Re: [RFC] Control dependencies

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Nov 22, 2013 at 12:59:41PM -0500, Peter Hurley wrote:
> On 11/22/2013 08:46 AM, Peter Zijlstra wrote:
> >How about the below version?
> >
> >---
> >--- a/kernel/events/ring_buffer.c
> >+++ b/kernel/events/ring_buffer.c
> >@@ -61,19 +61,20 @@ static void perf_output_put_handle(struc
> >  	 *
> >  	 *   kernel				user
> >  	 *
> >-	 *   READ ->data_tail			READ ->data_head
> >-	 *   smp_mb()	(A)			smp_rmb()	(C)
> >-	 *   WRITE $data			READ $data
> >-	 *   smp_wmb()	(B)			smp_mb()	(D)
> >-	 *   STORE ->data_head			WRITE ->data_tail
> >+	 *   if (LOAD ->data_tail) {		LOAD ->data_head
> >+	 *			(A)		smp_rmb()	(C)
> >+	 *	STORE $data			LOAD $data
> >+	 *	smp_wmb()	(B)		smp_mb()	(D)
> >+	 *	STORE ->data_head		STORE ->data_tail
> 
> 
> I wasn't subscribed to linux-arch so missed the smp_store_release()
> outcome, if there was one.
> 
> Are (B) and (D) still slated for changing to STORE.rel semantics,
> aka smp_store_release()?

The earlier proposal would have A and C be smp_load_acquire() and B and
D be smp_store_release().

> I realize that, for the perf ring buffer, (D) is in userspace but
> I'm also interested in non-perf situations where (D) would be in the
> kernel.

So we're still debating the exact semantics of smp_store_release(), it
now looks like it needs a heavier memory barrier than previously
thought. In which case using it wouldn't make sense for me anymore.

Note that C and D are in userspace and not in any hot path (usually)
They're only issued once to read an entire buffer backlog at once, so I
don't really care about them all that much.

A and B otoh are in kernel space and are issued for every single event
written, so I'm interested to get them as cheaply as possible.

With this proposed patch, we remove a full barrier, with the earlier
smp_load_acquire() / smp_store_release() patches we would only
downgrade the full barrier to an acquire barrier, which is still more
than no barrier at all.

And now it looks like the smp_store_release() would actually upgrade the
wmb to a full barrier on some systems at least.