On Thu, Feb 22, 2018 at 11:06:36AM +0100, Peter Zijlstra wrote: > On Wed, Feb 21, 2018 at 09:42:08PM -0800, Daniel Lustig 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. > > Excuse my ignorance (also jumping in the middle of things), but how can > this be? > > spin_unlock() is a store-release, this means the write to the lock word > must happen after any stores inside the critical section. > > spin_lock() is a load-acquire + test-and-set-ctrl-dep, we'll only > proceed with the critical section if we observe the lock 'unlocked', > which also means we must observe the stores prior to the unlock. > > And both the ctrl-dep and the ACQUIRE ensure future stores cannot happen > before. > > So while the lock store and subsequent critical section stores are > unordered, I don't see how it would be possible to not be ordered > against stores from a previous critical section. >
Or are we talking about a third party observing while not partaking in the lock-chain? Then I agree, the stores can be observed out of order by this 3rd actor.