- Fix #775: libunbound: subprocess reap causes parent process reap

to hang.

doc/Changelog

@@ -1,6 +1,8 @@
 8 November 2022: Wouter
 	- Fix to ignore tcp events for closed comm points.
 	- Fix to make sure to not read again after a tcp comm point is closed.
+	- Fix #775: libunbound: subprocess reap causes parent process reap
+	  to hang.
 
 21 October 2022: George
 	- Merge #767 from jonathangray: consistently use IPv4/IPv6 in

libunbound/context.c

@@ -70,6 +70,7 @@ context_finalize(struct ub_ctx* ctx)
 	} else {
 		log_init(cfg->logfile, cfg->use_syslog, NULL);
 	}
+	ctx->pipe_pid = getpid();
 	cfg_apply_local_port_policy(cfg, 65536);
 	config_apply(cfg);
 	if(!modstack_setup(&ctx->mods, cfg->module_conf, ctx->env))

libunbound/context.h

@@ -89,6 +89,12 @@ struct ub_ctx {
 	pid_t bg_pid;
 	/** tid of bg worker thread */
 	ub_thread_type bg_tid;
+	/** pid when pipes are created. This was the process when the
+	 * setup was called. Helps with clean up, so we can tell after a fork
+	 * which side of the fork the delete is on. */
+	pid_t pipe_pid;
+	/** when threaded, the worker that exists in the created thread. */
+	struct libworker* thread_worker;
 
 	/** do threading (instead of forking) for async resolution */
 	int dothread;

libunbound/libunbound.c

@@ -305,11 +305,29 @@ ub_ctx_delete(struct ub_ctx* ctx)
 	int do_stop = 1;
 	if(!ctx) return;
 
+	/* if the delete is called but it has forked, and before the fork
+	 * the context was finalized, then the bg worker is not stopped
+	 * from here. There is one worker, but two contexts that refer to
+	 * it and only one should clean up, the one with getpid == pipe_pid.*/
+	if(ctx->created_bg && ctx->pipe_pid != getpid()) {
+		do_stop = 0;
+		/* Stop events from getting deregistered, if the backend is
+		 * epoll, the epoll fd is the same as the other process.
+		 * That process should deregister them. */
+		if(ctx->qq_pipe->listen_com)
+			ctx->qq_pipe->listen_com->event_added = 0;
+		if(ctx->qq_pipe->res_com)
+			ctx->qq_pipe->res_com->event_added = 0;
+		if(ctx->rr_pipe->listen_com)
+			ctx->rr_pipe->listen_com->event_added = 0;
+		if(ctx->rr_pipe->res_com)
+			ctx->rr_pipe->res_com->event_added = 0;
+	}
 	/* see if bg thread is created and if threads have been killed */
 	/* no locks, because those may be held by terminated threads */
 	/* for processes the read pipe is closed and we see that on read */
 #ifdef HAVE_PTHREAD
-	if(ctx->created_bg && ctx->dothread) {
+	if(ctx->created_bg && ctx->dothread && do_stop) {
 		if(pthread_kill(ctx->bg_tid, 0) == ESRCH) {
 			/* thread has been killed */
 			do_stop = 0;
@@ -318,6 +336,23 @@ ub_ctx_delete(struct ub_ctx* ctx)
 #endif /* HAVE_PTHREAD */
 	if(do_stop)
 		ub_stop_bg(ctx);
+	if(ctx->created_bg && ctx->pipe_pid != getpid() && ctx->thread_worker) {
+		/* This delete is happening from a different process. Delete
+		 * the thread worker from this process memory space. The
+		 * thread is not there to do so, so it is freed here. */
+		struct ub_event_base* evbase = comm_base_internal(
+			ctx->thread_worker->base);
+		libworker_delete_event(ctx->thread_worker);
+		ctx->thread_worker = NULL;
+#ifdef USE_MINI_EVENT
+		ub_event_base_free(evbase);
+#else
+		/* cannot event_base_free, because the epoll_fd cleanup
+		 * in libevent could stop the original event_base in the
+		 * other process from working. */
+		free(evbase);
+#endif
+	}
 	libworker_delete_event(ctx->event_worker);
 
 	modstack_desetup(&ctx->mods, ctx->env);

libunbound/libworker.c

@@ -395,6 +395,7 @@ int libworker_bg(struct ub_ctx* ctx)
 		w = libworker_setup(ctx, 1, NULL);
 		if(!w) return UB_NOMEM;
 		w->is_bg_thread = 1;
+		ctx->thread_worker = w;
 #ifdef ENABLE_LOCK_CHECKS
 		w->thread_num = 1; /* for nicer DEBUG checklocks */
 #endif