Re: [PATCH tip/locking/core 0/6] compiler-context-analysis: Scoped init guards

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jan 22, 2026 at 07:30:42AM +0100, Christoph Hellwig wrote:

> That's better.  What would be even better for everyone would be:
> 
> 	mutex_prepare(&obj->mutex); /* acquire, but with a nice name */
> 	obj->data = FOO;
> 	mutex_init_prepared(&obj->mutex); /* release, barrier, actual init */
> 
> 	mutex_lock(&obj->mutex); /* IFF needed only */
> 

This is cannot work. There is no such thing is a release-barrier.
Furthermore, store-release, load-acquire needs an address dependency to
work.

When publishing an object, which is what we're talking about, we have
two common patterns:

 1) a locked data-structure

 2) RCU


The way 1) works is:

	Publish				Use

	lock(&structure_lock);
	insert(&structure, obj);
	unlock(&structure_lock);

					lock(&structure_lock)
					obj = find(&structure, key);
					...
					unlock(&structure_lock);

And here the Publish-unlock is a release which pairs with the Use-lock's
acquire and guarantees that Use sees both 'structure' in a coherent
state and obj as it was at the time of insertion. IOW we have
release-acquire through the &structure_lock pointer.

The way 2) works is:

	Publish				Use

	lock(&structure_lock);
	insert(&structure, obj)
	   rcu_assign_pointer(ptr, obj);
	unlock(&structure_lock);
	  	
					rcu_read_lock();
					obj = find_rcu(&structure, key);
					...
					rcu_read_unlock();


And here rcu_assign_pointer() is a store-release that pairs with an
rcu_dereference() inside find_rcu() on the same pointer.

There is no alternative way to order things, there must be a
release-acquire through a common address.

In both cases it is imperative the obj is fully (or full enough)
initialized before publication, because the consumer is only guaranteed
to see the state of the object it was in at publish time.