From 5a532b6873b6d619381c0188b91e48106e639808 Mon Sep 17 00:00:00 2001 From: Mathieu Desnoyers Date: Wed, 25 Jan 2012 09:39:28 -0500 Subject: [PATCH] Fix lttng sessiond spawn wait race If the child is scheduled before the parent, using "pause()" in the parent could be executed after the signal has been handled: that's entirely racy. Use a proper signal handler and flag to handle this racelessly. Signed-off-by: Mathieu Desnoyers --- src/bin/lttng/lttng.c | 22 +++++++++++++++++++--- 1 file changed, 19 insertions(+), 3 deletions(-) diff --git a/src/bin/lttng/lttng.c b/src/bin/lttng/lttng.c index 57927d4f3..aeaad1469 100644 --- a/src/bin/lttng/lttng.c +++ b/src/bin/lttng/lttng.c @@ -40,6 +40,7 @@ int opt_verbose; static int opt_no_sessiond; static char *opt_sessiond_path; static pid_t sessiond_pid; +static volatile int recv_child_signal; enum { OPT_SESSION_PATH, @@ -181,12 +182,14 @@ static void sighandler(int sig) case SIGCHLD: DBG("SIGCHLD caugth"); waitpid(sessiond_pid, &status, 0); + recv_child_signal = 1; /* Indicate that the session daemon died */ sessiond_pid = 0; ERR("Session daemon died (exit status %d)", WEXITSTATUS(status)); break; case SIGUSR1: /* Notify is done */ + recv_child_signal = 1; DBG("SIGUSR1 caugth"); break; default: @@ -294,6 +297,7 @@ static int spawn_sessiond(char *pathname) pid_t pid; MSG("Spawning a session daemon"); + recv_child_signal = 0; pid = fork(); if (pid == 0) { /* @@ -307,12 +311,24 @@ static int spawn_sessiond(char *pathname) } else { perror("execlp"); } - kill(getppid(), SIGTERM); /* unpause parent */ + kill(getppid(), SIGTERM); /* wake parent */ exit(EXIT_FAILURE); } else if (pid > 0) { sessiond_pid = pid; - /* Wait for lttng-sessiond to start */ - pause(); + /* + * Wait for lttng-sessiond to start. We need to use a + * flag to check if the signal has been sent to us, + * because the child can be scheduled before the parent, + * and thus send the signal before this check. In the + * signal handler, we set the recv_child_signal flag, so + * anytime we check it after the fork is fine. Note that + * sleep() is interrupted before the 1 second delay as + * soon as the signal is received, so it will not cause + * visible delay for the user. + */ + while (!recv_child_signal) { + sleep(1); + } if (!sessiond_pid) { exit(EXIT_FAILURE); } -- 2.34.1