Hi Paul,

On Mon, Feb 01, 2016 at 07:54:58PM -0800, Paul E. McKenney wrote:
> On Mon, Feb 01, 2016 at 01:56:22PM +0000, Will Deacon wrote:
> > On Fri, Jan 29, 2016 at 02:22:53AM -0800, Paul E. McKenney wrote:
> > > On Fri, Jan 29, 2016 at 09:59:59AM +0000, Will Deacon wrote:
> > > > On Thu, Jan 28, 2016 at 02:31:31PM -0800, Paul E. McKenney wrote:
> > > 
> > > [ . . . ]
> > > 
> > > > > For Linux in general, this is a question: How strict do we want to be
> > > > > about matching the type of write with the corresponding read?  My
> > > > > default approach is to initially be quite strict and loosen as needed.
> > > > > Here "quite strict" might mean requiring an rcu_assign_pointer() for
> > > > > the write and rcu_dereference() for the read, as opposed to (say)
> > > > > ACCESS_ONCE() for the read.  (I am guessing that this would be too
> > > > > tight, but it makes a good example.)
> > > > > 
> > > > > Thoughts?
> > > > 
> > > > That sounds broadly sensible to me and allows rcu_assign_pointer and
> > > > rcu_dereference to be used as drop-in replacements for release/acquire
> > > > where local transitivity isn't required. However, I don't think we can
> > > > rule out READ_ONCE/WRITE_ONCE interactions as they seem to be used
> > > > already in things like the osq_lock (albeit without the address
> > > > dependency).
> > > 
> > > Agreed.  So in the most strict case that I can imagine anyone putting
> > > up with, we have the following pairings:
> > 
> > I think we can group these up:
> > 
> > Locally transitive:
> > 
> > > o	smp_store_release() -> smp_load_acquire() (locally transitive)
> > 
> > Locally transitive chain termination:
> > 
> > (i.e. these can't be used to extend a chain)
> 
> Agreed.
> 
> > > o	smp_store_release() -> lockless_dereference() (???)
> > > o	rcu_assign_pointer() -> rcu_dereference()
> > > o	smp_store_release() -> READ_ONCE(); if

Just want to make sure, this one is actually:

o	smp_store_release() -> READ_ONCE(); if ;<WRITE_ONCE()>

right? Because control dependency only orders READ->WRITE.

If so, do we also need to take the following pairing into consideration?

o	smp_store_release() -> READ_ONCE(); if ;smp_rmb(); <ACCESS_ONCE()>

> 
> I am OK with the first and last, but I believe that the middle one
> has real use cases.  So the rcu_assign_pointer() -> rcu_dereference()
> case needs to be locally transitive.
> 

Hmm... I don't think we should differ rcu_dereference() and
lockless_dereference(). One reason: list_for_each_entry_rcu() are using
lockless_dereference() right now, which means we used to think
rcu_dereference() and lockless_dereference() are interchangeable, right?

Besides, Will, what's the reason of having a locally transitive chain
termination? Because on some architectures RELEASE->DEPENDENCY pairs may
not be locally transitive?

Regards,
Boqun

> > Globally transitive:
> > 
> > > o	smp_mb(); WRITE_ONCE() -> READ_ONCE(); (globally transitive)
> > > o	synchronize_rcu(); WRITE_ONCE() -> READ_ONCE(); (globally transitive)
> > 
> > RCU:
> > 
> > > o	synchronize_rcu(); WRITE_ONCE() -> rcu_read_lock(); READ_ONCE()
> > > 		(strange and wonderful properties)
> 
> Agreed.
> 
> > > Seem reasonable, or am I missing some?
> > 
> > Looks alright to me.
> 
> So I have some litmus tests to generate.  ;-)
> 
> 							Thnax, Paul
>