X-Git-Url: http://git.tuebingen.mpg.de/?p=paraslash.git;a=blobdiff_plain;f=server.c;h=9bf6c3f88004fb3c4a00a2d892b58a53e7a3df89;hp=79e67818c87a1573b312c00091c8199f41c286ed;hb=7ad440cc721a097ecff7aa3f38f22806c644a497;hpb=b27f3e8dd53af7209626a9f0e98eb4770602c2e7

diff --git a/server.c b/server.c
index 79e67818..9bf6c3f8 100644
--- a/server.c
+++ b/server.c
@@ -291,15 +291,32 @@ 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();
@@ -342,6 +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;
@@ -354,6 +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);
@@ -364,6 +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();