update x86 and ppc atomic ops

[urcu.git] / urcu-static.h

diff --git a/urcu-static.h b/urcu-static.h
index f1aab2f2dfed0123070e3d608fec00701a3b364d..0c6891e77ca0871f85cd1461333c0facd95bcf13 100644 (file)
--- a/urcu-static.h
+++ b/urcu-static.h
@@ -76,6 +76,15 @@
  * Inserts memory barriers on architectures that require them (currently only
  * Alpha) and documents which pointers are protected by RCU.
  *
+ * The compiler memory barrier in LOAD_SHARED() ensures that value-speculative
+ * optimizations (e.g. VSS: Value Speculation Scheduling) does not perform the
+ * data read before the pointer read by speculating the value of the pointer.
+ * Correct ordering is ensured because the pointer is read as a volatile access.
+ * This acts as a global side-effect operation, which forbids reordering of
+ * dependent memory operations. Note that such concern about dependency-breaking
+ * optimizations will eventually be taken care of by the "memory_order_consume"
+ * addition to forthcoming C++ standard.
+ *
  * Should match rcu_assign_pointer() or rcu_xchg_pointer().
  */
 
@@ -108,6 +117,12 @@
  */
 #define KICK_READER_LOOPS 10000
 
+#ifdef DEBUG_RCU
+#define rcu_assert(args...)    assert(args)
+#else
+#define rcu_assert(args...)
+#endif
+
 #ifdef DEBUG_YIELD
 #include <sched.h>
 #include <time.h>
@@ -117,8 +132,11 @@
 #define YIELD_READ     (1 << 0)
 #define YIELD_WRITE    (1 << 1)
 
-/* Updates without DEBUG_FULL_MB are much slower. Account this in the delay */
-#ifdef DEBUG_FULL_MB
+/*
+ * Updates without URCU_MB are much slower. Account this in
+ * the delay.
+ */
+#ifdef URCU_MB
 /* maximum sleep delay, in us */
 #define MAX_SLEEP 50
 #else
@@ -161,7 +179,7 @@ static inline void debug_yield_init(void)
 }
 #endif
 
-#ifdef DEBUG_FULL_MB
+#ifdef URCU_MB
 static inline void reader_barrier()
 {
        smp_mb();
@@ -212,22 +230,16 @@ static inline void _rcu_read_lock(void)
 
        tmp = urcu_active_readers;
        /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
-       /*
-        * The data dependency "read urcu_gp_ctr, write urcu_active_readers",
-        * serializes those two memory operations. The memory barrier in the
-        * signal handler ensures we receive the proper memory commit barriers
-        * required by _STORE_SHARED and _LOAD_SHARED whenever communication
-        * with the writer is needed.
-        */
-       if (likely(!(tmp & RCU_GP_CTR_NEST_MASK)))
+       if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
                _STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr));
-       else
+               /*
+                * Set active readers count for outermost nesting level before
+                * accessing the pointer. See force_mb_all_threads().
+                */
+               reader_barrier();
+       } else {
                _STORE_SHARED(urcu_active_readers, tmp + RCU_GP_COUNT);
-       /*
-        * Increment active readers count before accessing the pointer.
-        * See force_mb_all_threads().
-        */
-       reader_barrier();
+       }
 }
 
 static inline void _rcu_read_unlock(void)
@@ -236,6 +248,9 @@ static inline void _rcu_read_unlock(void)
        /*
         * Finish using rcu before decrementing the pointer.
         * See force_mb_all_threads().
+        * Formally only needed for outermost nesting level, but leave barrier
+        * in place for nested unlocks to remove a branch from the common case
+        * (no nesting).
         */
        _STORE_SHARED(urcu_active_readers, urcu_active_readers - RCU_GP_COUNT);
 }
@@ -263,7 +278,7 @@ static inline void _rcu_read_unlock(void)
 
 /**
  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
- * pointer to the data structure, which can be safely freed after waitin for a
+ * pointer to the data structure, which can be safely freed after waiting for a
  * quiescent state using synchronize_rcu().
  */