Re: can't oom-kill zap the victim's memory?

From: Michal Hocko
Date: Sat Sep 19 2015 - 11:58:44 EST


On Sat 19-09-15 17:03:16, Oleg Nesterov wrote:
> On 09/17, Kyle Walker wrote:
> >
> > Currently, the oom killer will attempt to kill a process that is in
> > TASK_UNINTERRUPTIBLE state. For tasks in this state for an exceptional
> > period of time, such as processes writing to a frozen filesystem during
> > a lengthy backup operation, this can result in a deadlock condition as
> > related processes memory access will stall within the page fault
> > handler.
>
> And there are other potential reasons for deadlock.
>
> Stupid idea. Can't we help the memory hog to free its memory? This is
> orthogonal to other improvements we can do.
>
> Please don't tell me the patch below is ugly, incomplete and suboptimal
> in many ways, I know ;) I am not sure it is even correct. Just to explain
> what I mean.

Unmapping the memory for the oom victim has been already mentioned as a
way to improve the OOM killer behavior. Nobody has implemented that yet
though unfortunately. I have that on my TODO list since we have
discussed it with Mel at LSF.

> Perhaps oom_unmap_func() should only zap the anonymous vmas... and there
> are a lot of other details which should be discussed if this can make any
> sense.

I have just returned from an internal conference so my head is
completely cabbaged. I will have a look on Monday. From a quick look
the idea is feasible. You cannot rely on the worker context because
workqueues might be completely stuck with at this stage. You also cannot
do take mmap_sem directly because that might be held already so you need
a try_lock instead. Focusing on anonymous vmas first sounds like a good
idea to me because that would be simpler I guess.

>
> Oleg.
> ---
>
> --- a/mm/oom_kill.c
> +++ b/mm/oom_kill.c
> @@ -493,6 +493,26 @@ void oom_killer_enable(void)
> up_write(&oom_sem);
> }
>
> +static struct mm_struct *oom_unmap_mm;
> +
> +static void oom_unmap_func(struct work_struct *work)
> +{
> + struct mm_struct *mm = xchg(&oom_unmap_mm, NULL);
> +
> + if (!atomic_inc_not_zero(&mm->mm_users))
> + return;
> +
> + // If this is not safe we can do use_mm() + unuse_mm()
> + down_read(&mm->mmap_sem);
> + if (mm->mmap)
> + zap_page_range(mm->mmap, 0, TASK_SIZE, NULL);
> + up_read(&mm->mmap_sem);
> +
> + mmput(mm);
> + mmdrop(mm);
> +}
> +static DECLARE_WORK(oom_unmap_work, oom_unmap_func);
> +
> #define K(x) ((x) << (PAGE_SHIFT-10))
> /*
> * Must be called while holding a reference to p, which will be released upon
> @@ -570,8 +590,8 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> victim = p;
> }
>
> - /* mm cannot safely be dereferenced after task_unlock(victim) */
> mm = victim->mm;
> + atomic_inc(&mm->mm_count);
> mark_tsk_oom_victim(victim);
> pr_err("Killed process %d (%s) total-vm:%lukB, anon-rss:%lukB, file-rss:%lukB\n",
> task_pid_nr(victim), victim->comm, K(victim->mm->total_vm),
> @@ -604,6 +624,10 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> rcu_read_unlock();
>
> do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
> + if (cmpxchg(&oom_unmap_mm, NULL, mm))
> + mmdrop(mm);
> + else
> + queue_work(system_unbound_wq, &oom_unmap_work);
> put_task_struct(victim);
> }
> #undef K

--
Michal Hocko
SUSE Labs
--
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/