Re: [PATCH 2/2] sched/wait: avoid abort_exclusive_wait() in __wait_on_bit_lock()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Sep 01, 2016 at 09:01:41PM +0200, Peter Zijlstra wrote:
> > test_and_set_bit() implies mb() so
> > the lockless list_empty_careful() case is fine, we can not miss the
> > condition if we race with unlock_page().
> 
> You're talking about this ordering?:
> 
> 	finish_wait()			clear_bit_unlock();
> 	  list_empty_careful()
> 
> 	/* MB implied */		smp_mb__after_atomic();
> 	test_and_set_bit()		wake_up_page()
> 					  ...
> 					    autoremove_wake_function()
> 					      list_del_init();
> 
> 
> That could do with spelling out I feel.. :-)

This ^^^

> > I am not sure we even want to conditionalize both finish_wait()'s,
> > we could simply call it unconditionally and once before test_and_set(),
> > the spurious wakeup is unlikely case.
> 
> 
> 	ret = 0;
> 
> 	for (;;) {
> 		prepare_to_wait_exclusive(wq, &q->wait, mode);
> 
> 		if (test_bit(&q->key.bit_nr, &q->key.flag))
> 			ret = action(&q->key, mode);
> 
> 		if (!test_and_set_bit(&q->key.bit_nr, &q->key.flag)) {
> 			/* we got the lock anyway, ignore the signal */
> 			ret = 0;
> 			break;
> 		}
> 
> 		if (ret)
> 			break;
> 	}
> 	finish_wait(wq, &q->wait);
> 
> 	return ret;
> 
> 
> Would not that work too?

Nope, because we need to do that finish_wait() before
test_and_set_bit()..

Also the problem with doing finish_wait() unconditionally would be
destroying the FIFO order. With a bit of bad luck you'd get starvation
cases :/