On Fri, Feb 10, 2017 at 06:23:40PM +0100, Vlastimil Babka wrote: > The migrate scanner in async compaction is currently limited to MIGRATE_MOVABLE > pageblocks. This is a heuristic intended to reduce latency, based on the > assumption that non-MOVABLE pageblocks are unlikely to contain movable pages. > > However, with the exception of THP's, most high-order allocations are not > movable. Should the async compaction succeed, this increases the chance that > the non-MOVABLE allocations will fallback to a MOVABLE pageblock, making the > long-term fragmentation worse. > > This patch attempts to help the situation by changing async direct compaction > so that the migrate scanner only scans the pageblocks of the requested > migratetype. If it's a non-MOVABLE type and there are such pageblocks that do > contain movable pages, chances are that the allocation can succeed within one > of such pageblocks, removing the need for a fallback. If that fails, the > subsequent sync attempt will ignore this restriction. > > Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx> Yes, IMO we should make the async compaction scanner decontaminate unmovable blocks. This is because we fall back to other-typed blocks before we reclaim, so any unmovable blocks that aren't perfectly occupied will fill with greedy page cache (and order-0 doesn't steal blocks back to make them compactable again). Subsequent unmovable higher-order allocations in turn are more likely to fall back and steal more movable blocks. As long as we have vastly more movable blocks than unmovable blocks, continuous page cache turnover will counteract this negative trend - pages are reclaimed mostly from movable blocks and some unmovable blocks, while new cache allocations are placed into the freed movable blocks - slowly moving cache out from unmovable blocks into movable ones. But that effect is independent of the rate of higher-order allocations and can be overwhelmed, so I think it makes sense to involve compaction directly in decontamination. The thing I'm not entirely certain about is the aggressiveness of this patch. Instead of restricting the async scanner to blocks of the same migratetype, wouldn't it be better (in terms of allocation latency) to simply let it compact *all* block types? Maybe changing it to look at unmovable blocks is enough to curb cross-contamination. Sure there will still be some, but now we're matching the decontamination rate to the rate of !movable higher-order allocations and don't just rely on the independent cache turnover rate, which during higher-order bursts might not be high enough to prevent an expansion of unmovable blocks. Does that make sense? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
