Using a Linux L2TP/IPsec VPN server

I heartily endorse 
this gigantic book!  

eXTReMe Tracker
Last update: May 24, 2010

1.1 Introduction

This webpage contains information on how to use L2TP/IPsec clients from Microsoft, Apple and other vendors in a 'Road Warrior' setup connecting to a Linux VPN server based on FreeS/WAN or its successors. FreeS/WAN is an IPsec implementation for Linux 2.x kernels, released under the GNU Public Licence. FreeS/WAN has been succeeded by Openswan and strongSwan. The main advantage of such a setup is cost: if you use the free clients that are included with Windows and Mac OS X, you don't have to spend money on third-party VPN clients or servers. Obviously you still need OS licences for those clients if you want to be legal.

IPsec is a network protocol for secure communication. It's an official Internet standard. Clients and devices from different vendors should interoperate (in theory), as long as they adhere to the IPsec standard. More background information on IPsec can be found in this illustrated guide to IPsec. The following L2TP/IPsec clients are supported by the setup described on this webpage:

Windows 2000 and Windows XP Using a Linux L2TP/IPsec VPN server with Windows 2000/XP
Windows Vista
Using a Linux L2TP/IPsec VPN server with Windows Vista
Windows 95 / 98 / Me / NT4 Using a Linux server with the Microsoft L2TP/IPSec VPN Client
SSH Sentinel, Forticlient and SafeNet SoftRemote Using a Linux server with third-party L2TP/IPsec clients
Windows Mobile 6, Windows Mobile 5.0 and Pocket PC 2003
Using a Linux L2TP/IPsec VPN server with Windows Mobile
Mac OS X v10.3 Panther and higher
Using a Linux L2TP/IPsec VPN server with Mac OS X
Using Linux as an L2TP/IPsec client

Windows 2000/XP/Vista, Pocket PC 2003, Windows Mobile and Mac OS X v10.3+ ship with a built-in L2TP/IPsec client. The "Microsoft L2TP/IPSec VPN Client" for Windows 95 / 98 / Me / NT4 is a free download from the Microsoft website. For brevity, I call it the "MSL2TP client" below. SSH Sentinel, Forticlient and SafeNet SoftRemote are third-party clients for Windows that support both IPsec and L2TP.

The information on this webpage is applicable to all L2TP/IPsec clients mentioned above. The Linux part of the configuration is basically the same for all these clients. Once you have read this page, check out the webpage for the particular client(s) that you want to use.

Nate Carlson has made an 'executive summary' for people who want just the facts.
There are several IPsec implementation available for Linux:
A Linux IPsec implementation typically consist of a kernel part and corresponding userland utilities. The kernel part of FreeS/WAN, Openswan and strongSwan is called KLIPS. The userland IKE daemon is called 'pluto'. Vanilla kernels (2.4 and older) do not ship with KLIPS by default. You will have to apply a KLIPS kernel patch or install loadable kernel modules for KLIPS. As mentioned above, kernels 2.6 and higher ship with a native IPsec implementation called NETKEY. Recent versions of FreeS/WAN c.s. support not only KLIPS but also NETKEY. To make things even more complex, there is also a NETKEY backport for kernel 2.4 and work is in progress to port KLIPS to kernel 2.6. This means that you have the following userland vs. kernel options on the Linux side:

Kernel 2.0 KLIPS
Kernel 2.2 KLIPS
Kernel 2.4 KLIPS
Kernel 2.4 NETKEY backport 1) 2)
Kernel 2.6 KLIPS Kernel 2.6 NETKEY 1)
FreeS/WAN 1.x

FreeS/WAN 2.x

Openswan 1.x X

Openswan 2.x

strongSwan 2.x


ipsec-tools utilities 3)


isakmpd Linux port



1) Linux 2.6+ contains NETKEY, a native IPsec implementation.
2) NETKEY has also been backported to kernel 2.4. This port is not included with the vanilla Linus kernel but some Linux distributions (Debian in particular) include the backport in their kernels.
3) The ipsec-tools utilities (including the IKE daemon 'racoon') are a Linux port of KAME. Ipsec-tools is included in most distributions.
4) There are issues with the heavily modified kernels of some distributions such as RHEL 3.

There is a feature comparison between FreeS/WAN and Openswan available and also a feature comparison between Openswan and strongSwan. Someone should make a good feature comparison between KLIPS and NETKEY but currently there isn't one. Each option has its pros and cons. I have not tested all combinations. Nowadays most people use Openswan. If you are interested in using ipsec-tools with kernel 2.6 and L2TP/IPsec, then you probably should check out this webpage by Chris Andrews. I have also not tested L2TP on top of other IPsec implementations such as those from OpenBSD (isakmpd) or FreeBSD. These BSD versions use KAME (or its fork ipsec-tools because development of KAME's racoon has ceased), so I assume that the procedure would be similar to what is described by Chris.

Disclaimer: I do not have experience with this setup in production use. But since the writing of these pages, commercial Linux products have started to support a similar (if not the same) L2TP/IPsec setup.

1.2 Results

The following Windows L2TP/IPsec clients have been tested:

The following L2TP/IPsec clients are available from Apple (for more info, see my other page):

Linux can connect as an L2TP/IPsec client to other L2TP/IPsec servers. See also this webpage.

All clients mentioned above support some form of NAT-Traversal. Note that you may need to obtain the latest version of your client to actually get the NAT-T support. Although NAT-T is supported by these clients, it may not always work when you connect to a Linux IPsec server. Some of the clients have been tested successfully, others do not work yet (this is discussed below).

1.3 Commercial products with L2TP/IPsec

Here is a list of L2TP/IPsec server products, in case you rather buy something off the shelf. Most of these are closed source, so you may have to pay for user licences. See also this feature chart on the VPNC website. (The list below does not imply that these products have been tested against Linux L2TP/IPsec).

1.4 Non-commercial and Open Source products with L2TP/IPsec

1.5 Author

The author of this document is Jacco de Leeuw. Corrections, additions, extra information etc. are much appreciated. A big thank you to everybody who has provided feedback!

2. Contents
3. Background information

Microsoft has included an IPsec client with Windows 2000 Professional / Server, Windows XP Home / Professional, Windows Vista, Pocket PC 2003 and Windows Mobile. The client is supplied with the base operating system so you do not have to download it. See also my other webpages.

