OpenVPN (64-bit)

What's new in this version:

Security fixes:
- openvpnserv (windows): fix DNS SearchList state pollution on (dis)connect.
- specific combinations of --dns config entries plus local DNS config could lead to corruption of pre-openvpn DNS config
- Fix use-after-free bug in ack_write_buf(), triggerable by a well-timed sequence of control channel + authentication packets

Bug found by multiple researchers:
- Fix use-after-free bug in tls_wrap_reneg(), triggerable by suitable sequence of dynamic tls-crypt control-channel packets
- Fix server crash on reception of suitably malformed auth-token, if auth-gen-token external-auth is active
- Fix memory-leak in tls-crypt-v2 client key handling that could lead to out-of-memory situations and subsequent server crashes
- Fix possible 1-byte buffer overrun on NTLMv2 proxy responses
- Fix another memory leak on reception of suitable tls-crypt-v2 packets that could lead to an out of memory situation and server crash

Fixed:
- Windows: fix plugin trusted-dir check prefix bypass this fixes a bug in the path checking logic we do on Windows for "is loading a plugin from this path allowed?", but since we could not find a way to exploit this unless starting with admin privs or a social engineering attack, not classified as a security fix)
- Windows: openvpnserv: rework ConvertItfDnsDomains and tests this fixes a buffer overread that is not exploitable and as such not classified as security fix)
- options: fix use-after-free of DNS options on client connect using suitable --dns or --dhcp-option DNS options in a server config - not pushed, but applying to the server itself - triggers a double free() and use-after-free condition, possibly crashing the server) (Github: OpenVPN/openvpn#1060)
- dns: Fix memory leak in dns_server_addr_parse, if too many server addresses are configured (Github: OpenVPN/openvpn#1055)
- improve multi-socket event handling further - multiple open UDP sockets with concurrent traffic could lead to inefficient processing, and theold code was also very hard to follow. This was initially triggered by a report from Joshua Rogers using ZeroPath, but turned out to be "just bad code" not a security vulnerability)
- Null-terminate tls-crypt client keys when testing - non-exploitable strlen() on a buffer that is not null-terminated mudp: send HMAC reset reply synchronously this fixes a bug where multiple incoming tls-crypt-v2 RESET packets on different sockets could end up overwriting each other's control structures, leading to initial handshake packets (HMAC reset reply) being sent to the wrong client IP, or on a non-suitable socket "v4 packet on a v6 socket"). Since the overall flow here is stateless by nature, do not artificially create state by creating elaborate queues, just send-or-drop.
- fix port-share and multi-socket interaction - port-share needs TCP listeners, but the check was wrong. So "as long as any of the listening sockets is TCP, port-share can be used" (Github: OpenVPN/openvpn#1027)
- Ensure pushed tun-mtu is no lower than TUN_MTU_MIN - this fixes a bug where a server can push a suitable combination of options and make the client ASSERT(). Reported as security issue by Haiyang Huang, but it was decided that the server always has means to make the client "not function properly", and it can not be exploited beyond that)
- Windows: socket: assert buffer length before reading prepended sockaddr family - a misbehaviour in the windows DCO driver could trigger an overread in the userland client. No such bug exists, which this was not treated as a security vulnerability

Documentation improvements:
- improve documentation for --float (Github: OpenVPN/openvpn#358)
- add documentation for --preresolve (Github: OpenVPN/openvpn#532)
- impove documentation around DNS config (Github: OpenVPN/openvpn#937)