Re: XDP socket rings, and LKMM litmus tests

[Paul E. McKenney]

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:
> > > On Wed, Mar 03, 2021 at 09:40:22AM -0800, Paul E. McKenney wrote:
> > > > On Wed, Mar 03, 2021 at 12:12:21PM -0500, Alan Stern wrote:
> > > 
> > > > > Local variables absolutely should be treated just like CPU registers, if 
> > > > > possible.  In fact, the compiler has the option of keeping local 
> > > > > variables stored in registers.
> > > > > 
> > > > > (Of course, things may get complicated if anyone writes a litmus test 
> > > > > that uses a pointer to a local variable,  Especially if the pointer 
> > > > > could hold the address of a local variable in one execution and a 
> > > > > shared variable in another!  Or if the pointer is itself a shared 
> > > > > variable and is dereferenced in another thread!)
> > > > 
> > > > Good point!  I did miss this complication.  ;-)
> > > 
> > > I suspect it wouldn't be so bad if herd7 disallowed taking addresses of 
> > > local variables.
> > > 
> > > > As you say, when its address is taken, the "local" variable needs to be
> > > > treated as is it were shared.  There are exceptions where the pointed-to
> > > > local is still used only by its process.  Are any of these exceptions
> > > > problematic?
> > > 
> > > Easiest just to rule out the whole can of worms.
> > 
> > Good point, given that a global can be used instead of a local for
> > any case where an address must be taken.
> 
> Another thing to consider: Almost all marked accesses involve using the 
> address of the storage location (for example, smp_load_acquire's first 
> argument must be a pointer).  As far as I can remember at the moment, 
> the only ones that don't are READ_ONCE and WRITE_ONCE.  So although we 
> might or might not want to allow READ_ONCE or WRITE_ONCE on a local 
> variable, we won't have to worry about any of the other kinds of marked 
> accesses.

Good point!

> > > > > But even if local variables are treated as non-shared storage locations, 
> > > > > we should still handle this correctly.  Part of the problem seems to lie 
> > > > > in the definition of the to-r dependency relation; the relevant portion 
> > > > > is:
> > > > > 
> > > > > 	(dep ; [Marked] ; rfi)
> > > > > 
> > > > > Here dep is the control dependency from the READ_ONCE to the 
> > > > > local-variable store, and the rfi refers to the following load of the 
> > > > > local variable.  The problem is that the store to the local variable 
> > > > > doesn't go in the Marked class, because it is notated as a plain C 
> > > > > assignment.  (And likewise for the following load.)
> > > > > 
> > > > > Should we change the model to make loads from and stores to local 
> > > > > variables always count as Marked?
> > > > 
> > > > As long as the initial (possibly unmarked) load would be properly
> > > > complained about.
> > > 
> > > Sorry, I don't understand what you mean.
> > 
> > I was thinking in terms of something like this in one of the processes:
> > 
> > 	p = gp; // Unmarked!
> > 	r1 = p;
> > 	q = r1; // Implicitly marked now?
> > 	if (q)
> > 		WRITE_ONCE(x, 1); // ctrl dep from gp???
> 
> I hope we won't have to worry about this!  :-)  Treating local variable 
> accesses as if they are always marked looks wrong.

Good, that is where I was also heading.  ;-)

> > > >  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?

> > > However, there are other aspects to consider, in particular, the 
> > > ordering relations on local-variable accesses.  But if, as Luc says, 
> > > local variables are treated just like registers then perhaps the issue 
> > > doesn't arise.
> > 
> > Here is hoping!
> > 
> > > > > What should have happened if the local variable were instead a shared 
> > > > > variable which the other thread didn't access at all?  It seems like a 
> > > > > weak point of the memory model that it treats these two things 
> > > > > differently.
> > > > 
> > > > But is this really any different than the situation where a global
> > > > variable is only accessed by a single thread?
> > > 
> > > Indeed; it is the _same_ situation.  Which leads to some interesting 
> > > questions, such as: What does READ_ONCE(r) mean when r is a local 
> > > variable?  Should it be allowed at all?  In what way is it different 
> > > from a plain read of r?
> > > 
> > > One difference is that the LKMM doesn't allow dependencies to originate 
> > > from a plain load.  Of course, when you're dealing with a local 
> > > variable, what matters is not the load from that variable but rather the 
> > > earlier loads which determined the value that had been stored there.  
> > > Which brings us back to the case of the
> > > 
> > > 	dep ; rfi
> > > 
> > > dependency relation, where the accesses in the middle are plain and 
> > > non-racy.  Should the LKMM be changed to allow this?
> > 
> > It would be nice, give or take the potential side effects.  ;-)
> > As in it would be nice, but might not be worthwhile.
> 
> Treating local variables like registers will automatically bring this 
> behavior.  So I think we'll be good.

Sounds good.

> > > There are other differences to consider.  For example:
> > > 
> > > 	r = READ_ONCE(x);
> > > 	smp_wmb();
> > > 	WRITE_ONCE(y, 1);
> > > 
> > > If the write to r were treated as a marked store, the smp_wmb would 
> > > order it (and consequently the READ_ONCE) before the WRITE_ONCE.  
> > > However we don't want to do this when r is a local variable.  Indeed, a 
> > > plain store wouldn't be ordered this way because the compiler might 
> > > optimize the store away entirely, leaving the smp_wmb nothing to act on.
> > 
> > Agreed, having smp_wmb() order things due to a write to a local variable
> > would not be what we want.
> > 
> > > So overall the situation is rather puzzling.  Treating local variables 
> > > as registers is probably the best answer.
> > 
> > That is sounding quite appealing at the moment.
> 
> Agreed.

							Thanx, Paul