Add support for x86 older than P4, with CONFIG_HAS_FENCE option

[urcu.git] / urcu.c

diff --git a/urcu.c b/urcu.c
index 9542b264a6dbf53d36d2fb63e7578b4e7ff2b2f5..f2aae3433eeb310c16b6ed82a45aacafea76135a 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -43,7 +43,9 @@ unsigned int __thread rand_yield;
 
 static struct reader_data *reader_data;
 static int num_readers, alloc_readers;
+#ifndef DEBUG_FULL_MB
 static int sig_done;
+#endif
 
 void internal_urcu_lock(void)
 {
@@ -75,38 +77,52 @@ static void switch_next_urcu_qparity(void)
 }
 
 #ifdef DEBUG_FULL_MB
+static void force_mb_single_thread(pthread_t tid)
+{
+       smp_mb();
+}
+
 static void force_mb_all_threads(void)
 {
-       mb();
+       smp_mb();
 }
 #else
+
+static void force_mb_single_thread(pthread_t tid)
+{
+       assert(reader_data);
+       sig_done = 0;
+       smp_mb();       /* write sig_done before sending the signals */
+       pthread_kill(tid, SIGURCU);
+       /*
+        * Wait for sighandler (and thus mb()) to execute on every thread.
+        * BUSY-LOOP.
+        */
+       while (sig_done < 1)
+               smp_rmb();      /* ensure we re-read sig-done */
+       smp_mb();       /* read sig_done before ending the barrier */
+}
+
 static void force_mb_all_threads(void)
 {
        struct reader_data *index;
        /*
-        * Ask for each threads to execute a mb() so we can consider the
+        * Ask for each threads to execute a smp_mb() so we can consider the
         * compiler barriers around rcu read lock as real memory barriers.
         */
        if (!reader_data)
                return;
-       debug_yield_write();
        sig_done = 0;
-       debug_yield_write();
-       mb();   /* write sig_done before sending the signals */
-       debug_yield_write();
-       for (index = reader_data; index < reader_data + num_readers; index++) {
+       smp_mb();       /* write sig_done before sending the signals */
+       for (index = reader_data; index < reader_data + num_readers; index++)
                pthread_kill(index->tid, SIGURCU);
-               debug_yield_write();
-       }
        /*
         * Wait for sighandler (and thus mb()) to execute on every thread.
         * BUSY-LOOP.
         */
        while (sig_done < num_readers)
-               barrier();
-       debug_yield_write();
-       mb();   /* read sig_done before ending the barrier */
-       debug_yield_write();
+               smp_rmb();      /* ensure we re-read sig-done */
+       smp_mb();       /* read sig_done before ending the barrier */
 }
 #endif
 
@@ -116,30 +132,35 @@ void wait_for_quiescent_state(void)
 
        if (!reader_data)
                return;
-       /* Wait for each thread urcu_active_readers count to become 0.
+       /*
+        * Wait for each thread urcu_active_readers count to become 0.
         */
        for (index = reader_data; index < reader_data + num_readers; index++) {
+               int wait_loops = 0;
                /*
-                * BUSY-LOOP.
+                * BUSY-LOOP. Force the reader thread to commit its
+                * urcu_active_readers update to memory if we wait for too long.
                 */
-               while (rcu_old_gp_ongoing(index->urcu_active_readers))
-                       barrier();
+               while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
+                       if (wait_loops++ == KICK_READER_LOOPS) {
+                               force_mb_single_thread(index->tid);
+                               wait_loops = 0;
+                       }
+               }
        }
 }
 
 void synchronize_rcu(void)
 {
+       internal_urcu_lock();
+
        /* All threads should read qparity before accessing data structure
-        * where new ptr points to. */
+        * where new ptr points to. Must be done within internal_urcu_lock
+        * because it iterates on reader threads.*/
        /* Write new ptr before changing the qparity */
        force_mb_all_threads();
-       debug_yield_write();
-
-       internal_urcu_lock();
-       debug_yield_write();
 
        switch_next_urcu_qparity();     /* 0 -> 1 */
-       debug_yield_write();
 
        /*
         * Must commit qparity update to memory before waiting for parity
@@ -147,13 +168,12 @@ void synchronize_rcu(void)
         * waiting forever while new readers are always accessing data (no
         * progress).
         */
-       mb();
+       smp_mb();
 
        /*
         * Wait for previous parity to be empty of readers.
         */
        wait_for_quiescent_state();     /* Wait readers in parity 0 */
-       debug_yield_write();
 
        /*
         * Must finish waiting for quiescent state for parity 0 before
@@ -161,10 +181,9 @@ void synchronize_rcu(void)
         * the writer waiting forever while new readers are always accessing
         * data (no progress).
         */
-       mb();
+       smp_mb();
 
        switch_next_urcu_qparity();     /* 1 -> 0 */
-       debug_yield_write();
 
        /*
         * Must commit qparity update to memory before waiting for parity
@@ -172,23 +191,19 @@ void synchronize_rcu(void)
         * waiting forever while new readers are always accessing data (no
         * progress).
         */
-       mb();
+       smp_mb();
 
        /*
         * Wait for previous parity to be empty of readers.
         */
        wait_for_quiescent_state();     /* Wait readers in parity 1 */
-       debug_yield_write();
-
-       internal_urcu_unlock();
-       debug_yield_write();
 
-       /* All threads should finish using the data referred to by old ptr
-        * before decrementing their urcu_active_readers count */
        /* Finish waiting for reader threads before letting the old ptr being
-        * freed. */
+        * freed. Must be done within internal_urcu_lock because it iterates on
+        * reader threads. */
        force_mb_all_threads();
-       debug_yield_write();
+
+       internal_urcu_unlock();
 }
 
 void urcu_add_reader(pthread_t id)
@@ -256,7 +271,12 @@ void urcu_unregister_thread(void)
 #ifndef DEBUG_FULL_MB
 void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
-       mb();
+       /*
+        * Executing this smp_mb() is the only purpose of this signal handler.
+        * It punctually promotes barrier() into smp_mb() on every thread it is
+        * executed on.
+        */
+       smp_mb();
        atomic_inc(&sig_done);
 }