Re: XDP socket rings, and LKMM litmus tests

["Paul E. McKenney" <paulmck@kernel.org>]

On Thu, Mar 04, 2021 at 04:27:53PM -0500, Alan Stern wrote:
> On Thu, Mar 04, 2021 at 11:05:15AM -0800, Paul E. McKenney wrote:
> > On Thu, Mar 04, 2021 at 10:35:24AM -0500, Alan Stern wrote:
> > > On Wed, Mar 03, 2021 at 09:04:07PM -0800, Paul E. McKenney wrote:
> > > > On Wed, Mar 03, 2021 at 10:21:01PM -0500, Alan Stern wrote:
> > > > > On Wed, Mar 03, 2021 at 02:03:48PM -0800, Paul E. McKenney wrote:
> > > > > > On Wed, Mar 03, 2021 at 03:22:46PM -0500, Alan Stern wrote:
> > > 
> > > > > > > >  And I cannot immediately think of a situation where
> > > > > > > > this approach would break that would not result in a data race being
> > > > > > > > flagged.  Or is this yet another failure of my imagination?
> > > > > > > 
> > > > > > > By definition, an access to a local variable cannot participate in a 
> > > > > > > data race because all such accesses are confined to a single thread.
> > > > > > 
> > > > > > True, but its value might have come from a load from a shared variable.
> > > > > 
> > > > > Then that load could have participated in a data race.  But the store to 
> > > > > the local variable cannot.
> > > > 
> > > > Agreed.  My thought was that if the ordering from the initial (non-local)
> > > > load mattered, then that initial load must have participated in a
> > > > data race.  Is that true, or am I failing to perceive some corner case?
> > > 
> > > Ordering can matter even when no data race is involved.  Just think
> > > about how much of the memory model is concerned with ordering of
> > > marked accesses, which don't participate in data races unless there is
> > > a conflicting plain access somewhere.
> > 
> > Fair point.  Should I have instead said "then that initial load must
> > have run concurrently with a store to that same variable"?
> 
> I'm losing track of the point you were originally trying to make.
> 
> Does ordering matter when there are no conflicting accesses?  Sure.  
> Consider this:
> 
> 	A: r1 = READ_ONCE(x);
> 	B: WRITE_ONCE(y, r1);
> 	   smp_wmb();
> 	C: WRITE_ONCE(z, 1);
> 
> Even if there are no other accesses to y at all (let alone any 
> conflicting ones), the mere existence of B forces A to be ordered before 
> C, and this is easily detectable by a litmus test.

Given that herd7 treats all local variables as registers (including
forbidding taking their addresses), and given that we are not thinking of
treating local-variable accesses as if they were marked, this is likely
all moot.

But just in case...

I was trying to figure out if there was a litmus test of the following
form where it might make a difference if local-variable accesses were
treated as if they were marked.  So is there something like this:

	r1 = x;
	if (r1)
	   	WRITE_ONCE(y, 1);

where implicitly treating the accesses to r1 as marked would make a
difference.  I was thinking that any such example would have to result
in LKMM flagging the load from x as a data race.  However, your example
inserting the smp_wmb() does shed some doubt on that theory.

This of course is moot unless we come back to treating local-variable
accesses as if they were marked.

							Thanx, Paul