X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdaemon.c;h=3c1e5c3b7817a63dfa6e0dffd8a290960868f999;hb=10a89ef04df5669c5cdd02f786150a7ab8454e01;hp=e78538cb9113e9ac4742c8cd2c43b5b47ca24825;hpb=ac718c9dbde6340a85d18c5c8d555d8e0ec88bb3;p=sliver-openvswitch.git

diff --git a/lib/daemon.c b/lib/daemon.c
index e78538cb9..3c1e5c3b7 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, 2012, 2013 Nicira, Inc.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -18,38 +18,66 @@
 #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 "ovs-thread.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? */
-static bool detach;
+VLOG_DEFINE_THIS_MODULE(daemon);
+
+/* --detach: Should we run in the background? */
+static bool detach;             /* Was --detach specified? */
+static bool detached;           /* Have we already detached? */
 
-/* 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 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;
+
+/* For each of the standard file descriptors, whether to replace it by
+ * /dev/null (if false) or keep it for the daemon to use (if true). */
+static bool save_fds[3];
+
+static void check_already_running(void);
+static int lock_pidfile(FILE *, int command);
+
 /* 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'.
@@ -61,6 +89,7 @@ make_pidfile_name(const char *name)
 void
 set_pidfile(const char *name)
 {
+    assert_single_threaded();
     free(pidfile);
     pidfile = make_pidfile_name(name);
 }
@@ -81,9 +110,16 @@ set_no_chdir(void)
     chdir_ = false;
 }
 
-/* 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. */
+/* Will we chdir to "/" as part of daemonizing? */
+bool
+is_chdir_enabled(void)
+{
+    return chdir_;
+}
+
+/* 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)
 {
@@ -98,144 +134,448 @@ set_detach(void)
     detach = true;
 }
 
-/* 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)
+/* Will daemonize() really detach? */
+bool
+get_detach(void)
+{
+    return detach;
+}
+
+/* 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)
+{
+    monitor = true;
+}
+
+/* A daemon doesn't normally have any use for the file descriptors for stdin,
+ * stdout, and stderr after it detaches.  To keep these file descriptors from
+ * e.g. holding an SSH session open, by default detaching replaces each of
+ * these file descriptors by /dev/null.  But a few daemons expect the user to
+ * redirect stdout or stderr to a file, in which case it is desirable to keep
+ * these file descriptors.  This function, therefore, disables replacing 'fd'
+ * by /dev/null when the daemon detaches. */
+void
+daemon_save_fd(int fd)
+{
+    ovs_assert(fd == STDIN_FILENO ||
+               fd == STDOUT_FILENO ||
+               fd == STDERR_FILENO);
+    save_fds[fd] = true;
+}
+
+/* Unregisters pidfile from being unlinked when the program terminates via
+* exit() or a fatal signal. */
+void
+remove_pidfile_from_unlink(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);
-        }
+        fatal_signal_remove_file_to_unlink(pidfile);
     }
-    return pid;
 }
 
-/* If a locked pidfile exists, issue a warning message and, unless
- * ignore_existing_pidfile() has been called, terminate the program. */
+/* Registers pidfile to be unlinked when the program terminates via exit() or a
+ * fatal signal. */
 void
-die_if_already_running(void)
+add_pidfile_to_unlink(void)
 {
-    pid_t pid = already_running();
-    if (pid) {
-        if (!overwrite_pidfile) {
-            ovs_fatal(0, "%s: already running as pid %ld",
-                      get_pidfile(), (long int) pid);
-        } else {
-            VLOG_WARN("%s: %s already running as pid %ld",
-                      get_pidfile(), program_name, (long int) pid);
-        }
+    if (pidfile) {
+        fatal_signal_add_file_to_unlink(pidfile);
     }
 }
 
-/* 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)
 {
-    if (pidfile) {
-        /* Create pidfile via temporary file, so that observers never see an
-         * empty pidfile or an unlocked pidfile. */
-        long int pid = getpid();
-        char *tmpfile;
-        int fd;
+    long int pid = getpid();
+    struct stat s;
+    char *tmpfile;
+    FILE *file;
+    int error;
 
