Messages in this thread | | | Date | Mon, 8 Jun 2009 16:54:57 +0900 | From | Daisuke Nishimura <> | Subject | Re: [PATCH mmotm] vmscan: fix may_swap handling for memcg |
| |
On Mon, 8 Jun 2009 15:53:50 +0900 (JST), KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> wrote: > > On Mon, 8 Jun 2009 12:20:54 +0900 (JST), KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> wrote: > > > Hi > > > > > Hi, thank you for your comment. > > > > > > From: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> > > > > > > > > Commit 2e2e425989080cc534fc0fca154cae515f971cf5 ("vmscan,memcg: reintroduce > > > > sc->may_swap) add may_swap flag and handle it at get_scan_ratio(). > > > > > > > > But the result of get_scan_ratio() is ignored when priority == 0, and this > > > > means, when memcg hits the mem+swap limit, anon pages can be swapped > > > > just in vain. Especially when memcg causes oom by mem+swap limit, > > > > we can see many and many pages are swapped out. > > > > > > > > Instead of not scanning anon lru completely when priority == 0, this patch adds > > > > a hook to handle may_swap flag in shrink_page_list() to avoid using useless swaps, > > > > and calls try_to_free_swap() if needed because it can reduce > > > > both mem.usage and memsw.usage if the page(SwapCache) is unused anymore. > > > > > > > > Such unused-but-managed-under-memcg SwapCache can be made in some paths, > > > > for example trylock_page() failure in free_swap_cache(). > > > > > > > > Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> > > > > > > I think root cause is following branch, right? > > yes. > > > > > if so, Why can't we handle this issue on shrink_zone()? > > > > > Just because priority==0 means oom is about to happen and I don't > > want to see oom if possible. > > So I thought it would be better to reclaim as much pages(memsw.usage) as possible > > in this case. > > hmmm.. > > In general, adding new branch to shrink_page_list() is not good idea. > it can cause performance degression. > > Plus, it is not big problem at all. it happen only when priority==0. > Definitely, priority==0 don't occur normally. But it happens under high memory pressure...
> and, too many recliaming pages is not only memcg issue. I don't think this > patch provide generic solution. > Ah, you're right. It's not only memcg issue.
> > Why your test environment makes oom so frequently? > Not so frequently :) But I can see almost all of pages are swapped-out when memcg causes oom by memsw.limit(it's a waste of cpu time). And even after Kamezawa-san's memcg-fix-behavior-under-memorylimit-equals-to-memswlimit.patch, I can sometimes see swap usage when mem.limit==memsw.limit(it's a waste of cpu time too).
Thanks, Daisuke Nishimura.
| |