Implement __OwningControllerProcess option

Implements part of feature 3049.

7 files changed, 473 insertions, 2 deletions

diff --git a/src/common/Makefile.am b/src/common/Makefile.am
index b1e03cd71..88892455e 100644
--- a/src/common/Makefile.am
+++ b/src/common/Makefile.am
@@ -12,11 +12,11 @@ libor_extra_source=
 endif
 
 libor_a_SOURCES = address.c log.c util.c compat.c container.c mempool.c \
-	memarea.c util_codedigest.c $(libor_extra_source)
+	memarea.c util_codedigest.c procmon.c $(libor_extra_source)
 libor_crypto_a_SOURCES = crypto.c aes.c tortls.c torgzip.c
 libor_event_a_SOURCES = compat_libevent.c
 
-noinst_HEADERS = address.h torlog.h crypto.h util.h compat.h aes.h torint.h tortls.h strlcpy.c strlcat.c torgzip.h container.h ht.h mempool.h memarea.h ciphers.inc compat_libevent.h tortls_states.h
+noinst_HEADERS = address.h torlog.h crypto.h util.h compat.h aes.h torint.h tortls.h strlcpy.c strlcat.c torgzip.h container.h ht.h mempool.h memarea.h procmon.h ciphers.inc compat_libevent.h tortls_states.h
 