+    /* Create a temporary pidfile. */
+    if (overwrite_pidfile) {
         tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
         fatal_signal_add_file_to_unlink(tmpfile);
-        fd = open(tmpfile, O_CREAT | O_WRONLY | O_TRUNC, 0666);
-        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_SETLK, &lck) != -1) {
-                char *text = xasprintf("%ld\n", pid);
-                if (write(fd, text, strlen(text)) == strlen(text)) {
-                    fatal_signal_add_file_to_unlink(pidfile);
-                    if (rename(tmpfile, pidfile) < 0) {
-                        VLOG_ERR("failed to rename \"%s\" to \"%s\": %s",
-                                 tmpfile, pidfile, strerror(errno));
-                        fatal_signal_remove_file_to_unlink(pidfile);
-                        close(fd);
-                    } else {
-                        /* Keep 'fd' open to retain the lock. */
-                    }
-                    free(text);
+    } else {
+        /* Everyone shares the same file which will be treated as a lock.  To
+         * avoid some uncomfortable race conditions, we can't set up the fatal
+         * signal unlink until we've acquired it. */
+        tmpfile = xasprintf("%s.tmp", pidfile);
+    }
+
+    file = fopen(tmpfile, "a+");
+    if (!file) {
+        VLOG_FATAL("%s: create failed (%s)", tmpfile, ovs_strerror(errno));
+    }
+
+    error = lock_pidfile(file, F_SETLK);
+    if (error) {
+        /* Looks like we failed to acquire the lock.  Note that, if we failed
+         * for some other reason (and '!overwrite_pidfile'), we will have
+         * left 'tmpfile' as garbage in the file system. */
+        VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile,
+                   ovs_strerror(error));
+    }
+
+    if (!overwrite_pidfile) {
+        /* We acquired the lock.  Make sure to clean up on exit, and verify
+         * that we're allowed to create the actual pidfile. */
+        fatal_signal_add_file_to_unlink(tmpfile);
+        check_already_running();
+    }
+
+    if (fstat(fileno(file), &s) == -1) {
+        VLOG_FATAL("%s: fstat failed (%s)", tmpfile, ovs_strerror(errno));
+    }
+
+    if (ftruncate(fileno(file), 0) == -1) {
+        VLOG_FATAL("%s: truncate failed (%s)", tmpfile, ovs_strerror(errno));
+    }
+
+    fprintf(file, "%ld\n", pid);
+    if (fflush(file) == EOF) {
+        VLOG_FATAL("%s: write failed (%s)", tmpfile, ovs_strerror(errno));
+    }
+
+    error = rename(tmpfile, pidfile);
+
+    /* Due to a race, 'tmpfile' may be owned by a different process, so we
+     * shouldn't delete it on exit. */
+    fatal_signal_remove_file_to_unlink(tmpfile);
+
+    if (error < 0) {
+        VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
+                   tmpfile, pidfile, ovs_strerror(errno));
+    }
+
+    /* Ensure that the pidfile will get deleted on exit. */
+    fatal_signal_add_file_to_unlink(pidfile);
+
+    /* Clean up.
+     *
+     * We don't close 'file' because its file descriptor must remain open to
