X-Git-Url: http://git.tuebingen.mpg.de/?p=paraslash.git;a=blobdiff_plain;f=server.c;h=13c8c85f98339023b90415125e9df275a28c98f3;hp=13082e7832c627765c810d993e7714632f918f96;hb=7f08e8b0eb9570f1eb787dbb4e10d56585b36bbf;hpb=c1d48060c104aecc99175155309ff2bca32494ca

diff --git a/server.c b/server.c
index 13082e78..13c8c85f 100644
--- a/server.c
+++ b/server.c
@@ -106,10 +106,30 @@ static struct signal_task *signal_task;
 /** The process id of the audio file selector process. */
 pid_t afs_pid = 0;
 
+/* The the main server process (parent of afs and the command handlers). */
+static pid_t server_pid;
+
+/**
+ * Tell whether the executing process is a command handler.
+ *
+ * Cleanup on exit must be performed differently for command handlers.
+ *
+ * \return True if the pid of the executing process is neither the server pid
+ * nor the afs pid.
+ */
+bool process_is_command_handler(void)
+{
+	pid_t pid = getpid();
+
+	return pid != afs_pid && pid != server_pid;
+}
+
 /** The task responsible for server command handling. */
 struct server_command_task {
 	/** TCP port on which para_server listens for connections. */
 	int listen_fd;
+	/* File descriptor for the accepted socket. */
+	int child_fd;
 	/** Copied from para_server's main function. */
 	int argc;
 	/** Argument vector passed to para_server's main function. */
@@ -274,8 +294,12 @@ static void handle_sighup(void)
 
 static int signal_post_select(struct sched *s, __a_unused void *context)
 {
-	int signum = para_next_signal(&s->rfds);
+	int ret, signum;
 
+	ret = task_get_notification(signal_task->task);
+	if (ret < 0)
+		return ret;
+	signum = para_next_signal(&s->rfds);
 	switch (signum) {
 	case 0:
 		return 0;
@@ -285,7 +309,7 @@ static int signal_post_select(struct sched *s, __a_unused void *context)
 	case SIGCHLD:
 		for (;;) {
 			pid_t pid;
-			int ret = para_reap_child(&pid);
+			ret = para_reap_child(&pid);
 			if (ret <= 0)
 				break;
 			if (pid != afs_pid)
@@ -301,27 +325,22 @@ static int signal_post_select(struct sched *s, __a_unused void *context)
 		PARA_EMERG_LOG("terminating on signal %d\n", signum);
 		kill(0, SIGTERM);
 		/*
-		 * We must wait for afs because afs catches SIGINT/SIGTERM.
-		 * Before reacting to the signal, afs might want to use the
+		 * We must wait for all of our children to die. For the afs
+		 * process or a command handler might want to use the
 		 * shared memory area and the mmd mutex.  If we destroy this
 		 * mutex too early and afs tries to lock the shared memory
 		 * area, the call to mutex_lock() will fail and terminate the
 		 * afs process. This leads to dirty osl tables.
-		 *
-		 * There's no such problem with the other children of the
-		 * server process (the command handlers) as these reset their
-		 * SIGINT/SIGTERM handlers to the default action, i.e.  these
-		 * processes get killed immediately by the above kill().
 		 */
-		PARA_INFO_LOG("waiting for afs (pid %d) to die\n",
-			(int)afs_pid);
-		waitpid(afs_pid, NULL, 0);
+		PARA_INFO_LOG("waiting for child processes to die\n");
+		mutex_unlock(mmd_mutex);
+		while (wait(NULL) != -1 || errno != ECHILD)
+			; /* still at least one child alive */
+		mutex_lock(mmd_mutex);
 cleanup:
 		free(mmd->afd.afhi.chunk_table);
-		close_listed_fds();
-		mutex_destroy(mmd_mutex);
-		shm_detach(mmd);
-		exit(EXIT_FAILURE);
+		task_notify_all(s, E_DEADLY_SIGNAL);
+		return -E_DEADLY_SIGNAL;
 	}
 	return 0;
 }
@@ -353,12 +372,14 @@ static void command_pre_select(struct sched *s, void *context)
 static int command_post_select(struct sched *s, void *context)
 {
 	struct server_command_task *sct = context;
-
 	int new_fd, ret, i;
 	char *peer_name;
 	pid_t child_pid;
 	uint32_t *chunk_table;
 