A separate IPsec client can be downloaded for free from the Microsoft website. This 'MSL2TP client' can be installed on Windows 95 / 98 / Me / NT4. Although the client is different from the one included with Windows 2000/XP/Vista, it is quite similar in functionality. For more information on the MSL2TP client, see my webpage "Using a Linux server with the Microsoft L2TP/IPSec VPN Client".

As far as I know, there is still no Microsoft client for Windows 3.x, Windows NT 4.0 Server and Pocket PC versions 2002 and older. But there are third-party IPsec clients on the market. For non-commercial use you can download the free PGPNet, but it is is limited to host-to-host connections.

There is however a snag with the free IPsec clients from Microsoft. You can use IPsec only in combination with another protocol called L2TP. It is fairly difficult (2000/XP/Vista) or probably even impossible (MSL2TP, Pocket PC) to get rid of this L2TP requirement. One might say that Microsoft "embraced and extended" the IPsec standard, in true Microsoft fashion. To be fair though, L2TP is currently a 'Proposed Internet Standard' (RFC 2661 ) and so is 'L2TP over IPsec' (RFC 3193). PPTP, on the other hand, is another widely used VPN protocol but it is not an official standard.

The use of the L2TP protocol means that you will have to use an L2TP daemon. Several L2TP daemons are available. I will come to that later. When a Windows L2TP/IPsec client connects to your Linux server, it first sets up an IPsec tunnel to Openswan. Then it uses the tunnel to connect to the L2TP daemon on the Linux server, after which the client can access machines on the internal network.

There is an alternative option if you happen to have a Windows 200x Server. With the L2TP/IPsec client you connect to the Linux IPsec server and then use that IPsec tunnel to connect to the L2TP server of Windows 200x, instead of Linux. Martin Köppe has written a Howto on this. The advantage would be that Microsoft's L2TP server is presumably more compatible with the Windows clients than the Open Source L2TP daemons. Note that with this setup the Windows 200x Server is not directly connected to the Internet: the Linux server is. One might regard this a security advantage. Users are authenticated through PPP against Windows 200x instead of Linux, so you will need Client Access Licences.

Microsoft apparently does not think that IPsec is a good protocol for authenticating teleworkers ("Roadwarriors"). This is a bit odd, because third-party clients (PGPNet, SoftRemote etc.) have absolutely no problem with it. Microsoft's rationale is stated on their website in a VPN FAQ. Basically, they claim that passwords are easier to use than certificates. But I suspect there might be another tactical decision behind this. The L2TP protocol supports additional authentication mechanisms which apparently suit Microsoft better (e.g. authenticating through the Windows logon credentials, which means selling more NT/200x client licences).

Back to Contents

4. Procedure overview Here is a schematic of the setup. The IP addresses in this picture are used in my sample configuration files as well.

The topology of a Linux L2TP/IPsec VPN server

This may look difficult (which it probably is) but if you already have a working Openswan system it should boil down to installing an extra RPM/Debian package for the L2TP daemon. You will also need to change the L2TP configuration files a little bit. My example configuration files use 192.168.1.x and you will need to change this to the subnet used on your internal network. One thing which complicates things a bit is NAT-Traversal support. I suggest you first try to get it working without NAT.

Back to Contents

5. Safety first: some security considerations

Read this part carefully. If you have just started reading this page you don't have to take immediate action. Just keep it in the back of your mind and remember to come back to it when you have got the whole lot working and you are about to hook it up to a hostile network such as the Internet.

5.1.1 Ports to open on the external interface: IPsec

You are probably using a firewall on your VPN server (or in front of it). In that case you will need to adjust your firewall so that the L2TP/IPsec protocol is allowed in.

There are many different types of firewalls on Linux. There are even differences between versions of the same Linux distribution. Many people prefer to use their own custom firewall rules instead of those of the distribution. I'm afraid it is impossible for me to provide specific instructions for each and every type of firewall. See also the FreeS/WAN documentation.

For L2TP/IPsec you need to open the same protocols and ports as for pure IPsec:

Note that IP 50 is a protocol, not a port. As you might know, IP has protocols such as ESP, UDP and TCP. The protocols TCP and UDP on their turn have ports such as TCP 80 (HTTP), UDP 500 (IKE) etc.

5.1.2 Ports to block on the external interface: L2TP

You should firewall the L2TP daemon on the external (hostile) interface so that it is not accessible. Or better said: you MUST firewall the L2TP daemon, otherwise you will be taking a huge risk.

So for L2TP/IPsec you need to firewall the following port on all interfaces except ipsec0 (or eth0 if you are using NETKEY instead of KLIPS):

5.2.1 The listen-addr parameter (KLIPS only)

In addition to firewalling L2TP, here is another suggestion which makes your setup even more secure. This step is described in the following paragraphs. By default the L2TP daemon listens on UDP port 1701. Should your firewall be down (for some reason), the L2TP daemon would be exposed on the external interface. So you do not want anyone to access the L2TP daemon through the VPN server's external interface. What you want is that only IPsec authenticated clients are able to access the L2TP daemon, i.e. L2TP packets should flow through the IPsec tunnel and not directly (unencrypted) between the server and the clients. However, by default the L2TP daemon listens on all interfaces, even on the external (hostile) interface. It binds to INADDR_ANY (for those in the know). Preferably you would like the L2TP daemon to bind to only the ipsec0 interface. Unfortunately, this is not possible. Unlike low-level network applications such as tcpdump and Ethereal you cannot bind the L2TP daemon to a particular interface.

Fortunately, there is a way to bind the L2TP daemon to a particular IP address. Patches are available for two of the major Open Source L2TP daemons (l2tpd and rp-l2tp) which allows the server to bind (listen) to one particular IP address. This listen-addr patch for l2tpd is included in my l2tpd RPM (from version 7jdl and up). To use this patch you add a parameter "listen-addr" to the L2TP daemon's configuration file (l2tpd.conf). The daemon will then listen on the IP address (the internal interface in my example configuration).

So now you have an L2TP daemon listening on the internal interface. The daemon cannot be accessed from the external interface, which is good. But the L2TP daemon should be accessible through the ipsec0 interface. This is done by configuring an iptables rule which forwards L2TP packets coming from the ipsec0 interface to the internal interface:

iptables -t nat --append PREROUTING -i ipsec0 -p udp --sport 1701 --dport 1701 -j DNAT --to-destination

(Where is again the IP address of the internal interface). The rule is deleted with:

iptables -t nat --delete PREROUTING -i ipsec0 -p udp --sport 1701 --dport 1701 -j DNAT --to-destination

Openswan must be running when you execute these lines, i.e. ipsec0 must exist. Alternatively, you could add these extra rules to a firewall script called by Openswan, namely the one specified by the leftfirewall= parameter. See also the FreeS/WAN documentation on this.

