lib/ovs-atomic-gcc4+: Use 'volatile' to enforce memory access.

Use 'volatile' to enforce a new memory access on each lockless atomic
store and read.  Without this a loop consisting of an atomic_read with
memory_order_relaxed would be simply optimized away.  Also, using
volatile is cheaper than adding a full compiler barrier (also) in that
case.

This use of a volatile cast mirrors the Linux kernel ACCESS_ONCE macro.

Without this change the more rigorous atomic test cases introduced in
a following patch will hang due to the atomic accesses being optimized
away.

Signed-off-by: Jarno Rajahalme <jrajahalme@nicira.com>
Acked-by: Ben Pfaff <blp@nicira.com>

diff --git a/lib/ovs-atomic-gcc4+.h b/lib/ovs-atomic-gcc4+.h
index 756696b..bb08ff9 100644 (file)
--- a/lib/ovs-atomic-gcc4+.h
+++ b/lib/ovs-atomic-gcc4+.h
@@ -83,7 +83,7 @@ atomic_signal_fence(memory_order order)
                                                         \
         if (IS_LOCKLESS_ATOMIC(*dst__)) {               \
             atomic_thread_fence(ORDER);                 \
-            *dst__ = src__;                             \
+            *(typeof(*DST) volatile *)dst__ = src__;    \
             atomic_thread_fence_if_seq_cst(ORDER);      \
         } else {                                        \
             atomic_store_locked(dst__, src__);          \
@@ -99,7 +99,7 @@ atomic_signal_fence(memory_order order)
                                                         \
         if (IS_LOCKLESS_ATOMIC(*src__)) {               \
             atomic_thread_fence_if_seq_cst(ORDER);      \
-            *dst__ = *src__;                            \
+            *dst__ = *(typeof(*SRC) volatile *)src__;   \
         } else {                                        \
             atomic_read_locked(src__, dst__);           \
         }                                               \