r20140: Make online/offline detection completely asynchronous.

Now I've done this I might be able to reduce the probe
timeout and reduce the backoff algorithm, going back
to checking every cache time seconds (5 mins by default),
as the parent or forked domain child will never block.
Jeremy.
(This used to be commit d0add5f946cf63ea43067e8e935876b5346d11de)

2 files changed, 181 insertions, 23 deletions

diff --git a/source3/nsswitch/winbindd_cm.c b/source3/nsswitch/winbindd_cm.c
index c0a6aeb85d..07360e85d5 100644
--- a/source3/nsswitch/winbindd_cm.c
+++ b/source3/nsswitch/winbindd_cm.c
@@ -67,31 +67,167 @@
 
 extern struct winbindd_methods reconnect_methods;
 
+struct dc_name_ip {
+	fstring name;
+	struct in_addr ip;
+};
+
 static NTSTATUS init_dc_connection_network(struct winbindd_domain *domain);
 static void set_dc_type_and_flags( struct winbindd_domain *domain );
+static BOOL get_dcs(TALLOC_CTX *mem_ctx, const struct winbindd_domain *domain,
+		    struct dc_name_ip **dcs, int *num_dcs);
 
 /****************************************************************
- If we're still offline, exponentially increase the timeout check.
+ Child failed to find DC's. Reschedule check.
 ****************************************************************/
 
-static void calc_new_online_timeout_check(struct winbindd_domain *domain)
+static void msg_failed_to_go_online(int msg_type, struct process_id src, void *buf, size_t len)
 {
-	int wbc = lp_winbind_cache_time();
+	struct winbindd_domain *domain;
+	const char *domainname = (const char *)buf;
 
-	if (domain->startup) {
-		domain->check_online_timeout = 10;
-	} else if (domain->check_online_timeout < wbc) {
-		domain->check_online_timeout = wbc;
-	} else {
-		uint32 new_to = domain->check_online_timeout * 3;
-		if (new_to > (3*60*60)) {
-			new_to = 3*60*60; /* 3 hours. */
+	if (buf == NULL || len == 0) {
+		return;
+	}
+
+	DEBUG(5,("msg_fail_to_go_online: received for domain %s.\n", domainname));
+
+	for (domain = domain_list(); domain; domain = domain->next) {
+		if (domain->internal) {
+			continue;
+		}
+
+		if (strequal(domain->name, domainname)) {
+			if (domain->online) {
+				/* We're already online, ignore. */
+				DEBUG(5,("msg_fail_to_go_online: domain %s "
+					"already online.\n", domainname));
+				continue;
+			}
+
+			/* Reschedule the online check. */
+			set_domain_offline(domain);
+			break;
+		}
+	}
+}
+
+/****************************************************************
+ Actually cause a reconnect from a message.
+****************************************************************/
+
+static void msg_try_to_go_online(int msg_type, struct process_id src, void *buf, size_t len)
+{
+	struct winbindd_domain *domain;
+	const char *domainname = (const char *)buf;
+
+	if (buf == NULL || len == 0) {
+		return;
+	}
+
+	DEBUG(5,("msg_try_to_go_online: received for domain %s.\n", domainname));
+
+	for (domain = domain_list(); domain; domain = domain->next) {
+		if (domain->internal) {
+			continue;
+		}
+
+		if (strequal(domain->name, domainname)) {
+
+			if (domain->online) {
+				/* We're already online, ignore. */
+				DEBUG(5,("msg_try_to_go_online: domain %s "
+					"already online.\n", domainname));
+				continue;
+			}
+
+			/* This call takes care of setting the online
+			   flag to true if we connected, or re-adding
+			   the offline handler if false. Bypasses online
+			   check so always does network calls. */
+
+			init_dc_connection_network(domain);
+			break;
 		}
-		domain->check_online_timeout = new_to;
 	}
 }
 