When the listen-addr parameter is used properly, the L2TP daemon will not listen on the external interface. So, should the firewall be down (shit happens), then the L2TP daemon will not be exposed on the external interface. It's still prudent to firewall incoming L2TP connections (UDP port 1701) on all interfaces except ipsec0. Use firewall blocking and the listen-addr parameter in tandem (a "belt and suspenders" approach).

5.2.2 The listen-addr parameter (with NETKEY)

Unfortunately the listen-addr trick mentioned above does not work with the native kernel 2.6 IPsec implementation (NETKEY). That is because NETKEY does not have ipsec0 style interfaces and NAT-after-IPsec is currently broken on vanilla kernel 2.6. There are 5 ways to solve (or work around) this problem on 2.6 kernels. Unfortunately I have not tested all of them yet. The first is to use KLIPS in the 2.6 kernel so that you will have ipsec0 style interfaces. Openswan 2.3+ supports KLIPS for 2.6 kernels but it is fairly new and might have issues. Another option is to use kernel 2.6.16 or higher which contains Patrick McHardy's ipsec hook netfilter patches (as mentioned here), with iptables >= 1.3.5. A third option is to build your own 2.6 kernel with these netfilter patches. A fourth (suboptimal) solution is use firewall rules on the IPsec server, i.e. have your L2TP daemon listen on all interfaces and then firewall all incoming L2TP connections on external interfaces. It is probably a good idea to use the "mark" option in netfilter (see Chris Andrews' page for an example). A final (suboptimal) solution is to place a separate firewall (possibly with NAT) in front of the VPN server. Then you can have the VPN server's L2TP daemon listen on all interfaces (you can even use firewall rules on the VPN server; having two firewalls should not be a problem). In these last two suboptimal cases you are relying on the firewall; should it be compromised, (accidentally) disabled or otherwise, the L2TP daemon is exposed. With potentially dire consequences...

5.3 ip_forward

One other security related point to notice is that people often set /proc/sys/net/ipv4/ip_forward to 1 for (VPN enabled) routers, so that packets coming from the IPsec tunnel are forwarded to the internal network. This can be done by adding forwardcontrol=yes to ipsec.conf. However, there are some security implications. Perhaps one or more iptables forward rules could do the same trick, when restricted to certain interfaces. Or you could use iproute2 (advanced routing). This is a bit outside the scope of this document.

5.4 chroot, UML, XEN, SELinux etc.

Both Openswan and l2tpd run as root. For extra security you could try to shoehorn them into a chroot jail or an SELinux policy. Or you could even virtualise your server with Usermode Linux, Xen, etc. I have not attempted to do this but apparently the people at Astaro have managed to run the L2TP/IPsec server in chroot. You could download an evaluation copy and check out how they did it. A commercial product that uses virtualisation to support multiple L2TP/IPsec tunnels is Stinghorn. Red Hat / Fedora appears to have a default SELinux policy but it is for racoon, not Openswan.

Back to Contents

6. VPN alternatives

Before I dig into the technical details of setting up Openswan with L2TP, let's take one step back. I assume that you are interested in providing remote access over the Internet to your users. Important factors in this are price, security and user friendliness and often you can only pick 2 out of these 3 factors. Several solutions are available, such as:

  1. A hardware device (or "appliance") at the client side.
  2. PPTP or SSTP, for instance using the clients included with Windows 95 and later, Mac OS and Linux/Unix.
  3. A remote desktop solution such as Citrix, Windows Terminal Server, pcAnywhere or VNC.
  4. An SSL-based VPN, such as SSL Explorer, HOB or Citrix Secure Gateway.
  5. Non-standards based Open Source solutions such as CIPE, vtun, tinc and OpenVPN.
  6. Non-standards based proprietary solutions such as Hamachi.
  7. Third-party IPsec clients such as NCP, PGP, SafeNet SoftRemote, SSH Sentinel or TheGreenBow VPN Client.
  8. The IPsec client included with Windows 2000/XP/Vista, manually configured to get rid of L2TP.
  9. The IPsec client included with Windows 2000/XP/Vista, configured with a free software tool to get rid of L2TP.
  10. L2TP/IPsec clients such as Windows 2000/XP/Vista, Pocket PC 2003, Windows Mobile, Mac OS X v10.3+ and the MSL2TP client (Win9x/Me/NT4).

Ad 6.1: These are hardware devices with PPTP/IPsec support such as those from Checkpoint, Cisco, Draytek or Netscreen. Hardware devices are easy to configure for the user (because the system administrator does it for them :-). The hardware box is then handed in person or shipped to the user. Another advantage is that most of these devices also contain a firewall. Security of hardware devices is fairly high. Some hardware devices can be centrally managed. A disadvantage is the price, which tends to be higher when the device has more VPN capabilities. Hardware boxes are also not very portable so they are less suitable for roaming users. Another problem (the classic hardware/software trade off) is that they may be difficult to extend with new features. Many hardware devices only support Preshared Keys (i.e. passwords) and not certificates. But Netscreen and SnapGear do support certificates. If the hardware device only supports Preshared Keys, the user is required to have a fixed IP address, unless the device also supports Aggressive Mode.

Ad 6.2: PPTP is free and easy to use but that's about it, really. PPTP clients are included with Windows 98 and higher. Updates (including a version for Windows 95) are available for free from the Microsoft website. PPTP clients are available for other operating systems as well, including BSD, Linux and Mac. An advantage of PPTP is that it supports NAT-Traversal, i.e. it can get through (most) devices that employ Network Address Translation (NAT). Microsoft claims that PPTP has a "good level of security that is suitable for most companies". Security professionals such as Bruce Schneier think otherwise. A German student managed to crack PPTP passwords in 4 hours. Asleap by Joshua Wright has been extended to crack weak PPTP passwords. The authors of the premier Open Source PPTP implementations PPTPClient and Poptop recommend against using PPTP. Others have found a flaw in Microsoft's PPTP implementation. Even Microsoft now admits that "as a VPN protocol, Microsoft considers PPTP to be non-strategic". Note also that PPTP is not an official Internet standard. It is an 'de-facto industry standard' set by Microsoft. From Microsoft's VPN FAQ mentioned above you might easily get the impression that their PPTP and MS-CHAPv2 protocols are official IETF standards. This is not the case. You can check that on the RFC website. It's even mentioned in the RFCs themselves: "[This RFC] does not specify an Internet standard of any kind". Those two RFCs have never been ratified as official standards. They expired. Quoting the Internet RFC FAQ: "Unscrupulous marketeers and a careless trade press sometimes falsely suggest that every RFC represents a standard, or that all standards have equal weight". PPTP also supports the MPPC compression protocol which is patent encumbered. So is PPTP proprietary? Microsoft says no, but they are not telling the whole story. Microsoft has developed yet another proprietary VPN protocol called SSTP. It is included with Windows Vista SP1 and Windows Server 2008. Other operating systems are not supported. SSTP is based on SSL and should be able to pass through most routers, unlike the GRE protocol used in PPTP which is sometimes blocked by routers.

