Re: perf events ring buffer memory barrier on powerpc

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, May 09, 2014 at 08:20:25AM -0400, Mikulas Patocka wrote:
> 
> 
> On Fri, 9 May 2014, Victor Kaplansky wrote:
> 
> > Mikulas Patocka <mpatocka@redhat.com> wrote on 05/08/2014 11:46:53 PM:
> > 
> > > > > BTW, it is why you also don't need ACCESS_ONCE() around @tail, but only
> > > > > around
> > > > > @head read.
> > > >
> > > > Agreed, the ACCESS_ONCE() around tail is superfluous since we're the one
> > > > updating tail, so there's no problem with the value changing
> > > > unexpectedly.
> > >
> > > You need ACCESS_ONCE even if you are the only process writing the value.
> > > Because without ACCESS_ONCE, the compiler may perform store tearing and
> > > split the store into several smaller stores. Search the file
> > > "Documentation/memory-barriers.txt" for the term "store tearing", it shows
> > > an example where one instruction storing 32-bit value may be split to two
> > > instructions, each storing 16-bit value.
> > >
> > > Mikulas
> > 
> > AFAIR, I was talking about redundant ACCESS_ONCE() around @tail *read* in
> > consumer code. As for ACCESS_ONCE() around @tail write in consumer code,
> > I see your point, but I don't think that volatile imposed by ACCESS_ONCE()
> > is appropriate, since:
> > 
> >     - compiler can generate several stores despite volatile if @tail
> >     is bigger in size than native machine data size, e.g. 64-bit on
> >     a 32-bit CPU.
> 
> That's true - so you should define data_head and data_tail as "unsigned 
> long", not "__u64".
> 
> >     - volatile imposed by ACCESS_ONCE() does nothing to prevent CPU from
> >     reordering, splitting or merging accesses. It can only mediate
> >     communication problems between processes running on same CPU.
> 
> That's why you need smp barrier in addition to ACCESS_ONCE. You need both 
> - the smp barrier (to prevent the CPU from reordering) and ACCESS_ONCE (to 
> prevent the compiler from splitting the write to smaller memory accesses).

IIRC the ring-buffer code uses the fact that one element remains
empty to make clever double use of a memory barrier.

> Since Linux 3.14, there are new macros smp_store_release and 
> smp_load_acquire that combine ACCESS_ONCE and memory barrier, so you can 
> use them. (they call compiletime_assert_atomic_type to make sure that you 
> don't use them on types that are not atomic, such as long long on 32-bit 
> architectures)

These are indeed useful and often simpler to use than raw barriers.

							Thanx, Paul

> > What you really want is to guarantee *atomicity* of @tail write on consumer
> > side.
> > 
> > -- Victor
> 
> Mikulas