Re: [RFC][PATCH 5/5] (experimental) chase and free cache only swap

[Johannes Weiner <hannes@cmpxchg.org>]

On Wed, May 27, 2009 at 09:08:13AM +0900, KAMEZAWA Hiroyuki wrote:
> On Tue, 26 May 2009 20:14:00 +0200
> Johannes Weiner <hannes@cmpxchg.org> wrote:
> 
> > On Tue, May 26, 2009 at 12:18:34PM +0900, KAMEZAWA Hiroyuki wrote:
> > > 
> > > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> > > 
> > > Just a trial/example patch.
> > > I'd like to consider more. Better implementation idea is welcome.
> > > 
> > > When the system does swap-in/swap-out repeatedly, there are 
> > > cache-only swaps in general.
> > > Typically,
> > >  - swapped out in past but on memory now while vm_swap_full() returns true
> > > pages are cache-only swaps. (swap_map has no references.)
> > > 
> > > This cache-only swaps can be an obstacles for smooth page reclaiming.
> > > Current implemantation is very naive, just scan & free.
> > 
> > I think we can just remove that vm_swap_full() check in do_swap_page()
> > and try to remove the page from swap cache unconditionally.
> > 
> I'm not sure why reclaim swap entry only at write fault.

How do you come to that conclusion?  Do you mean the current code does
that?  Did you understand that I suggested that?

> > If it's still mapped someplace else, we let it cached.  If not, there
> > is not much use for keeping it around and we free it.
> > 
> yes.
> 
> > When I removed it and did benchmarks, I couldn't spot any difference
> > in the timings, though.  Did you measure the benefits of your patch
> > somehow?
> My patche has no "performance benefit". (My patch description may be bad.)
> I just checked that cache-only-swap can be big.(by sysrq-m)
> 
> Even when we remove vm_swap_full() in do_swap_page(),
> swapin-readahead + trylock-at-zap + vmscan makes "unused" swap caches easily.
> It reaches 1M in 2hours test of heavy swap program.

Ouch.

> > According to the git history tree, vm_swap_full() was initially only
> > used to aggressively drop cache entries even when they are mapped.
> > 
> > Rik put it into vmscan to reclaim swap cache _at all_ for activated
> > pages.  But I think unconditionally dropping the cache entry makes
> > sense if the page gets shuffled around on the LRU list.  Better to
> > re-allocate a swap slot close to the new LRU buddies on the next scan.
> > 
> > And having this all covered, the need for the scanning your patch does
> > should be gone, unless I missed something.
> > 
> Considering memcg, global lru scanning is no help ;(
> And I'm writing this patch for memcg.

Oh, sorry.  That makes sense of course.