Ad 6.3: Remote desktop clients and servers often use proprietary protocols (VNC and NX are the exceptions). Effectively, this is "security through obscurity". Security experts generally don't like this concept. There is often little encryption (VNC) or the protocol is vulnerable to a Man-In-The-Middle Attack (VNC, RDP). Prices can also be fairly steep (Citrix, GoToMyPC, pcAnywhere). Again, VNC is the exception because it is open source. Windows Terminal Server (RDP) requires extra client licences (CALs). Remote desktop clients such as VNC and pcAnywhere only support one user per system.

Ad 6.4: SSL-based VPNs are very flexible to set up. In many cases the only requirement is that the user must have a recent browser. The user will have to surf to a certain website and then a special client will be downloaded automatically using Java or ActiveX. This means you will have to maintain and administrate a webserver connected to the Internet. Note that in principle the user can connect from anywhere, even dubious locations such as airport kiosks, Internet cafés etc. Once the client software has been downloaded, the user is presented with a remote desktop (similar to 6.3). In some products, the client will switch to a proprietary protocol (HOB, Windows TS ActiveX). Other products continue to use SSL for the remote desktop part (Citrix Secure Gateway?). SSL is used for downloading the applet and the remote desktop protocol itself. Proprietary protocols may be faster but SSL is a well-understood and open standard, so it is probably more secure than a vendor-specific protocol. Windows Vista SP1 ships with a SSL VPN called SSTP. SSL Explorer is the world's first open-source, browser-based SSL VPN solution. Commercial vendors are Aventail, F5 FirePass, Neoteris and Netilla (server appliance based on Tarantella). A disadvantage is that SSL based VPNs do not support all applications (compared to full blown VPNs such as PPTP/IPsec). For instance, one could argue that Outlook Web Access is also an SSL based VPN but you can use it only to access Exchange Server. Another disadvantage is that you will have to secure the special website and keep it locked down. Or, if you use an appliance, you better not forget to install security updates from the vendor.

Ad 6.5: There are also some other non-standards based Open Source solutions such as CIPE, vtun, tinc and OpenVPN. OpenVPN seems to be promising (based on OpenSSL, runs in user-space, is relatively lightweight, supports compression, simple-to-use and runs on 7 different OS's including Windows) but the others have serious security problems, according to security expert Peter Gutmann.

Ad 6.6: There are a few proprietary solutions of which Hamachi is an example. Hamachi is freeware and claims to be very easy to use ("zero-configuration"). It even works without change when both sides of the VPN are behind NAT. Although the makers of Hamachi have released some details about its protocol, it is still proprietary and a trade secret and thus has not seen any peer review by professional cryptographers. Source code is not available so you cannot fix bugs yourself, or add any new features and when there is no version for your operating system of choice, you are out of luck. Hamachi uses a central "mediation" server operated by the makers of Hamachi. All clients regularly communicate with this server.

Ad 6.7: A freeware IPsec client is available for Windows 2000/XP: the Shrew Software VPN Client. It has been tested with ipsec-tools based Linux VPN servers, but it works with Openswan/strongSwan as well. Alternatively, several vendors sell third-party IPsec clients: McAfee VPN client, NCP Secure Entry Client/Secure VPN/PKI Client (Openswan interop Wiki), PGP Mac VPN client, SafeNet SoftRemote, SSH Sentinel, Forticlient, TheGreenBow VPN Client. For a more complete list, see Openswan's interoperability overview. These clients generally have a lot of features, e.g. support for AES (a fast encryption standard). However, you will have to pay for these clients. This has the advantage that you receive support from the vendor, depending on your support contract. SSH Sentinel 1.2 "Internetpilot" (old and buggy) and PGP are free for non-commercial use.

Ad 6.8: By default, the IPsec clients included with Windows 2000/XP/Vista, Pocket PC and Mac OS X v10.3+ can only be used for tunnelling L2TP. You might want to get rid of L2TP, so that you don't have to install an L2TP daemon on your Linux server. This leaves you with 'pure' IPsec. Unfortunately, this is very difficult to do manually. You can find an example for Windows 2000 Professional on the Snapgear website. This is hardly user friendly by anyone's standards. As if Microsoft wanted to discourage pure IPsec without L2TP. Another drawback is that you will need Administrator privileges to change the IPsec settings using MMC. Note also that this VPN option is only available for Windows 2000, XP and Vista. The MSL2TP client and Pocket PC cannot be configured for IPsec without L2TP (as far as I know). In other words, if you have Windows 9x/Me/NT or Pocket PC, you will have to choose one of the other VPN options, possibly buying a third-party client (see option 5). This means you will have to support not one, but two different VPN options...

Ad 6.9: Fortunately, there are tools (wrappers around the built-in IPsec client). Two popular tools are available: the Linsys IPsec Tool (lsipsectool) and IPSEC.EXE by Marcus Müller. They do the configuring for you when you want to use Windows 2000/XP without L2TP. Like the previous VPN option, Administrator privileges are required to run the IPSEC.EXE utility, but you can get around this if you are prepared to run IPSEC.EXE as a Service (see this post by Uwe Knop). You will have to distribute the Linsys tool or the IPSEC.EXE program to your users, together with a configuration file and the user's private key and certificate. The configuration file resembles a Openswan config file, so you could even generate it on your Linux server. Nate Carlson has made an excellent webpage with instructions on using Windows 2000/XP with Openswan. There are also several graphical front-ends for the IPSEC.EXE utility available. There is also a free GUI that works similar to these graphical IPSEC.EXE wrappers: the Securepoint Personal Firewall & VPN Client. If you use Mac OS X v10.3+ instead of Windows, then there are similar programs available as well. Note that the MSL2TP client cannot be used with this option. It cannot be configured to get rid of L2TP, so you'll have the same MSL2TP related problem as with option 6.7.

