urcu-bp: use sys_membarrier when available

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

diff --git a/urcu-bp.c b/urcu-bp.c
index 4dc402865a026800ccef0bb35378be6241c81505..6b78af8bd8dfc77545a8a11adcfec997434e57bd 100644 (file)
--- a/urcu-bp.c
+++ b/urcu-bp.c
@@ -94,11 +94,32 @@ void *mremap_wrapper(void *old_address, size_t old_size,
 static
 int rcu_bp_refcount;
 
+/*
+ * RCU_MEMBARRIER is only possibly available on Linux.
+ */
+#ifdef __linux__
+#include <urcu/syscall-compat.h>
+#endif
+
+/* If the headers do not support SYS_membarrier, fall back on RCU_MB */
+#ifdef SYS_membarrier
+# define membarrier(...)               syscall(SYS_membarrier, __VA_ARGS__)
+#else
+# define membarrier(...)               -ENOSYS
+#endif
+
+enum membarrier_cmd {
+       MEMBARRIER_CMD_QUERY = 0,
+       MEMBARRIER_CMD_SHARED = (1 << 0),
+};
+
 static
 void __attribute__((constructor)) rcu_bp_init(void);
 static
 void __attribute__((destructor)) rcu_bp_exit(void);
 
+int urcu_bp_has_sys_membarrier;
+
 /*
  * rcu_gp_lock ensures mutual exclusion between threads calling
  * synchronize_rcu().
@@ -174,6 +195,14 @@ static void mutex_unlock(pthread_mutex_t *mutex)
                urcu_die(ret);
 }
 
+static void smp_mb_master(void)
+{
+       if (caa_likely(urcu_bp_has_sys_membarrier))
+               (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
+       else
+               cmm_smp_mb();
+}
+
 /*
  * Always called with rcu_registry lock held. Releases this lock between
  * iterations and grabs it again. Holds the lock when it returns.
@@ -254,7 +283,7 @@ void synchronize_rcu(void)
        /* All threads should read qparity before accessing data structure
         * where new ptr points to. */
        /* Write new ptr before changing the qparity */
-       cmm_smp_mb();
+       smp_mb_master();
 
        /*
         * Wait for readers to observe original parity or be quiescent.
@@ -303,7 +332,7 @@ void synchronize_rcu(void)
         * Finish waiting for reader threads before letting the old ptr being
         * freed.
         */
-       cmm_smp_mb();
+       smp_mb_master();
 out:
        mutex_unlock(&rcu_registry_lock);
        mutex_unlock(&rcu_gp_lock);
@@ -567,6 +596,10 @@ void rcu_bp_init(void)
                                urcu_bp_thread_exit_notifier);
                if (ret)
                        abort();
+               ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
+               if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED)) {
+                       urcu_bp_has_sys_membarrier = 1;
+               }
                initialized = 1;
        }
        mutex_unlock(&init_lock);

diff --git a/urcu/static/urcu-bp.h b/urcu/static/urcu-bp.h
index df25393d06d6a1a19b0b281f81cedf767bd2d470..182e5feb24f1b5d7022a89a30f342c38821a4211 100644 (file)
--- a/urcu/static/urcu-bp.h
+++ b/urcu/static/urcu-bp.h
@@ -102,6 +102,16 @@ struct rcu_reader {
  */
 extern DECLARE_URCU_TLS(struct rcu_reader *, rcu_reader);
 
+extern int urcu_bp_has_sys_membarrier;
+
+static inline void urcu_bp_smp_mb_slave(void)
+{
+       if (caa_likely(urcu_bp_has_sys_membarrier))
+               cmm_barrier();
+       else
+               cmm_smp_mb();
+}
+
 static inline enum rcu_state rcu_reader_state(unsigned long *ctr)
 {
        unsigned long v;
@@ -131,7 +141,7 @@ static inline void _rcu_read_lock_update(unsigned long tmp)
 {
        if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
                _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, _CMM_LOAD_SHARED(rcu_gp.ctr));
-               cmm_smp_mb();
+               urcu_bp_smp_mb_slave();
        } else
                _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp + RCU_GP_COUNT);
 }
@@ -170,7 +180,7 @@ static inline void _rcu_read_unlock(void)
        tmp = URCU_TLS(rcu_reader)->ctr;
        urcu_assert(tmp & RCU_GP_CTR_NEST_MASK);
        /* Finish using rcu before decrementing the pointer. */
-       cmm_smp_mb();
+       urcu_bp_smp_mb_slave();
        _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp - RCU_GP_COUNT);
        cmm_barrier();  /* Ensure the compiler does not reorder us with mutex */
 }