+     * hold the lock. */
+    pidfile_dev = s.st_dev;
+    pidfile_ino = s.st_ino;
+    free(tmpfile);
+}
+
+/* If configured with set_pidfile() or set_detach(), creates the pid file and
+ * detaches from the foreground session.  */
+void
+daemonize(void)
+{
+    daemonize_start();
+    daemonize_complete();
+}
+
+/* Calls fork() and on success returns its return value.  On failure, logs an
+ * error and exits unsuccessfully.
+ *
+ * Post-fork, but before returning, this function calls a few other functions
+ * that are generally useful if the child isn't planning to exec a new
+ * process. */
+pid_t
+fork_and_clean_up(void)
+{
+    pid_t pid = xfork();
+    if (pid > 0) {
+        /* Running in parent process. */
+        fatal_signal_fork();
+    } else if (!pid) {
+        /* Running in child process. */
+        time_postfork();
+        lockfile_postfork();
+    }
+    return pid;
+}
+
+/* Forks, then:
+ *
+ *   - In the parent, waits for the child to signal that it has completed its
+ *     startup sequence.  Then stores -1 in '*fdp' and returns the child's pid.
+ *
+ *   - In the child, stores a fd in '*fdp' and returns 0.  The caller should
+ *     pass the fd to fork_notify_startup() after it finishes its startup
+ *     sequence.
+ *
+ * If something goes wrong with the fork, logs a critical error and aborts the
+ * process. */
+static pid_t
+fork_and_wait_for_startup(int *fdp)
+{
+    int fds[2];
+    pid_t pid;
+
+    xpipe(fds);
+
+    pid = fork_and_clean_up();
+    if (pid > 0) {
+        /* Running in parent process. */
+        size_t bytes_read;
+        char c;
+
+        close(fds[1]);
+        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) {
+                if (WIFEXITED(status) && WEXITSTATUS(status)) {
+                    /* Child exited with an error.  Convey the same error
+                     * to our parent process as a courtesy. */
+                    exit(WEXITSTATUS(status));
                 } else {
-                    VLOG_ERR("%s: write failed: %s", tmpfile, strerror(errno));
-                    close(fd);
+                    char *status_msg = process_status_msg(status);
+                    VLOG_FATAL("fork child died before signaling startup (%s)",
+                               status_msg);
                 }
+            } else if (retval < 0) {
+                VLOG_FATAL("waitpid failed (%s)", ovs_strerror(errno));
             } else {
-                VLOG_ERR("%s: fcntl failed: %s", tmpfile, strerror(errno));
-                close(fd);
+                NOT_REACHED();
             }
-        } else {
-            VLOG_ERR("%s: create failed: %s", tmpfile, strerror(errno));
         }
-        fatal_signal_remove_file_to_unlink(tmpfile);
-        free(tmpfile);
+        close(fds[0]);
+        *fdp = -1;
+    } else if (!pid) {
+        /* Running in child process. */
+        close(fds[0]);
+        *fdp = fds[1];
     }
-    free(pidfile);
-    pidfile = NULL;
+
+    return pid;
 }
 