Ad 6.10: I will go into more details below, as L2TP over IPsec is the main topic of this page. One thing is certain, though: you will have to use an L2TP server on the Linux side.

Back to Contents

7. (Dis-)advantages of using IPsec with L2TP

Here are the advantages and disadvantages of using IPsec with L2TP.


Cons: As you can see, there are a lot of good reasons to not use L2TP over IPsec. But the single fact that it is reasonably secure and cost effective might make it interesting enough to consider it, until better solutions such as IKEv2 and DHCP over IPsec are more widely available.

Back to Contents

8. Road Warrior support

VPN users often have dynamic IP addresses: they can have a different IP address with every connection that they make. A good example of this are travellers who connect with a laptop from hotels or conference rooms ('Road Warriors'). Also, some cable/ADSL providers use DHCP to assign dynamic IP addresses which change regularly.

In IPsec there are several ways to support this scenario:

A Preshared Key is a secret password that is shared by both sides of the IPsec tunnel. Preferably, the PSK is distributed 'out-of-band', i.e. not over the hostile network (read: the Internet). For instance, you could meet the user face to face and hand him the PSK on a piece of paper. PSKs are fairly simple to use, but they do not scale very well with the number of users and if you want to use a PSK in a Road Warrior setup, all users with dynamic IP addresses will have to share the same PSK ("group secret"). This is of course a significant security risk: if one user leaves the company or loses his laptop, all the other users will have to get a new PSK. The alternative would be to give every user a different PSK, but this is only supported in IPsec if all users have fixed (= static) IP addresses. Because of this limitation PSKs are generally not used in Road Warrior setups, unless there is only one user or everyone has a static IP address (e.g. home workers with cable or ADSL).

With RSA authentication you specify the raw RSA public key of the user in the Openswan configuration. RSA authentication supports both static and dynamic IP addresses. RSA authentication is also relatively light-weight to implement and almost as easy to use as a password. RSA keys have the advantage that they are inherently more secure because passwords can be guessed. Unlike passwords, RSA keys cannot be remembered by the user. They will have to be cut and pasted into the configuration. Unfortunately, none of the IPsec clients that support RSA authentication seem to support L2TP/IPsec, and vice versa.

Some IPsec clients support Aggressive Mode. This allows them to use PSKs with dynamic IP addresses. Especially clients for devices with relatively little processing power such as Pocket PC and Palm use Aggressive Mode, because RSA encryption is much slower than symmetric key encryption. There is a patch for FreeS/WAN that adds support for Aggressive Mode. This patch is also included with Openswan 1.x and SuperFreeS/WAN. The trouble with Aggressive Mode is that security depends on the strength of the password itself (PPTP has the same problem). There are programs such as IKEcrack and Cain&Abel that attempt to crack the Preshared Key from intercepted sessions, as explained by Michael Thumann.

XAUTH (Hybrid Mode) is an extension to IPsec that was never ratified by IETF because it required changes to the (already complex) IKE standard. Cisco seems to be a big XAUTH proponent. XAUTH is not supported by FreeS/WAN or strongSwan, but there is an implementation in Openswan (disabled by default; requires a recompile). Philippe Sultan demonstrates that the XAUTH username and password can be obtained using a spoofed server if the Preshared Key is already known. That Preshared Key can be obtained through brute-force cracking (as mentioned in the previous paragraph) or by copying from the client itself (disk or memory). A solution to this problem is to switch to Cisco's Mutual Group Authentication (server certificates and client passwords; currently only available for Cisco VPN Concentrators, not IOS) or to use certificates (requires client passwords and both client and server certificates, but the client certificates can be 'junk' so a full PKI can be avoided).

DHCP-over-IPsec is supported by Openswan but there is very little client support. The only Windows client is currently SSH Sentinel but this product has been discontinued. (Note: this is not the same as DHCP Inform, which is essentially DHCP-over-PPP-over-L2TP-over-IPsec).

IKEv2 is a successor to the current IKE. It will support 'legacy' authentication modes such as passwords through EAP. IKEv2 is expected to become the main standard. StrongSwan already supports IKEv2 and other implementations are under development.

X.509 certificates are supported by almost all L2TP/IPsec clients. Openswan supports it too, courtesy of a patch by Strongsec. Certificates are generally considered the way to go for Road Warriors. The disadvantage is that you will need to set up some kind of Public Key Infrastructure (PKI), which may be an administrative burden. You need to generate, distribute, revoke etc. the X.509 certificates for the Openswan host and the L2TP/IPsec clients. More information can be found in this section on certificates.

Other IPsec authentication methods such as AuthIP, CRACK, HYBRID, Kerberos and PIC exist as well, but currently there are no implementations available for Openswan or any other Linux IPsec implementation (as far as I know).

Back to Contents

9. Installation (Linux kernel etc.)

Here are the components to be used at the server side.

9.1 Obtaining Openswan / strongSwan / FreeS/WAN

You may prefer to use the kernel supplied with your Linux distribution if it already contains KLIPS or NETKEY. That should give you a head start. Later on, if you decide you need the latest versions, extra patches or more features, you can decide to compile your own kernel and userland utilities. This is not a tutorial on how to compile and install Openswan et al. Please refer to FreeS/WAN's documentation and Strongsec's instructions. Openswan is supported on many distributions, sometimes by the vendor, sometimes by third parties. I'll discuss the specifics for some of these distributions below. Openswan RPMs for SuSE, Red Hat, RHEL, and Fedora are available on the Openswan website. The corresponding SRPMs can be found elsewhere on the site.

Red Hat 9 and previous versions are no longer supported by Red Hat. Fedora 4 and previous version are also  no longer supported by the Fedora Project. The Fedora Legacy Project has been discontinued so there are no more security updates for these old versions. There are also other alternatives if you use an old Red Hat version. MandrakeLinux 10.0 and older are no longer supported by Mandriva. SuSE 8.0 and older are no longer supported. Obviously it does not make much sense to run a VPN server on an old distribution with unpatched vulnerabilities. The VPN part may be highly secure but it won't help much if you are using, say, a vulnerable kernel. This means that I won't be releasing updated RPMs for Mandrake 8.x and Mandrake 9.0. (Also from a practical standpoint: they simply had to make room for newer distributions on my systems). Of course you should be able to compile the Source RPM yourself if there is no RPM available for your distribution.

