Survive remote connection failures in the bridge accept loop

Test-driving the bridge with a wrong --token showed a connecting
client hanging forever: the failed WebSocket handshake threw inside
the accept loop, killing it silently and leaving the local socket
open, with all subsequent connections ignored. Now a failed relay
setup closes the client socket immediately (the client errors out
instead of hanging), logs the cause to stderr, and the bridge keeps
accepting. The remote side is opened before any local transport
resources so a failure has nothing to leak.

Author: bbatsov

src/drawbridge/bridge.clj

@@ -70,8 +70,11 @@
   endpoint at `url` until either side disconnects. Calls `on-close`
   once when the connection winds down."
   [^Socket sock url http-headers poll-opts on-close]
-  (let [local (transport/bencode sock)
-        remote (remote-transport url http-headers)
+  ;; Open the remote side first: it's the fallible one (unreachable
+  ;; endpoint, auth rejection), and at this point no local transport
+  ;; resources exist yet for a throw to leak.
+  (let [remote (remote-transport url http-headers)
+        local (transport/bencode sock)
         {:keys [active-poll-ms idle-poll-ms idle-after-ms]} poll-opts
         open? (atom true)
         last-activity (atom (System/currentTimeMillis))
@@ -141,8 +144,19 @@
         (loop []
           (let [^Socket sock (.accept server)]
             (swap! connections conj sock)
-            (relay sock url http-headers poll-opts
-                   #(swap! connections disj sock))
+            ;; A failure to reach the remote endpoint (e.g. a 401 on
+            ;; the first request) must not kill the accept loop, and
+            ;; the local client should see its connection drop right
+            ;; away rather than hang waiting for a handshake.
+            (try
+              (relay sock url http-headers poll-opts
+                     #(swap! connections disj sock))
+              (catch Exception e
+                (swap! connections disj sock)
+                (.close sock)
+                (binding [*out* *err*]
+                  (println "drawbridge.bridge: could not connect to" url "-"
+                           (str e)))))
             (recur)))
         ;; Closing the server socket unblocks accept with an exception.
         (catch Exception _)))

test/drawbridge/auth_test.clj

@@ -92,3 +92,20 @@
                     responses (nrepl/message client {:op "eval" :code "(* 6 7)"})]
                 (is (some #(= "42" (:value %)) responses))))
             (finally (bridge/stop-bridge b))))))))
+
+(deftest bridge-rejected-by-endpoint
+  (testing "clients of a misconfigured bridge fail fast instead of hanging"
+    (with-secure-server "s3cret"
+      (fn [port]
+        (let [b (bridge/start-bridge
+                 {:url (str "http://localhost:" port "/")
+                  :http-headers {"Authorization" "Bearer WRONG"}})]
+          (try
+            ;; Two rounds: the accept loop must also survive the
+            ;; first failed relay and keep serving new connections.
+            (dotimes [_ 2]
+              (is (thrown? Exception
+                           (with-open [conn (nrepl/connect :port (:port b))]
+                             (let [client (nrepl/client conn 20000)]
+                               (doall (nrepl/message client {:op "eval" :code "(+ 1 2)"})))))))
+            (finally (bridge/stop-bridge b))))))))