 /****************************************************************
+ Fork a child to try and contact a DC. Do this as contacting a
+ DC requires blocking lookups and we don't want to block our
+ parent.
+****************************************************************/
+
+static BOOL fork_child_dc_connect(struct winbindd_domain *domain)
+{
+	extern BOOL override_logfile;
+	struct dc_name_ip *dcs = NULL;
+	int num_dcs = 0;
+	TALLOC_CTX *mem_ctx = NULL;
+	pid_t child_pid;
+	pid_t parent_pid = sys_getpid();
+
+	/* Stop zombies */
+	CatchChild();
+
+	message_block();
+
+	child_pid = sys_fork();
+
+	if (child_pid == -1) {
+		DEBUG(0, ("fork_child_dc_connect: Could not fork: %s\n", strerror(errno)));
+		message_unblock();
+		return False;
+	}
+
+	if (child_pid != 0) {
+		/* Parent */
+		message_register(MSG_WINBIND_TRY_TO_GO_ONLINE,msg_try_to_go_online);
+		message_register(MSG_WINBIND_FAILED_TO_GO_ONLINE,msg_failed_to_go_online);
+		message_unblock();
+		return True;
+	}
+
+	/* Child. */
+
+	/* Leave messages blocked - we will never process one. */
+
+	/* tdb needs special fork handling */
+	if (tdb_reopen_all(1) == -1) {
+		DEBUG(0,("tdb_reopen_all failed.\n"));
+		_exit(0);
+	}
+
+	close_conns_after_fork();
+
+	if (!override_logfile) {
+		reopen_logs();
+	}
+
+	mem_ctx = talloc_init("fork_child_dc_connect");
+	if (!mem_ctx) {
+		DEBUG(0,("talloc_init failed.\n"));
+		_exit(0);
+	}
+
+	if ((!get_dcs(mem_ctx, domain, &dcs, &num_dcs)) || (num_dcs == 0)) {
+		/* Still offline ? Can't find DC's. */
+		message_send_pid(pid_to_procid(parent_pid), MSG_WINBIND_FAILED_TO_GO_ONLINE,
+				domain->name,
+				strlen(domain->name)+1, False);
+		_exit(0);
+	}
+
+	/* We got a DC. Send a message to our parent to get it to
+	   try and do the same. */
+
+	message_send_pid(pid_to_procid(parent_pid), MSG_WINBIND_TRY_TO_GO_ONLINE,
+				domain->name,
+				strlen(domain->name)+1, False);
+	_exit(0);
+}
+
+/****************************************************************
  Handler triggered if we're offline to try and detect a DC.
 ****************************************************************/
 
@@ -127,12 +263,32 @@ static void check_domain_online_handler(struct timed_event *te,
 		return;
 	}
 
-	/* This call takes care of setting the online
-	   flag to true if we connected, or re-adding
-	   the offline handler if false. Bypasses online
-	   check so always does network calls. */
+	/* Fork a child to test if it can contact a DC. 
+	   If it can then send ourselves a message to
+	   cause a reconnect. */
 
-        init_dc_connection_network(domain);
+	fork_child_dc_connect(domain);
+}
+
+/****************************************************************
+ If we're still offline, exponentially increase the timeout check.
+****************************************************************/
+
+static void calc_new_online_timeout_check(struct winbindd_domain *domain)
+{
+	int wbc = lp_winbind_cache_time();
+
+	if (domain->startup) {
+		domain->check_online_timeout = 10;
+	} else if (domain->check_online_timeout < wbc) {
+		domain->check_online_timeout = wbc;
+	} else {
+		uint32 new_to = domain->check_online_timeout * 3;
+		if (new_to > (3*60*60)) {
+			new_to = 3*60*60; /* 3 hours. */
+		}
+		domain->check_online_timeout = new_to;
+	}
 }
 
 /****************************************************************
@@ -244,6 +400,10 @@ static void set_domain_online(struct winbindd_domain *domain)
 		TALLOC_FREE(domain->check_online_event);
 	}
 
+	/* Ensure we ignore any pending child messages. */
+	message_deregister(MSG_WINBIND_TRY_TO_GO_ONLINE);
+	message_deregister(MSG_WINBIND_FAILED_TO_GO_ONLINE);
+
 	domain->online = True;
 }
 
@@ -662,11 +822,6 @@ static NTSTATUS cm_prepare_connection(const struct winbindd_domain *domain,
 	return result;
 }
 
-struct dc_name_ip {
-	fstring name;
-	struct in_addr ip;
-};
-
 static BOOL add_one_dc_unique(TALLOC_CTX *mem_ctx, const char *domain_name,
 			      const char *dcname, struct in_addr ip,
 			      struct dc_name_ip **dcs, int *num)
diff --git a/source3/nsswitch/winbindd_dual.c b/source3/nsswitch/winbindd_dual.c
index 70af8333c5..a61b158c24 100644
--- a/source3/nsswitch/winbindd_dual.c
+++ b/source3/nsswitch/winbindd_dual.c
@@ -769,7 +769,10 @@ static BOOL fork_domain_child(struct winbindd_child *child)
 	}
 
 	ZERO_STRUCT(state);
-	state.pid = getpid();
+	state.pid = sys_getpid();
+
+	/* Stop zombies */
+	CatchChild();
 
 	/* Ensure we don't process messages whilst we're
 	   changing the disposition for the child. */