9.1.1 Red Hat, Fedora and RHEL

Red Hat does not include Openswan's KLIPS in Red Hat Linux or Fedora. They prefer to use the native IPsec implementation (NETKEY). This implementation is included in Fedora Core 2 and higher and RHEL. You will also need Openswan userland utilities to use NETKEY. Openswan RPMs are included in Fedora Core 3 and higher. RPMs are also available from the Openswan website, Axel Thimm's website and Dag Wieërs website.

Fedora Core 1 and Red Hat Linux up to version 9 do not ship with KLIPS or NETKEY. Axel Thimm, Dag Wieërs and the Openswan team itself provide KLIPS enabled kernel RPMs and Openswan userland RPMs for these older distributions. Source and binary RPMs from the Openswan team are available here. You need two RPMs: the Openswan kernel-module RPM and the Openswan userland RPM. The advantage of using an Openswan kernel-module RPM is that you don't have to install a new kernel. The kernel-module RPM is installed as an addition to the current (stock) kernel. You don't even have to reboot. The downside is that that this kernel-module RPM does not support NAT-T. If you do need NAT-T, you will have to install Axel's KLIPS enabled kernel instead of the Openswan kernel-module RPM. If you do not need NAT-T (or if you don't want to bother with it as this moment), simply use the Openswan kernel-module RPM.

If you want to use Openswan, do not install the ipsec-tools RPM that is also included with Fedora. It contains another IKE daemon, racoon, which conflicts with the IKE daemon of Openswan, pluto. If you prefer to use ipsec-tools (racoon) instead of Openswan, you probably want to check out Chris Andrews' webpage.

FC 2 and later (and quite possibly also RHEL) do not contain 'Legacy (BSD) PTY support' (CONFIG_LEGACY_PTYS=y), in contrast with previous FC and Red Hat versions. You will need to use l2tpd RPM version 11jdl or later. These versions use the new(er) Unix98-style pty system instead of the legacy BSD-style pty system. L2tpd version before 11jdl will not be able to start pppd and abort with the error message "getPtyMaster: No more free pseudo-tty's". An alternative is to switch from l2tpd to another L2TP daemon such as rp-l2tp.

If you use Red Hat Enterprise (RHEL) or clones such as CentOSLineox, Tao Linux, Whitebox Linux or X/OS Linux then I don't have much advice because I do not have these installed myself. It seems that they use the backported NETKEY implementation for kernel 2.4. Openswan RPMs for RHEL are available on the Openswan website.

According to Paul Wouters of the Openswan team, RHEL 3 (and its derivatives such as CentOS 3) "is the worst choice for a kernel for IPsec related matters" because its kernel is a 2.4 / 2.6 hybrid. You are better off with RHEL 4+ or Fedora.

9.1.2 SuSE / Novell

Novell (SuSE) was a sponsor of Openswan development. SuSE kernels already contain IPsec support, either through NETKEY or KLIPS. But SuSE employee Kurt Garloff has made source and binary RPMs for several versions of SuSE. They contain more IPsec related patches than SuSE's official kernels. Some of Kurt's KLIPS based kernels also contain the NAT-T patch. Alternatively, Openswan RPMs are available on the Openswan website.

9.1.3 Mandriva / Mandrake

The 2.6 kernels included with Mandriva 2005-2006 and Mandrake 10.x contain the native NETKEY implementation. FreeS/WAN 2.04 is included with Mandrake 10.0 and FreeS/WAN 2.06 is included with Mandrake 10.1, Mandrive 2005 (=10.2) and 2006. Unfortunately, FreeS/WAN version 2.06 cannot be used for L2TP/IPsec because Transport Mode has been removed. Openswan and strongSwan RPMs are available in the Mandriva Cooker / contrib directory. If you don't want to use FreeS/WAN or one of its successors, you can use racoon from ipsec-tools instead. Mandrake 10.0--10.2 also ship with a SuperFreeS/WAN RPM but it is based on FreeS/WAN 1.99 so it is utterly useless because that version of FreeS/WAN does not support NETKEY. Unfortunately the openswan RPM in Cooker has a dependency on the ipsec-tools RPM. That RPM will install and run (the competing IKE daemon) racoon by default. That means you will have to stop racoon (service racoon stop) and remove it from the startup scripts (chkconfig --del racoon) before you install the Openswan RPM.

The stock kernel included with Mandrakelinux 10.1 (and probably Mandriva 10.2 as well) does not contain 'Legacy (BSD) PTY support'. This is the same problem as with Fedora Core 2 and 3. This means that you will have to download l2tpd RPM version 11jdl or later. Or you will have to switch to another L2TP daemon such as rp-l2tp.

The stock kernels included with Mandrake 9.x contain KLIPS and the X.509 patch. A FreeS/WAN RPM is also included. This should in theory be sufficient to get you started. In practice however, there are some problems. The current kernel for Mandrake 9.2 at the time of this writing is kernel- Unfortunately that kernel contains FreeS/WAN 1.99.8 but the userland utilities are FreeS/WAN 2.01 (freeswan-2.01-1mdk.i586.rpm). So there is a mismatch and it will not work. You could try to compile the Openswan RPM from Mandrake Cooker. Mandrake 9.1's kernels for x86 are broken (kernel- There is a workaround, however. Kernel- for x86 is OK. And Mandrake 9.1's kernels for PowerPC are also OK. The current kernel for Mandrake 9.1 at the time of this writing is kernel- That kernel contains FreeS/WAN 2.00 but unfortunately the userland utilities are still FreeS/WAN 1.99, so it does not work. Mandrake 9.0 uses a fairly old version of the X.509 patch (be sure to avoid using special characters in certificate names!).

Mandrake 8.1 and 8.2 contain older versions of FreeS/WAN without the X.509 patch. The stock kernels included with 8.1 and 8.2 will not work with L2TP/IPsec because the X.509 patch is required. That means you should either upgrade your kernel or upgrade your Mandrake. Upgrading to a more recent Mandrake version is highly recommended. Should you still want to use Mandrake 8.x, you can find newer kernels here. You will also need an update for the FreeS/WAN user mode utilities (normally called freeswan-x.xx.rpm). I could not find these in the official updates but I did find a version 1.98 on Mandrake Cooker at Unfortunately (again), the latest Mandrake 8.x kernels contain older FreeS/WAN versions than 1.98 so there is a mismatch.

Only Mandrake 9.2 contains the "Delete/Notification" patch and the 'malformed payload' patch. But if you use an older Mandrake version it can be fixed.

