lkml.org 
[lkml]   [2015]   [Feb]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v3 3/3] mm/compaction: enhance compaction finish condition
Hello,

At 2015/2/2 18:20, Vlastimil Babka wrote:
> On 02/02/2015 08:15 AM, Joonsoo Kim wrote:
>> Compaction has anti fragmentation algorithm. It is that freepage
>> should be more than pageblock order to finish the compaction if we don't
>> find any freepage in requested migratetype buddy list. This is for
>> mitigating fragmentation, but, there is a lack of migratetype
>> consideration and it is too excessive compared to page allocator's anti
>> fragmentation algorithm.
>>
>> Not considering migratetype would cause premature finish of compaction.
>> For example, if allocation request is for unmovable migratetype,
>> freepage with CMA migratetype doesn't help that allocation and
>> compaction should not be stopped. But, current logic regards this
>> situation as compaction is no longer needed, so finish the compaction.
>
> This is only for order >= pageblock_order, right? Perhaps should be told explicitly.

I might be wrong. If we applied patch1, so after the system runs for some time,
there must be no MIGRATE_CMA free pages in the system, right? If so, the
example above doesn't exist anymore.

>
>> Secondly, condition is too excessive compared to page allocator's logic.
>> We can steal freepage from other migratetype and change pageblock
>> migratetype on more relaxed conditions in page allocator. This is designed
>> to prevent fragmentation and we can use it here. Imposing hard constraint
>> only to the compaction doesn't help much in this case since page allocator
>> would cause fragmentation again.
>>
>> To solve these problems, this patch borrows anti fragmentation logic from
>> page allocator. It will reduce premature compaction finish in some cases
>> and reduce excessive compaction work.
>>
>> stress-highalloc test in mmtests with non movable order 7 allocation shows
>> considerable increase of compaction success rate.
>>
>> Compaction success rate (Compaction success * 100 / Compaction stalls, %)
>> 31.82 : 42.20
>>
>> Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
>> ---
>> mm/compaction.c | 14 ++++++++++++--
>> mm/internal.h | 2 ++
>> mm/page_alloc.c | 12 ++++++++----
>> 3 files changed, 22 insertions(+), 6 deletions(-)
>>
>> diff --git a/mm/compaction.c b/mm/compaction.c
>> index 782772d..d40c426 100644
>> --- a/mm/compaction.c
>> +++ b/mm/compaction.c
>> @@ -1170,13 +1170,23 @@ static int __compact_finished(struct zone *zone, struct compact_control *cc,
>> /* Direct compactor: Is a suitable page free? */
>> for (order = cc->order; order < MAX_ORDER; order++) {
>> struct free_area *area = &zone->free_area[order];
>> + bool can_steal;
>>
>> /* Job done if page is free of the right migratetype */
>> if (!list_empty(&area->free_list[migratetype]))
>> return COMPACT_PARTIAL;
>>
>> - /* Job done if allocation would set block type */
>> - if (order >= pageblock_order && area->nr_free)
>> + /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
>> + if (migratetype == MIGRATE_MOVABLE &&
>> + !list_empty(&area->free_list[MIGRATE_CMA]))
>> + return COMPACT_PARTIAL;
>
> The above AFAICS needs #ifdef CMA otherwise won't compile without CMA.
>
>> +
>> + /*
>> + * Job done if allocation would steal freepages from
>> + * other migratetype buddy lists.
>> + */
>> + if (find_suitable_fallback(area, order, migratetype,
>> + true, &can_steal) != -1)
>> return COMPACT_PARTIAL;
>> }
>>
>> diff --git a/mm/internal.h b/mm/internal.h
>> index c4d6c9b..9640650 100644
>> --- a/mm/internal.h
>> +++ b/mm/internal.h
>> @@ -200,6 +200,8 @@ isolate_freepages_range(struct compact_control *cc,
>> unsigned long
>> isolate_migratepages_range(struct compact_control *cc,
>> unsigned long low_pfn, unsigned long end_pfn);
>> +int find_suitable_fallback(struct free_area *area, unsigned int order,
>> + int migratetype, bool only_stealable, bool *can_steal);
>>
>> #endif
>>
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 6cb18f8..0a150f1 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -1177,8 +1177,8 @@ static void steal_suitable_fallback(struct zone *zone, struct page *page,
>> set_pageblock_migratetype(page, start_type);
>> }
>>
>> -static int find_suitable_fallback(struct free_area *area, unsigned int order,
>> - int migratetype, bool *can_steal)
>> +int find_suitable_fallback(struct free_area *area, unsigned int order,
>> + int migratetype, bool only_stealable, bool *can_steal)
>> {
>> int i;
>> int fallback_mt;
>> @@ -1198,7 +1198,11 @@ static int find_suitable_fallback(struct free_area *area, unsigned int order,
>> if (can_steal_fallback(order, migratetype))
>> *can_steal = true;
>>
>> - return i;
>> + if (!only_stealable)
>> + return i;
>> +
>> + if (*can_steal)
>> + return i;
>
> So I've realized that this problaby won't always work as intended :/ Because we
> still differ from what page allocator does.
> Consider we compact for UNMOVABLE allocation. First we try RECLAIMABLE fallback.
> Turns out we could fallback, but not steal, hence we skip it due to
> only_stealable == true. So we try MOVABLE, and turns out we can steal, so we
> finish compaction.
> Then the allocation attempt follows, and it will fallback to RECLAIMABLE,
> without extra stealing. The compaction decision for MOVABLE was moot.
> Is it a big problem? Probably not, the compaction will still perform some extra
> anti-fragmentation on average, but we should consider it.
>
> I've got another idea for small improvement. We should only test for fallbacks
> when migration scanner has scanned (and migrated) a whole pageblock. Should be a
> simple alignment test of cc->migrate_pfn.
> Advantages:
> - potentially less checking overhead
> - chances of stealing increase if we created more free pages for migration
> - thus less fragmentation
> The cost is a bit more time spent compacting, but it's bounded and worth it
> (especially the less fragmentation) IMHO.

This seems to make the compaction a little compicated... I kind of
don't know why there is more anti-fragmentation by using this approach.

Thanks.

>
>> }
>>
>> return -1;
>> @@ -1220,7 +1224,7 @@ __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
>> --current_order) {
>> area = &(zone->free_area[current_order]);
>> fallback_mt = find_suitable_fallback(area, current_order,
>> - start_migratetype, &can_steal);
>> + start_migratetype, false, &can_steal);
>> if (fallback_mt == -1)
>> continue;
>>
>>
>
> --
> 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/
>


\
 
 \ /
  Last update: 2015-02-02 15:01    [W:0.084 / U:0.564 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site