X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Fdaemon.c;h=ef1a24ef1e350a0a65e8a93052d468cae1bced1e;hb=c1a543a8d6d2847983b6b0defd1e19777da85715;hp=46c9a88e750bc554cda1c92393a6412bd7eba633;hpb=c69ee87c10818267f991236201150b1fa51ae519;p=sliver-openvswitch.git

diff --git a/lib/daemon.c b/lib/daemon.c
index 46c9a88e7..ef1a24ef1 100644
--- a/lib/daemon.c
+++ b/lib/daemon.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2008, 2009, 2010 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.
@@ -18,9 +18,12 @@
 #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"
@@ -30,20 +33,25 @@
 #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 descriptor used by daemonize_start() and daemonize_complete(). */
@@ -53,14 +61,17 @@ static int daemonize_fd = -1;
  * it dies due to an error signal? */
 static bool monitor;
 
+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'.
@@ -99,9 +110,9 @@ 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)
 {
@@ -131,96 +142,77 @@ daemon_set_monitor(void)
     monitor = 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)
+/* 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)
 {
-    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);
-        }
+    long int pid = getpid();
+    struct stat s;
+    char *tmpfile;
+    FILE *file;
+    int error;
+
+    /* Create a temporary pidfile. */
+    tmpfile = xasprintf("%s.tmp%ld", pidfile, pid);
+    fatal_signal_add_file_to_unlink(tmpfile);
+    file = fopen(tmpfile, "w+");
+    if (!file) {
+        VLOG_FATAL("%s: create failed (%s)", tmpfile, strerror(errno));
     }
-    return pid;
-}
 
-/* If a locked pidfile exists, issue a warning message and, unless
- * ignore_existing_pidfile() has been called, terminate the program. */
-void
-die_if_already_running(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 (fstat(fileno(file), &s) == -1) {
+        VLOG_FATAL("%s: fstat failed (%s)", tmpfile, strerror(errno));
     }
-}
 
-/* 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. */
-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;
-
-        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 {
-                    VLOG_ERR("%s: write failed: %s", tmpfile, strerror(errno));
-                    close(fd);
-                }
-            } else {
-                VLOG_ERR("%s: fcntl failed: %s", tmpfile, strerror(errno));
-                close(fd);
+    fprintf(file, "%ld\n", pid);
+    if (fflush(file) == EOF) {
+        VLOG_FATAL("%s: write failed (%s)", tmpfile, strerror(errno));
+    }
+
+    error = lock_pidfile(file, F_SETLK);
+    if (error) {
+        VLOG_FATAL("%s: fcntl(F_SETLK) failed (%s)", tmpfile, strerror(error));
+    }
+
+    /* Rename or link it to the correct name. */
+    if (overwrite_pidfile) {
+        if (rename(tmpfile, pidfile) < 0) {
+            VLOG_FATAL("failed to rename \"%s\" to \"%s\" (%s)",
+                       tmpfile, pidfile, strerror(errno));
+        }
+    } else {
+        do {
+            error = link(tmpfile, pidfile) == -1 ? errno : 0;
+            if (error == EEXIST) {
+                check_already_running();
             }
-        } else {
-            VLOG_ERR("%s: create failed: %s", tmpfile, strerror(errno));
+        } while (error == EINTR || error == EEXIST);
+        if (error) {
+            VLOG_FATAL("failed to link \"%s\" as \"%s\" (%s)",
+                       tmpfile, pidfile, strerror(error));
+        }
+    }
+
+    /* Ensure that the pidfile will get deleted on exit. */
+    fatal_signal_add_file_to_unlink(pidfile);
+
+    /* Delete the temporary pidfile if it still exists. */
+    if (!overwrite_pidfile) {
+        error = fatal_signal_unlink_file_now(tmpfile);
+        if (error) {
+            VLOG_FATAL("%s: unlink failed (%s)", tmpfile, strerror(error));
         }
-        fatal_signal_remove_file_to_unlink(tmpfile);
-        free(tmpfile);
     }
+
+    /* 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);
     free(pidfile);
     pidfile = NULL;
 }
@@ -240,18 +232,17 @@ fork_and_wait_for_startup(int *fdp)
     int fds[2];
     pid_t pid;
 
