Re: [RFC] locking/rwsem: Avoid issuing wakeup before setting the reader waiter to nil

[Peter Zijlstra]

On Thu, Nov 29, 2018 at 12:02:19PM -0500, Waiman Long wrote:
> On 11/29/2018 11:06 AM, Peter Zijlstra wrote:

> > Why; at that point we know the wakeup will happen after, which is all we
> > require.
> >

> Thread 1                                  Thread 2      Thread 3
> 
>     rwsem_down_read_failed()
>  raw_spin_lock_irq(&sem->wait_lock);
>  list_add_tail(&waiter.list, &wait_list);
>  raw_spin_unlock_irq(&sem->wait_lock);
>                                                         __rwsem_mark_wake();
>                                                          wake_q_add();
>                                           wake_up_q();
>                                                          waiter->task =
> NULL; --+
>  while (true)
> {                                                                 |
>  
> set_current_state(TASK_UNINTERRUPTIBLE);                                     
> |
>   if (!waiter.task) //
> false                                                    |
>      
> break;                                                                    |
>  
> schedule();                                                                  
> |
>  }                                                                       
> <-----+
>                                                         wake_up_q(&wake_q);

I think that thing is horribly whitespace damanaged. At least, it's not
making sense to me.

> OK, I got confused by the thread racing chart shown in the patch. It
> will be clearer if the clearing of waiter->task is moved down as shown.
> Otherwise, moving the clearing of waiter->task before wake_q_add() won't
> make a difference. So the patch can be a possible fix.
> 
> Still we are talking about 3 threads racing with each other. The
> clearing of wake_q.next in wake_up_q() is not atomic and it is hard to
> predict the racing result of the concurrent wake_q operations between
> threads 2 and 3. The essence of my tentative patch is to prevent the
> concurrent wake_q operations in the first place.

wake_up_q() should, per the barriers in wake_up_process, ensure that if
wake_a_add() fails, there will be a wakeup of that task after that
point.

So if we put wake_up_q() at the location where wake_up_process() should
be, it should all work.

The bug in question is that it can happen at any time after
wake_q_add(), not necessarily at wake_up_q().