X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdaemon.c;h=a9cada86646ddae0d6b9acbc98fea54c1d596238;hb=00c08589876b7c1cd8f57e5ebb3e66bb164c5a3d;hp=9a1be55dbfc868d14274a138e9bd28c1491dc521;hpb=b8781ff08d9981258e75789c6f4ed18a56991577;p=sliver-openvswitch.git

diff --git a/lib/daemon.c b/lib/daemon.c
index 9a1be55db..a9cada866 100644
--- a/lib/daemon.c
+++ b/lib/daemon.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2009 Nicira Networks.
+ * Copyright (c) 2008, 2009, 2010, 2011 Nicira Networks.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,43 +18,57 @@
 #include "daemon.h"
 #include <errno.h>
 #include <fcntl.h>
+#include <signal.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/resource.h>
 #include <sys/wait.h>
+#include <sys/stat.h>
 #include <unistd.h>
+#include "command-line.h"
 #include "fatal-signal.h"
 #include "dirs.h"
 #include "lockfile.h"
+#include "process.h"
 #include "socket-util.h"
 #include "timeval.h"
 #include "util.h"
-
-#define THIS_MODULE VLM_daemon
 #include "vlog.h"
 
-/* Should we run in the background? */
+VLOG_DEFINE_THIS_MODULE(daemon);
+
+/* --detach: Should we run in the background? */
 static bool detach;
 
-/* Name of pidfile (null if none). */
+/* --pidfile: Name of pidfile (null if none). */
 static char *pidfile;
 
-/* Create pidfile even if one already exists and is locked? */
+/* Device and inode of pidfile, so we can avoid reopening it. */
+static dev_t pidfile_dev;
+static ino_t pidfile_ino;
+
+/* --overwrite-pidfile: Create pidfile even if one already exists and is
+   locked? */
 static bool overwrite_pidfile;
 
-/* Should we chdir to "/"? */
+/* --no-chdir: Should we chdir to "/"? */
 static bool chdir_ = true;
 
-/* File descriptors used by daemonize_start() and daemonize_complete(). */
-static int daemonize_fds[2];
+/* File descriptor used by daemonize_start() and daemonize_complete(). */
+static int daemonize_fd = -1;
+
+/* --monitor: Should a supervisory process monitor the daemon and restart it if
+ * it dies due to an error signal? */
+static bool monitor;
 
 /* Returns the file name that would be used for a pidfile if 'name' were
  * provided to set_pidfile().  The caller must free the returned string. */
 char *
-make_pidfile_name(const char *name) 
+make_pidfile_name(const char *name)
 {
-    return (!name ? xasprintf("%s/%s.pid", ovs_rundir, program_name)
-            : *name == '/' ? xstrdup(name)
-            : xasprintf("%s/%s", ovs_rundir, name));
+    return (!name
+            ? xasprintf("%s/%s.pid", ovs_rundir(), program_name)
+            : abs_file_name(ovs_rundir(), name));
 }
 
 /* Sets up a following call to daemonize() to create a pidfile named 'name'.
@@ -93,9 +107,9 @@ is_chdir_enabled(void)
     return chdir_;
 }
 
-/* Normally, die_if_already_running() will terminate the program with a message
- * if a locked pidfile already exists.  If this function is called,
- * die_if_already_running() will merely log a warning. */
+/* Normally, daemonize() or damonize_start() will terminate the program with a
+ * message if a locked pidfile already exists.  If this function is called, an
+ * existing pidfile will be replaced, with a warning. */
 void
 ignore_existing_pidfile(void)
 {
@@ -117,38 +131,28 @@ get_detach(void)
     return detach;
 }
 
-/* If a pidfile has been configured and that pidfile already exists and is
- * locked by a running process, returns the pid of the running process.
- * Otherwise, returns 0. */
-static pid_t
-already_running(void)
+/* Sets up a following call to daemonize() to fork a supervisory process to
+ * monitor the daemon and restart it if it dies due to an error signal.  */
+void
+daemon_set_monitor(void)
 {
-    pid_t pid = 0;
-    if (pidfile) {
-        int fd = open(pidfile, O_RDWR);
-        if (fd >= 0) {
-            struct flock lck;
-            lck.l_type = F_WRLCK;
-            lck.l_whence = SEEK_SET;
-            lck.l_start = 0;
-            lck.l_len = 0;
-            if (fcntl(fd, F_GETLK, &lck) != -1 && lck.l_type != F_UNLCK) {
-                pid = lck.l_pid;
-            }
-            close(fd);
-        }
-    }
-    return pid;
+    monitor = true;
 }
 
