Re: [PATCH] refcount_t: documentation for memory ordering differences

[Andrea Parri]

On Fri, Dec 01, 2017 at 12:34:23PM -0800, Randy Dunlap wrote:
> On 11/29/2017 04:36 AM, Elena Reshetova wrote:
> > Some functions from refcount_t API provide different
> > memory ordering guarantees that their atomic counterparts.
> > This adds a document outlining these differences.
> > 
> > Signed-off-by: Elena Reshetova <elena.reshetova@xxxxxxxxx>
> > ---
> >  Documentation/core-api/index.rst              |   1 +
> >  Documentation/core-api/refcount-vs-atomic.rst | 129 ++++++++++++++++++++++++++
> >  2 files changed, 130 insertions(+)
> >  create mode 100644 Documentation/core-api/refcount-vs-atomic.rst
> 
> > diff --git a/Documentation/core-api/refcount-vs-atomic.rst b/Documentation/core-api/refcount-vs-atomic.rst
> > new file mode 100644
> > index 0000000..5619d48
> > --- /dev/null
> > +++ b/Documentation/core-api/refcount-vs-atomic.rst
> > @@ -0,0 +1,129 @@
> > +===================================
> > +refcount_t API compared to atomic_t
> > +===================================
> > +
> > +The goal of refcount_t API is to provide a minimal API for implementing
> > +an object's reference counters. While a generic architecture-independent
> > +implementation from lib/refcount.c uses atomic operations underneath,
> > +there are a number of differences between some of the refcount_*() and
> > +atomic_*() functions with regards to the memory ordering guarantees.
> > +This document outlines the differences and provides respective examples
> > +in order to help maintainers validate their code against the change in
> > +these memory ordering guarantees.
> > +
> > +memory-barriers.txt and atomic_t.txt provide more background to the
> > +memory ordering in general and for atomic operations specifically.
> > +
> > +Relevant types of memory ordering
> > +=================================
> > +
> > +**Note**: the following section only covers some of the memory
> > +ordering types that are relevant for the atomics and reference
> > +counters and used through this document. For a much broader picture
> > +please consult memory-barriers.txt document.
> > +
> > +In the absence of any memory ordering guarantees (i.e. fully unordered)
> > +atomics & refcounters only provide atomicity and
> > +program order (po) relation (on the same CPU). It guarantees that
> > +each atomic_*() and refcount_*() operation is atomic and instructions
> > +are executed in program order on a single CPU.
> > +This is implemented using READ_ONCE()/WRITE_ONCE() and
> > +compare-and-swap primitives.
> > +
> > +A strong (full) memory ordering guarantees that all prior loads and
> > +stores (all po-earlier instructions) on the same CPU are completed
> > +before any po-later instruction is executed on the same CPU.
> > +It also guarantees that all po-earlier stores on the same CPU
> > +and all propagated stores from other CPUs must propagate to all
> > +other CPUs before any po-later instruction is executed on the original
> > +CPU (A-cumulative property). This is implemented using smp_mb().
> 
> I don't know what "A-cumulative property" means, and google search didn't
> either.

The description above seems to follow the (informal) definition given in:

  https://github.com/aparri/memory-model/blob/master/Documentation/explanation.txt
  (c.f., in part., Sect. 13-14)

and formalized by the LKMM. (The notion of A-cumulativity also appears, in
different contexts, in some memory consistency literature, e.g.,

  http://www.cl.cam.ac.uk/~pes20/ppc-supplemental/index.html
  http://www.cl.cam.ac.uk/~pes20/armv8-mca/
  https://arxiv.org/abs/1308.6810 )

A typical illustration of A-cumulativity (for smp_store_release(), say) is
given with the following program:

int x = 0;
int y = 0;

void thread0()
{
	WRITE_ONCE(x, 1);
}

void thread1()
{
	int r0;

	r0 = READ_ONCE(x);
	smp_store_release(&y, 1);
}

void thread2()
{
	int r1;
	int r2;

	r1 = READ_ONCE(y);
	smp_rmb();
	r2 = READ_ONCE(x);
}

(This is a variation of the so called "message-passing" pattern, where the
 stores are "distributed" over two threads; see also

  https://github.com/aparri/memory-model/blob/master/litmus-tests/WRC%2Bpooncerelease%2Brmbonceonce%2BOnce.litmus )

The question we want to address is whether the final state

  (r0 == 1 && r1 == 1 && r2 == 0)

can be reached/is allowed, and the answer is no (due to the A-cumulativity
of the store-release).

By contrast, dependencies provides no (A-)cumulativity; for example, if we
modify the previous program by replacing the store-release with a data dep.
as follows:

int x = 0;
int y = 0;

void thread0()
{
	WRITE_ONCE(x, 1);
}

void thread1()
{
	int r0;

	r0 = READ_ONCE(x);
	WRITE_ONCE(x, r0);
}

void thread2()
{
	int r1;
	int r2;

	r1 = READ_ONCE(y);
	smp_rmb();
	r2 = READ_ONCE(x);
}

then that same final state is allowed (and observed on some PPC machines).

  Andrea


> 
> Is it non-cumulative, similar to typical vs. atypical, where atypical
> roughly means non-typical.  Or is it accumlative (something being
> accumulated, summed up, gathered up)?
> 
> Or is it something else.. TBD?
> 
> > +A RELEASE memory ordering guarantees that all prior loads and
> > +stores (all po-earlier instructions) on the same CPU are completed
> > +before the operation. It also guarantees that all po-earlier
> > +stores on the same CPU and all propagated stores from other CPUs
> > +must propagate to all other CPUs before the release operation
> > +(A-cumulative property). This is implemented using smp_store_release().
> 
> thanks.
> -- 
> ~Randy