Re: [PATCH] mm,oom: fix oom invocation issues

From: Michal Hocko
Date: Thu May 18 2017 - 10:29:11 EST


On Thu 18-05-17 22:57:10, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > It is racy and it basically doesn't have any allocation context so we
> > might kill a task from a different domain. So can we do this instead?
> > There is a slight risk that somebody might have returned VM_FAULT_OOM
> > without doing an allocation but from my quick look nobody does that
> > currently.
>
> I can't tell whether it is safe to remove out_of_memory() from
> pagefault_out_of_memory(). There are VM_FAULT_OOM users in fs/
> directory. What happens if pagefault_out_of_memory() was called as a
> result of e.g. GFP_NOFS allocation failure?

Then we would bypass GFP_NOFS oom protection and could trigger a
premature OOM killer invocation.

> Is it guaranteed that all memory allocations that might occur from
> page fault event (or any action that might return VM_FAULT_OOM)
> are allowed to call oom_kill_process() from out_of_memory() before
> reaching pagefault_out_of_memory() ?

The same applies here.

> Anyway, I want
>
> /* Avoid allocations with no watermarks from looping endlessly */
> - if (test_thread_flag(TIF_MEMDIE))
> + if (alloc_flags == ALLOC_NO_WATERMARKS && test_thread_flag(TIF_MEMDIE))
> goto nopage;
>
> so that we won't see similar backtraces and memory information from both
> out_of_memory() and warn_alloc().

I do not think this is an improvement and it is unrelated to the
discussion here.

--
Michal Hocko
SUSE Labs