-/* If configured with set_pidfile() or set_detach(), creates the pid file and
- * detaches from the foreground session.  */
-void
-daemonize(void)
+static void
+fork_notify_startup(int fd)
 {
-    if (detach) {
-        char c = 0;
-        int fds[2];
-        if (pipe(fds) < 0) {
-            ovs_fatal(errno, "pipe failed");
+    if (fd != -1) {
+        size_t bytes_written;
+        int error;
+
+        error = write_fully(fd, "", 1, &bytes_written);
+        if (error) {
+            VLOG_FATAL("pipe write failed (%s)", ovs_strerror(error));
         }
 
-        switch (fork()) {
-        default:
-            /* Parent process: wait for child to create pidfile, then exit. */
-            close(fds[1]);
-            fatal_signal_fork();
-            if (read(fds[0], &c, 1) != 1) {
-                ovs_fatal(errno, "daemon child failed to signal startup");
+        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;
             }
-            exit(0);
+        }
+    }
+    return false;
+}
+
+static void
+monitor_daemon(pid_t daemon_pid)
+{
+    /* XXX Should log daemon's stderr output at startup time. */
+    time_t last_restart;
+    char *status_msg;
+    int crashes;
 
-        case 0:
-            /* Child process. */
-            close(fds[0]);
-            make_pidfile();
-            write(fds[1], &c, 1);
-            close(fds[1]);
-            setsid();
-            if (chdir_) {
-                chdir("/");
+    subprogram_name = "monitor";
+    status_msg = xstrdup("healthy");
+    last_restart = TIME_MIN;
+    crashes = 0;
+    for (;;) {
+        int retval;
+        int status;
+
+        proctitle_set("monitoring pid %lu (%s)",
+                      (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)", ovs_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",
+                                  ovs_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);
             }
-            time_postfork();
-            lockfile_postfork();
-            break;
-
-        case -1:
-            /* Error. */
-            ovs_fatal(errno, "could not fork");
-            break;
         }
-    } else {
+    }
+    free(status_msg);
+
+    /* Running in new daemon process. */
+    proctitle_restore();
+    subprogram_name = "";
+}
+
+/* Close standard file descriptors (except any that the client has requested we
+ * leave open by calling daemon_save_fd()).  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) {
+        int fd;
+
+        for (fd = 0; fd < 3; fd++) {
+            if (!save_fds[fd]) {
+                dup2(null_fd, fd);
+            }
+        }
+    }
+
+    /* Disable logging to stderr to avoid wasting CPU time. */
+    vlog_set_levels(NULL, VLF_CONSOLE, VLL_OFF);
+}
+
+/* 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
+ * daemon_complete()) or that it failed to start up (by exiting with a nonzero
+ * exit code). */
+void
+daemonize_start(void)
+{
+    assert_single_threaded();
+    daemonize_fd = -1;
+
+    if (detach) {
+        if (fork_and_wait_for_startup(&daemonize_fd) > 0) {
+            /* Running in parent process. */
+            exit(0);
+        }
+
+        /* Running in daemon or monitor process. */
+        setsid();
+    }
+
+    if (monitor) {
+        int saved_daemonize_fd = daemonize_fd;
+        pid_t daemon_pid;
+
+        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);
+        }
+        /* Running in daemon process. */
+    }
+
+    if (pidfile) {
         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.  It also
+ * call daemonize_post_detach().
+ *
+ * Calling this function more than once has no additional effect. */
+void
+daemonize_complete(void)
+{
+    if (pidfile) {
+        free(pidfile);
+        pidfile = NULL;
+    }
+
+    if (!detached) {
+        detached = true;
+
+        fork_notify_startup(daemonize_fd);
+        daemonize_fd = -1;
+        daemonize_post_detach();
+    }
+}
+
+/* If daemonization is configured, then this function does traditional Unix
+ * daemonization behavior: join a new session, chdir to the root (if not
+ * disabled), and close the standard file descriptors.
+ *
+ * It only makes sense to call this function as part of an implementation of a
+ * special daemon subprocess.  A normal daemon should just call
+ * daemonize_complete(). */
+void
+daemonize_post_detach(void)
+{
+    if (detach) {
+        if (chdir_) {
+            ignore(chdir("/"));
+        }
+        close_standard_fds();
+    }
 }
 
 void
@@ -248,45 +588,117 @@ 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 int
+lock_pidfile__(FILE *file, int command, struct flock *lck)
+{
+    int error;
+
+    lck->l_type = F_WRLCK;
+    lck->l_whence = SEEK_SET;
+    lck->l_start = 0;
+    lck->l_len = 0;
+    lck->l_pid = 0;
+
+    do {
+        error = fcntl(fileno(file), command, lck) == -1 ? errno : 0;
+    } while (error == EINTR);
+    return error;
+}
+
+static int
+lock_pidfile(FILE *file, int command)
 {
-    char line[128];
     struct flock lck;
+
+    return lock_pidfile__(file, command, &lck);
+}
+
+static pid_t
+read_pidfile__(const char *pidfile, bool delete_if_stale)
+{
+    struct stat s, s2;
+    struct flock lck;
+    char line[128];
     FILE *file;
     int error;
 
-    file = fopen(pidfile, "r");
+    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 && delete_if_stale) {
+            return 0;
+        }
         error = errno;
-        VLOG_WARN("%s: open: %s", pidfile, strerror(error));
+        VLOG_WARN("%s: open: %s", pidfile, ovs_strerror(error));
         goto error;
     }
 
