#include <stdlib.h>
#include <pthread.h>
+#include <syscall.h>
+#include <unistd.h>
#include <compiler.h>
#include <arch.h>
* 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().
*/
(_________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
*/
#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>
#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
}
#endif
-#ifdef DEBUG_FULL_MB
+#ifdef URCU_MB
static inline void reader_barrier()
{
smp_mb();
extern long __thread urcu_active_readers;
+extern int gp_futex;
+
+/*
+ * Wake-up waiting synchronize_rcu(). Called from many concurrent threads.
+ */
+static inline void wake_up_gp(void)
+{
+ if (unlikely(atomic_read(&gp_futex) == -1)) {
+ atomic_set(&gp_futex, 0);
+ futex(&gp_futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0);
+ }
+}
+
static inline int rcu_old_gp_ongoing(long *value)
{
long v;
static inline void _rcu_read_unlock(void)
{
- reader_barrier();
+ long tmp;
+
+ tmp = urcu_active_readers;
/*
* 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_active_readers,
+ urcu_active_readers - RCU_GP_COUNT);
+ /* write urcu_active_readers before read futex */
+ reader_barrier();
+ wake_up_gp();
+ } else {
+ _STORE_SHARED(urcu_active_readers,
+ urcu_active_readers - RCU_GP_COUNT);
+ }
}
/**
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(); \
+ 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