Re: [RFC v2 3/7] powerpc: atomic: Implement atomic{,64}_{add,sub}_return_* variants

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

On Wed, Sep 23, 2015 at 08:07:55AM +0800, Boqun Feng wrote:
> On Tue, Sep 22, 2015 at 08:25:40AM -0700, Paul E. McKenney wrote:
> > On Tue, Sep 22, 2015 at 07:37:04AM +0800, Boqun Feng wrote:
> > > On Tue, Sep 22, 2015 at 07:26:56AM +0800, Boqun Feng wrote:
> > > > On Mon, Sep 21, 2015 at 11:24:27PM +0100, Will Deacon wrote:
> > > > > Hi Boqun,
> > > > > 
> > > > > On Sun, Sep 20, 2015 at 09:23:03AM +0100, Boqun Feng wrote:
> > > > > > On Sat, Sep 19, 2015 at 11:33:10PM +0800, Boqun Feng wrote:
> > > > > > > On Fri, Sep 18, 2015 at 05:59:02PM +0100, Will Deacon wrote:
> > > > > > > > On Wed, Sep 16, 2015 at 04:49:31PM +0100, Boqun Feng wrote:
> > > > > > > > > On powerpc, we don't need a general memory barrier to achieve acquire and
> > > > > > > > > release semantics, so __atomic_op_{acquire,release} can be implemented
> > > > > > > > > using "lwsync" and "isync".
> > > > > > > > 
> > > > > > > > I'm assuming isync+ctrl isn't transitive, so we need to get to the bottom
> > > > > > > 
> > > > > > > Actually the transitivity is still guaranteed here, I think ;-)
> > > > > 
> > > > > The litmus test I'm thinking of is:
> > > > > 
> > > > > 
> > > > > {
> > > > > 0:r2=x;
> > > > > 1:r2=x; 1:r5=z;
> > > > > 2:r2=z; 2:r4=x;
> > > > > }
> > > > >  P0           | P1            | P2           ;
> > > > >  li r1,1      | lwz r1,0(r2)  | lwz r1,0(r2) ;
> > > > >  stw r1,0(r2) | cmpw r1,r1    | cmpw r1,r1   ;
> > > > >               | beq LC00      | beq  LC01    ;
> > > > >               | LC00:         | LC01:        ;
> > > > >               | isync         | isync        ;
> > > > >               | li r4,1       | lwz r3,0(r4) ;
> > > > >               | stw r4,0(r5)  |              ;
> > > > > exists
> > > > > (1:r1=1 /\ 2:r1=1 /\ 2:r3=0)
> > > > > 
> > > > > 
> > > > > Which appears to be allowed. I don't think you need to worry about backwards
> > > > > branches for the ctrl+isync construction (none of the current example do,
> > > > > afaict).
> > > > > 
> > > > 
> > > > Yes.. my care of backwards branches is not quite related to the topic, I
> > > > concerned that mostly because my test is using atomic operation, and I
> > > > just want to test the exact asm code.
> > > > 
> > > > > Anyway, all the problematic cases seem to arise when we start mixing
> > > > > ACQUIRE/RELEASE accesses with relaxed accesses (i.e. where an access from
> > > > > one group reads from an access in the other group). It would be simplest
> > > > > to say that this doesn't provide any transitivity guarantees, and that
> > > > > an ACQUIRE must always read from a RELEASE if transitivity is required.
> > > > > 
> > > > 
> > > > Agreed. RELEASE alone doesn't provide transitivity and transitivity is
> > >           ^^^^^^^
> > > This should be ACQUIRE...
> > > 
> > > > guaranteed only if an ACQUIRE read from a RELEASE. That's exactly the
> > > > direction which the link (https://lkml.org/lkml/2015/9/15/836) is
> > > > heading to. So I think we are fine here to use ctrl+isync here, right?
> > 
> > We are going to have to err on the side of strictness, that is, having
> > the documentation place more requirements on the developer than the union
> > of the hardware does.  Besides, I haven't heard any recent complaints
> > that memory-barriers.txt is too simple.  ;-)
> 
> Agreed ;-)
> 
> For atomic operations, using isync in ACQUIRE operations does gaurantee
> that a pure RELEASE/ACQUIRE chain provides transitivity. So, again, I
> think we are fine here to use isync in ACQUIRE atomic operations,
> unless you think we need to be more strict, i.e, making ACQUIRE itself
> provide transitivy?

As I understand it, either isync or lwsync suffices, with the choice
depending on the hardware.  The kernel will rewrite itself at boot time
if you use the appropriate macro.  ;-)

							Thanx, Paul