9.1.4 Debian / Ubuntu / Knoppix etc.

Debian's "unstable" and "testing" distribution contain freeswan-modules-source and openswan. Simply run the command 'apt-get install openswan' in install the Openswan .deb file. Most of Debian's 2.4 and 2.6 kernels ship with NETKEY support. If you want to build a kernel yourself, you may need to install the kernel-headers, kernel-package, debianutils and fakeroot packages.

Debian ships with l2tpd, xl2tpd and l2tpns. There is a bug in recent versions of xl2tpd which results in a kernel Oops on Ubuntu 7.10 if AppArmor is running (which is the default). You'll have to either disable AppArmor (perhaps only for xl2tpd?), downgrade xl2tpd, change to another L2TP daemon or wait for this bug to be fixed.

9.1.5 Slackware

Derek T. Murphy has released (unofficial) kernels with FreeS/WAN support.

9.1.6 Gentoo

Openswan and strongSwan packages are available for Gentoo. Gentoo specific information can be found on the Gentoo forum and on this webpage by Brett Curtis. See also the Gentoo Wiki on the Openswan website.

9.1.7 Other distributions

You probably have to compile the kernel yourself, if there are none available on or other sites.

9.2.1 The X.509 certificate patch

The X.509 certificate patch by Strongsec adds support for variable IP addresses to FreeS/WAN. Openswan and strongSwan already contain this patch. In addition to X.509 support, the patch provides another feature which is vitally important to L2TP/IPsec. L2TP/IPsec clients want to restrict the IPsec tunnel so that only UDP packets (=IP protocol 17) are tunnelled. The standard version of FreeS/WAN can not make such restrictions. Vanilla FreeS/WAN wants to encrypt all protocols between itself and the other party, not just the L2TP traffic. Openswan, strongSwan and the X.509 certificate patch for FreeS/WAN support two parameters that solve this problem: leftprotoport and rightprotoport.

Most L2TP/IPsec clients (including the updated clients included with Windows 2000/XP, Windows Vista and XP SP2+) require the following parameters:


However the original (non-updated) L2TP/IPsec client included with Windows 2000/XP requires these parameters which seem to be a mistake (or at least an inconsistency) by Microsoft because no other vendor uses them):


Mac OS X 10.3.x+ clients require yet another configuration because they use a 'floating' UDP port. Fortunately, this configuration also covers the rightprotoport=17/1701 mentioned above:


On Openswan 2.4.10+ you use this instead:


So if you have a mixed environment with different clients, it is probably best to use leftprotoport=17/1701 and rightprotoport=17/0 (or 17/%any on older versions) and simply drop support for non-updated Windows clients. Those clients would have to install the KB Q818043 update or install XP SP2. (As an alternative, the Openswan development team suggests to use leftprotoport=17/%any and rightprotoport=17/%any but this allows access to all UDP daemons that may run on the Openswan server. This is a bit too lenient for my taste. Another solution would be an ugly hack where leftprotoport=17/0 is accepted as a side-effect if leftprotoport=17/1701 is specified in the configuration).

Only recent versions of the X.509 patch (e.g. 1.4.8+ for FreeS/WAN 2.x and 0.9.37+ for FreeS/WAN 1.99) can actually restrict the IPsec connection to the specified protocol and port. Older versions of the X.509 do support the left/rightprotoport parameters but they can not enforce those restrictions. They will accept the protocol 17 (UDP) / port 1701 (L2TP) restriction but still happily send all other traffic (ICMP, HTTP, SSH etc.) through the IPsec tunnel, ignoring the restriction.

9.2.2 Protocol tunnelling problem

If you don't use the X.509 patch there will be an interoperability problem between the L2TP/IPsec clients and FreeS/WAN (reported by Jason A. Pattie, among others). An IPsec connection can not be set up because FreeS/WAN complains:

 "peer client ID payload ID_IPV4_ADDR specifies protocol 17; we only support 0"

If you get this error then you will have to apply the X.509 patch to the FreeS/WAN userland programs.

9.2.3 "Delete/Notification" support

FreeS/WAN versions below 2.00 ignore "Delete/Notification" messages. Mathieu Lafon has made a "Delete/Notification" patch for FreeS/WAN which contains support for these messages. Openswan, strongSwan and FreeS/WAN versions 2.00 and higher already contain this patch.

A client may want to inform the server that it has taken down the IPsec connection. It may do so by sending a "Delete SA" message. If the server does support "Delete SA" messages, it simply ignores them and the IPsec connection may time out. In the mean time however, a client could of course reconnect to the server. The server then gets confused about the packets it receives. It thinks the packets are for the old connection, whereas the client is convinced it has a new IPsec connection. Mathieu Lafon's "Delete/Notification" patch fixes this. Note that if the client crashes or if the Internet connection breaks down, the client does not get a chance to send a "Delete SA" message. In that case, the server will time out, with or without the "Delete/Notification" patch. For more information on this issue, see this document by FreeS/WAN team member Henry Spencer. The "Delete/Notification" patch seems especially useful for the MSL2TP client (see this paragraph). (Note: the MSL2TP client seems to send each "Delete SA" message twice. The first message will normally remove the SA so the second one will result in a harmless "ignoring Delete SA payload" message).

Dominique Cressatti wrote in private correspondence that the "Delete/Notification" support is required if you want the connection to last for more than a few hours.

Mandrake RPMs lack "Delete/Notification" support. Also missing is the 'malformed payload' patch which is recommended for use with the MSL2TP client. I added these two patches to the original RPM (the modified RPMs have been signed with my PGP key). Again, I do not recommend using outdated and unmaintained versions of Mandrake, but here they are anyway:

   Mandrake Cooker original:  freeswan-1.98b-1mdk.src.rpm
   Modified 8.1/9.0 source:   freeswan-1.98b-2jdl.src.rpm
   Mandrake 8.1 binary (x86): freeswan-1.98b-2jdl.i586.rpm (please read this)
   Mandrake 9.0 binary (x86): freeswan-1.98b-2jdl.i586.rpm (please read this)
   Mandrake 9.1 original:     freeswan-1.99-3mdk.src.rpm
   Modified 9.1 source:       freeswan-1.99-4jdl.src.rpm
   Mandrake 9.1 binary (x86): freeswan-1.99-4jdl.i586.rpm
   Mandrake 9.1 binary (PPC): freeswan-1.99-4jdl.ppc.rpm

   You may need to hold the 'Shift' key while clicking these links!

Back to Contents

10. Configuration of Openswan

Once you have a kernel with IPsec support, you will need to configure one or more IPsec connections. (If you use FreeS/WAN, don't forget to disable Opportunistic Encryption first).

