fix race - set_closed() needs to be moved further out now

previously, we called it way too early from within vdp_pesi_fini() /
vdp_pesi_bytes() - see 6632dc5d

but the call from vped_deliver() was still from within the FSM, and we
need to finish additional cleanup in ved_task() before we can let the
topreq thread start delivering and possibly fini'ing / freeing our node.

src/vdp_pesi.c

@@ -920,7 +920,9 @@ ved_task(struct worker *wrk, void *priv)
 		task_fini(pesi_tree, req->transport_priv);
 		break;
 	case T_NEXUS:
-		assert(node->req == req); // XXX THIS HAS FAILED
+		assert(node->req == req);
+		assert(node->state == ST_OPEN ||
+		       node->state == ST_PRIVATE);
 		/*
 		 * XXX rethink: T_NEXUS' req gets fini't in
 		 * bytes_push_worklist(), but not if it had no nodes below,
@@ -928,7 +930,6 @@ ved_task(struct worker *wrk, void *priv)
 		 * "zero object" case here
 		 */
 		if (VSTAILQ_EMPTY(&node->nexus.children)) {
-			assert(node->state == ST_DELIVERED);
 			assert(pesi == req->transport_priv);
 			assert(req == node->req);
 
@@ -936,11 +937,22 @@ ved_task(struct worker *wrk, void *priv)
 			req = NULL;
 			pesi_destroy(&pesi);
 			req_fini(&node->req, wrk);
-			task_fini(pesi_tree, pesi);
 		} else {
+			// pesi destroyed within the FSM via vdp_pesi_fini
+			AZ(req->transport_priv);
 			pesi = NULL;
-			task_fini(pesi_tree, req->transport_priv);
 		}
+
+		Lck_Lock(&pesi_tree->tree->tree_lock);
+		set_closed(pesi_tree->tree, node);
+		Lck_Unlock(&pesi_tree->tree->tree_lock);
+
+		if (VSTAILQ_EMPTY(&node->nexus.children))
+			assert(node->state == ST_DELIVERED);
+		else
+			assert(node->state == ST_CLOSED);
+
+		task_fini(pesi_tree, pesi);
 		break;
 	default:
 		INCOMPL();
@@ -2612,9 +2624,6 @@ clean:
 noesi:
 	vped_close_first_vdp(req, &VDP_pesi_buf);
 out:
-	Lck_Lock(&tree->tree_lock);
-	set_closed(tree, node);
-	Lck_Unlock(&tree->tree_lock);
 	return;
 close:
 	VDP_close(req);