Re: [PATCH] oom killer (Core)

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, 2004-12-04 at 18:52 -0800, William Lee Irwin III wrote:
> On Wed, Dec 01, 2004 at 10:49:03AM +0100, tglx@linutronix.de wrote:
> > The oom killer has currently some strange effects when triggered.
> > It gets invoked multiple times and the selection of the task to kill
> > does not take processes into account which fork a lot of child processes.
> > The patch solves this by
> > - Preventing reentrancy
> > - Checking for memory threshold before selection and kill.
> > - Taking child processes into account when selecting the process to kill
> 
> Hmm, this thread seems to be serious. 

:)

> I'll audit the policy adjustments
> for issues with the mechanisms (e.g. killing kernel threads, races with
> timeouts).

Andrea provided a quite good fix for the invokation problem. I'm just
tracking down a different problem which was revieled by his changes.

The selection mechanism is currently taking a couple of things into
account:

 - VM size
 - nice value
 - CPU time
 - owner == root ?
 - CAP_SYS_RAWIO

I found out, that it is neccecary to take the child processes into
account to detect processes which fork a lot of childs.

The current mechanism rather kills sshd or portmap as they happen to
have a bigger VM size than the process which forked a lot of childs.

So I added a check for the child processes with an own VM. This solved
the problem quite well. Of course it does not count the childs of PID <
2.

tglx