login-common: Fix connection hanging if ostream write fails

If the ostream write failed, the istream was closed. This prevented its IO
callback from being called, so the connection was just stuck.

This happaned at least with haproxy health check connections.

diff --git a/src/login-common/client-common.c b/src/login-common/client-common.c
index 1cc1b2dcbb6e6716b13bd93a654ef46864d5acb7..ec954d63f76f07dce140627c9cce464311c4f969 100644 (file)
--- a/src/login-common/client-common.c
+++ b/src/login-common/client-common.c
@@ -1649,6 +1649,11 @@ void client_notify_status(struct client *client, bool bad, const char *text)
                client->v.notify_status(client, bad, text);
 }
 
                client->v.notify_status(client, bad, text);
 }
 

+static void client_send_raw_data_destroy(struct client *client)
+{
+       client_destroy(client, "Disconnected: Output error");
+}
+

 void client_common_send_raw_data(struct client *client,
                                 const void *data, size_t size)
 {
 void client_common_send_raw_data(struct client *client,
                                 const void *data, size_t size)
 {


@@ -1659,8 +1664,18 @@ void client_common_send_raw_data(struct client *client,
                /* either disconnection or buffer full. in either case we want
                   this connection destroyed. however destroying it here might
                   break things if client is still tried to be accessed without
                /* either disconnection or buffer full. in either case we want
                   this connection destroyed. however destroying it here might
                   break things if client is still tried to be accessed without

-                  being referenced.. */
+                  being referenced, so do it lazily from a timeout. */
+               io_remove(&client->io);

                i_stream_close(client->input);
                i_stream_close(client->input);

+               /* If the client is already being destroyed, don't schedule
+                  another destroy. Its to_disconnect was already removed and
+                  wouldn't be cleaned up again, leaking the timeout. */
+               if (!client->destroyed) {
+                       timeout_remove(&client->to_disconnect);
+                       client->to_disconnect =
+                               timeout_add_short(0,
+                                       client_send_raw_data_destroy, client);
+               }

        }
 }
 
        }
 }