 /* If a locked pidfile exists, issue a warning message and, unless
  * ignore_existing_pidfile() has been called, terminate the program. */
-void
+static void
 die_if_already_running(void)
 {
-    pid_t pid = already_running();
-    if (pid) {
+    pid_t pid;
+    if (!pidfile) {
+        return;
+    }
+    pid = read_pidfile_if_exists(pidfile);
+    if (pid > 0) {
         if (!overwrite_pidfile) {
+            VLOG_ERR("%s: %s already running as pid %ld, aborting",
+                      get_pidfile(), program_name, (long int) pid);
             ovs_fatal(0, "%s: already running as pid %ld",
                       get_pidfile(), (long int) pid);
         } else {
@@ -158,9 +162,9 @@ die_if_already_running(void)
     }
 }
 
-/* If a pidfile has been configured, creates it and stores the running process'
- * pid init.  Ensures that the pidfile will be deleted when the process
- * exits. */
+/* If a pidfile has been configured, creates it and stores the running
+ * process's pid in it.  Ensures that the pidfile will be deleted when the
+ * process exits. */
 static void
 make_pidfile(void)
 {
@@ -191,12 +195,21 @@ make_pidfile(void)
                         close(fd);
                     } else {
                         /* Keep 'fd' open to retain the lock. */
+                        struct stat s;
+
+                        if (!fstat(fd, &s)) {
+                            pidfile_dev = s.st_dev;
+                            pidfile_ino = s.st_ino;
+                        } else {
+                            VLOG_ERR("%s: fstat failed: %s",
+                                     pidfile, strerror(errno));
+                        }
                     }
-                    free(text);
                 } else {
                     VLOG_ERR("%s: write failed: %s", tmpfile, strerror(errno));
                     close(fd);
                 }
+                free(text);
             } else {
                 VLOG_ERR("%s: fcntl failed: %s", tmpfile, strerror(errno));
                 close(fd);
@@ -220,6 +233,190 @@ daemonize(void)
     daemonize_complete();
 }
 
