On Mon, May 24, 2010 at 12:35:21PM +0300, Gleb Natapov wrote:
On Thu, May 20, 2010 at 04:36:32PM +0300, Michael S. Tsirkin wrote:
diff --git a/src/virtio-ring.h b/src/virtio-ring.h index 3fb86fe..8b546f4 100644 --- a/src/virtio-ring.h +++ b/src/virtio-ring.h @@ -9,8 +9,9 @@
#define virt_to_phys(v) (unsigned long)(v) #define phys_to_virt(p) (void*)(p) -#define wmb() barrier() -#define mb() barrier() +/* Compiler barrier is enough as an x86 CPU does not reorder reads or writes */ +#define smp_rmb() barrier() +#define smp_wmb() barrier()
I thought you were going to use real memory barriers (although we concluded compiler barrier should be enough, but it's better to be safe then sorry).
-- Gleb.
On x86, ops of the same kind are ordered by CPU, so all we need to do is 1. avoid depending on order between reads and writes 2. prevent the compiler from reordering memory accesses