On Thu, Sep 06, 2012 at 01:17:25PM +0100, Mel Gorman wrote: > On Thu, Sep 06, 2012 at 06:44:04PM +0800, Shaohua Li wrote: > > > > isolate_migratepages_range() might isolate none pages, for example, when > > zone->lru_lock is contended and compaction is async. In this case, we should > > abort compaction, otherwise, compact_zone will run a useless loop and make > > zone->lru_lock is even contended. > > > > It might also isolate no pages because the range was 100% allocated and > there were no free pages to isolate. This is perfectly normal and I suspect > this patch effectively disables compaction. What problem did you observe > that this patch is aimed at? I'm running a random swapin/out workload. When memory is fragmented enough, I saw 100% cpu usage. perf shows zone->lru_lock is heavily contended in isolate_migratepages_range. I'm using slub(I didn't see the problem with slab), the allocation is for radix_tree_node slab, which needs 4 pages. Even If I just apply the second patch, the system is still in 100% cpu usage. The spin_is_contended check can't cure the problem completely. Trace shows compact_zone will run a useless loop and each loop contend the lru_lock. With this patch, the cpu usage becomes normal (about 20% utilization). -- 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>