On Mon, Sep 14, 2015 at 04:38:48PM +0100, Will Deacon wrote:

On Mon, Sep 14, 2015 at 01:11:56PM +0100, Peter Zijlstra wrote:

quoted

On Mon, Sep 14, 2015 at 02:01:53PM +0200, Peter Zijlstra wrote:

quoted

The scenario is:

	CPU0			CPU1

				unlock(x)
				  smp_store_release(&x->lock, 0);

	unlock(y)
	  smp_store_release(&next->lock, 1); /* next == &y */

				lock(y)
				  while (!(smp_load_acquire(&y->lock))
					cpu_relax();


Where the lock does _NOT_ issue a store to acquire the lock at all. Now
I don't think any of our current primitives manage this, so we should be
good, but it might just be possible.

So with a bit more through this seems fundamentally impossible, you
always needs some stores in a lock() implementation, the above for
instance needs to queue itself, otherwise CPU0 will not be able to find
it etc..

Which brings us back round to separating LOCK/UNLOCK from ACQUIRE/RELEASE.

I believe that we do need to do this, unless we decide to have unlock-lock
continue to imply only acquire and release, rather than full ordering.
I believe that Mike Ellerman is working up additional benchmarking
on this.

							Thanx, Paul

If we say that UNLOCK(foo) -> LOCK(bar) is ordered but RELEASE(baz) ->
ACQUIRE(boz) is only ordered by smp_mb__release_acquire(), then I think
we're in a position where we can at least build arbitrary locks portably
out of ACQUIRE/RELEASE operations, even though I don't see any users of
that macro in the imminent future.

I'll have a crack at some documentation.

Will