-    if (pipe(fds) < 0) {
-        ovs_fatal(errno, "pipe failed");
-    }
+    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(fds[0], &c, 1) != 1) {
+        if (read_fully(fds[0], &c, 1, &bytes_read) != 0) {
             int retval;
             int status;
 
@@ -267,7 +258,8 @@ fork_and_wait_for_startup(int *fdp)
                 exit(WEXITSTATUS(status));
             }
 
-            ovs_fatal(errno, "fork child failed to signal startup");
+            VLOG_FATAL("fork child failed to signal startup (%s)",
+                       strerror(errno));
         }
         close(fds[0]);
         *fdp = -1;
@@ -278,7 +270,7 @@ fork_and_wait_for_startup(int *fdp)
         lockfile_postfork();
         *fdp = fds[1];
     } else {
-        ovs_fatal(errno, "could not fork");
+        VLOG_FATAL("fork failed (%s)", strerror(errno));
     }
 
     return pid;
@@ -293,7 +285,7 @@ fork_notify_startup(int fd)
 
         error = write_fully(fd, "", 1, &bytes_written);
         if (error) {
-            ovs_fatal(error, "could not write to pipe");
+            VLOG_FATAL("pipe write failed (%s)", strerror(error));
         }
 
         close(fd);
@@ -323,14 +315,17 @@ should_restart(int status)
 static void
 monitor_daemon(pid_t daemon_pid)
 {
-    /* XXX Should limit the rate at which we restart the daemon. */
     /* 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;
@@ -344,22 +339,52 @@ monitor_daemon(pid_t daemon_pid)
         } while (retval == -1 && errno == EINTR);
 
         if (retval == -1) {
-            ovs_fatal(errno, "waitpid failed");
+            VLOG_FATAL("waitpid failed (%s)", strerror(errno));
         } else if (retval == daemon_pid) {
             char *s = process_status_msg(status);
-            free(status_msg);
-            status_msg = xasprintf("pid %lu died, %s",
-                                   (unsigned long int) daemon_pid, s);
-            free(s);
-
             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("%s, exiting", status_msg);
+                VLOG_INFO("pid %lu died, %s, exiting",
+                          (unsigned long int) daemon_pid, s);
+                free(s);
                 exit(0);
             }
         }
@@ -383,6 +408,9 @@ close_standard_fds(void)
         dup2(null_fd, STDOUT_FILENO);
         dup2(null_fd, STDERR_FILENO);
     }
+
+    /* 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
@@ -417,15 +445,24 @@ daemonize_start(void)
         /* Running in daemon process. */
     }
 
-    make_pidfile();
+    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. */
+ * process that the child process has completed startup successfully.
+ *
+ * Calling this function more than once has no additional effect. */
 void
 daemonize_complete(void)
 {
     fork_notify_startup(daemonize_fd);
+    daemonize_fd = -1;
 
     if (detach) {
         setsid();
@@ -433,6 +470,7 @@ daemonize_complete(void)
             ignore(chdir("/"));
         }
         close_standard_fds();
+        detach = false;
     }
 }
 
@@ -446,39 +484,111 @@ 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));
         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;
+    error = lock_pidfile__(file, F_GETLK, &lck);
+    if (error) {
         VLOG_WARN("%s: fcntl: %s", pidfile, 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, strerror(error));
+            goto error;
+        }
+        VLOG_DBG("%s: deleted stale pidfile", pidfile);
+        fclose(file);
+        return 0;
     }
 
     if (!fgets(line, sizeof line, file)) {
@@ -493,9 +603,12 @@ 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;
     }
 
@@ -508,3 +621,25 @@ 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, strerror(-pid));
+    }
+}