Re: [PATCH] oom: skip frozen tasks

[Michal Hocko <mhocko@suse.cz>]

On Fri 26-08-11 09:09:46, Michal Hocko wrote:
> On Thu 25-08-11 14:14:20, David Rientjes wrote:
> > On Thu, 25 Aug 2011, Michal Hocko wrote:
> > 
> > > > > > That's obviously false since we call oom_killer_disable() in 
> > > > > > freeze_processes() to disable the oom killer from ever being called in the 
> > > > > > first place, so this is something you need to resolve with Rafael before 
> > > > > > you cause more machines to panic.
> > > > >
> > > > > I didn't mean suspend/resume path (that is protected by oom_killer_disabled)
> > > > > so the patch doesn't make any change.
> > > > 
> > > > Confused... freeze_processes() does try_to_freeze_tasks() before
> > > > oom_killer_disable() ?
> > > 
> > > Yes you are right, I must have been blind. 
> > > 
> > > Now I see the point. We do not want to panic while we are suspending and
> > > the memory is really low just because all the userspace is already in
> > > the the fridge.
> > > Sorry for confusion.
> > > 
> > > I still do not follow the oom_killer_disable note from David, though.
> > > 
> > 
> > oom_killer_disable() was added to that path for a reason when all threads 
> > are frozen: memory allocations still occur in the suspend path in an oom 
> > condition and adding the oom_killer_disable() will cause those 
> > allocations to fail rather than sending pointless SIGKILLs to frozen 
> > threads.
> > 
> > Now consider if the only _eligible_ threads for oom kill (because of 
> > cpusets or mempolicies) are those that are frozen.  We certainly do not 
> > want to panic because other cpusets are still getting work done.  We'd 
> > either want to add a mem to the cpuset or thaw the processes because the 
> > cpuset is oom.
> 
> Sure.
> 
> > 
> > You can't just selectively skip certain threads when their state can be 
> > temporary without risking a panic.  That's why this patch is a 
> > non-starter.
> > 
> > A much better solution would be to lower the badness score that the oom 
> > killer uses for PF_FROZEN threads so that they aren't considered a 
> > priority for kill unless there's nothing else left to kill.
> 
> Yes, sounds better.

.. but still not sufficient. We also have to thaw the process
as well. Just a quick hacked up patch (not tested, just for an
illustration).
Would something like this work?
--- 
>From 305a8139a72b20709e6b59ff8f4d322a9e04ab19 Mon Sep 17 00:00:00 2001
From: Michal Hocko <mhocko@suse.cz>
Date: Fri, 26 Aug 2011 10:39:35 +0200
Subject: [PATCH] oom: do not live lock on frozen tasks

[WARNING untested]

OOM can end up in a live lock if select_bad_process picks up a frozen
task. On the other hand we cannot mark such processes as unkillable
because we could panic the system even though there is a chance that
somebody could thaw the process so we can make a forward process (e.g.
a process from another cpuset or with a different nodemask).

Let's give all frozen tasks a bonus (OOM_SCORE_ADJ_MAX/2) so that we do
not consider them unless really necessary and if we really pick up one
then thaw its threads before we try to kill it.

TODO
- given bonus might be too big?
- aren't we racing with try_to_freeze_tasks?
---
 mm/oom_kill.c |   13 +++++++++++++
 1 files changed, 13 insertions(+), 0 deletions(-)

diff --git a/mm/oom_kill.c b/mm/oom_kill.c
index 626303b..fd194bc 100644
--- a/mm/oom_kill.c
+++ b/mm/oom_kill.c
@@ -32,6 +32,7 @@
 #include <linux/mempolicy.h>
 #include <linux/security.h>
 #include <linux/ptrace.h>
+#include <linux/freezer.h>
 
 int sysctl_panic_on_oom;
 int sysctl_oom_kill_allocating_task;
@@ -214,6 +215,14 @@ unsigned int oom_badness(struct task_struct *p, struct mem_cgroup *mem,
 	points += p->signal->oom_score_adj;
 
 	/*
+	 * Do not try to kill frozen tasks unless there is nothing else to kill.
+	 * We do not want to give it 1 point because we still want to select a good
+	 * candidate among all frozen tasks. Let's give it a reasonable bonus.
+	 */
+	if (frozen(p))
+		points -= OOM_SCORE_ADJ_MAX/2;
+
+	/*
 	 * Never return 0 for an eligible task that may be killed since it's
 	 * possible that no single user task uses more than 0.1% of memory and
 	 * no single admin tasks uses more than 3.0%.
@@ -450,6 +459,10 @@ static int oom_kill_task(struct task_struct *p, struct mem_cgroup *mem)
 			pr_err("Kill process %d (%s) sharing same memory\n",
 				task_pid_nr(q), q->comm);
 			task_unlock(q);
+
+			if (frozen(q))
+				thaw_process(q);
+
 			force_sig(SIGKILL, q);
 		}
 
-- 
1.7.5.4

-- 
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9    
Czech Republic