Add missing cpu_relax in loop

[urcu.git] / urcu.c

diff --git a/urcu.c b/urcu.c
index f2aae3433eeb310c16b6ed82a45aacafea76135a..b04b121344c5e0d425c0d2cfaee34a662ecb8cbf 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -92,14 +92,21 @@ 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 has a smp_mb(). But beware, we assume it performs
+        * a cache flush on architectures with non-coherent cache. Let's play
+        * safe and don't assume anything : we use smp_mc() to make sure the
+        * cache flush is enforced.
+        * smp_mb();    write sig_done before sending the signals
+        */
+       smp_mc();       /* 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 */
+       while (LOAD_REMOTE(sig_done) < 1)
+               cpu_relax();
        smp_mb();       /* read sig_done before ending the barrier */
 }
 
@@ -113,15 +120,22 @@ static void force_mb_all_threads(void)
        if (!reader_data)
                return;
        sig_done = 0;
-       smp_mb();       /* write sig_done before sending the signals */
+       /*
+        * pthread_kill has a smp_mb(). But beware, we assume it performs
+        * a cache flush on architectures with non-coherent cache. Let's play
+        * safe and don't assume anything : we use smp_mc() to make sure the
+        * cache flush is enforced.
+        * smp_mb();    write sig_done before sending the signals
+        */
+       smp_mc();       /* write sig_done before sending the signals */
        for (index = reader_data; index < reader_data + num_readers; index++)
                pthread_kill(index->tid, SIGURCU);
        /*
         * Wait for sighandler (and thus mb()) to execute on every thread.
         * BUSY-LOOP.
         */
-       while (sig_done < num_readers)
-               smp_rmb();      /* ensure we re-read sig-done */
+       while (LOAD_REMOTE(sig_done) < num_readers)
+               cpu_relax();
        smp_mb();       /* read sig_done before ending the barrier */
 }
 #endif
@@ -145,6 +159,8 @@ void wait_for_quiescent_state(void)
                        if (wait_loops++ == KICK_READER_LOOPS) {
                                force_mb_single_thread(index->tid);
                                wait_loops = 0;
+                       } else {
+                               cpu_relax();
                        }
                }
        }
@@ -168,7 +184,7 @@ void synchronize_rcu(void)
         * waiting forever while new readers are always accessing data (no
         * progress).
         */
-       smp_mb();
+       smp_mc();
 
        /*
         * Wait for previous parity to be empty of readers.
@@ -181,7 +197,7 @@ void synchronize_rcu(void)
         * the writer waiting forever while new readers are always accessing
         * data (no progress).
         */
-       smp_mb();
+       smp_mc();
 
        switch_next_urcu_qparity();     /* 1 -> 0 */
 
@@ -191,7 +207,7 @@ void synchronize_rcu(void)
         * waiting forever while new readers are always accessing data (no
         * progress).
         */
-       smp_mb();
+       smp_mc();
 
        /*
         * Wait for previous parity to be empty of readers.