Re: v3.4-rc2 out-of-memory problems (was Re: 3.4-rc1sticks-and-crashs)

[David Rientjes]

On Mon, 9 Apr 2012, Linus Torvalds wrote:

> > You could that if you also turned the check for "ret == NOTIFY_OK" in
> > profile_handoff_task() into "ret & NOTIFY_OK" in your patch, otherwise you
> > get a double free from __put_task_struct() and oprofile.
> 
> Why? NOTIFY_DONE is zero.
> 

Oops, right.

> >  (1)  fix the lowmemorykiller so it doesn't need to use these notifiers at
> >      all, which isn't difficult, for 3.4, then
> 
> I do think that that makes sense. Fixing people to not use notifiers
> is always a good idea. Why would anybody sane even care about the
> process going away anyway? If some lowmemorykiller decides to kill off
> a process that no longer exists, kill() should happily return ENOSRCH,
> and we're all good
> 

It's apparently waiting for a killed thread to exit before selecting 
another victim or the one second timeout expires.  (And you only get to 
prevent needless kills if you have CONFIG_PROFILING, otherwise it doesn't 
care.)

> At the same time, the *only* user of that stupid handoff thing is
> oprofile, afaik, and if we use a refcount, why the hell doesn't
> oprofile just use a refcount to begin with, instead of using that
> notifier?: IOW, *both* users of the notifier seem to be just retarded.
> 

Agreed and since the current implementation relies on CONFIG_PROFILING I 
think it's safe to remove the notifier and add a hook only for oprofile so 
it can do free_task() when it wants to.  No refcounting required.

I've already proposed a patch that removes the notifier for 
lowmemorykiller with the added benefit that it doesn't rely on 
CONFIG_PROFILING at all.  If that's merged for 3.4, I'll remove the task 
handoff callchain entirely for 3.5 since oprofile is the only user.