 common_sha1.i: $(libor_SOURCES) $(libor_crypto_a_SOURCES) $(noinst_HEADERS)
 	if test "@SHA1SUM@" != none; then \
diff --git a/src/common/procmon.c b/src/common/procmon.c
new file mode 100644
index 000000000..cee956a52
--- /dev/null
+++ b/src/common/procmon.c
@@ -0,0 +1,331 @@
+
+/**
+ * \file procmon.c
+ * \brief Process-termination monitor functions
+ **/
+
+#include "procmon.h"
+
+#include "util.h"
+
+#ifdef HAVE_EVENT2_EVENT_H
+#include <event2/event.h>
+#else
+#include <event.h>
+#endif
+
+#ifdef HAVE_SIGNAL_H
+#include <signal.h>
+#endif
+#ifdef HAVE_ERRNO_H
+#include <errno.h>
+#endif
+
+#ifdef MS_WINDOWS
+#include <windows.h>
+
+/* Windows does not define pid_t, but _getpid() returns an int. */
+typedef int pid_t;
+#endif
+
+/* Define to 1 if process-termination monitors on this OS and Libevent
+   version must poll for process termination themselves. */
+#define PROCMON_POLLS 1
+/* Currently we need to poll in some way on all systems. */
+
+#ifdef PROCMON_POLLS
+static void tor_process_monitor_poll_cb(int unused1, short unused2,
+                                        void *procmon_);
+#endif
+
+/* This struct may contain pointers into the original process
+ * specifier string, but it should *never* contain anything which
+ * needs to be freed. */
+struct parsed_process_specifier_t {
+  pid_t pid;
+};
+
+/** Parse the process specifier given in <b>process_spec</b> into
+ * *<b>ppspec</b>.  Return 0 on success; return -1 and store an error
+ * message into *<b>msg</b> on failure.  The caller must not free the
+ * returned error message. */
+static int
+parse_process_specifier(const char *process_spec,
+                        struct parsed_process_specifier_t *ppspec,
+                        const char **msg)
+{
+  long pid_l;
+  int pid_ok = 0;
+  char *pspec_next;
+
+  /* If we're lucky, long will turn out to be large enough to hold a
+   * PID everywhere that Tor runs. */
+  pid_l = tor_parse_long(process_spec, 0, 1, LONG_MAX, &pid_ok, &pspec_next);
+
+  /* Reserve room in the ‘process specifier’ for additional
+   * (platform-specific) identifying information beyond the PID, to
+   * make our process-existence checks a bit less racy in a future
+   * version. */
+  if ((*pspec_next != 0) && (*pspec_next != ' ') && (*pspec_next != ':')) {
+    pid_ok = 0;
+  }
+
+  ppspec->pid = (pid_t)(pid_l);
+  if (!pid_ok || (pid_l != (long)(ppspec->pid))) {
+    *msg = "invalid PID";
+    goto err;
+  }
+
+  return 0;
+ err:
+  return -1;
+}
+
+struct tor_process_monitor_t {
+  /** Log domain for warning messages. */
+  log_domain_mask_t log_domain;
+
+  /** All systems: The best we can do in general is poll for the
+   * process's existence by PID periodically, and hope that the kernel
+   * doesn't reassign the same PID to another process between our
+   * polls. */
+  pid_t pid;
+
+#ifdef MS_WINDOWS
+  /** Windows-only: Should we poll hproc?  If false, poll pid
+   * instead. */
+  int poll_hproc;
+
+  /** Windows-only: Get a handle to the process (if possible) and
+   * periodically check whether the process we have a handle to has
+   * ended. */
+  HANDLE hproc;
+  /* XXX023 We can and should have Libevent watch hproc for us,
+   * if/when some version of Libevent 2.x can be told to do so. */
+#endif
+
+  /* XXX023 On Linux, we can and should receive the 22nd
+   * (space-delimited) field (‘starttime’) of /proc/$PID/stat from the
+   * owning controller and store it, and poll once in a while to see
+   * whether it has changed -- if so, the kernel has *definitely*
+   * reassigned the owning controller's PID and we should exit.  On
+   * FreeBSD, we can do the same trick using either the 8th
+   * space-delimited field of /proc/$PID/status on the seven FBSD
+   * systems whose admins have mounted procfs, or the start-time field
+   * of the process-information structure returned by kvmgetprocs() on
+   * any system.  The latter is ickier. */
+  /* XXX023 On FreeBSD (and possibly other kqueue systems), we can and
+   * should arrange to receive EVFILT_PROC NOTE_EXIT notifications for
+   * pid, so we don't have to do such a heavyweight poll operation in
+   * order to avoid the PID-reassignment race condition.  (We would
+   * still need to poll our own kqueue periodically until some version
+   * of Libevent 2.x learns to receive these events for us.) */
+
+  /** A Libevent event structure, to either poll for the process's
+   * existence or receive a notification when the process ends. */
+  struct event *e;
+
+  /** A callback to be called when the process ends. */
+  tor_procmon_callback_t cb;
+  void *cb_arg; /**< A user-specified pointer to be passed to cb. */
+};
+
+/** Verify that the process specifier given in <b>process_spec</b> is
+ * syntactically valid.  Return 0 on success; return -1 and store an
+ * error message into *<b>msg</b> on failure.  The caller must not
+ * free the returned error message. */
+int
+tor_validate_process_specifier(const char *process_spec,
+                               const char **msg)
+{
+  struct parsed_process_specifier_t ppspec;
+
+  tor_assert(msg != NULL);
+  *msg = NULL;
+
+  return parse_process_specifier(process_spec, &ppspec, msg);
+}
+
+#ifdef HAVE_EVENT2_EVENT_H
+#define PERIODIC_TIMER_FLAGS EV_PERSIST
+#else
+#define PERIODIC_TIMER_FLAGS (0)
+#endif
+
+static const struct timeval poll_interval_tv = {15, 0};
+
+/** Create a process-termination monitor for the process specifier
+ * given in <b>process_spec</b>.  Return a newly allocated
+ * tor_process_monitor_t on success; return NULL and store an error
+ * message into *<b>msg</b> on failure.  The caller must not free
+ * the returned error message.
+ *
+ * When the monitored process terminates, call
+ * <b>cb</b>(<b>cb_arg</b>).
+ */
+tor_process_monitor_t *
+tor_process_monitor_new(struct event_base *base,
+                        const char *process_spec,
+                        log_domain_mask_t log_domain,
+                        tor_procmon_callback_t cb, void *cb_arg,
+                        const char **msg)
+{
+  tor_process_monitor_t *procmon = tor_malloc(sizeof(tor_process_monitor_t));
