On Thu, Feb 22, 2018 at 11:15:04AM +0100, Peter Zijlstra wrote:
> On Thu, Feb 22, 2018 at 02:58:47PM +0800, Boqun Feng wrote:
> > > And yes, if we go with a purely RCpc interpretation of acquire and
> > > release, then I don't believe the writes in the previous critical
> > > section would be ordered with the writes in the subsequent critical
> > > section.  That's really all the argument boils down to.  We'd like
> > 
> > I think atomics in Linux kernel(and in LKMM) are purely RCpc, right?
> > Alan and Andrea? 
> > 
> > And we are not going to change it, are we?
> > 
> > If atomics in Linux kernel are purely RCpc, then it cerntainly makes
> > sense for riscv to provide purely RCpc atomics.
> 
> So there's 3 things:
> 
>   smp_load_acquire(), smp_store_release()
> 
>   atomic*_{acquire,release}()
> 
>   *_{lock,unlock}();
> 
> Which are all quite distinct.
> 
> smp_load_acquire() and smp_store_release() are RCpc, and there is no
> discussion of ever wanting to change that.
> 
> The atomics also follow this and are RCpc, in fact the RELEASE only
> applies to the STORE and the ACQUIRE only applies to the LOAD of the
> atomics.
> 
> The locking primitives OTOH we would really rather like to be RCsc, and
> everybody except PPC has them as such, PPC being the only one having
> RCpc locks.
> 
> I read the part you quoted from Daniel as being purely about spinlocks,
> the 'critical section' wording being a dead give away, so I'm then
> somewhat confused why you talk about atomics.

Maybe it's me who misunderstand Daniel's words. But my understanding is
that riscv people are on a debate about whether their "RCpc" atomic
instructions need to be more strict: release+acquire pair orders two
writes. And I thought that atomics(including RmW atomics) in kernel only
have purely RCpc semantics, which I needed to check with you guy. And if
I'm right, it's cerntainly fine for riscv "RCpc" instruction to be
purely RCpc.

Note that even on PPC, the release+acquire pair of atomics orders writes
before and after, and on x86, writes are ordered since it's TSO. So
strictly speaking, I think our current implementation of atomics are a
little more strict than purely RCpc. If we think this is an requirement
for implementation of atomic primitives, than the current version of
riscv's "RCpc" atomics don't suffice.

Regards,
Boqun