Re: [RFC] locking: rwbase: Take care of ordering guarantee for fastpath reader

[Boqun Feng <boqun.feng@gmail.com>]

On Thu, Sep 02, 2021 at 01:55:29PM +0200, Peter Zijlstra wrote:
[...]
> >  	raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
> >  	rwbase_rtmutex_unlock(rtm);
> >  }
> > @@ -216,8 +229,14 @@ static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
> >  	 */
> >  	rwbase_set_and_save_current_state(state);
> >  
> > -	/* Block until all readers have left the critical section. */
> > -	for (; atomic_read(&rwb->readers);) {
> > +	/*
> > +	 * Block until all readers have left the critical section.
> > +	 *
> > +	 * _acqurie() is needed in case that the reader side runs in the fast
> > +	 * path, pairing with the atomic_dec_and_test() in rwbase_read_unlock(),
> > +	 * provides ACQUIRE.
> > +	 */
> > +	for (; atomic_read_acquire(&rwb->readers);) {
> >  		/* Optimized out for rwlocks */
> >  		if (rwbase_signal_pending_state(state, current)) {
> >  			__set_current_state(TASK_RUNNING);
> 
> I think we can restructure things to avoid this one, but yes. Suppose we
> do:
> 
> 	readers = atomic_sub_return_relaxed(READER_BIAS, &rwb->readers);
> 
> 	/*
> 	 * These two provide either an smp_mb() or an UNLOCK+LOCK

By "UNLOCK+LOCK", you mean unlock(->pi_lock) + lock(->wait_lock), right?
This may be unrelated, but in our memory model only unlock+lock pairs on
the same lock provide TSO-like ordering ;-) IOW, unlock(->pi_lock) +
lock(->wait_lock) on the same CPU doesn't order reads before and after.
Consider the following litmus:


	C unlock-lock
	{
	}

	P0(spinlock_t *s, spinlock_t *p, int *x, int *y)
	{
		int r1;
		int r2;

		spin_lock(s);
		r1 = READ_ONCE(*x);
		spin_unlock(s);
		spin_lock(p);
		r2 = READ_ONCE(*y);
		spin_unlock(p);
	}

	P1(int *x, int *y)
	{
		WRITE_ONCE(*y, 1);
		smp_wmb();
		WRITE_ONCE(*x, 1);
	}

	exists (0:r1=1 /\ 0:r2=0)

herd result:

	Test unlock-lock Allowed
	States 4
	0:r1=0; 0:r2=0;
	0:r1=0; 0:r2=1;
	0:r1=1; 0:r2=0;
	0:r1=1; 0:r2=1;
	Ok
	Witnesses
	Positive: 1 Negative: 3
	Condition exists (0:r1=1 /\ 0:r2=0)
	Observation unlock-lock Sometimes 1 3
	Time unlock-lock 0.01
	Hash=a8b772fd25f963f73a0d8e70e36ee255


> 	 * ordering, either is strong enough to provide ACQUIRE order
> 	 * for the above load of @readers.
> 	 */
> 	rwbase_set_and_save_current_state(state);
> 	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
> 
> 	while (readers) {
> 		...
> 		readers = atomic_read(&rwb->readers);

The above should be _acquire(), right? Pairs with the last reader
exiting the critical section and dec ->readers to 0. If so, it
undermines the necessity of the restructure?

Regards,
Boqun

> 		if (readers)
> 			rwbase_schedule();
> 		...
> 	}
> 
> 
> > @@ -229,6 +248,9 @@ static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
> >  		/*
> >  		 * Schedule and wait for the readers to leave the critical
> >  		 * section. The last reader leaving it wakes the waiter.
> > +		 *
> > +		 * _acquire() is not needed, because we can rely on the smp_mb()
> > +		 * in set_current_state() to provide ACQUIRE.
> >  		 */
> >  		if (atomic_read(&rwb->readers) != 0)
> >  			rwbase_schedule();
> > @@ -253,7 +275,11 @@ static inline int rwbase_write_trylock(struct rwbase_rt *rwb)
> >  	atomic_sub(READER_BIAS, &rwb->readers);
> >  
> >  	raw_spin_lock_irqsave(&rtm->wait_lock, flags);
> > -	if (!atomic_read(&rwb->readers)) {
> > +	/*
> > +	 * _acquire() is needed in case reader is in the fast path, pairing with
> > +	 * rwbase_read_unlock(), provides ACQUIRE.
> > +	 */
> > +	if (!atomic_read_acquire(&rwb->readers)) {
> 
> Moo; the alternative is using dec_and_lock instead of dec_and_test, but
> that's not going to be worth it.
> 
> >  		atomic_set(&rwb->readers, WRITER_BIAS);
> >  		raw_spin_unlock_irqrestore(&rtm->wait_lock, flags);
> >  		return 1;
> > -- 
> > 2.32.0
> >