Re: abnormal OOM killer message

[Mel Gorman <mel@csn.ul.ie>]

On Wed, Aug 19, 2009 at 07:52:42PM +0900, Minchan Kim wrote:
> Thanks for good comment, Mel. 
> 
> On Wed, 19 Aug 2009 11:36:11 +0100
> Mel Gorman <mel@csn.ul.ie> wrote:
> 
> > On Wed, Aug 19, 2009 at 03:49:58PM +0900, Minchan Kim wrote:
> > > On Wed, 19 Aug 2009 15:24:54 +0900
> > > ????????? <chungki.woo@gmail.com> wrote:
> > > 
> > > > Thank you very much for replys.
> > > > 
> > > > But I think it seems not to relate with stale data problem in compcache.
> > > > My question was why last chance to allocate memory was failed.
> > > > When OOM killer is executed, memory state is not a condition to
> > > > execute OOM killer.
> > > > Specially, there are so many pages of order 0. And allocating order is zero.
> > > > I think that last allocating memory should have succeeded.
> > > > That's my worry.
> > > 
> > > Yes. I agree with you.
> > > Mel. Could you give some comment in this situation ?
> > > Is it possible that order 0 allocation is failed 
> > > even there are many pages in buddy ?
> > > 
> > 
> > Not ordinarily. If it happens, I tend to suspect that the free list data
> > is corrupted and would put a check in __rmqueue() that looked like
> > 
> > 	BUG_ON(list_empty(&area->free_list) && area->nr_free);
> 
> If memory is corrupt, it would be not satisfied with both condition. 
> It would be better to ORed condition.
> 
> BUG_ON(list_empty(&area->free_list) || area->nr_free);
> 

But it's perfectly reasonable to have nr_free a positive value. The
point of the check is ensure the counters make sense. If nr_free > 0 and
the list is empty, it means accounting is all messed up and the values
reported for "free" in the OOM message are fiction.

> > The second question is, why are we in direct reclaim this far above the
> > watermark? It should only be kswapd that is doing any reclaim at that
> > point. That makes me wonder again are the free lists corrupted.
> 
> It does make sense!
> 
> > The other possibility is that the zonelist used for allocation in the
> > troubled path contains no populated zones. I would put a BUG_ON check in
> > get_page_from_freelist() to check if the first zone in the zonelist has no
> > pages. If that bug triggers, it might explain why OOMs are triggering for
> > no good reason.
> 
> Yes. Chungki. Could you put the both BUG_ON in each function and
> try to reproduce the problem ?
> 
> > I consider both of those possibilities abnormal though.
> > 
> > > > 
> > > > -----------------------------------------------------------------------------------------------------------------------------------------------
> > > >       page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order,
> > > > <== this is last chance
> > > >                            zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET);
> > > > <== uses ALLOC_WMARK_HIGH
> > > >       if (page)
> > > >       goto got_pg;
> > > > 
> > > >       out_of_memory(zonelist, gfp_mask, order);
> > > >       goto restart;
> > > > -----------------------------------------------------------------------------------------------------------------------------------------------
> > > > 
> > > > > Let me have a question.
> > > > > Now the system has 79M as total swap.
> > > > > It's bigger than system memory size.
> > > > > Is it possible in compcache?
> > > > > Can we believe the number?
> > > > 
> > > > Yeah, It's possible. 79Mbyte is data size can be swap.
> > > > It's not compressed data size. It's just original data size.
> > > 
> > > You means your pages with 79M are swap out in compcache's reserved
> > > memory?
> > > 
> > 
> > -- 
> > Mel Gorman
> > Part-time Phd Student                          Linux Technology Center
> > University of Limerick                         IBM Dublin Software Lab
> 
> 
> -- 
> Kind regards,
> Minchan Kim
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab