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

[David Rientjes]

On Tue, 15 Mar 2011, Oleg Nesterov wrote:

> When I did this change I looked at 81236810226f71bd9ff77321c8e8276dae7efc61
> and the changelog says:
> 
> 	__oom_kill_task() is called to elevate the task's timeslice and give it
> 	access to memory reserves so that it may quickly exit.
> 
> 	This privilege is unnecessary, however, if the task has already detached
> 	its mm.
> 
> Now you are saing this is pointless.
> 

If you have the commit id, do a "git blame 8123681022", because I see a:

5081dde3 (Nick Piggin        2006-09-25 23:31:32 -0700 222)             if (!p->mm)
5081dde3 (Nick Piggin        2006-09-25 23:31:32 -0700 223)                     continue;

in select_bad_process() and it's also iterating over every thread:

a49335cc (Paul Jackson       2005-09-06 15:18:09 -0700 215)     do_each_thread(g, p) {

It's pointless since oom-skip-zombies-when-iterating-tasklist.patch in 
-mm reintroduced the filter for !p->mm in select_bad_process() which was 
still there when 81236810 was merged; it's a small optimization, though, 
to avoid races where the mm becomes detached between the process' 
selection in select_bad_process() and its kill in oom_kill_process().

> The problem is, we can't trust per-thread PF_EXITING checks. But I guess
> we will discuss this more anyway.
> 

My approach, as you saw with 
oom-avoid-deferring-oom-killer-if-exiting-task-is-being-traced.patch in 
-mm is to add exceptions to the oom killer when we can't trust that 
PF_EXITING will soon be exiting.  I think that's a much more long-term 
maintainable solution instead of inferring the status of a thread based on 
external circumstances (such as number of threads in the thread group) 
that could easily change out from under us and once again break the oom 
killer.
--
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/