X-Git-Url: http://git.tuebingen.mpg.de/?p=paraslash.git;a=blobdiff_plain;f=server.c;h=95ff25cf77afb9f772a4336626e2bf340a2da868;hp=9c06991378512172e95b70165118ae1ea873bdb1;hb=0dd69d3988a677aeb8d0d3aea8364c664ac35fb9;hpb=48d037e4ff6f055f1b01dc7ae3d0c778c30821bf

diff --git a/server.c b/server.c
index 9c069913..95ff25cf 100644
--- a/server.c
+++ b/server.c
@@ -59,7 +59,7 @@
  *	- Ring buffer: \ref ringbuffer.c, \ref ringbuffer.h,
  *	- Hashing: \ref hash.h, \ref sha1.h, \ref sha1.c,
  *	- Crypto: \ref crypt.c.
- *
+ *	- Forward error correction: \ref fec.c
  */
 
 #include <signal.h>
@@ -291,15 +291,35 @@ static void signal_post_select(struct sched *s, struct task *t)
 			if (pid != mmd->afs_pid)
 				continue;
 			PARA_EMERG_LOG("fatal: afs died\n");
-			goto genocide;
+			kill(0, SIGTERM);
+			goto cleanup;
 		}
 		break;
 	/* die on sigint/sigterm. Kill all children too. */
 	case SIGINT:
 	case SIGTERM:
 		PARA_EMERG_LOG("terminating on signal %d\n", st->signum);
-genocide:
 		kill(0, SIGTERM);
+		/*
+		 * We must wait for afs because afs catches SIGINT/SIGTERM.
+		 * Before reacting to the signal, afs 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)mmd->afs_pid);
+		waitpid(mmd->afs_pid, NULL, 0);
+cleanup:
+		free(mmd->afd.afhi.chunk_table);
+		free(mmd->afd.afhi.info_string);
+		close_listed_fds();
 		mutex_destroy(mmd_mutex);
 		shm_detach(mmd);
 		exit(EXIT_FAILURE);
@@ -315,25 +335,15 @@ static void init_signal_task(void)
 	st->task.post_select = signal_post_select;
 	sprintf(st->task.status, "signal task");
 
+	PARA_NOTICE_LOG("setting up signal handling\n");
 	st->fd = para_signal_init(); /* always successful */
-
-	PARA_NOTICE_LOG("setting up signal handlers\n");
-	if (para_install_sighandler(SIGINT) < 0)
-		goto err;
-	if (para_install_sighandler(SIGTERM) < 0)
-		goto err;
-	if (para_install_sighandler(SIGHUP) < 0)
-		goto err;
-	if (para_install_sighandler(SIGCHLD) < 0)
-		goto err;
-	if (signal(SIGPIPE, SIG_IGN) == SIG_ERR)
-		goto err;
+	para_install_sighandler(SIGINT);
+	para_install_sighandler(SIGTERM);
+	para_install_sighandler(SIGHUP);
+	para_install_sighandler(SIGCHLD);
+	para_sigaction(SIGPIPE, SIG_IGN);
 	add_close_on_fork_list(st->fd);
 	register_task(&st->task);
-	return;
-err:
-	PARA_EMERG_LOG("could not install signal handlers\n");
-	exit(EXIT_FAILURE);
 }
 
 static void command_pre_select(struct sched *s, struct task *t)
@@ -349,6 +359,8 @@ static void command_post_select(struct sched *s, struct task *t)
 	int new_fd, ret, i;
 	char *peer_name;
 	pid_t child_pid;
+	uint32_t *chunk_table;
+	char *info_string;
 
 	if (!FD_ISSET(sct->listen_fd, &s->rfds))
 		return;
@@ -361,6 +373,16 @@ static void command_post_select(struct sched *s, struct task *t)
 	mmd->num_connects++;
 	mmd->active_connections++;
 	random();
+	/* The chunk table and the info_string are pointers located in the
+	 * mmd struct that point to dynamically allocated memory that must be
+	 * freed by the parent and the child. However, as the mmd struct is in
+	 * a shared memory area, there's no guarantee that after the fork these
+	 * pointers are still valid in child context. As these two pointers are
+	 * not used in the child anyway, we save them to local variables and
+	 * free the memory via that copy in the child.
+	 */
+	info_string = mmd->afd.afhi.info_string;
+	chunk_table = mmd->afd.afhi.chunk_table;
 	child_pid = fork();
 	if (child_pid < 0) {
 		ret = -ERRNO_TO_PARA_ERROR(errno);
@@ -371,6 +393,9 @@ static void command_post_select(struct sched *s, struct task *t)
 		/* parent keeps accepting connections */
 		return;
 	}
+	/* mmd might already have changed at this point */
+	free(info_string);
+	free(chunk_table);
 	alarm(ALARM_TIMEOUT);
 	close_listed_fds();
 	para_signal_shutdown();
@@ -445,18 +470,20 @@ err:
 static int init_afs(void)
 {
 	int ret, afs_server_socket[2];
+	pid_t afs_pid;
 
 	ret = socketpair(PF_UNIX, SOCK_DGRAM, 0, afs_server_socket);
 	if (ret < 0)
 		exit(EXIT_FAILURE);
 	afs_socket_cookie = para_random((uint32_t)-1);
-	mmd->afs_pid = fork();
-	if (mmd->afs_pid < 0)
+	afs_pid = fork();
+	if (afs_pid < 0)
 		exit(EXIT_FAILURE);
-	if (!mmd->afs_pid) { /* child (afs) */
+	if (afs_pid == 0) { /* child (afs) */
 		close(afs_server_socket[0]);
 		afs_init(afs_socket_cookie, afs_server_socket[1]);
 	}
+	mmd->afs_pid = afs_pid;
 	close(afs_server_socket[1]);
 	ret = mark_fd_nonblocking(afs_server_socket[0]);
 	if (ret < 0)
@@ -467,6 +494,12 @@ static int init_afs(void)
 	return afs_server_socket[0];
 }
 
+__noreturn static void tmp_sigchld_handler(__a_unused int s)
+{
+	PARA_EMERG_LOG("caught early SIGCHLD\n");
+	exit(EXIT_FAILURE);
+}
+
 static void server_init(int argc, char **argv)
 {
 	struct server_cmdline_parser_params params = {
@@ -505,10 +538,13 @@ static void server_init(int argc, char **argv)
 	 * before the child gets a chance to ignore this signal -- only the
 	 * good die young.
 	 */
-	if (signal(SIGUSR1, SIG_IGN) == SIG_ERR) {
-		PARA_EMERG_LOG("failed to ignore SIGUSR1\n");
-		exit(EXIT_FAILURE);
-	}
+	para_sigaction(SIGUSR1, SIG_IGN);
+	/*
+	 * We have to install a SIGCHLD handler before the afs process is being
+	 * forked off. Otherwise, para_server does not notice if afs dies before
+	 * the SIGCHLD handler has been installed by init_signal_task() below.
+	 */
+	para_sigaction(SIGCHLD, tmp_sigchld_handler);
 	PARA_NOTICE_LOG("initializing the audio file selector\n");
 	afs_socket = init_afs();
 	init_signal_task();