Re: [PATCH 03/11] locking, rwsem: introduce basis for down_write_killable

[Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>]

Michal Hocko wrote:
> On Tue 10-05-16 19:43:20, Tetsuo Handa wrote:
> > I hit "allowing the OOM killer to select the same thread again" problem
> > ( http://lkml.kernel.org/r/20160408113425.GF29820@dhcp22.suse.cz ), but
> > I think that there is a bug in down_write_killable() series (at least
> > "locking, rwsem: introduce basis for down_write_killable" patch).
> > 
> > Complete log is at http://I-love.SAKURA.ne.jp/tmp/serial-20160510-sem.txt.xz .
> [...]
> > 2 threads (PID: 1314 and 1443) are sleeping at rwsem_down_read_failed()
> > but no thread is sleeping at rwsem_down_write_failed_killable().
> > If there is no thread waiting for write lock, threads waiting for read
> > lock must be able to run. This suggests that one of threads which was
> > waiting for write lock forgot to wake up reader threads.
> 
> Or that the write lock holder is still keeping the lock held. I do not
> see such a process in your list though. Is it possible that the
> debug_show_all_locks would just miss it as it is not sleeping?

I don't think it is possible. This reproducer
( http://lkml.kernel.org/r/201605061958.HHG48967.JVFtSLFQOFOOMH@I-love.SAKURA.ne.jp )
creates a thread group with two threads, and two of these two threads are
sleeping at rwsem_down_read_failed() waiting for mmap_sem.
SysRq-t suggests that PID 1443 called rwsem_down_write_failed_killable()
before calling rwsem_down_read_failed().

By the way, I suggested you to show traces of threads which are using the OOM victim's mm
( http://lkml.kernel.org/r/201603172200.CIE52148.QOVSOHJFMLOFtF@I-love.SAKURA.ne.jp ),
but you said that showing all locks held by !TASK_RUNNING threads would be useful
( http://lkml.kernel.org/r/20160323120716.GE7059@dhcp22.suse.cz ).
Do you admit that debug_show_all_locks() is not always useful by suspecting
the possibility of debug_show_all_locks() failing to report a thread which
held mmap_sem for write? (This is a kmallocwd topic, so I stop here.)

> 
> > Looking at rwsem_down_read_failed(), reader threads waiting for the
> > writer thread to release the lock are waiting on sem->wait_list list.
> > Looking at __rwsem_down_write_failed_common(), when the writer thread
> > escaped the
> > 
> >                  /* Block until there are no active lockers. */
> >                  do {
> >                          if (signal_pending_state(state, current)) {
> >                                  raw_spin_lock_irq(&sem->wait_lock);
> >                                  ret = ERR_PTR(-EINTR);
> >                                  goto out;
> >                          }
> >                          schedule();
> >                          set_current_state(state);
> >                  } while ((count = sem->count) & RWSEM_ACTIVE_MASK);
> > 
> > loop due to SIGKILL, I think that the writer thread needs to check for
> > remaining threads on sem->wait_list list and wake up reader threads
> > before rwsem_down_write_failed_killable() returns -EINTR.
> 
> I am not sure I understand. The rwsem counter is not write locked while
> the thread is sleeping and when we fail on the signal pending so readers
> should be able to proceed, no?
> 
I guess __rwsem_do_wake() is needed for waking up the readers because
I guess the sequence occurred was

  (1) PID 1314 requested down_read() and succeeded.
  (2) PID 1443 requested down_write_killable() and blocked.
  (3) The OOM killer sent SIGKILL to PID 1314 and PID 1443.
  (4) PID 1443 left down_write_killable() with -EINTR.
  (5) PID 1314 called up_read() and down_read()
      while PID 1443 called down_read().

.

> Or are you suggesting that the failure path should call rwsem_wake?

I don't know how rwsem works. Please consult maintainers.

Peter Zijlstra wrote:
> Mutex is much simpler; it doesn't have to do the reader-vs-writer
> fairness thing.
> 
> However, at the time I was thinking that if we have:
> 
> 	reader (owner)
> 	writer (pending)
> 	reader (blocked on writer)
> 
> and writer would get cancelled, the up_read() would do a wakeup and kick
> the blocked reader.
> 
> But yes, immediately kicking further pending waiters might be better.
> 
> Also, looking at it again; I think we're forgetting to re-adjust the
> BIAS for the cancelled writer.

Yes, I think so.
> 
> Davidlohr, Waiman, can you look at this?
>