Virtualization - Dec 2017 - [PATCH tip/core/rcu 21/21] drivers/vhost: Remove now-redundant read_barrier_depends()

On Tue, Dec 05, 2017 at 09:24:21PM +0200, Michael S. Tsirkin
wrote:
> On Tue, Dec 05, 2017 at 08:17:33PM +0100, Peter Zijlstra wrote:
> > On Tue, Dec 05, 2017 at 08:57:46PM +0200, Michael S. Tsirkin wrote:
> > 
> > > I don't see WRITE_ONCE inserting any barriers, release or
> > > write.
> > 
> > Correct, never claimed there was.
> > 
> > Just saying that:
> > 
> > 	obj = READ_ONCE(*foo);
> > 	val = READ_ONCE(obj->val);
> > 
> > Never needs a barrier (except on Alpha and we want to make that go
> > away). Simply because a CPU needs to complete the load of @obj before
it
> > can compute the address &obj->val. Thus the second load _must_
come
> > after the first load and we get LOAD-LOAD ordering.
> > 
> > Alpha messing that up is a royal pain, and Alpha not being an
> > active/living architecture is just not worth the pain of keeping this
in
> > the generic model.
> > 
> 
> Right. What I am saying is that for writes you need
> 
> WRITE_ONCE(obj->val, 1);
> smp_wmb();
> WRITE_ONCE(*foo, obj);
I believe Peter was instead suggesting:

WRITE_ONCE(obj->val, 1);
smp_store_release(foo, obj);
> and this barrier is no longer paired with anything until
> you realize there's a dependency barrier within READ_ONCE.
> 
> Barrier pairing was a useful tool to check code validity,
> maybe there are other, better tools now.
There are quite a few people who say that smp_store_release() is
easier for the tools to analyze than is smp_wmb().  My experience with
smp_read_barrier_depends() and rcu_dereference() leads me to believe
that they are correct.

							Thanx, Paul

On Tue, Dec 05, 2017 at 11:33:39AM -0800, Paul E. McKenney
wrote:
> On Tue, Dec 05, 2017 at 09:24:21PM +0200, Michael S. Tsirkin wrote:
> > On Tue, Dec 05, 2017 at 08:17:33PM +0100, Peter Zijlstra wrote:
> > > On Tue, Dec 05, 2017 at 08:57:46PM +0200, Michael S. Tsirkin
wrote:
> > > 
> > > > I don't see WRITE_ONCE inserting any barriers, release
or
> > > > write.
> > > 
> > > Correct, never claimed there was.
> > > 
> > > Just saying that:
> > > 
> > > 	obj = READ_ONCE(*foo);
> > > 	val = READ_ONCE(obj->val);
> > > 
> > > Never needs a barrier (except on Alpha and we want to make that
go
> > > away). Simply because a CPU needs to complete the load of @obj
before it
> > > can compute the address &obj->val. Thus the second load
_must_ come
> > > after the first load and we get LOAD-LOAD ordering.
> > > 
> > > Alpha messing that up is a royal pain, and Alpha not being an
> > > active/living architecture is just not worth the pain of keeping
this in
> > > the generic model.
> > > 
> > 
> > Right. What I am saying is that for writes you need
> > 
> > WRITE_ONCE(obj->val, 1);
> > smp_wmb();
> > WRITE_ONCE(*foo, obj);
> 
> I believe Peter was instead suggesting:
> 
> WRITE_ONCE(obj->val, 1);
> smp_store_release(foo, obj);
Isn't that more expensive though?

> > and this barrier is no longer paired with anything until
> > you realize there's a dependency barrier within READ_ONCE.
> > 
> > Barrier pairing was a useful tool to check code validity,
> > maybe there are other, better tools now.
> 
> There are quite a few people who say that smp_store_release() is
> easier for the tools to analyze than is smp_wmb().  My experience with
> smp_read_barrier_depends() and rcu_dereference() leads me to believe
> that they are correct.
> 
> 							Thanx, Paul
OK, but smp_store_release is still not paired with anything since we
rely on READ_ONCE to include the implicit dpendendency barrier.

-- 
MST

On Tue, Dec 05, 2017 at 09:51:48PM +0200, Michael S. Tsirkin
wrote:
> > > WRITE_ONCE(obj->val, 1);
> > > smp_wmb();
> > > WRITE_ONCE(*foo, obj);
> > 
> > I believe Peter was instead suggesting:
> > 
> > WRITE_ONCE(obj->val, 1);
> > smp_store_release(foo, obj);
> 
> Isn't that more expensive though?
Depends on the architecture. The only architecture where it is more
expensive and people actually still care about is ARM I think.

On Tue, Dec 05, 2017 at 09:51:48PM +0200, Michael S. Tsirkin
wrote:
> On Tue, Dec 05, 2017 at 11:33:39AM -0800, Paul E. McKenney wrote:
> > On Tue, Dec 05, 2017 at 09:24:21PM +0200, Michael S. Tsirkin wrote:
[ . . . ]
> > > and this barrier is no longer paired with anything until
> > > you realize there's a dependency barrier within READ_ONCE.
> > > 
> > > Barrier pairing was a useful tool to check code validity,
> > > maybe there are other, better tools now.
> > 
> > There are quite a few people who say that smp_store_release() is
> > easier for the tools to analyze than is smp_wmb().  My experience with
> > smp_read_barrier_depends() and rcu_dereference() leads me to believe
> > that they are correct.
> 
> OK, but smp_store_release is still not paired with anything since we
> rely on READ_ONCE to include the implicit dpendendency barrier.
Why wouldn't you consider the smp_store_release() to be paired with
the new improved READ_ONCE()?

							Thanx, Paul