Re: [PATCH 1/3] LKMM: Add litmus test for RCU GP guarantee where updater frees object

[Joel Fernandes]

On Fri, Mar 20, 2020 at 11:07:10AM -0400, Alan Stern wrote:
> On Fri, 20 Mar 2020, Andrea Parri wrote:
> 
> > On Fri, Mar 20, 2020 at 02:55:50AM -0400, Joel Fernandes (Google) wrote:
> > > This adds an example for the important RCU grace period guarantee, which
> > > shows an RCU reader can never span a grace period.
> > > 
> > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > ---
> > >  .../litmus-tests/RCU+sync+free.litmus         | 40 +++++++++++++++++++
> > >  1 file changed, 40 insertions(+)
> > >  create mode 100644 tools/memory-model/litmus-tests/RCU+sync+free.litmus
> > > 
> > > diff --git a/tools/memory-model/litmus-tests/RCU+sync+free.litmus b/tools/memory-model/litmus-tests/RCU+sync+free.litmus
> > > new file mode 100644
> > > index 0000000000000..c4682502dd296
> > > --- /dev/null
> > > +++ b/tools/memory-model/litmus-tests/RCU+sync+free.litmus
> > > @@ -0,0 +1,40 @@
> > > +C RCU+sync+free
> > > +
> > > +(*
> > > + * Result: Never
> > > + *
> > > + * This litmus test demonstrates that an RCU reader can never see a write after
> > > + * the grace period, if it saw writes that happen before the grace period. This
> > > + * is a typical pattern of RCU usage, where the write before the grace period
> > > + * assigns a pointer, and the writes after destroy the object that the pointer
> > > + * points to.
> > > + *
> > > + * This guarantee also implies, an RCU reader can never span a grace period and
> > > + * is an important RCU grace period memory ordering guarantee.
> > > + *)
> > > +
> > > +{
> > > +x = 1;
> > > +y = x;
> > > +z = 1;
> > 
> > FYI, this could become a little more readable if we wrote it as follows:
> > 
> > int x = 1;
> > int *y = &x;
> > int z = 1;
> 
> Also, the test won't work with klitmus7 unless you do this.

Will do.

> > The LKMM tools are happy either way, just a matter of style/preference;
> > and yes, MP+onceassign+derefonce isn't currently following mine...  ;-/
> > 
> > 
> > > +}
> > > +
> > > +P0(int *x, int *z, int **y)
> > > +{
> > > +	int r0;
> > 
> > This would need to be "int *r0;" in order to make klitmus7(+gcc) happy.

Sorry fixed it now, my version of herd did not complain on this so I missed it.

> > > +	int r1;
> > > +
> > > +	rcu_read_lock();
> > > +	r0 = rcu_dereference(*y);
> > > +	r1 = READ_ONCE(*r0);
> > > +	rcu_read_unlock();
> > > +}
> > > +
> > > +P1(int *x, int *z, int **y)
> > > +{
> > > +	rcu_assign_pointer(*y, z);
> > 
> > AFAICT, you don't need this "RELEASE"; e.g., compare this test with the
> > example in:
> > 
> >   https://www.kernel.org/doc/Documentation/RCU/Design/Requirements/Requirements.html#Grace-Period%20Guarantee
> > 
> > What am I missing?
> 
> If z were not a simple variable but a more complicated structure, the
> RELEASE would be necessary to ensure that all P1's prior changes to z
> became visible before the write to y.
> 
> Besides, it's good form always to match rcu_dereference() with 
> rcu_assign_pointer(), for code documentation if nothing else.

Yes, adding to what Alan said, you can see the effect of not using
rcu_assign_pointer() in: MP+onceassign+derefonce.litmus

Alan and Andrea, may I add your Reviewed-by or Acked-by tags on the v2?

thanks,

 - Joel