On Fri, 2 Jul 2010 12:33:15 -0700 Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > On Tue, 29 Jun 2010 12:34:34 +0100 > Mel Gorman <mel@xxxxxxxxx> wrote: > > > Here is V3 that depends again on flusher threads to do writeback in > > direct reclaim rather than stack switching which is not something I'm > > likely to get done before xfs/btrfs are ignoring writeback in mainline > > (phd sucking up time). > > IMO, implemetning stack switching for this is not a good idea. We > _already_ have a way of doing stack-switching. It's called > "schedule()". > Sure. > The only reason I can see for implementing an in-place stack switch > would be if schedule() is too expensive. And if we were to see > excessive context-switch overheads in this code path (and we won't) > then we should get in there and try to reduce the contect switch rate > first. > Maybe a concern of in-place stack exchange lovers is that it's difficult to guarantee when the pageout() will be issued. I'd like to try to add a call as - pageout_request(page) .... request to pageout a page to a daemon(kswapd). - pageout_barrier(zone? node?) .... wait until all writebacks ends. Implementation dilemmna: Because page->lru is very useful link to implement calls like above, but there is a concern that using it will hide pages from vmscan unnecessarily. Avoding to use of page->lru means to use another structure like pagevec, but it means page_count()+1 and pins pages unnecessarily. I'm now considering how to implement safe&scalable way to pageout-in-another-stack(thread)... I wonder it will require some throttling method for pageout, anyway. And my own problem is that I should add per-memcg threads or using some thread-pool ;( But it will be an another topic. Thanks, -Kame -- 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>