-    lck.l_type = F_WRLCK;
-    lck.l_whence = SEEK_SET;
-    lck.l_start = 0;
-    lck.l_len = 0;
-    if (fcntl(fileno(file), F_GETLK, &lck)) {
-        error = errno;
-        VLOG_WARN("%s: fcntl: %s", pidfile, strerror(error));
+    error = lock_pidfile__(file, F_GETLK, &lck);
+    if (error) {
+        VLOG_WARN("%s: fcntl: %s", pidfile, ovs_strerror(error));
         goto error;
     }
     if (lck.l_type == F_UNLCK) {
-        error = ESRCH;
-        VLOG_WARN("%s: pid file is not locked", pidfile);
-        goto error;
+        /* pidfile exists but it isn't locked by anyone.  We need to delete it
+         * so that a new pidfile can go in its place.  But just calling
+         * unlink(pidfile) makes a nasty race: what if someone else unlinks it
+         * before we do and then replaces it by a valid pidfile?  We'd unlink
+         * their valid pidfile.  We do a little dance to avoid the race, by
+         * locking the invalid pidfile.  Only one process can have the invalid
+         * pidfile locked, and only that process has the right to unlink it. */
+        if (!delete_if_stale) {
+            error = ESRCH;
+            VLOG_DBG("%s: pid file is stale", pidfile);
+            goto error;
+        }
+
+        /* Get the lock. */
+        error = lock_pidfile(file, F_SETLK);
+        if (error) {
+            /* We lost a race with someone else doing the same thing. */
+            VLOG_WARN("%s: lost race to lock pidfile", pidfile);
+            goto error;
+        }
+
+        /* Is the file we have locked still named 'pidfile'? */
+        if (stat(pidfile, &s) || fstat(fileno(file), &s2)
+            || s.st_ino != s2.st_ino || s.st_dev != s2.st_dev) {
+            /* No.  We lost a race with someone else who got the lock before
+             * us, deleted the pidfile, and closed it (releasing the lock). */
+            error = EALREADY;
+            VLOG_WARN("%s: lost race to delete pidfile", pidfile);
+            goto error;
+        }
+
+        /* We won the right to delete the stale pidfile. */
+        if (unlink(pidfile)) {
+            error = errno;
+            VLOG_WARN("%s: failed to delete stale pidfile (%s)",
+                      pidfile, ovs_strerror(error));
+            goto error;
+        }
+        VLOG_DBG("%s: deleted stale pidfile", pidfile);
+        fclose(file);
+        return 0;
     }
 
     if (!fgets(line, sizeof line, file)) {
         if (ferror(file)) {
             error = errno;
-            VLOG_WARN("%s: read: %s", pidfile, strerror(error));
+            VLOG_WARN("%s: read: %s", pidfile, ovs_strerror(error));
         } else {
             error = ESRCH;
             VLOG_WARN("%s: read: unexpected end of file", pidfile);
@@ -295,9 +707,12 @@ read_pidfile(const char *pidfile)
     }
 
     if (lck.l_pid != strtoul(line, NULL, 10)) {
+        /* The process that has the pidfile locked is not the process that
+         * created it.  It must be stale, with the process that has it locked
+         * preparing to delete it. */
         error = ESRCH;
-        VLOG_WARN("l_pid (%ld) != %s pid (%s)",
-                   (long int) lck.l_pid, pidfile, line);
+        VLOG_WARN("%s: stale pidfile for pid %s being deleted by pid %ld",
+                  pidfile, line, (long int) lck.l_pid);
         goto error;
     }
 
@@ -310,3 +725,25 @@ 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, false);
+}
+
+/* Checks whether a process with the given 'pidfile' is already running and,
+ * if so, aborts.  If 'pidfile' is stale, deletes it. */
+static void
+check_already_running(void)
+{
+    long int pid = read_pidfile__(pidfile, true);
+    if (pid > 0) {
+        VLOG_FATAL("%s: already running as pid %ld, aborting", pidfile, pid);
+    } else if (pid < 0) {
+        VLOG_FATAL("%s: pidfile check failed (%s), aborting",
+                   pidfile, ovs_strerror(-pid));
+    }
+}