Re: [Ksummit-discuss] [TECH TOPIC] asm-generic implementations of low-level synchronisation constructs

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

On Thu, May 08, 2014 at 04:13:31PM +0100, Will Deacon wrote:
> On Thu, May 08, 2014 at 03:27:34PM +0100, Peter Zijlstra wrote:
> > On Thu, May 08, 2014 at 11:13:12AM +0200, Peter Zijlstra wrote:
> > > ATOMIC_RET(ptr, __ret, stmt)
> > > ({
> > > 	typeof(*ptr) __new, __val;
> > > 
> > > 	smp_mb__before_llsc();
> > > 
> > > 	do {
> > > 		__val = load_locked(ptr);
> > > 		stmt;
> > > 	} while (!store_conditional(ptr, __new));
> > > 
> > > 	smp_mb__after_llsc();
> > > 
> > > 	__ret;
> > > })
> > 
> > So the most common constraint (which you've confirmed is true for ARM as
> > well) is that we should not have memory accesses in between an LL/SC.
> 
> Yup.
> 
> > Making sure GCC doesn't do any is tricky, the best I can come up with is
> > tagging all variables with the register qualifier, like:
> > 
> > ATOMIC_RET(ptr, __ret, stmt)
> > ({
> > 	register typeof(*ptr) __new, __val;
> > 
> > 	smp_mb__before_llsc();
> > 
> > 	do {
> > 		__val = load_locked(ptr);
> > 		stmt;
> > 	} while (!store_conditional(ptr, __new));
> > 
> > 	smp_mb__after_llsc();
> > 
> > 	__ret;
> > })
> > 
> > Now, I'm not at all sure if register still means anything to GCC, but in
> > the faint hope that it still sees it as a hint this might just work.
> 
> I just ran this past our compiler guys and they threw rocks at me. Even if
> it happens to work today, I don't think we can rely on it tomorrow,
> unfortunately.
> 
> I think that makes the case for extended the fine-grained atomics
> implemented by each architecture, but it would still be great to have a way
> to compose them.

I suppose that we could attempt to apply the same structure to the asms,
but it is not clear that this would be a win.  There is some benefit to
having the assembly for each primitive laid out in one place, even if
it does mean quite a bit of duplicate code.

But yes, it sure feels like there should be some better way to do this...

							Thanx, Paul