+static pid_t
+fork_and_wait_for_startup(int *fdp)
+{
+    int fds[2];
+    pid_t pid;
+
+    xpipe(fds);
+
+    pid = fork();
+    if (pid > 0) {
+        /* Running in parent process. */
+        size_t bytes_read;
+        char c;
+
+        close(fds[1]);
+        fatal_signal_fork();
+        if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
+            int retval;
+            int status;
+
+            do {
+                retval = waitpid(pid, &status, 0);
+            } while (retval == -1 && errno == EINTR);
+
+            if (retval == pid
+                && WIFEXITED(status)
+                && WEXITSTATUS(status)) {
+                /* Child exited with an error.  Convey the same error to
+                 * our parent process as a courtesy. */
+                exit(WEXITSTATUS(status));
+            }
+
+            VLOG_FATAL("fork child failed to signal startup (%s)",
+                       strerror(errno));
+        }
+        close(fds[0]);
+        *fdp = -1;
+    } else if (!pid) {
+        /* Running in child process. */
+        close(fds[0]);
+        time_postfork();
+        lockfile_postfork();
+        *fdp = fds[1];
+    } else {
+        VLOG_FATAL("fork failed (%s)", strerror(errno));
+    }
+
+    return pid;
+}
+
+static void
+fork_notify_startup(int fd)
+{
+    if (fd != -1) {
+        size_t bytes_written;
+        int error;
+
+        error = write_fully(fd, "", 1, &bytes_written);
+        if (error) {
+            VLOG_FATAL("pipe write failed (%s)", strerror(error));
+        }
+
+        close(fd);
+    }
+}
+
+static bool
+should_restart(int status)
+{
+    if (WIFSIGNALED(status)) {
+        static const int error_signals[] = {
+            SIGABRT, SIGALRM, SIGBUS, SIGFPE, SIGILL, SIGPIPE, SIGSEGV,
+            SIGXCPU, SIGXFSZ
+        };
+
+        size_t i;
+
+        for (i = 0; i < ARRAY_SIZE(error_signals); i++) {
+            if (error_signals[i] == WTERMSIG(status)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+static void
+monitor_daemon(pid_t daemon_pid)
+{
+    /* XXX Should log daemon's stderr output at startup time. */
+    const char *saved_program_name;
+    time_t last_restart;
+    char *status_msg;
+    int crashes;
+
+    saved_program_name = program_name;
+    program_name = xasprintf("monitor(%s)", program_name);
+    status_msg = xstrdup("healthy");
+    last_restart = TIME_MIN;
+    crashes = 0;
+    for (;;) {
+        int retval;
+        int status;
+
+        proctitle_set("%s: monitoring pid %lu (%s)",
+                      saved_program_name, (unsigned long int) daemon_pid,
+                      status_msg);
+
+        do {
+            retval = waitpid(daemon_pid, &status, 0);
+        } while (retval == -1 && errno == EINTR);
+
+        if (retval == -1) {
+            VLOG_FATAL("waitpid failed (%s)", strerror(errno));
+        } else if (retval == daemon_pid) {
+            char *s = process_status_msg(status);
+            if (should_restart(status)) {
+                free(status_msg);
+                status_msg = xasprintf("%d crashes: pid %lu died, %s",
+                                       ++crashes,
+                                       (unsigned long int) daemon_pid, s);
+                free(s);
+
+                if (WCOREDUMP(status)) {
+                    /* Disable further core dumps to save disk space. */
+                    struct rlimit r;
+
+                    r.rlim_cur = 0;
+                    r.rlim_max = 0;
+                    if (setrlimit(RLIMIT_CORE, &r) == -1) {
+                        VLOG_WARN("failed to disable core dumps: %s",
+                                  strerror(errno));
+                    }
+                }
+
+                /* Throttle restarts to no more than once every 10 seconds. */
+                if (time(NULL) < last_restart + 10) {
+                    VLOG_WARN("%s, waiting until 10 seconds since last "
+                              "restart", status_msg);
+                    for (;;) {
+                        time_t now = time(NULL);
+                        time_t wakeup = last_restart + 10;
+                        if (now >= wakeup) {
+                            break;
+                        }
+                        sleep(wakeup - now);
+                    }
+                }
+                last_restart = time(NULL);
+
+                VLOG_ERR("%s, restarting", status_msg);
+                daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
+                if (!daemon_pid) {
+                    break;
+                }
+            } else {
+                VLOG_INFO("pid %lu died, %s, exiting",
+                          (unsigned long int) daemon_pid, s);
+                free(s);
+                exit(0);
+            }
+        }
+    }
+    free(status_msg);
+
+    /* Running in new daemon process. */
+    proctitle_restore();
+    free((char *) program_name);
+    program_name = saved_program_name;
+}
+
+/* Close stdin, stdout, stderr.  If we're started from e.g. an SSH session,
+ * then this keeps us from holding that session open artificially. */
+static void
+close_standard_fds(void)
+{
+    int null_fd = get_null_fd();
+    if (null_fd >= 0) {
+        dup2(null_fd, STDIN_FILENO);
+        dup2(null_fd, STDOUT_FILENO);
+        dup2(null_fd, STDERR_FILENO);
+    }
+}
+
 /* If daemonization is configured, then starts daemonization, by forking and
  * returning in the child process.  The parent process hangs around until the
  * child lets it know either that it completed startup successfully (by calling
@@ -228,72 +425,55 @@ daemonize(void)
 void
 daemonize_start(void)
 {
-    if (detach) {
-        pid_t pid;
+    daemonize_fd = -1;
 
-        if (pipe(daemonize_fds) < 0) {
-            ovs_fatal(errno, "pipe failed");
+    if (detach) {
+        if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
+            /* Running in parent process. */
+            exit(0);
         }
+        /* Running in daemon or monitor process. */
+    }
 
-        pid = fork();
-        if (pid > 0) {
-            /* Running in parent process. */
-            char c;
-
-            close(daemonize_fds[1]);
-            fatal_signal_fork();
-            if (read(daemonize_fds[0], &c, 1) != 1) {
-                int retval;
-                int status;
-
-                do {
-                    retval = waitpid(pid, &status, 0);
-                } while (retval == -1 && errno == EINTR);
-
-                if (retval == pid
-                    && WIFEXITED(status)
-                    && WEXITSTATUS(status)) {
-                    /* Child exited with an error.  Convey the same error to
-                     * our parent process as a courtesy. */
-                    exit(WEXITSTATUS(status));
-                }
+    if (monitor) {
+        int saved_daemonize_fd = daemonize_fd;
+        pid_t daemon_pid;
 
