Re: [PATCH 2/2] memcg fix stale swap cache account leak v6

[Balbir Singh]

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2009-05-08 14:09:10]:

> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> 
> In general, Linux's swp_entry handling is done by combination of lazy techniques
> and global LRU. It works well but when we use mem+swap controller, some more
> strict control is appropriate. Otherwise, swp_entry used by a cgroup will be
> never freed until global LRU works. In a system where memcg is well-configured,
> global LRU doesn't work frequently.
> 
>   Example A) Assume a swap cache which is not mapped.
>               CPU0                            CPU1
> 	   zap_pte()....                  shrink_page_list()
> 	    free_swap_and_cache()           lock_page()
> 		page seems busy.
> 
>   Example B) Assume swapin-readahead.
> 	      CPU0			      CPU1
> 	   zap_pte()			  read_swap_cache_async()
> 					  swap_duplicate().
>            swap_entry_free() = 1
> 	   find_get_page()=> NULL.
> 					  add_to_swap_cache().
> 					  issue swap I/O. 
> 
> There are many patterns of this kind of race (but no problems).
> 
> free_swap_and_cache() is called for freeing swp_entry. But it is a best-effort
> function. If the swp_entry/page seems busy, swp_entry is not freed.
> This is not a problem because global-LRU will find SwapCache at page reclaim.
> 
> If memcg is used, on the other hand, global LRU may not work. Then, above
> unused SwapCache will not be freed.
> (unmapped SwapCache occupy swp_entry but never be freed if not on memcg's LRU)
> 
> So, even if there are no tasks in a cgroup, swp_entry usage still remains.
> In bad case, OOM by mem+swap controller is triggered by this "leak" of
> swp_entry as Nishimura reported.
> 
> Considering this issue, swapin-readahead itself is not very good for memcg.
> It read swap cache which will not be used. (and _unused_ swapcache will
> not be accounted.) Even if we account swap cache at add_to_swap_cache(),
> we need to account page to several _unrelated_ memcg. This is bad.
> 
> This patch tries to fix racy case of free_swap_and_cache() and page status.
> 
> After this patch applied, following test works well.
> 
>   # echo 1-2M > ../memory.limit_in_bytes
>   # run tasks under memcg.
>   # kill all tasks and make memory.tasks empty
>   # check memory.memsw.usage_in_bytes == memory.usage_in_bytes and
>     there is no _used_ swp_entry.
> 
> What this patch does is
>  - avoid swapin-readahead when memcg is activated.
>  - try to free swapcache immediately after Writeback is done.
>  - Handle racy case of __remove_mapping() in vmscan.c
> 
> TODO:
>  - tmpfs should use real readahead rather than swapin readahead...
> 
> Changelog: v5 -> v6
>  - works only when memcg is activated.
>  - check after I/O works only after writeback.
>  - avoid swapin-readahead when memcg is activated.
>  - fixed page refcnt issue.
> Changelog: v4->v5
>  - completely new design.
> 
> Reported-by: Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx>
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>

I know we discussed readahead changes this in the past

1. the memcg_activated() check should be memcg_swap_activated(), no?
   In type 1, the problem can be solved by unaccounting the pages
   in swap_entry_free
   Type 2 is not a problem, since the accounting is already correct
   Hence my assertion that this problem occurs only when swapaccount
   is enabled.
2. I don't mind adding space overhead to swap_cgroup, if this problem
   can be fought that way. The approaches so far have made my head go
   round.
3. Disabling readahead is a big decision and will need loads of
   review/data before we can decide to go this route.


-- 
	Balbir
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/