Re: [RFC][PATCH 3/3] account swapcache

[Daisuke Nishimura <d-nishimura@mtf.biglobe.ne.jp>]

KAMEZAWA Hiroyuki wrote:
> On Fri, 11 Apr 2008 21:20:55 +0900
> Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
>> IMO, not charging swap caches as memory occasionally causes a problem
>> that swap caches are not freed even when a process that owns
>> those pages try to free them(e.g. task exit).
>>
>> For example:
>>
>>   Some pages are being reclaimed via memcg memory reclaim.
>>
>>   Assume that shrink_page_list() has already moved those pages
>>   to swap cache, unmapped them from ptes, removed from mz->lru,
>>   and is working on other pages on page_list.
>>   Those swap cache pages are unlocked and
>>   page_count of them are 2(swap cache, isolate_page).
>>
>>   At the same time on other CPU, if the process that owns those
>>   pages are trying to free them, free_swap_and_cache() cannot
>>   free those pages unless vm_swap_full, because find_get_pages()
>>   increases page_count.
>>
>> I think this rare case itself also exists on global memory reclaim,
>> but global memory reclaim does not assume that those pases have
>> been freed, so, if it need to free more memory, those pases
>> will be freed later because they remain on global inactive list.
>>
> yes.
> 
>> The problem here is that those swap cache pages are uncharged
>> from memcg, so memcg can never reclaim those pages that belonged
>> to the group.
>>
> why "never" uncharged ? 
> 
> Assume "page" is SwapCache and unmapped and clean. 
> ==
>  shrink_page_list()
> 	-> PageSwapCache() == true
> 	-> PageWriteback() == false
> 	-> PageDirty()     == false
> 	-> PagePrivate()   == true or false
> 	-> remove_mapping()
> 		-> page_count() == 2
> 		-> PageDirty()  == false
> 		-> PageSwapCache() == true
> 			-> __delete_from_swap_cache()
> 			-> true
> 	-> page will be freed
> ==
> 
You are right.

I was thinking the case below.
Assume some anonymous pages(mapped, referenced, !SwapCache)
are being reclaimed.

shrink_page_list()
	-> add_to_swap() <- makes the page dirty.
	->  try_to_unmap() <- uncharged from memcg and removed from mz->lru.
	-> PageDirty() == true
		sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced
			goto keep_locked
	-> unlocks the page and will work on other pages on page_list.

And, if on other CPU the process that owns those pages is exiting
at the timing of my example above, those pages remain only on
global lru, and are never charged(mapped) because the process exits.

I said "never" because once they are removed from mz->lru,
mem_cgroup_isolate_pages() doesn't select those pages
unless they are charged(mapped) again.


> page shirinking can free SwapCache regardless of vm_swap_full() result.
> Of course, my patch handles __delete_from_swap_cache().
>  
Yes.
I think your patch can handle the case what I'm saying.


Thanks,
Daisuke Nishimura.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>