[PATCH] rbtree: use READ_ONCE in RB_EMPTY_ROOT

[PATCH] rbtree: use READ_ONCE in RB_EMPTY_ROOT

[Davidlohr Bueso]

With d72da4a4d97 (rbtree: Make lockless searches non-fatal) our rbtrees
provide weak guarantees that allows us to do lockless (and very speculative)
reads of the tree. Such readers cannot see partial stores on nodes, ie
left/right as well as root. As such, similar to the WRITE_ONCE semantics when
doing rotations, use READ_ONCE when checking the root node in RB_EMPTY_ROOT.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
---
 include/linux/rbtree.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/rbtree.h b/include/linux/rbtree.h
index a5aa7ae..b690009 100644
--- a/include/linux/rbtree.h
+++ b/include/linux/rbtree.h
@@ -50,7 +50,7 @@ struct rb_root {
 #define RB_ROOT	(struct rb_root) { NULL, }
 #define	rb_entry(ptr, type, member) container_of(ptr, type, member)
 
-#define RB_EMPTY_ROOT(root)  ((root)->rb_node == NULL)
+#define RB_EMPTY_ROOT(root)  (READ_ONCE((root)->rb_node) == NULL)
 
 /* 'empty' nodes are nodes that are known not to be inserted in an rbtree */
 #define RB_EMPTY_NODE(node)  \
-- 
2.1.4

Re: [PATCH] rbtree: use READ_ONCE in RB_EMPTY_ROOT

[Peter Zijlstra]

On Mon, Jan 11, 2016 at 10:22:43AM -0800, Davidlohr Bueso wrote:
> With d72da4a4d97 (rbtree: Make lockless searches non-fatal) our rbtrees
> provide weak guarantees that allows us to do lockless (and very speculative)
> reads of the tree. Such readers cannot see partial stores on nodes, ie
> left/right as well as root. As such, similar to the WRITE_ONCE semantics when
> doing rotations, use READ_ONCE when checking the root node in RB_EMPTY_ROOT.

No objection, but is this actually used anywhere?

Or is this a just-in-case completeness thing?

Re: [PATCH] rbtree: use READ_ONCE in RB_EMPTY_ROOT

[Davidlohr Bueso]

On Tue, 12 Jan 2016, Peter Zijlstra wrote:

>On Mon, Jan 11, 2016 at 10:22:43AM -0800, Davidlohr Bueso wrote:
>> With d72da4a4d97 (rbtree: Make lockless searches non-fatal) our rbtrees
>> provide weak guarantees that allows us to do lockless (and very speculative)
>> reads of the tree. Such readers cannot see partial stores on nodes, ie
>> left/right as well as root. As such, similar to the WRITE_ONCE semantics when
>> doing rotations, use READ_ONCE when checking the root node in RB_EMPTY_ROOT.
>
>No objection, but is this actually used anywhere?

I found this because I wanted to use the waiter check in rtmutexes in a lockless
fashion (ie rt_mutex_has_waiters).

>
>Or is this a just-in-case completeness thing?

This too.

Thanks,
Davidlohr

Re: [PATCH] rbtree: use READ_ONCE in RB_EMPTY_ROOT

[Peter Zijlstra]

On Tue, Jan 12, 2016 at 07:09:46AM -0800, Davidlohr Bueso wrote:
> On Tue, 12 Jan 2016, Peter Zijlstra wrote:
> 
> >On Mon, Jan 11, 2016 at 10:22:43AM -0800, Davidlohr Bueso wrote:
> >>With d72da4a4d97 (rbtree: Make lockless searches non-fatal) our rbtrees
> >>provide weak guarantees that allows us to do lockless (and very speculative)
> >>reads of the tree. Such readers cannot see partial stores on nodes, ie
> >>left/right as well as root. As such, similar to the WRITE_ONCE semantics when
> >>doing rotations, use READ_ONCE when checking the root node in RB_EMPTY_ROOT.
> >
> >No objection, but is this actually used anywhere?
> 
> I found this because I wanted to use the waiter check in rtmutexes in a lockless
> fashion (ie rt_mutex_has_waiters).

OK, fair enough.

Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>