On Sat, Apr 08, 2023 at 08:57:57PM +0200, Jonas Oberhauser wrote:
>
> On 4/8/2023 6:49 PM, Joel Fernandes wrote:
> > On Fri, Apr 07, 2023 at 05:49:02PM -0700, Paul E. McKenney wrote:
> > > On Fri, Apr 07, 2023 at 03:05:01PM +0200, Jonas Oberhauser wrote:
> > > >
> > > > On 4/7/2023 2:12 AM, Joel Fernandes wrote:
> > > > >
> > > > > > On Apr 6, 2023, at 6:34 PM, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Thu, Apr 06, 2023 at 05:36:13PM -0400, Alan Stern wrote:
> > > > > > > Paul:
> > > > > > >
> > > > > > > I just saw that two of the files in
> > > > > > > tools/memory-model/litmus-tests have
> > > > > > > almost identical names:
> > > > > > >
> > > > > > > Z6.0+pooncelock+pooncelock+pombonce.litmus
> > > > > > > Z6.0+pooncelock+poonceLock+pombonce.litmus
> > > > > > >
> > > > > > > They differ only by a lower-case 'l' vs. a capital 'L'. It's
> > > > > > > not at all
> > > > > > > easy to see, and won't play well in case-insensitive filesystems.
> > > > > > >
> > > > > > > Should one of them be renamed?
> > > > > >
> > FWIW, if I move that smp_mb_after..() a step lower, that also makes the test
> > work (see below).
> >
> > If you may look over quickly my analysis of why this smp_mb_after..() is
> > needed, it is because what I marked as a and d below don't have an hb
> > relation right?
>
> I think a and d have an hb relation due to the
> a ->po-rel X ->rfe Y ->acq-po d
> edges (where X and Y are the unlock/lock events I annotated in your example
> below).
I kind of disagree with that, because if I understand correctly, a ->hb d
means ALL CPUs agree as a universal fact that a happened before d.
Clearly, without the smp_mb(), CPU P2 disagrees that a happened before d.
So the po-rel acq-po doesn't imply a->hb d, IMHO. Correct me if I'm wrong
though with any counter example. ;-)
>
> Generally, an mb_unlock_lock isn't used to give you hb, but to turn some
> (coe/fre) ; hb* edges into pb edges
>
> In this case, that would probably be
> f ->fre a ->hb* f (where a ->hb* f comes from a ->hb* d ->hb e ->hb f)
> By adding the mb_unlock_lock_po in one of the right places, this becomes f
> ->pb f,
> thus forbidden.
This I fully agree with. I observed this litmus is actually the R-pattern
with P0 split into 2 CPUs by spltting the thread of execution using a lock
and ordering them with an ->rfe and the exists() clause.
Otherwise it is identical.
In the R-pattern also, you need an smp_mb() between the pair of accesses.
Using the same annotations but instead applying them to the R-pattern, it
looks like:
P0(int *x, int *y)
{
WRITE_ONCE(*x, 1); // a
// Here we need an smp_mb() to order the stores to x and z.
WRITE_ONCE(*z, 1); // d
}
P2(int *x, int *z)
{
int r1;
WRITE_ONCE(*z, 2); // e
smp_mb();
r1 = READ_ONCE(*x); // f
exists (z=2 /\ 2:r1=0)
thanks,
- Joel
>
> Have fun,
> jonas
>
>
> >
> > (*
> > b ->rf c
> >
> > d ->co e
> >
> > e ->hb f
> >
> > basically the issue is a ->po b ->rf c ->po d does not imply a ->hb d
> > *)
> >
> > P0(int *x, int *y, spinlock_t *mylock)
> > {
> > spin_lock(mylock);
> > WRITE_ONCE(*x, 1); // a
> > WRITE_ONCE(*y, 1); // b
> > spin_unlock(mylock); // X
> > }
> >
> > P1(int *y, int *z, spinlock_t *mylock)
> > {
> > int r0;
> >
> > spin_lock(mylock); // Y
> > r0 = READ_ONCE(*y); // c
> > smp_mb__after_spinlock(); // moving this a bit lower also works fwiw.
> > WRITE_ONCE(*z, 1); // d
> > spin_unlock(mylock);
> > }
> >
> > P2(int *x, int *z)
> > {
> > int r1;
> >
> > WRITE_ONCE(*z, 2); // e
> > smp_mb();
> > r1 = READ_ONCE(*x); // f
> > }
> >
> > exists (1:r0=1 /\ z=2 /\ 2:r1=0)
> >
> >
> > > Would someone like to to a "git mv" send the resulting patch?
> > Yes I can do that in return as I am thankful in advance for the above
> > discussion. ;)
> >
> > thanks,
> >
> > - Joel
> >
>
