Re: [PATCH 1/3 for 2.6.38] oom: oom_kill_process: don't set TIF_MEMDIEif !p->mm

[David Rientjes]

On Mon, 14 Mar 2011, Oleg Nesterov wrote:

> oom_kill_process() simply sets TIF_MEMDIE and returns if PF_EXITING.
> This is very wrong by many reasons. In particular, this thread can
> be the dead group leader. Check p->mm != NULL.
> 

This is true only for the oom_kill_allocating_task sysctl where it is 
required in all cases to kill current; current won't be triggering the oom 
killer if it's dead.

oom_kill_process() is called with the thread selected by 
select_bad_process() and that function will not return any thread if any 
eligible task is found to be PF_EXITING and is not current, or any 
eligible task is found to have TIF_MEMDIE.

In other words, for this conditional to be true in oom_kill_process(), 
then p must be current and so it cannot be the dead group leader as 
specified in your changelog unless PF_EXITING gets set between 
select_bad_process() and the oom_kill_process() call: we don't care about 
that since it's in the exit path and we therefore want to give it access 
to memory reserves to quickly exit anyway and the check for PF_EXITING in 
select_bad_process() prevents any infinite loop of that task getting 
constantly reselected if it's dead.

> Note: this is _not_ enough. Just a minimal fix.
> 
> Signed-off-by: Oleg Nesterov <oleg@xxxxxxxxxx>
> ---
> 
>  mm/oom_kill.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> --- 38/mm/oom_kill.c~1_kill_fix_pf_exiting	2011-03-14 17:53:05.000000000 +0100
> +++ 38/mm/oom_kill.c	2011-03-14 18:51:49.000000000 +0100
> @@ -470,7 +470,7 @@ static int oom_kill_process(struct task_
>  	 * If the task is already exiting, don't alarm the sysadmin or kill
>  	 * its children or threads, just set TIF_MEMDIE so it can die quickly
>  	 */
> -	if (p->flags & PF_EXITING) {
> +	if (p->flags & PF_EXITING && p->mm) {
>  		set_tsk_thread_flag(p, TIF_MEMDIE);
>  		boost_dying_task_prio(p, mem);
>  		return 0;
> 
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/