Date: Fri, 5 Jun 2009 11:19:02 +0900
From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
To: balbir@linux.vnet.ibm.com
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
       "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
       "nishimura@mxp.nes.nec.co.jp" <nishimura@mxp.nes.nec.co.jp>
Subject: Re: [PATCH] remove memory.limit v.s. memsw.limit comparison.
Message-Id: <20090605111902.69d10470.kamezawa.hiroyu@jp.fujitsu.com>
In-Reply-To: <20090605011019.GG7504@balbir.in.ibm.com>
References: <20090604141043.9a1064fd.kamezawa.hiroyu@jp.fujitsu.com>
	<20090604123625.GE7504@balbir.in.ibm.com>
	<0921392c77890fc84fa69653ae4f31d9.squirrel@webmail-b.css.fujitsu.com>
	<20090605011019.GG7504@balbir.in.ibm.com>
Organization: FUJITSU Co. LTD.
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3180
Lines: 80

On Fri, 5 Jun 2009 09:10:19 +0800
Balbir Singh <balbir@linux.vnet.ibm.com> wrote:

> * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-06-05 00:45:03]:
> 
> > Balbir Singh wrote:
> > > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-06-04
> > > 14:10:43]:
> > >
> > >> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> > >>
> > >> Removes memory.limit < memsw.limit at setting limit check completely.
> > >>
> > >> The limitation "memory.limit <= memsw.limit" was added just because
> > >> it seems sane ...if memory.limit > memsw.limit, only memsw.limit works.
> > >>
> > >> But To implement this limitation, we needed to use private mutex and
> > >> make
> > >> the code a bit complated.
> > >> As Nishimura pointed out, in real world, there are people who only want
> > >> to use memsw.limit.
> > >>
> > >> Then, this patch removes the check. user-land library or middleware can
> > >> check
> > >> this in userland easily if this really concerns.
> > >>
> > >> And this is a good change to charge-and-reclaim.
> > >>
> > >> Now, memory.limit is always checked before memsw.limit
> > >> and it may do swap-out. But, if memory.limit == memsw.limit, swap-out is
> > >> finally no help and hits memsw.limit again. So, let's allow the
> > >> condition
> > >> memory.limit > memsw.limit. Then we can skip unnecesary swap-out.
> > >>
> > >> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> > >> ---
> > >
> > > There is one other option, we could set memory.limit_in_bytes ==
> > > memory.memsw.limit_in_bytes provided it is set to LONG_LONG_MAX. I am
> > > not convinced that we should allow memsw.limit_in_bytes to be less
> > > that limit_in_bytes, it will create confusion and the API is already
> > > exposed.
> > >
> > Ahhhh, I get your point.
> >  if memory.memsw.limit_in_bytes < memory.limit_in_bytes, no swap will
> >  be used bacause currnet try_to_free_pages() for memcg skips swap-out.
> >  Then, only global-LRU will use swap.
> >  This behavior is not easy to understand.
> > 
> > Sorry, I don't push this patch as this is. But adding documentation about
> > "What happens when you set memory.limit == memsw.limit" will be necessary.
> > 
> > ...maybe give all jobs to user-land and keep the kernel as it is now
> > is a good choice.
> 
> Yes, probably and with libcgroup and configuration, defaults should
> not be hard to setup. Worst case we can use a script to setup both the
> values.
> 
> > 
> > BTW, I'd like to avoid useless swap-out in memory.limit == memsw.limit case.
> > If someone has good idea, please :(
> >
> 
> Are you seeing swap even with memory.limit == memory.memsw.limit? Only
> global pressure should cause swapout, no? 
> 
yes, only global pressure should cause swapout.
Then, this means processes can be killed by OOM even while no swap usage.
This is a _new_ case under the condition memcg is used. Then, documentation will
be good.

Thanks,
-Kame

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