Re: [RFC][PATCH] fix swap entries is not reclaimed in proper way for mem+swap controller

[Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>]

On Fri, 24 Apr 2009 13:33:06 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> On Wed, 22 Apr 2009 14:38:33 +0900
> Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> 
> > >  #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
> > >  extern void mem_cgroup_uncharge_swap(swp_entry_t ent);
> > > +extern void mem_cgroup_mark_swapcache_stale(struct page *page);
> > > +extern void mem_cgroup_fixup_swapcache(struct page *page);
> > >  #else
> > >  static inline void mem_cgroup_uncharge_swap(swp_entry_t ent)
> > >  {
> > >  }
> > > +static void mem_cgroup_check_mark_swapcache_stale(struct page *page)
> > > +{
> > > +}
> > > +static void mem_cgroup_fixup_swapcache(struct page *page)
> > > +{
> > > +}
> > >  #endif
> > >  
> > I think they should be defined in MEM_RES_CTLR case.
> > Exhausting swap entries problem is not depend on MEM_RES_CTLR_SWAP.
> > 
> Could you explain this more ? I can't understand.
> 
STALE(!PageCgroupUsed) SwapCache *without owner process* can be created by
the race between exit()..free_swap_and_cache() and read_swap_cache_async()(type-1)
or between exit()..page_remove_rmap() and shrink_page_list()(type-2).
(I don't think STALE SwapCache itself is problematic as long as there is
an actual user of the SwapCache.)

Those NOUSER STALE SwapCache are NOT depend on MEM_RES_CTLR_SWAP.

If total_swap_size is small enough not to trigger global LRU scan,
all swap space can be used up.
I confirmed before that all swap space were used up(and caused oom)
with mem.limit=32M/total_swap_size=50M.


Thanks,
Daisuke Nishimura.