Fix: Wrong poll events on UST application socket

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

diff --git a/src/bin/lttng-sessiond/main.c b/src/bin/lttng-sessiond/main.c
index 41cdf112687c86fc60af8d832b07684dfc4c49bd..6bc8b00e914b1d1d6a15c6e82d2e0f520609afcc 100644 (file)
--- a/src/bin/lttng-sessiond/main.c
+++ b/src/bin/lttng-sessiond/main.c
@@ -412,6 +412,7 @@ static void cleanup(void)
                ERR("Unable to clean %s", rundir);
        }
        free(cmd);
+       free(rundir);
 
        DBG("Cleaning up all sessions");
 
@@ -837,6 +838,25 @@ static void *thread_manage_consumer(void *data)
 
        DBG("[thread] Manage consumer started");
 
+       /*
+        * Since the consumer thread can be spawned at any moment in time, we init
+        * the health to a poll status (1, which is a valid health over time).
+        * When the thread starts, we update here the health to a "code" path being
+        * an even value so this thread, when reaching a poll wait, does not
+        * trigger an error with an even value.
+        *
+        * Here is the use case we avoid.
+        *
+        * +1: the first poll update during initialization (main())
+        * +2 * x: multiple code update once in this thread.
+        * +1: poll wait in this thread (being a good health state).
+        * == even number which after the wait period shows as a bad health.
+        *
+        * In a nutshell, the following poll update to the health state brings back
+        * the state to an even value meaning a code path.
+        */
+       health_poll_update(&consumer_data->health);
+
        health_code_update(&consumer_data->health);
 
        ret = lttcomm_listen_unix_sock(consumer_data->err_sock);
@@ -1172,12 +1192,12 @@ static void *thread_manage_apps(void *data)
                                                ust_app_unregister(ust_cmd.sock);
                                        } else {
                                                /*
-                                                * We just need here to monitor the close of the UST
-                                                * socket and poll set monitor those by default.
-                                                * Listen on POLLIN (even if we never expect any
-                                                * data) to ensure that hangup wakes us.
+                                                * We only monitor the error events of the socket. This
+                                                * thread does not handle any incoming data from UST
+                                                * (POLLIN).
                                                 */
-                                               ret = lttng_poll_add(&events, ust_cmd.sock, LPOLLIN);
+                                               ret = lttng_poll_add(&events, ust_cmd.sock,
+                                                               LPOLLERR & LPOLLHUP & LPOLLRDHUP);
                                                if (ret < 0) {
                                                        goto error;
                                                }
@@ -2114,7 +2134,7 @@ static int process_client_msg(struct command_ctx *cmd_ctx, int sock,
        case LTTNG_LIST_DOMAINS:
        case LTTNG_START_TRACE:
        case LTTNG_STOP_TRACE:
-       case LTTNG_DATA_AVAILABLE:
+       case LTTNG_DATA_PENDING:
                need_domain = 0;
                break;
        default:
@@ -2809,9 +2829,9 @@ skip_domain:
                                bytecode);
                break;
        }
-       case LTTNG_DATA_AVAILABLE:
+       case LTTNG_DATA_PENDING:
        {
-               ret = cmd_data_available(cmd_ctx->session);
+               ret = cmd_data_pending(cmd_ctx->session);
                break;
        }
        default:
@@ -3976,7 +3996,9 @@ int main(int argc, char **argv)
        /*
         * Init health counters of the consumer thread. We do a quick hack here to
         * the state of the consumer health is fine even if the thread is not
-        * started.  This is simply to ease our life and has no cost what so ever.
+        * started. Once the thread starts, the health state is updated with a poll
+        * value to set a health code path. This is simply to ease our life and has
+        * no cost what so ever.
         */
        health_init(&kconsumer_data.health);
        health_poll_update(&kconsumer_data.health);
@@ -4073,7 +4095,25 @@ exit_dispatch:
                goto error;     /* join error, exit without cleanup */
        }
 
+       ret = join_consumer_thread(&ustconsumer32_data);
+       if (ret != 0) {
+               PERROR("join_consumer ust32");
+               goto error;     /* join error, exit without cleanup */
+       }
+
+       ret = join_consumer_thread(&ustconsumer64_data);
+       if (ret != 0) {
+               PERROR("join_consumer ust64");
+               goto error;     /* join error, exit without cleanup */
+       }
+
 exit_client:
+       ret = pthread_join(health_thread, &status);
+       if (ret != 0) {
+               PERROR("pthread_join health thread");
+               goto error;     /* join error, exit without cleanup */
+       }
+
 exit_health:
 exit:
        /*