Fix powerpc uatomic 4-byte cmpxchg (cmpd -> cmpw)

[urcu.git] / urcu-qsbr.c

diff --git a/urcu-qsbr.c b/urcu-qsbr.c
index 1ac553faefdf20ba1f775fe3af1fd49fd2e698f2..0e1252fdbba707e4515cbbabe8c8a837a6408c79 100644 (file)
--- a/urcu-qsbr.c
+++ b/urcu-qsbr.c
@@ -106,14 +106,35 @@ static void wait_gp(void)
                      NULL, NULL, 0);
 }
 
-static void wait_for_quiescent_state(void)
+static void update_counter_and_wait(void)
 {
        LIST_HEAD(qsreaders);
        int wait_loops = 0;
        struct rcu_reader *index, *tmp;
 
-       if (list_empty(&registry))
-               return;
+#if (BITS_PER_LONG < 64)
+       /* Switch parity: 0 -> 1, 1 -> 0 */
+       STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
+#else  /* !(BITS_PER_LONG < 64) */
+       /* Increment current G.P. */
+       STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
+#endif /* !(BITS_PER_LONG < 64) */
+
+       /*
+        * Must commit rcu_gp_ctr update to memory before waiting for quiescent
+        * state. Failure to do so could result in the writer waiting forever
+        * while new readers are always accessing data (no progress). Enforce
+        * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
+        */
+       barrier();
+
+       /*
+        * 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 each thread rcu_reader_qs_gp count to become 0.
         */
@@ -125,9 +146,9 @@ static void wait_for_quiescent_state(void)
                        smp_mb();
                }
 
-               list_for_each_entry_safe(index, tmp, &registry, head) {
+               list_for_each_entry_safe(index, tmp, &registry, node) {
                        if (!rcu_gp_ongoing(&index->ctr))
-                               list_move(&index->head, &qsreaders);
+                               list_move(&index->node, &qsreaders);
                }
 
                if (list_empty(&registry)) {
@@ -159,14 +180,6 @@ static void wait_for_quiescent_state(void)
  */
 
 #if (BITS_PER_LONG < 64)
-/*
- * called with rcu_gp_lock held.
- */
-static void switch_next_rcu_qparity(void)
-{
-       STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
-}
-
 void synchronize_rcu(void)
 {
        unsigned long was_online;
@@ -189,44 +202,35 @@ void synchronize_rcu(void)
 
        mutex_lock(&rcu_gp_lock);
 
-       switch_next_rcu_qparity();      /* 0 -> 1 */
-
-       /*
-        * Must commit qparity update to memory before waiting for parity
-        * 0 quiescent state. Failure to do so could result in the writer
-        * waiting forever while new readers are always accessing data (no
-        * progress).
-        * Ensured by STORE_SHARED and LOAD_SHARED.
-        */
+       if (list_empty(&registry))
+               goto out;
 
        /*
         * Wait for previous parity to be empty of readers.
         */
-       wait_for_quiescent_state();     /* Wait readers in parity 0 */
+       update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 
        /*
         * Must finish waiting for quiescent state for parity 0 before
-        * committing qparity update to memory. Failure to do so could result in
-        * the writer waiting forever while new readers are always accessing
-        * data (no progress).
-        * Ensured by STORE_SHARED and LOAD_SHARED.
+        * committing next rcu_gp_ctr update to memory. Failure to do so could
+        * result in the writer waiting forever while new readers are always
+        * accessing data (no progress).  Enforce compiler-order of load
+        * rcu_reader ctr before store to rcu_gp_ctr.
         */
-
-       switch_next_rcu_qparity();      /* 1 -> 0 */
+       barrier();
 
        /*
-        * Must commit qparity update to memory before waiting for parity
-        * 1 quiescent state. Failure to do so could result in the writer
-        * waiting forever while new readers are always accessing data (no
-        * progress).
-        * Ensured by STORE_SHARED and LOAD_SHARED.
+        * 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.
         */
-       wait_for_quiescent_state();     /* Wait readers in parity 1 */
-
+       update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
+out:
        mutex_unlock(&rcu_gp_lock);
 
        /*
@@ -254,8 +258,10 @@ void synchronize_rcu(void)
                STORE_SHARED(rcu_reader.ctr, 0);
 
        mutex_lock(&rcu_gp_lock);
-       STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
-       wait_for_quiescent_state();
+       if (list_empty(&registry))
+               goto out;
+       update_counter_and_wait();
+out:
        mutex_unlock(&rcu_gp_lock);
 
        if (was_online)
@@ -299,7 +305,7 @@ void rcu_register_thread(void)
        assert(rcu_reader.ctr == 0);
 
        mutex_lock(&rcu_gp_lock);
-       list_add(&rcu_reader.head, &registry);
+       list_add(&rcu_reader.node, &registry);
        mutex_unlock(&rcu_gp_lock);
        _rcu_thread_online();
 }
@@ -312,7 +318,7 @@ void rcu_unregister_thread(void)
         */
        _rcu_thread_offline();
        mutex_lock(&rcu_gp_lock);
-       list_del(&rcu_reader.head);
+       list_del(&rcu_reader.node);
        mutex_unlock(&rcu_gp_lock);
 }