Re: [RFC][PATCH 7/7] kref: Implement using refcount_t

[Peter Zijlstra]

On Tue, Nov 15, 2016 at 10:19:09PM +0800, Boqun Feng wrote:
> On Tue, Nov 15, 2016 at 02:01:54PM +0100, Peter Zijlstra wrote:
> > On Tue, Nov 15, 2016 at 08:33:37PM +0800, Boqun Feng wrote:
> > > Hi Peter,
> > > 
> > > On Mon, Nov 14, 2016 at 06:39:53PM +0100, Peter Zijlstra wrote:
> > > [...]
> > > > +/*
> > > > + * Similar to atomic_dec_and_test(), it will BUG on underflow and fail to
> > > > + * decrement when saturated at UINT_MAX.
> > > > + *
> > > > + * Provides release memory ordering, such that prior loads and stores are done
> > > > + * before a subsequent free.
> > > 
> > > I'm not sure this is correct, the RELEASE semantics is for the STORE
> > > part of cmpxchg, and semantically it will guarantee that memory
> > > operations after cmpxchg won't be reordered upwards, for example, on
> > > ARM64, the following code:
> > > 
> > > 	WRITE_ONCE(x, 1)
> > > 	
> > > 	atomic_cmpxchg_release(&a, 1, 2);
> > > 	  r1 = ll(&a)
> > > 	  if (r1 == 1) {
> > > 	    sc_release(&a, 2);
> > > 	  }
> > > 	
> > > 	free()
> > > 
> > > could be reordered as, I think:
> > > 
> > > 	atomic_cmpxchg_release(&a, 1, 2);
> > > 	  r1 = ll(&a)
> > > 	  if (r1 == 1) {
> > > 	    free()
> > > 	    WRITE_ONCE(x, 1)
> > > 	    sc_release(&a, 2);
> > > 	  }
> > > 
> > > Of course, we need to wait for Will to confirm about this. But if this
> > > could happen, we'd better to use a smp_mb()+atomic_cmpxchg_relaxed()
> > > here and for other refcount_dec_and_*().
> > 
> > Can't happen I think because of the control dependency between
> > dec_and_test() and free().
> > 
> > That is, the cmpxchg_release() must complete to determine if it was
> > successful or it needs a retry. The success, combined with the state of
> > the variable will then determine if we call free().
> > 
> 
> The thing is that determination of the variable's state(i.e.
> store_release() succeeds) and the actual writeback to memory are two
> separate events. So yes, free() won't execute before store_release()
> commits successfully, but there is no barrier here to order the memory
> effects of store_release() and free().

Doesn't matter. If we dropped the refcount to 0, nobody else will be
observing this memory anymore (unless ill-formed program). The only
thing we need is that the free() will not be speculated.

This is because all RmW on a specific variable, irrespective of their
memory ordering on other loads/stores, are totally ordered against one
another.

> But as I said, we actually only need the pairing of orderings:
> 
> 1) load part of cmpxchg -> free() 
> 2) object accesses -> store part of cmpxchg
> 
> Ordering #1 can be achieved via control dependency as you pointed out
> that free()s very much includes stores. And ordering #2 can be achieved
> with RELEASE.
> 
> So the code is right, I just thought the comment may be misleading. The
> reason we use cmpxchg_release() is just for achieving ordering #2, and
> not to order "prior loads and stores" with "a subsequent free".
> 
> Am I missing some subtle orderings here?

I would want to further quality 1), it must be no earlier than the load
of the last / successful ll/sc round.

At that point we're guaranteed a reference count of 1 that _will_ drop
to 0, and thus nobody else (should) reference that memory anymore.

If we agree on this, I'll update the comment :-) Will, do you too agree?