+  struct parsed_process_specifier_t ppspec;
+
+  tor_assert(msg != NULL);
+  *msg = NULL;
+
+  if (procmon == NULL) {
+    *msg = "out of memory";
+    goto err;
+  }
+
+  procmon->log_domain = log_domain;
+
+  if (parse_process_specifier(process_spec, &ppspec, msg))
+    goto err;
+
+  procmon->pid = ppspec.pid;
+
+#ifdef MS_WINDOWS
+  procmon->hproc = OpenProcess(PROCESS_QUERY_INFORMATION | SYNCHRONIZE,
+                               FALSE,
+                               procmon->pid);
+
+  if (procmon->hproc != NULL) {
+    procmon->poll_hproc = 1;
+    log_info(procmon->log_domain, "Successfully opened handle to process %d; "
+             "monitoring it.",
+             (int)(procmon->pid));
+  } else {
+    /* If we couldn't get a handle to the process, we'll try again the
+     * first time we poll. */
+    log_info(procmon->log_domain, "Failed to open handle to process %d; will "
+             "try again later.",
+             (int)(procmon->pid));
+  }
+#endif
+
+  procmon->cb = cb;
+  procmon->cb_arg = cb_arg;
+
+#ifdef PROCMON_POLLS
+  procmon->e = tor_event_new(base, -1 /* no FD */, PERIODIC_TIMER_FLAGS,
+                             tor_process_monitor_poll_cb, procmon);
+  /* Note: If you port this file to plain Libevent 2, check that
+   * procmon->e is non-NULL.  We don't need to here because
+   * tor_evtimer_new never returns NULL. */
+
+  evtimer_add(procmon->e, &poll_interval_tv);
+#else
+#error OOPS?
+#endif
+
+  return procmon;
+ err:
+  tor_process_monitor_free(procmon);
+  return NULL;
+}
+
+#ifdef PROCMON_POLLS
+/** Libevent callback to poll for the existence of the process
+ * monitored by <b>procmon_</b>. */
+static void
+tor_process_monitor_poll_cb(int unused1, short unused2, void *procmon_)
+{
+  tor_process_monitor_t *procmon = (tor_process_monitor_t *)(procmon_);
+  int its_dead_jim;
+
+  (void)unused1; (void)unused2;
+
+  tor_assert(procmon != NULL);
+
+#ifdef MS_WINDOWS
+  if (procmon->poll_hproc) {
+    DWORD exit_code;
+    if (!GetExitCodeProcess(procmon->hproc, &exit_code)) {
+      char *errmsg = format_win32_error(GetLastError());
+      log_warn(procmon->log_domain, "Error \"%s\" occurred while polling "
+               "handle for monitored process %d; assuming it's dead."
+               errmsg, procmon->pid);
+      tor_free(errmsg);
+      its_dead_jim = 1;
+    } else {
+      its_dead_jim = (exit_code != STILL_ACTIVE);
+    }
+  } else {
+    /* All we can do is try to open the process, and look at the error
+     * code if it fails again. */
+    procmon->hproc = OpenProcess(PROCESS_QUERY_INFORMATION | SYNCHRONIZE,
+                                 FALSE,
+                                 procmon->pid);
+
+    if (procmon->hproc != NULL) {
+      log_info(procmon->log_domain, "Successfully opened handle to monitored "
+               "process %d.",
+               procmon->pid);
+      its_dead_jim = 0;
+      procmon->poll_hproc = 1;
+    } else {
+      DWORD err_code = GetLastError();
+      char *errmsg = format_win32_error(err_code);
+
+      /* When I tested OpenProcess's error codes on Windows 7, I
+       * received error code 5 (ERROR_ACCESS_DENIED) for PIDs of
+       * existing processes that I could not open and error code 87
+       * (ERROR_INVALID_PARAMETER) for PIDs that were not in use.
+       * Since the nonexistent-process error code is sane, I'm going
+       * to assume that all errors other than ERROR_INVALID_PARAMETER
+       * mean that the process we are monitoring is still alive. */
+      its_dead_jim = (err_code == ERROR_INVALID_PARAMETER);
+
+      if (!its_dead_jim)
+        log_info(procmon->log_domain, "Failed to open handle to monitored "
+                 "process %d, and error code %d (%s) is not 'invalid "
+                 "parameter' -- assuming the process is still alive.",
+                 procmon->pid,
+                 err_code, err_msg);
+
+      tor_free(err_msg);
+    }
+  }
+#else
+  /* Unix makes this part easy, if a bit racy. */
+  its_dead_jim = kill(procmon->pid, 0);
+  its_dead_jim = its_dead_jim && (errno == ESRCH);
+#endif
+
+  log(its_dead_jim ? LOG_NOTICE : LOG_INFO,
+      procmon->log_domain, "Monitored process %d is %s.",
+      (int)procmon->pid,
+      its_dead_jim ? "dead" : "still alive");
+
+  if (its_dead_jim) {
+    procmon->cb(procmon->cb_arg);
+#ifndef HAVE_EVENT2_EVENT_H
+  } else {
+    evtimer_add(procmon->e, &poll_interval_tv);
+#endif
+  }
+}
+#endif
+
+/** Free the process-termination monitor <b>procmon</b>. */
+void
+tor_process_monitor_free(tor_process_monitor_t *procmon)
+{
+  if (procmon == NULL)
+    return;
+
+#ifdef MS_WINDOWS
+  if (procmon->hproc != NULL)
+    CloseHandle(procmon->hproc);
+#endif
+
+  if (procmon->e != NULL)
+    tor_event_free(procmon->e);
+
+  tor_free(procmon);
+}
+
diff --git a/src/common/procmon.h b/src/common/procmon.h
new file mode 100644
index 000000000..f8c7444d5
--- /dev/null
+++ b/src/common/procmon.h
@@ -0,0 +1,29 @@
+
+/**
+ * \file procmon.h
+ * \brief Headers for procmon.c
+ **/
+
+#ifndef TOR_PROCMON_H
+#define TOR_PROCMON_H
+
+#include "compat.h"
+#include "compat_libevent.h"
+
+#include "torlog.h"
+
+typedef struct tor_process_monitor_t tor_process_monitor_t;
+
+typedef void (*tor_procmon_callback_t)(void *);
+
+int tor_validate_process_specifier(const char *process_spec,
+                                   const char **msg);
+tor_process_monitor_t *tor_process_monitor_new(struct event_base *base,
+                                               const char *process_spec,
+                                               log_domain_mask_t log_domain,
+                                               tor_procmon_callback_t cb,
+                                               void *cb_arg,
+                                               const char **msg);
+void tor_process_monitor_free(tor_process_monitor_t *procmon);
+
+#endif
diff --git a/src/or/config.c b/src/or/config.c
index 34208e85b..b2bc9f3e9 100644
--- a/src/or/config.c
+++ b/src/or/config.c
@@ -38,6 +38,8 @@
 #include <shlobj.h>
 #endif
 
