diff --git a/broker/src/tunneldigger_broker/eventloop.py b/broker/src/tunneldigger_broker/eventloop.py
index d31461a..0c1af21 100644
--- a/broker/src/tunneldigger_broker/eventloop.py
+++ b/broker/src/tunneldigger_broker/eventloop.py
@@ -54,10 +54,10 @@ def start(self):
 
                     pollable, file_object = mapping
 
-                    if event & select.EPOLLIN:
+                    if event & select.EPOLLIN or event & select.EPOLLERR or event & select.EPOLLHUP:
+                        # If there's anything new, we read. If there was an error, the read will
+                        # return that error so we can handle it.
                         pollable.read(file_object)
-                    elif event & select.EPOLLERR or event & select.EPOLLHUP:
-                        pollable.error()
             except IOError:
                 # IOError get produced by signal even. in version 3.5 this is fixed an the poll retries
                 # TODO: in py3 it's InterruptedError
diff --git a/broker/src/tunneldigger_broker/hooks.py b/broker/src/tunneldigger_broker/hooks.py
index 7170104..43d5b59 100644
--- a/broker/src/tunneldigger_broker/hooks.py
+++ b/broker/src/tunneldigger_broker/hooks.py
@@ -50,9 +50,6 @@ def register(self, event_loop):
         event_loop.register(self, self.process.stderr, select.EPOLLIN)
         self.event_loop = event_loop
 
-    def error(self):
-        self.close()
-
     def close(self):
         """
         Closes the hook process.
@@ -163,9 +160,6 @@ def run_hook(self, name, *args):
         except OSError as e:
             logger.error("Error while executing script '%s': %s" % (script, e))
 
-    def error(self):
-        self.close()
-
     def close(self):
         os.close(self.sigchld_fd)
 
diff --git a/broker/src/tunneldigger_broker/network.py b/broker/src/tunneldigger_broker/network.py
index 2c98503..7225192 100644
--- a/broker/src/tunneldigger_broker/network.py
+++ b/broker/src/tunneldigger_broker/network.py
@@ -55,8 +55,11 @@ def register(self, event_loop):
         event_loop.register(self, self.socket, select.EPOLLIN)
         self.event_loop = event_loop
 
-    def error(self):
-        self.close()
+    def error(self, direction, e):
+        """
+        `direction` can be "reading" or "writing"
+        """
+        logger.warning("{}: error {} socket: {}".format(self.name, direction, e))
 
     def close(self):
         """
@@ -106,9 +109,6 @@ def read(timer_self, file_object):
                     if interval is None:
                         timer_self.close()
 
-            def error(self):
-                self.close()
-
             def close(timer_self):
                 self.event_loop.unregister(timer)
                 self.timers.remove(timer_self)
@@ -130,6 +130,7 @@ def write(self, address, data):
         try:
             self.socket.sendto(data, address)
         except socket.error:
+            self.error("writing", e)
             return
 
     def write_message(self, address, msg_type, msg_data=b''):
@@ -167,11 +168,7 @@ def read(self, file_object):
         try:
             data, address = self.socket.recvfrom(2048)
         except socket.error as e:
-            if e.errno == errno.EMSGSIZE:
-                # silence these, they occur during PMTU probing
-                pass
-            else:
-                logger.warning("{}: error reading from socket: {}".format(self.name, e))
+            self.error("reading", e)
             return
 
         msg_type, msg_data = protocol.parse_message(data)
diff --git a/broker/src/tunneldigger_broker/tunnel.py b/broker/src/tunneldigger_broker/tunnel.py
index 8b32cfd..6c9792c 100644
--- a/broker/src/tunneldigger_broker/tunnel.py
+++ b/broker/src/tunneldigger_broker/tunnel.py
@@ -239,16 +239,13 @@ def keepalive(self):
             logger.warning("%s: timed out", self.name)
             self.close(reason=protocol.ERROR_REASON_TIMEOUT)
 
-    def error(self):
-        # Read from the socket, to "consume" the error (and show it in the log)
-        self.read(None)
-        # Here we have a problem. This can indicate permanent connection failure
-        # (https://github.com/wlanslovenija/tunneldigger/issues/143), or it can
-        # indicate that we sent a packet that was too big (e.g. the PMTU probe
-        # reply, see https://github.com/wlanslovenija/tunneldigger/issues/171).
-        # To distinguish the two, we count how many consecutive errors we see without a proper message in between.
-        # If that reaches a threshold, we consider this error permanent and close the connection.
-        # PMTU discovery sends 6 probes, so 10 should be safe as a threshold.
+    def error(self, direction, e):
+        if e.errno == errno.EMSGSIZE:
+            # ignore these, they occur during PMTU probing
+            return
+        super(Tunnel, self).error(direction, e)
+        # This connection is probably dead, but we've seen connections in the wild that occasionally raise
+        # an error. So only kill the connection if this happens again and again.
         # We could just rely on the timeout, but when there's a lot of errors it seems better to
         # kill the connection early rather than waiting for 2 whole minutes.
         self.error_count += 1
@@ -347,8 +344,9 @@ def message(self, address, msg_type, msg_data, raw_length):
 
         # Update keepalive indicator.
         self.last_alive = time.time()
-        # Remember that we got a message -- reset error count for transient error tolerance.
-        self.error_count = 0
+        # We got a message, so *something* is working. Reduce error count for transient error tolerance.
+        if self.error_count > 0:
+            self.error_count -= 1
 
         if msg_type == protocol.CONTROL_TYPE_ERROR:
             # Error notification from the remote side.