add rcu_cmpxchg_pointer

[urcu.git] / urcu.c

diff --git a/urcu.c b/urcu.c
index efbb4b7890a71524fa8264877f3da63565291591..a8f85c6f02199d4ecdfb20b1a0098cef213c138c 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -36,12 +36,23 @@
 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu.h"
 
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+#ifndef URCU_MB
+void __attribute__((constructor)) urcu_init(void);
+void __attribute__((destructor)) urcu_exit(void);
+#else
+void urcu_init(void)
+{
+}
+#endif
+
+static int init_done;
+
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 /*
  * Global grace period counter.
  * Contains the current RCU_GP_CTR_BIT.
- * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
+ * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  * Written to only by writer with mutex taken. Read by both writer and readers.
  */
 long urcu_gp_ctr = RCU_GP_COUNT;
@@ -70,7 +81,7 @@ static struct reader_registry *registry;
 static char __thread need_mb;
 static int num_readers, alloc_readers;
 
-void internal_urcu_lock(void)
+static void internal_urcu_lock(void)
 {
        int ret;
 
@@ -97,7 +108,7 @@ void internal_urcu_lock(void)
 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
-void internal_urcu_unlock(void)
+static void internal_urcu_unlock(void)
 {
        int ret;
 
@@ -116,7 +127,7 @@ static void switch_next_urcu_qparity(void)
        STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 }
 
-#ifdef DEBUG_FULL_MB
+#ifdef URCU_MB
 #ifdef HAS_INCOHERENT_CACHES
 static void force_mb_single_thread(struct reader_registry *index)
 {
@@ -128,7 +139,7 @@ static void force_mb_all_threads(void)
 {
        smp_mb();
 }
-#else /* #ifdef DEBUG_FULL_MB */
+#else /* #ifdef URCU_MB */
 #ifdef HAS_INCOHERENT_CACHES
 static void force_mb_single_thread(struct reader_registry *index)
 {
@@ -195,7 +206,7 @@ static void force_mb_all_threads(void)
        }
        smp_mb();       /* read ->need_mb before ending the barrier */
 }
-#endif /* #else #ifdef DEBUG_FULL_MB */
+#endif /* #else #ifdef URCU_MB */
 
 void wait_for_quiescent_state(void)
 {
@@ -249,12 +260,11 @@ void synchronize_rcu(void)
         */
 
        /*
-        * Current RCU formal verification model assumes sequential execution of
-        * the write-side. Add core synchronization instructions. Can be removed
-        * if the formal model is extended to prove that reordering is still
-        * correct.
+        * Adding a smp_mb() which is _not_ formally required, but makes the
+        * model easier to understand. It does not have a big performance impact
+        * anyway, given this is the write-side.
         */
-       sync_core();    /* Formal model assumes serialized execution */
+       smp_mb();
 
        /*
         * Wait for previous parity to be empty of readers.
@@ -269,7 +279,12 @@ void synchronize_rcu(void)
         * Ensured by STORE_SHARED and LOAD_SHARED.
         */
 
-       sync_core();    /* Formal model assumes serialized execution */
+       /*
+        * Adding a smp_mb() which is _not_ formally required, but makes the
+        * model easier to understand. It does not have a big performance impact
+        * anyway, given this is the write-side.
+        */
+       smp_mb();
 
        switch_next_urcu_qparity();     /* 1 -> 0 */
 
@@ -281,7 +296,12 @@ void synchronize_rcu(void)
         * Ensured by STORE_SHARED and LOAD_SHARED.
         */
 
-       sync_core();    /* Formal model assumes serialized execution */
+       /*
+        * Adding a smp_mb() which is _not_ formally required, but makes the
+        * model easier to understand. It does not have a big performance impact
+        * anyway, given this is the write-side.
+        */
+       smp_mb();
 
        /*
         * Wait for previous parity to be empty of readers.
@@ -327,6 +347,12 @@ void *rcu_xchg_pointer_sym(void **p, void *v)
        return xchg(p, v);
 }
 
+void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
+{
+       wmb();
+       return cmpxchg(p, old, _new);
+}
+
 void *rcu_publish_content_sym(void **p, void *v)
 {
        void *oldptr;
@@ -388,6 +414,7 @@ static void rcu_remove_reader(pthread_t id)
 void rcu_register_thread(void)
 {
        internal_urcu_lock();
+       urcu_init();    /* In case gcc does not support constructor attribute */
        rcu_add_reader(pthread_self());
        internal_urcu_unlock();
 }
@@ -399,7 +426,7 @@ void rcu_unregister_thread(void)
        internal_urcu_unlock();
 }
 
-#ifndef DEBUG_FULL_MB
+#ifndef URCU_MB
 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
        /*
@@ -412,12 +439,26 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
        smp_mb();
 }
 
-void __attribute__((constructor)) urcu_init(void)
+/*
+ * urcu_init constructor. Called when the library is linked, but also when
+ * reader threads are calling rcu_register_thread().
+ * Should only be called by a single thread at a given time. This is ensured by
+ * holing the internal_urcu_lock() from rcu_register_thread() or by running at
+ * library load time, which should not be executed by multiple threads nor
+ * concurrently with rcu_register_thread() anyway.
+ */
+void urcu_init(void)
 {
        struct sigaction act;
        int ret;
 
+       if (init_done)
+               return;
+       init_done = 1;
+
        act.sa_sigaction = sigurcu_handler;
+       act.sa_flags = SA_SIGINFO | SA_RESTART;
+       sigemptyset(&act.sa_mask);
        ret = sigaction(SIGURCU, &act, NULL);
        if (ret) {
                perror("Error in sigaction");
@@ -425,7 +466,7 @@ void __attribute__((constructor)) urcu_init(void)
        }
 }
 
-void __attribute__((destructor)) urcu_exit(void)
+void urcu_exit(void)
 {
        struct sigaction act;
        int ret;
@@ -438,4 +479,4 @@ void __attribute__((destructor)) urcu_exit(void)
        assert(act.sa_sigaction == sigurcu_handler);
        free(registry);
 }
-#endif /* #ifndef DEBUG_FULL_MB */
+#endif /* #ifndef URCU_MB */