+#include "procmon.h"
+
 /** Enumeration of types which option values can take */
 typedef enum config_type_t {
   CONFIG_TYPE_STRING = 0,   /**< An arbitrary string. */
@@ -393,6 +395,7 @@ static config_var_t _option_vars[] = {
   VAR("__LeaveStreamsUnattached",BOOL,  LeaveStreamsUnattached,   "0"),
   VAR("__HashedControlSessionPassword", LINELIST, HashedControlSessionPassword,
       NULL),
+  VAR("__OwningControllerProcess",STRING,OwningControllerProcess, NULL),
   V(MinUptimeHidServDirectoryV2, INTERVAL, "24 hours"),
   V(_UsingTestNetworkDefaults,   BOOL,     "0"),
 
@@ -1229,6 +1232,11 @@ options_act(or_options_t *old_options)
     return -1;
   }
 
+  if (monitor_owning_controller_process(options->OwningControllerProcess)) {
+    log_warn(LD_CONFIG, "Error monitoring owning controller process");
+    return -1;
+  }
+
   /* reload keys as needed for rendezvous services. */
   if (rend_service_load_keys()<0) {
     log_warn(LD_GENERAL,"Error loading rendezvous service keys");
@@ -3446,6 +3454,16 @@ options_validate(or_options_t *old_options, or_options_t *options,
     }
   }
 
+  if (options->OwningControllerProcess) {
+    const char *validate_pspec_msg = NULL;
+    if (tor_validate_process_specifier(options->OwningControllerProcess,
+                                       &validate_pspec_msg)) {
+      tor_asprintf(msg, "Bad OwningControllerProcess: %s",
+                   validate_pspec_msg);
+      return -1;
+    }
+  }
+
   if (options->ControlListenAddress) {
     int all_are_local = 1;
     config_line_t *ln;
diff --git a/src/or/control.c b/src/or/control.c
index 926a46520..3a9690443 100644
--- a/src/or/control.c
+++ b/src/or/control.c
@@ -32,6 +32,8 @@
 #include "routerlist.h"
 #include "routerparse.h"
 
+#include "procmon.h"
+
 /** Yield true iff <b>s</b> is the state of a control_connection_t that has
  * finished authentication and is accepting commands. */
 #define STATE_IS_OPEN(s) ((s) == CONTROL_CONN_STATE_OPEN)
@@ -3779,6 +3781,90 @@ init_cookie_authentication(int enabled)
   return 0;
 }
 
