On Mon, Feb 10, 2014 at 07:12:03PM +0100, Nicholas Mc Guire wrote:

On Mon, 10 Feb 2014, Peter Zijlstra wrote:

quoted

On Mon, Feb 10, 2014 at 06:17:12PM +0100, Nicholas Mc Guire wrote:

quoted

maybe I'm missing/missunderstanding something here but
pi_unlock -> arch_spin_unlock is a full mb() 

Nope, arch_spin_unlock() on x86 is a single add[wb] without LOCK prefix.

The lock and unlock primitives are in general specified to have ACQUIRE
resp. RELEASE semantics.

See Documentation/memory-barriers.txt for far too much head-hurting
details.

I did check that - but from the code check it seems to me to be using a
lock prefix in the fast __add() path and an explicit smp_add() in the slow
path (arch/x86/include/asm/spinlock.h:arch_spin_unlock) the __add from 
arch/x86/include/asm/cmpxchg.h does lock or am I missinterpreting this ?
the other archs I believe were all doing explicit mb()/smp_mb() in the 
arch_spin_unlock - will go check this again.

It uses UNLOCK_LOCK_PREFIX, which if you look carefully, is normally
always "". Only some 'broken' i386 chips require a LOCK there.