On Tue, Apr 12, 2011 at 7:34 PM, KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> wrote: >> > No good stat_threshold might makes performance hurt. >> >> Yes. That's I want it. >> My intention is that if you write down log fully, it can help much >> newbies to understand the patch in future and it would be very clear >> Andrew to merge it. >> >> What I want is following as. >> == >> >> Currently, memory hotplug doesn't updates pcp->stat_threashold. >> Then, It ends up making the wrong stat_threshold and percpu_driftmark. >> >> It could make confusing zoneinfo or overhead by frequent draining. >> Even when memory is low and kswapd is awake, it can mismatch between >> the number of real free pages and vmstat NR_FREE_PAGES so that it can >> result in the livelock. Please look at aa4548403 for more. >> >> This patch solves the issue. >> == > > Now, wakeup_kswapd() are using zone_watermark_ok_safe(). (ie avoid to use > per-cpu stat jiffies). Then, I don't think we have livelock chance. > Am I missing something? > I have no idea. I just referenced the description in aa4548403. As I look code, zone_watermark_ok_safe works well if percpu_drift_mark is set rightly. but if memory hotplug happens, zone->present_pages would be changed so that it can affect wmarks. It means it can affect percpu_drift_mark, I think. My point is to write down the description clear. -- Kind regards, Minchan Kim -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>