+/** A copy of the process specifier of Tor's owning controller, or
+ * NULL if this Tor instance is not currently owned by a process. */
+static char *owning_controller_process_spec = NULL;
+
+/** A process-termination monitor for Tor's owning controller, or NULL
+ * if this Tor instance is not currently owned by a process. */
+static tor_process_monitor_t *owning_controller_process_monitor = NULL;
+
+/** Process-termination monitor callback for Tor's owning controller
+ * process. */
+static void
+owning_controller_procmon_cb(void *unused)
+{
+  int shutdown_slowly = server_mode(get_options());
+
+  (void)unused;
+
+  log_notice(LD_CONTROL, "Owning controller process has vanished -- "
+             "%s.",
+             shutdown_slowly ? "shutting down" : "exiting now");
+
+  /* XXXX This chunk of code should be a separate function, called
+   * here and by process_signal(SIGINT). */
+
+  if (!shutdown_slowly) {
+    tor_cleanup();
+    exit(0);
+  }
+  /* XXXX This will close all listening sockets except control-port
+   * listeners.  Perhaps we should close those too. */
+  hibernate_begin_shutdown();
+}
+
+/** Set <b>process_spec</b> as Tor's owning controller process.
+ * Return 0 on success, -1 on failure. */
+int
+monitor_owning_controller_process(const char *process_spec)
+{
+  const char *msg;
+
+  tor_assert((owning_controller_process_spec == NULL) ==
+             (owning_controller_process_monitor == NULL));
+
+  if (owning_controller_process_spec != NULL) {
+    if ((process_spec != NULL) && !strcmp(process_spec,
+                                          owning_controller_process_spec)) {
+      /* Same process -- return now, instead of disposing of and
+       * recreating the process-termination monitor. */
+      return 0;
+    }
+
+    /* We are currently owned by a process, and we should no longer be
+     * owned by it.  Free the process-termination monitor. */
+    tor_process_monitor_free(owning_controller_process_monitor);
+    owning_controller_process_monitor = NULL;
+
+    tor_free(owning_controller_process_spec);
+    owning_controller_process_spec = NULL;
+  }
+
+  tor_assert((owning_controller_process_spec == NULL) &&
+             (owning_controller_process_monitor == NULL));
+
+  if (process_spec == NULL)
+    return 0;
+
+  owning_controller_process_spec = tor_strdup(process_spec);
+  owning_controller_process_monitor =
+    tor_process_monitor_new(tor_libevent_get_base(),
+                            owning_controller_process_spec,
+                            LD_CONTROL,
+                            owning_controller_procmon_cb, NULL,
+                            &msg);
+
+  if (owning_controller_process_monitor == NULL) {
+    log_warn(LD_CONTROL, "Couldn't create process-termination monitor for "
+             "owning controller: %s",
+             msg);
+    return -1;
+  }
+
+  return 0;
+}
+
 /** Convert the name of a bootstrapping phase <b>s</b> into strings
  * <b>tag</b> and <b>summary</b> suitable for display by the controller. */
 static int
diff --git a/src/or/control.h b/src/or/control.h
index 2ae96b4b8..81c23010d 100644
--- a/src/or/control.h
+++ b/src/or/control.h
@@ -70,6 +70,8 @@ smartlist_t *decode_hashed_passwords(config_line_t *passwords);
 void disable_control_logging(void);
 void enable_control_logging(void);
 
+int monitor_owning_controller_process(const char *process_spec);
+
 void control_event_bootstrap(bootstrap_status_t status, int progress);
 void control_event_bootstrap_problem(const char *warn, int reason);
 
diff --git a/src/or/or.h b/src/or/or.h
index d667358eb..546c38a9a 100644
--- a/src/or/or.h
+++ b/src/or/or.h
@@ -2669,6 +2669,11 @@ typedef struct {
   int DisablePredictedCircuits; /**< Boolean: does Tor preemptively
                                  * make circuits in the background (0),
                                  * or not (1)? */
+
+  /** Process specifier for a controller that ‘owns’ this Tor
+   * instance.  Tor will terminate if its owning controller does. */
+  char *OwningControllerProcess;
+
   int ShutdownWaitLength; /**< When we get a SIGINT and we're a server, how
                            * long do we wait before exiting? */
   char *SafeLogging; /**< Contains "relay", "1", "0" (meaning no scrubbing). */