Re: [PATCH] oom_reaper: close race without using oom_lock

[Michal Hocko]

On Fri 21-07-17 06:47:11, Tetsuo Handa wrote:
> Michal Hocko wrote:
> > On Wed 19-07-17 05:51:03, Tetsuo Handa wrote:
> > > Michal Hocko wrote:
> > > > On Tue 18-07-17 23:06:50, Tetsuo Handa wrote:
> > > > > Commit e2fe14564d3316d1 ("oom_reaper: close race with exiting task")
> > > > > guarded whole OOM reaping operations using oom_lock. But there was no
> > > > > need to guard whole operations. We needed to guard only setting of
> > > > > MMF_OOM_REAPED flag because get_page_from_freelist() in
> > > > > __alloc_pages_may_oom() is called with oom_lock held.
> > > > > 
> > > > > If we change to guard only setting of MMF_OOM_SKIP flag, the OOM reaper
> > > > > can start reaping operations as soon as wake_oom_reaper() is called.
> > > > > But since setting of MMF_OOM_SKIP flag at __mmput() is not guarded with
> > > > > oom_lock, guarding only the OOM reaper side is not sufficient.
> > > > > 
> > > > > If we change the OOM killer side to ignore MMF_OOM_SKIP flag once,
> > > > > there is no need to guard setting of MMF_OOM_SKIP flag, and we can
> > > > > guarantee a chance to call get_page_from_freelist() in
> > > > > __alloc_pages_may_oom() without depending on oom_lock serialization.
> > > > > 
> > > > > This patch makes MMF_OOM_SKIP act as if MMF_OOM_REAPED, and adds a new
> > > > > flag which acts as if MMF_OOM_SKIP, in order to close both race window
> > > > > (the OOM reaper side and __mmput() side) without using oom_lock.
> > > > 
> > > > Why do we need this patch when
> > > > http://lkml.kernel.org/r/20170626130346.26314-1-mhocko@xxxxxxxxxx
> > > > already removes the lock and solves another problem at once?
> > > 
> > > We haven't got an answer from Hugh and/or Andrea whether that patch is safe.
> > 
> > So what? I haven't see anybody disputing the correctness. And to be
> > honest I really dislike your patch. Yet another round kind of solutions
> > are just very ugly hacks usually because they are highly timing
> > sensitive.
> 
> Yes, OOM killer is highly timing sensitive.
> 
> > 
> > > Even if that patch is safe, this patch still helps with CONFIG_MMU=n case.
> > 
> > Could you explain how?
> 
> Nothing prevents sequence below.
> 
>     Process-1              Process-2
> 
>     Takes oom_lock.
>     Fails get_page_from_freelist().
>     Enters out_of_memory().
>     Gets SIGKILL.
>     Gets TIF_MEMDIE.
>     Leaves out_of_memory().
>     Releases oom_lock.
>     Enters do_exit().
>     Calls __mmput().
>                            Takes oom_lock.
>                            Fails get_page_from_freelist().
>     Releases some memory.
>     Sets MMF_OOM_SKIP.
>                            Enters out_of_memory().
>                            Selects next victim because there is no !MMF_OOM_SKIP mm.
>                            Sends SIGKILL needlessly.
> 
> If we ignore MMF_OOM_SKIP once, we can avoid sequence above.

But we set MMF_OOM_SKIP _after_ the process lost its address space (well
after the patch which allows to race oom reaper with the exit_mmap).

> 
>     Process-1              Process-2
> 
>     Takes oom_lock.
>     Fails get_page_from_freelist().
>     Enters out_of_memory().
>     Get SIGKILL.
>     Get TIF_MEMDIE.
>     Leaves out_of_memory().
>     Releases oom_lock.
>     Enters do_exit().
>     Calls __mmput().
>                            Takes oom_lock.
>                            Fails get_page_from_freelist().
>     Releases some memory.
>     Sets MMF_OOM_SKIP.
>                            Enters out_of_memory().
>                            Ignores MMF_OOM_SKIP mm once.
>                            Leaves out_of_memory().
>                            Releases oom_lock.
>                            Succeeds get_page_from_freelist().

OK, so let's say you have another task just about to jump into
out_of_memory and ... end up in the same situation. This race is just
unavoidable.

> Strictly speaking, this patch is independent with OOM reaper.
> This patch increases possibility of succeeding get_page_from_freelist()
> without sending SIGKILL. Your patch is trying to drop it silently.
> 
> Serializing setting of MMF_OOM_SKIP with oom_lock is one approach,
> and ignoring MMF_OOM_SKIP once without oom_lock is another approach.

Or simply making sure that we only set the flag _after_ the address
space is gone, which is what I am proposing.

-- 
Michal Hocko
SUSE Labs