Fix: thread exit vs futex wait/wakeup race

[lttng-tools.git] / src / bin / lttng-relayd / main.c

diff --git a/src/bin/lttng-relayd/main.c b/src/bin/lttng-relayd/main.c
index dc19a69c77a548df81fffc2aa7e44affd12279b4..8a183db28d226a0b25d924889a4ac9ed2715e11a 100644 (file)
--- a/src/bin/lttng-relayd/main.c
+++ b/src/bin/lttng-relayd/main.c
@@ -969,12 +969,16 @@ static void *relay_thread_dispatcher(void *data)
 
        health_code_update();
 
-       while (!CMM_LOAD_SHARED(dispatch_thread_exit)) {
+       for (;;) {
                health_code_update();
 
                /* Atomically prepare the queue futex */
                futex_nto1_prepare(&relay_conn_queue.futex);
 
+               if (CMM_LOAD_SHARED(dispatch_thread_exit)) {
+                       break;
+               }
+
                do {
                        health_code_update();
 
@@ -2680,6 +2684,11 @@ error:
                        destroy_conn,
                        sock_n.node) {
                health_code_update();
+
+               if (session_abort(destroy_conn->session)) {
+                       assert(0);
+               }
+
                /*
                 * No need to grab another ref, because we own
                 * destroy_conn.
@@ -2932,6 +2941,12 @@ exit_init_data:
        health_app_destroy(health_relayd);
 exit_health_app_create:
 exit_options:
+       /*
+        * Wait for all pending call_rcu work to complete before tearing
+        * down data structures. call_rcu worker may be trying to
+        * perform lookups in those structures.
+        */
+       rcu_barrier();
        relayd_cleanup();
 
        /* Ensure all prior call_rcu are done. */