Re: [PATCH 24/25] mm, compaction: Capture a page under direct compaction

[Vlastimil Babka]

On 1/4/19 1:50 PM, Mel Gorman wrote:
> Compaction is inherently race-prone as a suitable page freed during
> compaction can be allocated by any parallel task. This patch uses a
> capture_control structure to isolate a page immediately when it is freed
> by a direct compactor in the slow path of the page allocator. The intent
> is to avoid redundant scanning.
> 
>                                         4.20.0                 4.20.0
>                                selective-v2r15          capture-v2r15
> Amean     fault-both-1         0.00 (   0.00%)        0.00 *   0.00%*
> Amean     fault-both-3      2624.85 (   0.00%)     2594.49 (   1.16%)
> Amean     fault-both-5      3842.66 (   0.00%)     4088.32 (  -6.39%)
> Amean     fault-both-7      5459.47 (   0.00%)     5936.54 (  -8.74%)
> Amean     fault-both-12     9276.60 (   0.00%)    10160.85 (  -9.53%)
> Amean     fault-both-18    14030.73 (   0.00%)    13908.92 (   0.87%)
> Amean     fault-both-24    13298.10 (   0.00%)    16819.86 * -26.48%*
> Amean     fault-both-30    17648.62 (   0.00%)    17901.74 (  -1.43%)
> Amean     fault-both-32    19161.67 (   0.00%)    18621.32 (   2.82%)
> 
> Latency is only moderately affected but the devil is in the details.
> A closer examination indicates that base page fault latency is much
> reduced but latency of huge pages is increased as it takes creater care
> to succeed. Part of the "problem" is that allocation success rates
> are close to 100% even when under pressure and compaction gets harder
> 
>                                    4.20.0                 4.20.0
>                           selective-v2r15          capture-v2r15
> Percentage huge-1         0.00 (   0.00%)        0.00 (   0.00%)
> Percentage huge-3        99.95 (   0.00%)       99.98 (   0.03%)
> Percentage huge-5        98.83 (   0.00%)       98.01 (  -0.84%)
> Percentage huge-7        96.78 (   0.00%)       98.30 (   1.58%)
> Percentage huge-12       98.85 (   0.00%)       97.76 (  -1.10%)
> Percentage huge-18       97.52 (   0.00%)       99.05 (   1.57%)
> Percentage huge-24       97.07 (   0.00%)       99.34 (   2.35%)
> Percentage huge-30       96.59 (   0.00%)       99.08 (   2.58%)
> Percentage huge-32       95.94 (   0.00%)       99.03 (   3.22%)
> 
> And scan rates are reduced as expected by 10% for the migration
> scanner and 37% for the free scanner indicating that there is
> less redundant work.
> 
> Compaction migrate scanned    20338945.00    18133661.00
> Compaction free scanned       12590377.00     7986174.00
> 
> The impact on 2-socket is much larger albeit not presented. Under
> a different workload that fragments heavily, the allocation latency
> is reduced by 26% while the success rate goes from 63% to 80%
> 
> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>

Great, you crossed off this old TODO item, and didn't need pageblock isolation
to do that :D

I have just one worry...

> @@ -837,6 +873,12 @@ static inline void __free_one_page(struct page *page,
>  
>  continue_merging:
>  	while (order < max_order - 1) {
> +		if (compaction_capture(capc, page, order)) {
> +			if (likely(!is_migrate_isolate(migratetype)))
> +				__mod_zone_freepage_state(zone, -(1 << order),
> +								migratetype);
> +			return;

What about MIGRATE_CMA pageblocks and compaction for non-movable allocation,
won't that violate CMA expecteations?
And less critically, this will avoid the migratetype stealing decisions and
actions, potentially resulting in worse fragmentation avoidance?

> +		}
>  		buddy_pfn = __find_buddy_pfn(pfn, order);
>  		buddy = page + (buddy_pfn - pfn);
>