On Mon, Apr 30, 2012 at 8:34 PM, Nick Piggin <npiggin@xxxxxxxxx> wrote: > On 25 April 2012 04:37, Ying Han <yinghan@xxxxxxxxxx> wrote: >> On Mon, Apr 23, 2012 at 10:36 PM, Nick Piggin <npiggin@xxxxxxxxx> wrote: >>> On 24 April 2012 06:56, Ying Han <yinghan@xxxxxxxxxx> wrote: >>>> This is not a patch targeted to be merged at all, but trying to understand >>>> a logic in global direct reclaim. >>>> >>>> There is a logic in global direct reclaim where reclaim fails on priority 0 >>>> and zone->all_unreclaimable is not set, it will cause the direct to start over >>>> from DEF_PRIORITY. In some extreme cases, we've seen the system hang which is >>>> very likely caused by direct reclaim enters infinite loop. >>> >>> Very likely, or definitely? Can you reproduce it? What workload? >> >> No, we don't have reproduce workload for that yet. Everything is based >> on the watchdog dump file :( >> >>> >>>> >>>> There have been serious patches trying to fix similar issue and the latest >>>> patch has good summary of all the efforts: >>>> >>>> commit 929bea7c714220fc76ce3f75bef9056477c28e74 >>>> Author: KOSAKI Motohiro <kosaki.motohiro@xxxxxxxxxxxxxx> >>>> Date: Thu Apr 14 15:22:12 2011 -0700 >>>> >>>> vmscan: all_unreclaimable() use zone->all_unreclaimable as a name >>>> >>>> Kosaki explained the problem triggered by async zone->all_unreclaimable and >>>> zone->pages_scanned where the later one was being checked by direct reclaim. >>>> However, after the patch, the problem remains where the setting of >>>> zone->all_unreclaimable is asynchronous with zone is actually reclaimable or not. >>>> >>>> The zone->all_unreclaimable flag is set by kswapd by checking zone->pages_scanned in >>>> zone_reclaimable(). Is that possible to have zone->all_unreclaimable == false while >>>> the zone is actually unreclaimable? >>>> >>>> 1. while kswapd in reclaim priority loop, someone frees a page on the zone. It >>>> will end up resetting the pages_scanned. >>>> >>>> 2. kswapd is frozen for whatever reason. I noticed Kosaki's covered the >>>> hibernation case by checking oom_killer_disabled, but not sure if that is >>>> everything we need to worry about. The key point here is that direct reclaim >>>> relies on a flag which is set by kswapd asynchronously, that doesn't sound safe. >>>> >>>> Instead of keep fixing the problem, I am wondering why we have the logic >>>> "not oom but keep trying reclaim w/ priority 0 reclaim failure" at the first place: >>>> >>>> Here is the patch introduced the logic initially: >>>> >>>> commit 408d85441cd5a9bd6bc851d677a10c605ed8db5f >>>> Author: Nick Piggin <npiggin@xxxxxxx> >>>> Date: Mon Sep 25 23:31:27 2006 -0700 >>>> >>>> [PATCH] oom: use unreclaimable info >>>> >>>> However, I didn't find detailed description of what problem the commit trying >>>> to fix and wondering if the problem still exist after 5 years. I would be happy >>>> to see the later case where we can consider to revert the initial patch. >>> >>> The problem we were having is that processes would be killed at seemingly >>> random points of time, under heavy swapping, but long before all swap was >>> used. >>> >>> The particular problem IIRC was related to testing a lot of guests on an s390 >>> machine. I'm ashamed to have not included more information in the >>> changelog -- I suspect it was probably in a small batch of patches with a >>> description in the introductory mail and not properly placed into patches :( >>> >>> There are certainly a lot of changes in the area since then, so I couldn't be >>> sure of what will happen by taking this out. >>> >>> I don't think the page allocator "try harder" logic was enough to solve the >>> problem, and I think it was around in some form even back then. >>> >>> The biggest problem is that it's not an exact science. It will never do the >>> right thing for everybody, sadly. Even if it is able to allocate pages at a >>> very slow rate, this is effectively as good as a hang for some users. For >>> others, they want to be able to manually intervene before anything is killed. >>> >>> Sorry if this isn't too helpful! Any ideas would be good. Possibly need to have >>> a way to describe these behaviours in an abstract way (i.e., not just magic >>> numbers), and allow user to tune it. >> >> Thank you Nick and this is helpful. I looked up on the patches you >> mentioned, and I can see what problem they were trying to solve by >> that time. However things have been changed a lot, and it is hard to >> tell if the problem still remains on the current kernel or not. By >> spotting each by each, I see either the patch has been replaced by >> different logic or the same logic has been implemented differently. >> >> For this particular one patch, we now have code which does page alloc >> retry before entering OOM. So I am wondering if that will help the OOM >> situation by that time. > > Well it's not doing exactly the same thing, actually. And note that the > problem was not about parallel OOM-killing. The fact that the page > reclaim has not made any progress when we last called in does not > actually mean that it cannot make _any_ progress. > > My patch is more about detecting the latter case. I don't see there > is equivalent logic in page allocator to replace it. > > But again: this is not a question of correct or incorrect as far as I > can see, simply a matter of where you define "hopeless"! I could > easily see the need for way to bias that (kill quickly, medium, try to > never kill). That is right. We ( at google) seems to be on the other end of the bias where we prefer to oom kill instead of hopelessly looping in the reclaim path. Normally if the application gets into that state, the performance would be sucks already and they might prefer to be restarted :) Unfortunately, We weren't being able to reproduce the issue with synthetic workload. So far it only happens in production with a particular workload when the memory runs really really tight. The current logic seems perfer to reclaim more than going oom kill, and that might not fit all user's expectation. However, I guess it is hard to convince for any changes since different users has different bias as you said.... Thanks --Ying -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href