Add missing cpu_relax in loop

[urcu.git] / urcu.c

diff --git a/urcu.c b/urcu.c
index 8381f7b6f61bc892cdf25d671d21ef5b05aae9a5..b04b121344c5e0d425c0d2cfaee34a662ecb8cbf 100644 (file)
--- a/urcu.c
+++ b/urcu.c
@@ -92,7 +92,14 @@ 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.
@@ -115,9 +122,12 @@ static void force_mb_all_threads(void)
        sig_done = 0;
        /*
         * pthread_kill has a smp_mb(). But beware, we assume it performs
-        * a cache flush on architectures with non-coherent cache.
+        * 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);
        /*
@@ -149,6 +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();
                        }
                }
        }