Re: [PATCH] memcg: fix stale swap cache leak v5

[Balbir Singh <balbir@linux.vnet.ibm.com>]

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-04-30 18:47:38]:

> On Thu, 30 Apr 2009 15:12:52 +0530
> Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> 
> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-04-30 18:04:26]:
> > 
> > > On Thu, 30 Apr 2009 16:35:39 +0900
> > > KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > > 
> > > > On Thu, 30 Apr 2009 16:16:27 +0900
> > > > KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > > > 
> > > > > This is v5 but all codes are rewritten.
> > > > > 
> > > > > After this patch, when memcg is used,
> > > > >  1. page's swapcount is checked after I/O (without locks). If the page is
> > > > >     stale swap cache, freeing routine will be scheduled.
> > > > >  2. vmscan.c calls try_to_free_swap() when __remove_mapping() fails.
> > > > > 
> > > > > Works well for me. no extra resources and no races.
> > > > > 
> > > > > Because my office will be closed until May/7, I'll not be able to make a
> > > > > response. Posting this for showing what I think of now.
> > > > > 
> > > > I found a hole immediately after posted this...sorry. plz ignore this patch/
> > > > see you again in the next month.
> > > > 
> > > I'm now wondering to disable "swapin readahed" completely when memcg is used...
> > > Then, half of the problems will go away immediately.
> > > And it's not so bad to try to free swapcache if swap writeback ends. Then, another
> > > half will go away...
> > >
> > 
> > Could you clarify? Will memcg not account for swapin readahead pages?
> >  
> swapin-readahead pages are _not_ accounted now. (And I think _never_)
> But has race and leak swp_entry account until global LRU runs.
> 
> "Don't do swapin-readahead, at all" will remove following race completely.
> ==
>          CPU0                  CPU1
>  free_swap_and_cache()
>                         read_swapcache_async()
> ==
> swp_entry to be freed will not be read-in.
> 
> I think there will no performance regression in _usual_ case even if no readahead.
> But has no number yet.
>

Kamezawa, Daisuke,

Can't we just correct the accounting and leave the page on the global
LRU?

Daisuke in the race conditions mentioned is (2) significant? Since the
accounting is already fixed during mem_cgroup_uncharge_page()?

 

-- 
	Balbir