Avoid dirty afs tables on shutdown.

The server process must wait for afs because afs catches SIGINT/SIGTERM
and may thus not respond immediately to these SIGNALS.

Before reacting to the signal, afs might want to use the shared memory
area and the mmd mutex.  If the server process destroys 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().

diff --git a/server.c b/server.c
index a4115eb65734391128481f6fa421d548aba78e20..da55b99923621975335a86a7dfbbad41e3e6a573 100644 (file)
--- 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");
                        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);
                }
                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);
                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();
                free(mmd->afd.afhi.chunk_table);
                free(mmd->afd.afhi.info_string);
                close_listed_fds();