X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.h;h=92b31df603d157ffa942534af1d5731d912c2342;hp=444d9cdba3346d987070dfa8fc68b78df4fe923f;hb=b715b99eb0d52e0744bfea46c62849f260a5ec58;hpb=6e32665beaad7fef48e9ab3c8c4cda2e9c232dde

diff --git a/urcu.h b/urcu.h
index 444d9cd..92b31df 100644
--- a/urcu.h
+++ b/urcu.h
@@ -31,6 +31,19 @@
 #define rmb()   asm volatile("lfence":::"memory")
 #define wmb()   asm volatile("sfence" ::: "memory")
 
+/* Assume SMP machine, given we don't have this information */
+#define CONFIG_SMP 1
+
+#ifdef CONFIG_SMP
+#define smp_mb()	mb()
+#define smp_rmb()	rmb()
+#define smp_wmb()	wmb()
+#else
+#define smp_mb()	barrier()
+#define smp_rmb()	barrier()
+#define smp_wmb()	barrier()
+#endif
+
 static inline void atomic_inc(int *v)
 {
 	asm volatile("lock; incl %0"
@@ -121,10 +134,19 @@ static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
 #include <sched.h>
 #include <time.h>
 #include <pthread.h>
+#include <unistd.h>
 
 #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
+/* maximum sleep delay, in us */
+#define MAX_SLEEP 50
+#else
+#define MAX_SLEEP 30000
+#endif
+
 extern unsigned int yield_active;
 extern unsigned int __thread rand_yield;
 
@@ -132,14 +154,14 @@ static inline void debug_yield_read(void)
 {
 	if (yield_active & YIELD_READ)
 		if (rand_r(&rand_yield) & 0x1)
-			sched_yield();
+			usleep(rand_r(&rand_yield) % MAX_SLEEP);
 }
 
 static inline void debug_yield_write(void)
 {
 	if (yield_active & YIELD_WRITE)
 		if (rand_r(&rand_yield) & 0x1)
-			sched_yield();
+			usleep(rand_r(&rand_yield) % MAX_SLEEP);
 }
 
 static inline void debug_yield_init(void)
@@ -161,6 +183,18 @@ static inline void debug_yield_init(void)
 }
 #endif
 
+#ifdef DEBUG_FULL_MB
+static inline void read_barrier()
+{
+	smp_mb();
+}
+#else
+static inline void read_barrier()
+{
+	barrier();
+}
+#endif
+
 /*
  * The trick here is that RCU_GP_CTR_BIT must be a multiple of 8 so we can use a
  * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits.
@@ -179,17 +213,21 @@ extern long urcu_gp_ctr;
 
 extern long __thread urcu_active_readers;
 
-static inline int rcu_old_gp_ongoing(int *value)
+static inline int rcu_old_gp_ongoing(long *value)
 {
 	long v;
 
 	if (value == NULL)
 		return 0;
 	debug_yield_write();
+	/*
+	 * Make sure both tests below are done on the same version of *value
+	 * to insure consistency.
+	 */
 	v = ACCESS_ONCE(*value);
 	debug_yield_write();
 	return (v & RCU_GP_CTR_NEST_MASK) &&
-		 ((v ^ ACCESS_ONCE(urcu_gp_ctr)) & RCU_GP_CTR_BIT);
+		 ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT);
 }
 
 static inline void rcu_read_lock(void)
@@ -199,8 +237,9 @@ static inline void rcu_read_lock(void)
 	debug_yield_read();
 	tmp = urcu_active_readers;
 	debug_yield_read();
+	/* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 	if (likely(!(tmp & RCU_GP_CTR_NEST_MASK)))
-		urcu_active_readers = urcu_gp_ctr + RCU_GP_COUNT;
+		urcu_active_readers = urcu_gp_ctr;
 	else
 		urcu_active_readers = tmp + RCU_GP_COUNT;
 	debug_yield_read();
@@ -208,14 +247,14 @@ static inline void rcu_read_lock(void)
 	 * Increment active readers count before accessing the pointer.
 	 * See force_mb_all_threads().
 	 */
-	barrier();
+	read_barrier();
 	debug_yield_read();
 }
 
 static inline void rcu_read_unlock(void)
 {
 	debug_yield_read();
-	barrier();
+	read_barrier();
 	debug_yield_read();
 	/*
 	 * Finish using rcu before decrementing the pointer.