s3: Fix bug 7844: Race in winbind

If a child dies, the parent process right away closes the socket.
This is wrong, with tevent we still have events pending. This works
fine for epoll but does not for at least the FreeBSD select variant.
Tevent sticks a closed socket into the select masks. This then
returns an error EBADF. When this happens, the parent winbind dies
instead of forking a new child.

This moves the socket close from the SIGCHLD cleanup function to
the socket receiver. I could not reproduce the parent death anymore
and it did not create an obvious fd leak.

1 files changed, 8 insertions, 4 deletions

diff --git a/source3/winbindd/winbindd_dual.c b/source3/winbindd/winbindd_dual.c
index 117d55d2dac..360f1b2dbb2 100644
--- a/source3/winbindd/winbindd_dual.c
+++ b/source3/winbindd/winbindd_dual.c
@@ -134,7 +134,7 @@ static void wb_child_request_trigger(struct tevent_req *req,
 		req, struct wb_child_request_state);
 	struct tevent_req *subreq;
 
-	if ((state->child->pid == 0) && (!fork_domain_child(state->child))) {
+	if ((state->child->sock == -1) && (!fork_domain_child(state->child))) {
 		tevent_req_error(req, errno);
 		return;
 	}
@@ -164,6 +164,12 @@ static void wb_child_request_done(struct tevent_req *subreq)
 	ret = wb_simple_trans_recv(subreq, state, &state->response, &err);
 	TALLOC_FREE(subreq);
 	if (ret == -1) {
+		/*
+		 * The basic parent/child communication broke, close
+		 * our socket
+		 */
+		close(state->child->sock);
+		state->child->sock = -1;
 		tevent_req_error(req, err);
 		return;
 	}
@@ -502,6 +508,7 @@ void setup_child(struct winbindd_domain *domain, struct winbindd_child *child,
 			  "logname == NULL");
 	}
 
+	child->sock = -1;
 	child->domain = domain;
 	child->table = table;
 	child->queue = tevent_queue_create(NULL, "winbind_child");
@@ -530,9 +537,6 @@ void winbind_child_died(pid_t pid)
 	/* This will be re-added in fork_domain_child() */
 
 	DLIST_REMOVE(children, child);
-
-	close(child->sock);
-	child->sock = -1;
 	child->pid = 0;
 }