Index: /trunk/src/VBox/NetworkServices/NAT/proxy.c
===================================================================
--- /trunk/src/VBox/NetworkServices/NAT/proxy.c	(revision 54899)
+++ /trunk/src/VBox/NetworkServices/NAT/proxy.c	(revision 54900)
@@ -358,4 +358,35 @@
             closesocket(s);
             return INVALID_SOCKET;
+        }
+    }
+#endif
+
+#if defined(RT_OS_WINDOWS)
+    /*
+     * lwIP only holds one packet of "refused data" for us.  Proxy
+     * relies on OS socket send buffer and doesn't do its own
+     * buffering.  Unfortunately on Windows send buffer is very small
+     * (8K by default) and is not dynamically adpated by the OS it
+     * seems.  So a single large write will fill it up and that will
+     * make lwIP drop segments, causing guest TCP into pathologic
+     * resend patterns.  As a quick and dirty fix just bump it up.
+     */
+    if (stype == SOCK_STREAM) {
+        int sndbuf;
+        socklen_t optlen = sizeof(sndbuf);
+
+        status = getsockopt(s, SOL_SOCKET, SO_SNDBUF, (char *)&sndbuf, &optlen);
+        if (status == 0) {
+            if (sndbuf < 64 * 1024) {
+                sndbuf = 64 * 1024;
+                status = setsockopt(s, SOL_SOCKET, SO_SNDBUF,
+                                    (char *)&sndbuf, optlen);
+                if (status != 0) {
+                    DPRINTF(("SO_SNDBUF: setsockopt: %R[sockerr]\n", SOCKERRNO()));
+                }
+            }
+        }
+        else {
+            DPRINTF(("SO_SNDBUF: getsockopt: %R[sockerr]\n", SOCKERRNO()));
         }
     }