-                ovs_fatal(errno, "daemon child failed to signal startup");
-            }
-            exit(0);
-        } else if (!pid) {
-            /* Running in child process. */
-            close(daemonize_fds[0]);
-            make_pidfile();
-            time_postfork();
-            lockfile_postfork();
-        } else {
-            ovs_fatal(errno, "could not fork");
+        daemon_pid = fork_and_wait_for_startup(&daemonize_fd);
+        if (daemon_pid > 0) {
+            /* Running in monitor process. */
+            fork_notify_startup(saved_daemonize_fd);
+            close_standard_fds();
+            monitor_daemon(daemon_pid);
         }
-    } else {
-        make_pidfile();
+        /* Running in daemon process. */
     }
+
+    die_if_already_running();
+    make_pidfile();
+
+    /* Make sure that the unixctl commands for vlog get registered in a
+     * daemon, even before the first log message. */
+    vlog_init();
 }
 
 /* If daemonization is configured, then this function notifies the parent
- * process that the child process has completed startup successfully. */
+ * process that the child process has completed startup successfully.
+ *
+ * Calling this function more than once has no additional effect. */
 void
 daemonize_complete(void)
 {
-    if (detach) {
-        size_t bytes_written;
-        int error;
-
-        error = write_fully(daemonize_fds[1], "", 1, &bytes_written);
-        if (error) {
-            ovs_fatal(error, "could not write to pipe");
-        }
+    fork_notify_startup(daemonize_fd);
+    daemonize_fd = -1;
 
-        close(daemonize_fds[1]);
+    if (detach) {
         setsid();
         if (chdir_) {
             ignore(chdir("/"));
         }
+        close_standard_fds();
+        detach = false;
     }
 }
 
@@ -307,21 +487,34 @@ daemon_usage(void)
         "  --pidfile[=FILE]        create pidfile (default: %s/%s.pid)\n"
         "  --overwrite-pidfile     with --pidfile, start even if already "
                                    "running\n",
-        ovs_rundir, program_name);
+        ovs_rundir(), program_name);
 }
 
-/* Opens and reads a PID from 'pidfile'.  Returns the nonnegative PID if
- * successful, otherwise a negative errno value. */
-pid_t
-read_pidfile(const char *pidfile)
+static pid_t
+read_pidfile__(const char *pidfile, bool must_exist)
 {
     char line[128];
     struct flock lck;
+    struct stat s;
     FILE *file;
     int error;
 
+    if ((pidfile_ino || pidfile_dev)
+        && !stat(pidfile, &s)
+        && s.st_ino == pidfile_ino && s.st_dev == pidfile_dev) {
+        /* It's our own pidfile.  We can't afford to open it, because closing
+         * *any* fd for a file that a process has locked also releases all the
+         * locks on that file.
+         *
+         * Fortunately, we know the associated pid anyhow: */
+        return getpid();
+    }
+
     file = fopen(pidfile, "r");
     if (!file) {
+        if (errno == ENOENT && !must_exist) {
+            return 0;
+        }
         error = errno;
         VLOG_WARN("%s: open: %s", pidfile, strerror(error));
         goto error;
@@ -331,6 +524,7 @@ read_pidfile(const char *pidfile)
     lck.l_whence = SEEK_SET;
     lck.l_start = 0;
     lck.l_len = 0;
+    lck.l_pid = 0;
     if (fcntl(fileno(file), F_GETLK, &lck)) {
         error = errno;
         VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
@@ -369,3 +563,21 @@ error:
     }
     return -error;
 }
+
+/* Opens and reads a PID from 'pidfile'.  Returns the positive PID if
+ * successful, otherwise a negative errno value. */
+pid_t
+read_pidfile(const char *pidfile)
+{
+    return read_pidfile__(pidfile, true);
+}
+
+
+/* Opens and reads a PID from 'pidfile', if it exists.  Returns 0 if 'pidfile'
+ * doesn't exist, the positive PID if successful, otherwise a negative errno
+ * value. */
+pid_t
+read_pidfile_if_exists(const char *pidfile)
+{
+    return read_pidfile__(pidfile, false);
+}