lkml.org 
[lkml]   [2010]   [Feb]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] memcg: handle panic_on_oom=always case
On Wed, Feb 17, 2010 at 03:04:45PM +0900, KAMEZAWA Hiroyuki wrote:
> tested on mmotm-Feb11.
>
> Balbir-san, Nishimura-san, I want review from both of you.
>
> ==
>
> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
>
> Now, if panic_on_oom=2, the whole system panics even if the oom happend
> in some special situation (as cpuset, mempolicy....).
> Then, panic_on_oom=2 means painc_on_oom_always.
>
> Now, memcg doesn't check panic_on_oom flag. This patch adds a check.
>
> Maybe someone doubts how it's useful. kdump+panic_on_oom=2 is the
> last tool to investigate what happens in oom-ed system. If a task is killed,
> the sysytem recovers and used memory were freed, there will be few hint
> to know what happnes. In mission critical system, oom should never happen.
> Then, investigation after OOM is very important.
> Then, panic_on_oom=2+kdump is useful to avoid next OOM by knowing
> precise information via snapshot.

No I don't doubt it is useful, and I think this probably is the simplest
and most useful semantic. So thanks for doing this.

I hate to pick nits in a trivial patch but I will anyway:


> TODO:
> - For memcg, it's for isolate system's memory usage, oom-notiifer and
> freeze_at_oom (or rest_at_oom) should be implemented. Then, management
> daemon can do similar jobs (as kdump) in safer way or taking snapshot
> per cgroup.
>
> CC: Balbir Singh <balbir@linux.vnet.ibm.com>
> CC: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
> CC: David Rientjes <rientjes@google.com>
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> ---
> Documentation/cgroups/memory.txt | 2 ++
> Documentation/sysctl/vm.txt | 5 ++++-
> mm/oom_kill.c | 2 ++
> 3 files changed, 8 insertions(+), 1 deletion(-)
>
> Index: mmotm-2.6.33-Feb11/Documentation/cgroups/memory.txt
> ===================================================================
> --- mmotm-2.6.33-Feb11.orig/Documentation/cgroups/memory.txt
> +++ mmotm-2.6.33-Feb11/Documentation/cgroups/memory.txt
> @@ -182,6 +182,8 @@ list.
> NOTE: Reclaim does not work for the root cgroup, since we cannot set any
> limits on the root cgroup.
>
> +Note2: When panic_on_oom is set to "2", the whole system will panic.
> +

Maybe:

NOTE2: When panic_on_oom is set to "2", the whole system will panic in
case of an oom event in any cgroup.

> 2. Locking
>
> The memory controller uses the following hierarchy
> Index: mmotm-2.6.33-Feb11/Documentation/sysctl/vm.txt
> ===================================================================
> --- mmotm-2.6.33-Feb11.orig/Documentation/sysctl/vm.txt
> +++ mmotm-2.6.33-Feb11/Documentation/sysctl/vm.txt
> @@ -573,11 +573,14 @@ Because other nodes' memory may be free.
> may be not fatal yet.
>
> If this is set to 2, the kernel panics compulsorily even on the
> -above-mentioned.
> +above-mentioned. Even oom happens under memoyr cgroup, the whole
> +system panics.
memory

>
> The default value is 0.
> 1 and 2 are for failover of clustering. Please select either
> according to your policy of failover.
> +2 seems too strong but panic_on_oom=2+kdump gives you very strong
> +tool to investigate a system which should never cause OOM.

I don't think you need say 2 seems too strong because as you rightfully
say, it has real uses. The hint about using it to investigate OOM
conditions is good though.

>
> =============================================================
>
> Index: mmotm-2.6.33-Feb11/mm/oom_kill.c
> ===================================================================
> --- mmotm-2.6.33-Feb11.orig/mm/oom_kill.c
> +++ mmotm-2.6.33-Feb11/mm/oom_kill.c
> @@ -471,6 +471,8 @@ void mem_cgroup_out_of_memory(struct mem
> unsigned long points = 0;
> struct task_struct *p;
>
> + if (sysctl_panic_on_oom == 2)
> + panic("out of memory(memcg). panic_on_oom is selected.\n");
> read_lock(&tasklist_lock);
> retry:
> p = select_bad_process(&points, mem);


\
 
 \ /
  Last update: 2010-02-17 09:47    [W:0.059 / U:0.108 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site