Re: [v3 1/6] mm, oom: use oom_victims counter to synchronize oom victim selection

[Tetsuo Handa]

Roman Gushchin wrote:
> On Fri, Jun 23, 2017 at 06:52:20AM +0900, Tetsuo Handa wrote:
> > Tetsuo Handa wrote:
> > Oops, I misinterpreted. This is where a multithreaded OOM victim with or without
> > the OOM reaper can get stuck forever. Think about a process with two threads is
> > selected by the OOM killer and only one of these two threads can get TIF_MEMDIE.
> > 
> >   Thread-1                 Thread-2                 The OOM killer           The OOM reaper
> > 
> >                            Calls down_write(&current->mm->mmap_sem).
> >   Enters __alloc_pages_slowpath().
> >                            Enters __alloc_pages_slowpath().
> >   Takes oom_lock.
> >   Calls out_of_memory().
> >                                                     Selects Thread-1 as an OOM victim.
> >   Gets SIGKILL.            Gets SIGKILL.
> >   Gets TIF_MEMDIE.
> >   Releases oom_lock.
> >   Leaves __alloc_pages_slowpath() because Thread-1 has TIF_MEMDIE.
> >                                                                              Takes oom_lock.
> >                                                                              Will do nothing because down_read_trylock() fails.
> >                                                                              Releases oom_lock.
> >                                                                              Gives up and sets MMF_OOM_SKIP after one second.
> >                            Takes oom_lock.
> >                            Calls out_of_memory().
> >                            Will not check MMF_OOM_SKIP because Thread-1 still has TIF_MEMDIE. // <= get stuck waiting for Thread-1.
> >                            Releases oom_lock.
> >                            Will not leave __alloc_pages_slowpath() because Thread-2 does not have TIF_MEMDIE.
> >                            Will not call up_write(&current->mm->mmap_sem).
> >   Reaches do_exit().
> >   Calls down_read(&current->mm->mmap_sem) in exit_mm() in do_exit(). // <= get stuck waiting for Thread-2.
> >   Will not call up_read(&current->mm->mmap_sem) in exit_mm() in do_exit().
> >   Will not clear TIF_MEMDIE in exit_oom_victim() in exit_mm() in do_exit().
> 
> That's interesting... Does it mean, that we have to give an access to the reserves
> to all threads to guarantee the forward progress?

Yes, for we don't have __GFP_KILLABLE flag.

> 
> What do you think about Michal's approach? He posted a link in the thread.

Please read that thread.