On Thu, 1 Dec 2011, Dave Chinner wrote:

> > /*
> >  * /proc/<pid>/oom_score_adj set to OOM_SCORE_ADJ_MIN disables oom killing for
> >  * pid.
> >  */
> > #define OOM_SCORE_ADJ_MIN       (-1000)
> > 
> > 
> > IIUC, this task cannot be killed by oom-killer because of oom_score_adj settings.
> 
> It's not me or the test suite that setting this, so it's something
> the kernel must be doing automagically.
> 

The kernel does not set oom_score_adj to ever disable oom killing for a 
thread.  The only time the kernel touches oom_score_adj is when setting it 
to "1000" in ksm and swap to actually prefer a memory allocator for oom 
killing.

It's also possible to change this value via the deprecated 
/proc/pid/oom_adj interface until it is removed next year.  Check your 
dmesg for warnings about using the deprecated oom_adj interface or change 
the printk_once() in oom_adjust_write() to a normal printk() to catch it.