+	ret = task_get_notification(sct->task);
+	if (ret < 0)
+		return ret;
 	ret = para_accept(sct->listen_fd, &s->rfds, NULL, 0, &new_fd);
 	if (ret <= 0)
 		goto out;
@@ -391,9 +412,7 @@ static int command_post_select(struct sched *s, void *context)
 	PARA_INFO_LOG("accepted connection from %s\n", peer_name);
 	/* mmd might already have changed at this point */
 	free(chunk_table);
-	alarm(ALARM_TIMEOUT);
-	close_listed_fds();
-	signal_shutdown(signal_task);
+	sct->child_fd = new_fd;
 	/*
 	 * put info on who we are serving into argv[0] to make
 	 * client ip visible in top/ps
@@ -402,8 +421,21 @@ static int command_post_select(struct sched *s, void *context)
 		memset(sct->argv[i], 0, strlen(sct->argv[i]));
 	i = sct->argc - 1 - lls_num_inputs(cmdline_lpr);
 	sprintf(sct->argv[i], "para_server (serving %s)", peer_name);
-	handle_connect(new_fd);
-	/* never reached*/
+	/* ask other tasks to terminate */
+	task_notify_all(s, E_CHILD_CONTEXT);
+	/*
+	 * After we return, the scheduler calls server_select() with a minimal
+	 * timeout value, because the remaining tasks have a notification
+	 * pending. Next it calls the ->post_select method of these tasks,
+	 * which will return negative in view of the notification. This causes
+	 * schedule() to return as there are no more runnable tasks.
+	 *
+	 * Note that semaphores are not inherited across a fork(), so we don't
+	 * hold the lock at this point. Since server_select() drops the lock
+	 * prior to calling para_select(), we need to acquire it here.
+	 */
+	mutex_lock(mmd_mutex);
+	return -E_CHILD_CONTEXT;
 out:
 	if (ret < 0)
 		PARA_CRIT_LOG("%s\n", para_strerror(-ret));
@@ -416,6 +448,7 @@ static void init_server_command_task(struct server_command_task *sct,
 	int ret;
 
 	PARA_NOTICE_LOG("initializing tcp command socket\n");
+	sct->child_fd = -1;
 	sct->argc = argc;
 	sct->argv = argv;
 	ret = para_listen_simple(IPPROTO_TCP, OPT_UINT32_VAL(PORT));
@@ -512,6 +545,7 @@ static void server_init(int argc, char **argv, struct server_command_task *sct)
 	/* become daemon */
 	if (OPT_GIVEN(DAEMON))
 		daemon_pipe = daemonize(true /* parent waits for SIGTERM */);
+	server_pid = getpid();
 	init_random_seed_or_die();
 	daemon_log_welcome("server");
 	init_ipc_or_die(); /* init mmd struct and mmd->lock */
@@ -610,11 +644,27 @@ int main(int argc, char *argv[])
 	server_init(argc, argv, sct);
 	mutex_lock(mmd_mutex);
 	ret = schedule(&sched);
+	/*
+	 * We hold the mmd lock: it was re-acquired in server_select()
+	 * after the select call.
+	 */
+	mutex_unlock(mmd_mutex);
 	sched_shutdown(&sched);
+	signal_shutdown(signal_task);
+	if (!process_is_command_handler()) { /* parent (server) */
+		mutex_destroy(mmd_mutex);
+		shm_detach(mmd);
+		if (ret < 0)
+			PARA_EMERG_LOG("%s\n", para_strerror(-ret));
+	} else {
+		alarm(ALARM_TIMEOUT);
+		close_listed_fds();
+		ret = handle_connect(sct->child_fd);
+	}
+	vss_shutdown();
+	shm_detach(mmd);
 	lls_free_parse_result(server_lpr, CMD_PTR);
 	if (server_lpr != cmdline_lpr)
 		lls_free_parse_result(cmdline_lpr, CMD_PTR);
-	if (ret < 0)
-		PARA_EMERG_LOG("%s\n", para_strerror(-ret));
 	exit(ret < 0? EXIT_FAILURE : EXIT_SUCCESS);
 }