On Thu, Apr 08, 2010 at 06:27:06PM +0300, Avi Kivity wrote: > On 04/08/2010 06:23 PM, Andrea Arcangeli wrote: > > On Thu, Apr 08, 2010 at 02:44:01PM +0300, Avi Kivity wrote: > > > >> Results here are less than stellar. While khugepaged is pulling pages > >> together, something is breaking them apart. Even after memory pressure > >> is removed, this behaviour continues. Can it be that compaction is > >> tearing down huge pages? > >> > > migrate will split hugepages, but memory compaction shouldn't migrate > > hugepages... If it does I agree it needs fixing. > > > > > > Well, khugepaged was certainly fighting with something. Perhaps ftrace > will point the finger. It's likely splitting and migrating hugepages really, like you just said. > It ran stably for me FWIW. It's not easily reproducible (the tiny race in memcg prepare_migration also was hard to reproduce but that one was easy to fix by reproducing it just once). I suspect something's wrong with isolate_migratepages, I don't see where it skips tail pages, and it should skip all compound pages to avoid migrating them which would also solve your issue. something like that... ;). I think migration splits the hugepages good enough and while migration has a pin on the page (to keep khugepaged away). -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>