Re: [PATCH] fix leak of swap accounting as stale swap cache under memcg

[KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

On Mon, 27 Apr 2009 21:08:56 +0900
Daisuke Nishimura <d-nishimura@mtf.biglobe.ne.jp> wrote:

> > Index: mmotm-2.6.30-Apr24/mm/vmscan.c
> > ===================================================================
> > --- mmotm-2.6.30-Apr24.orig/mm/vmscan.c
> > +++ mmotm-2.6.30-Apr24/mm/vmscan.c
> > @@ -661,6 +661,9 @@ static unsigned long shrink_page_list(st
> >  		if (PageAnon(page) && !PageSwapCache(page)) {
> >  			if (!(sc->gfp_mask & __GFP_IO))
> >  				goto keep_locked;
> > +			/* avoid making more stale swap caches */
> > +			if (memcg_stale_swap_congestion())
> > +				goto keep_locked;
> >  			if (!add_to_swap(page))
> >  				goto activate_locked;
> >  			may_enter_fs = 1;
> > 
> Well, as I mentioned before(http://marc.info/?l=linux-kernel&m=124066623510867&w=2),
> this cannot avoid type-2(set !PageCgroupUsed by the owner process via
> page_remove_rmap()->mem_cgroup_uncharge_page() before being added to swap cache).
> If these swap caches go through shrink_page_list() without beeing freed
> for some reason, these swap caches doesn't go back to memcg's LRU.
> 
> Type-2 doesn't pressure memsw.usage, but you can see it by plotting
> "grep SwapCached /proc/meminfo".
> 
> And I don't think it's a good idea to add memcg_stale_swap_congestion() here.
> This means less possibility to reclaim pages.
> 
Hmm. maybe adding congestion_wait() ?

> Do you dislike the patch I attached in the above mail ?
> 
I doubt whether the patch covers all type-2 case.

> If not, please merge it(I tested your prvious version with some fixes and
> my patch, and it worked well). Or shall I send is as a separate patch
> to fix type-2 after your patch(yours looks good to me for type-1)?
> (to tell the truth, I want reuse memcg_free_unused_swapcache() in another patch)
> 
> 
I'll consider again and post v3.
But I'll go into a series of holidays, so, may not come back until May/6.

Thanks,
-Kame


> Thanks,
> Daisuke Nishimura.
>