++++++++++++++++++++++++++++++- > >> mm/page_alloc.c | 4 ---- > >> 2 files changed, 40 insertions(+), 5 deletions(-) > >> > >>diff --git a/mm/compaction.c b/mm/compaction.c > >>index 8fa888d..41b49d7 100644 > >>--- a/mm/compaction.c > >>+++ b/mm/compaction.c > >>@@ -1179,6 +1179,7 @@ static int compact_zone(struct zone *zone, struct compact_control *cc) > >> while ((ret = compact_finished(zone, cc, migratetype)) == > >> COMPACT_CONTINUE) { > >> int err; > >>+ unsigned long last_migrated_pfn = 0; > > > >I think that this definition looks odd. > >In every iteration, last_migrated_pfn is re-defined as 0. > >Maybe, it is on outside of the loop. > > Oops you're right, that's a mistake and it makes the code miss some > of the drain points (a minority I think but anyway). > > >> > >> switch (isolate_migratepages(zone, cc)) { > >> case ISOLATE_ABORT: > >>@@ -1187,7 +1188,12 @@ static int compact_zone(struct zone *zone, struct compact_control *cc) > >> cc->nr_migratepages = 0; > >> goto out; > >> case ISOLATE_NONE: > >>- continue; > >>+ /* > >>+ * We haven't isolated and migrated anything, but > >>+ * there might still be unflushed migrations from > >>+ * previous cc->order aligned block. > >>+ */ > >>+ goto check_drain; > >> case ISOLATE_SUCCESS: > >> ; > >> } > >>@@ -1212,6 +1218,39 @@ static int compact_zone(struct zone *zone, struct compact_control *cc) > >> goto out; > >> } > >> } > >>+ > >>+ /* > >>+ * Record where we have freed pages by migration and not yet > >>+ * flushed them to buddy allocator. Subtract 1, because often > >>+ * we finish a pageblock and migrate_pfn points to the first > >>+ * page* of the next one. In that case we want the drain below > >>+ * to happen immediately. > >>+ */ > >>+ if (!last_migrated_pfn) > >>+ last_migrated_pfn = cc->migrate_pfn - 1; > > > >And, I wonder why last_migrated_pfn is set after isolate_migratepages(). > > Not sure I understand your question. With the mistake above, it > cannot currently be set at the point isolate_migratepages() is > called, so you might question the goto check_drain in the > ISOLATE_NONE case, if that's what you are wondering about. > > When I correct that, it might be set when COMPACT_CLUSTER_MAX pages > are isolated and migrated the middle of a pageblock, and then the > rest of the pageblock contains no pages that could be isolated, so > the last isolate_migratepages() attempt in the pageblock returns > with ISOLATE_NONE. Still there were some migrations that produced > free pages that should be drained at that point. To clarify my question, I attach psuedo code that I thought correct. static int compact_zone() { unsigned long last_migrated_pfn = 0; ... compaction_suitable(); ... while (compact_finished()) { if (!last_migrated_pfn) last_migrated_pfn = cc->migrate_pfn - 1; isolate_migratepages(); switch case migrate_pages(); ... check_drain: (at the end of loop) do flush and reset last_migrated_pfn if needed } } We should record last_migrated_pfn before isolate_migratepages() and then compare it with cc->migrate_pfn after isolate_migratepages() to know if we moved away from the previous cc->order aligned block. Am I missing something? 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/