lkml.org 
[lkml]   [2009]   [Jun]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] remove memory.limit v.s. memsw.limit comparison.
On Fri, 5 Jun 2009 00:45:03 +0900 (JST), "KAMEZAWA Hiroyuki" <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> 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.
>
I agree.

> ...maybe give all jobs to user-land and keep the kernel as it is now
> is a good choice.
>
> BTW, I'd like to avoid useless swap-out in memory.limit == memsw.limit case.
> If someone has good idea, please :(
>
I think so too.

From my simple thoughts, how about changing __mem_cgroup_try_charge() like:

1. initialize "noswap" as "bool noswap = !!(mem->res.limit == mem->memsw.limit)".
2. add check "if (mem->res.limit == mem->memsw.limit)" on charge failure to mem->res
and set "noswap" to true if needed.
3. charge mem->memsw before mem->res.
There would be other ideas, but I prefer 1 among these choices.


Thanks,
Daisuke Nishimura.


\
 
 \ /
  Last update: 2009-06-05 02:43    [W:0.308 / U:0.448 seconds]
©2003-2011 Jasper Spaans. Advertise on this site