Messages in this thread | | | Date | Mon, 29 Aug 2011 16:17:55 -0700 (PDT) | From | David Rientjes <> | Subject | Re: mm->oom_disable_count is broken |
| |
On Mon, 29 Aug 2011, Oleg Nesterov wrote:
> > IIRC, I did pointed out this issue. But nobody replied. > > I think ->oom_disable_count is currently broken. but now I have no time to > > audit this stuff. So, I'd suggest to revert this code if nobody don't fix it. > > I tend to agree, of course we can fix oom_disable_count but I don't > really understand why do we want it. >
I'd rather just remove it entirely, we'll have to ask it's author. Ying, do you see a reason to keep oom_disable_count around?
The only thing that I can see it doing is preventing a thread that shares an ->mm with an unkillable thread from being killed itself since it won't lead to future memory freeing. It prevents the second tasklist iteration after a task has been chosen to check if another thread sharing the memory cannot be killed.
I'd rather just kill the thread anyway because there's a chance that the OOM_DISABLE thread is waiting on it and may free its memory as well and there's no guarantee that when you set a thread to be OOM_DISABLE that all threads sharing the same memory are disabled as well.
> And. personally I dislike it because ->oom_disable_count is just another > proof that ->oom_score_adj should be in ->mm, not per-process. IIRC, > you already explained me why we can't do this, but - sorry - I forgot. > May be something with vfork... Could you explain this again? >
I actually really wanted oom_score_adj to be in the ->mm, it would simplify a lot of the code :) The problem was the inheritance property: we expect a job scheduler that is OOM_DISABLE to be able to vfork, change the oom_score_adj of the child, and then exec so that it is not oom disabled before starting to allocate memory. If this were in the mm, then setting the oom_score_adj of the child prior to exec would change the job scheduler's oom score as well.
| |