X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-static.h;h=7d6d7ac25fbd915e8191f7ece2cad27eddbb9251;hp=91d37470fe785ffffd13dc7ebbebeb1357170cd7;hb=1ccf9cbe71d7392fb87e4b0046bd2eb94a75b541;hpb=809f4fdeada96d22ac2d91603f3dfb68282f4071

diff --git a/urcu-static.h b/urcu-static.h
index 91d3747..7d6d7ac 100644
--- a/urcu-static.h
+++ b/urcu-static.h
@@ -31,9 +31,13 @@
 
 #include <stdlib.h>
 #include <pthread.h>
+#include <syscall.h>
+#include <unistd.h>
 
-#include <compiler.h>
-#include <arch.h>
+#include <urcu/compiler.h>
+#include <urcu/arch.h>
+#include <urcu/arch_uatomic.h>
+#include <urcu/list.h>
 
 /*
  * Identify a shared load. A smp_rmc() or smp_mc() should come before the load.
@@ -81,7 +85,9 @@
  * 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.
+ * 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().
  */
@@ -92,6 +98,10 @@
 				(_________p1);				\
 				})
 
+#define futex(...)		syscall(__NR_futex, __VA_ARGS__)
+#define FUTEX_WAIT		0
+#define FUTEX_WAKE		1
+
 /*
  * This code section can only be included in LGPL 2.1 compatible source code.
  * See below for the function call wrappers which can be used in code meant to
@@ -115,6 +125,17 @@
  */
 #define KICK_READER_LOOPS 10000
 
+/*
+ * Active attempts to check for reader Q.S. before calling futex().
+ */
+#define RCU_QS_ACTIVE_ATTEMPTS 100
+
+#ifdef DEBUG_RCU
+#define rcu_assert(args...)	assert(args)
+#else
+#define rcu_assert(args...)
+#endif
+
 #ifdef DEBUG_YIELD
 #include <sched.h>
 #include <time.h>
@@ -125,10 +146,10 @@
 #define YIELD_WRITE	(1 << 1)
 
 /*
- * Updates without CONFIG_URCU_AVOID_SIGNALS are much slower. Account this in
+ * Updates without URCU_MB are much slower. Account this in
  * the delay.
  */
-#ifdef CONFIG_URCU_AVOID_SIGNALS
+#ifdef URCU_MB
 /* maximum sleep delay, in us */
 #define MAX_SLEEP 50
 #else
@@ -171,7 +192,7 @@ static inline void debug_yield_init(void)
 }
 #endif
 
-#ifdef CONFIG_URCU_AVOID_SIGNALS
+#ifdef URCU_MB
 static inline void reader_barrier()
 {
 	smp_mb();
@@ -199,7 +220,30 @@ static inline void reader_barrier()
  */
 extern long urcu_gp_ctr;
 
-extern long __thread urcu_active_readers;
+struct urcu_reader {
+	/* Data used by both reader and synchronize_rcu() */
+	long ctr;
+	char need_mb;
+	/* Data used for registry */
+	struct list_head head __attribute__((aligned(CACHE_LINE_SIZE)));
+	pthread_t tid;
+};
+
+extern struct urcu_reader __thread urcu_reader;
+
+extern int gp_futex;
+
+/*
+ * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
+ */
+static inline void wake_up_gp(void)
+{
+	if (unlikely(uatomic_read(&gp_futex) == -1)) {
+		uatomic_set(&gp_futex, 0);
+		futex(&gp_futex, FUTEX_WAKE, 1,
+		      NULL, NULL, 0);
+	}
+}
 
 static inline int rcu_old_gp_ongoing(long *value)
 {
@@ -220,31 +264,38 @@ static inline void _rcu_read_lock(void)
 {
 	long tmp;
 
-	tmp = urcu_active_readers;
+	tmp = urcu_reader.ctr;
 	/* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */
 	if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
-		_STORE_SHARED(urcu_active_readers, _LOAD_SHARED(urcu_gp_ctr));
+		_STORE_SHARED(urcu_reader.ctr, _LOAD_SHARED(urcu_gp_ctr));
 		/*
 		 * 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);
+		_STORE_SHARED(urcu_reader.ctr, tmp + RCU_GP_COUNT);
 	}
 }
 
 static inline void _rcu_read_unlock(void)
 {
-	reader_barrier();
+	long tmp;
+
+	tmp = urcu_reader.ctr;
 	/*
 	 * 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);
+	if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) {
+		reader_barrier();
+		_STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
+		/* write urcu_reader.ctr before read futex */
+		reader_barrier();
+		wake_up_gp();
+	} else {
+		_STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT);
+	}
 }
 
 /**
@@ -268,6 +319,22 @@ static inline void _rcu_read_unlock(void)
 		STORE_SHARED(p, v);			\
 	})
 
+/**
+ * _rcu_cmpxchg_pointer - same as rcu_assign_pointer, but tests if the pointer
+ * is as expected by "old". If succeeds, returns the previous pointer to the
+ * data structure, which can be safely freed after waiting for a quiescent state
+ * using synchronize_rcu(). If fails (unexpected value), returns old (which
+ * should not be freed !).
+ */
+
+#define _rcu_cmpxchg_pointer(p, old, _new)		\
+	({						\
+		if (!__builtin_constant_p(_new) ||	\
+		    ((_new) != NULL))			\
+			wmb();				\
+		uatomic_cmpxchg(p, old, _new);		\
+	})
+
 /**
  * _rcu_xchg_pointer - same as rcu_assign_pointer, but returns the previous
  * pointer to the data structure, which can be safely freed after waiting for a
@@ -279,7 +346,7 @@ static inline void _rcu_read_unlock(void)
 		if (!__builtin_constant_p(v) ||		\
 		    ((v) != NULL))			\
 			wmb();				\
-		xchg(p, v);				\
+		uatomic_xchg(p, v);			\
 	})
 
 /*