On Thu, Mar 26, 2015 at 04:58:46PM +0100, Michal Hocko wrote: > On Wed 25-03-15 02:17:12, Johannes Weiner wrote: > > There is not much point in rushing back to the freelists and burning > > CPU cycles in direct reclaim when somebody else is in the process of > > OOM killing, or right after issuing a kill ourselves, because it could > > take some time for the OOM victim to release memory. > > Yes this makes sense and it is better than what we have now. The > question is how long we should wait. I can see you have gone with HZ. > What is the value based on? Have your testing shown that the OOM victim > manages to die within a second most of the time? > > I do not want to get into which value is the best discussion but I would > expect a larger value. Most OOM victims are not blocked so they would > wake up soon. This is a safety net for those who are blocked and I do > not think we have to expedite those rare cases and rather optimize for > "regular" OOM situations. How about 10-30s? Yup, I agree with that reasoning. We can certainly go higher than HZ. However, we should probably try to stay within the thresholds of any lock/hang detection watchdogs, which on a higher level includes the user itself, who might get confused if the machine hangs for 30s. As I replied to Vlastimil, once the OOM victim hangs for several seconds without a deadlock, failing the allocation wouldn't seem entirely unreasonable, either. But yes, something like 5-10s would still sound good to me. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