The examples in the Openswan documentation assume that "Roadwarrior" clients want to make "host-to-subnet" IPsec connections. I assume you want to do this as well. In other words, you want users to have access to an (internal) subnet protected by a Openswan gateway. In 'pure' IPsec (i.e. without L2TP) you do this using the left/rightsubnet keyword.

The Openswan documentation does not cover L2TP/IPsec connections. In L2TP/IPsec the user also wants to access the subnet, but the mechanism to achieve this is different. First, you specify a host-to-host IPsec tunnel between the Windows/Macintosh client and the Openswan server. Once this IPsec connection is up, the client connects to the L2TP server on the Openswan server. The L2TP server then forwards the packets to the internal subnet. In pure IPsec, Openswan does the forwarding itself.

10.1 Preshared Keys overview

Most L2TP/IPsec clients support two kinds of authentication methods: X.509 certificates and Preshared Keys (PSK). I suggest you try PSKs first, get an understanding of how IPsec works with the L2TP/IPsec clients and then switch to certificates if you need them.

10.1.1 Preshared Key: configuration an IPsec connection

Here is an example Openswan configuration file (included in my l2tpd RPM). It defines an IPsec connection for one user. Note that if you want to use this configuration you will have to add it to your ipsec.conf file (or add the following parameter at the end of your ipsec.conf file: include L2TP*.conf).

# Configuration supporting multiple users with any type of
# IPsec/L2TP client. This includes the updated Windows 2000/XP
# (MS KB Q818043), Vista and Mac OS X 10.3+ but excludes the
# non-updated Windows 2000/XP.
# Authenticates through a Pre-Shared Key. Supports clients that
# are not behind NAT. Does not support clients that are behind NAT.

conn L2TP-PSK
        # ----------------------------------------------------------
        # The VPN server.
        # Allow incoming connections on the external network interface.
        # If you want to use a different interface or if there is no
        # defaultroute, you can use:   left=your.ip.addr.ess
        # If you insist on supporting non-updated Windows clients,
        # you can use:    leftprotoport=17/%any
        # ----------------------------------------------------------
        # The remote user(s).
        # Allow incoming connections only from this IP address.
        # If you want to allow multiple connections from any IP address,
        # you can use:    right=%any
        # ----------------------------------------------------------
        # Change 'ignore' to 'add' to enable this configuration.

As you can see, the configuration is relatively simple. Note the absence of leftsubnet and rightsubnet, as explained above. A convention followed by many people is to use "left" for the server ("local") and "right" for Road Warriors. Recent versions of FreeS/WAN, Openswan and strongSwan support the use of "right=%any" for Preshared Keys as well as certificates.

Note: by default my example configuration files are not activated by default (for security reasons). You will have to change auto=ignore to auto=add if you want to enable them.

10.1.2 Specifying the Preshared Key

You enter the Preshared Key in /etc/ipsec.secrets following the FreeS/WAN guidelines. On Mandrake the path can be /etc/freeswan, /etc/ openswan or /etc/strongswan. Nothing much to it. For instance:

# Sample /etc/ipsec.secrets file
# The Openswan server has an IP address of
# Preshared Keys for two clients with fixed IP addresses: PSK "keyforoneclient" PSK "keyforanotherclient"

# Preshared Key for clients connecting from any IP address: %any: PSK "keysharedbyallclients"
# (Line above only works on recent versions of Openswan).

# There is a subtle difference with the following
# (see also 'man ipsec.secrets') which affects NATed
# clients that use a PSK:
# : PSK "keysharedbyallclients"

10.2 Perfect Forward Secrecy

Perfect Forward Secrecy (PFS) provides extra security. When you enable PFS, your adversaries (hackers, competitors, law enforcement, the mob etc.) cannot decipher packets previously sent through the IPsec connection, even if they eavesdropped on the encrypted connection and they have your secret key (through hacking, court order, escrow etc.). This property of PFS is also known as "escrow-foilage".

As you can see in the example above, there is a line:


This parameter is required because Apple's and Microsoft's L2TP/IPsec clients do not enable PFS. Openswan, on the other hand, enables PFS by default. (One could only speculate why PFS is not used by Apple and Microsoft as a default. Is it because of the <insert your favourite 3-letter government agency>?).

The easiest way to solve this interoperability problem is to disable it explicitly in Free/SWAN. But this is what the FreeS/WAN team says about it:

"The FreeS/WAN default is [pfs=yes]. We see no reason not to; this is more secure and costs little. The PFS settings on the two ends must match. You can turn PFS off in FreeS/WAN with the  pfs=no   setting in ipsec.conf(5), but if possible we suggest you enable PFS on the other end instead. That is more secure".
They may be right but in this case it is much more work. 'The other end' means the Microsoft client or Mac OS X v10.3+. Enabling PFS on the MSL2TP client is done by editing the Windows registry. This could be a bit dangerous and perhaps even too advanced if you let your users do it. So the alternatives are changing the PFS setting on each and every client, or disabling PFS by adding one line at the server side. Unfortunately, I don't know how to enable PFS for L2TP/IPsec on Windows 2000/XP/Vista or Mac OS X 10.3+ using their graphical interfaces.

Note that Openswan will use PFS when the client asks for it, even when pfs=no is specified in the Openswan configuration. Openswan will ignore the pfs=no for that particular client because PFS is more secure and if the client supports it, why not use it? So pfs=no allows you to connect with clients that are configured with or without PFS.

In short, if you see the following error in your Openswan logs:

     "we require PFS but Quick I1 SA specifies no GROUP_DESCRIPTION"

...the simplest solution would be to add the line   pfs=no  to your Openswan configuration.

Back to Contents

11. Configuring the L2TP/IPsec clients

I assume that you have configured Openswan at this stage. Now it is time to install and configure the L2TP/IPsec clients. This depends on the type(s) of clients you use (Windows 2000/XP/Vista, SoftRemote, Mac OS X 10.3+, Pocket PC etc.). See the list of clients at the top of this page.

After you have configured your client(s), you can use it to initiate the VPN connection. It will first try to set up an IPsec connection and then an L2TP connection.

Back to Contents

12. Initiating the IPsec connection

Initiate ('dial') the VPN connection. The procedure for this depends on the type of client (see the previous section). The client will report an error ("The computer you are dialling is not responding" or something similar). The error is correct: you do not have an L2TP server yet so just ignore the error for the moment.

The IPsec connection should come up successfully, though. You can check that in the Openswan logfile (normally /var/log/secure) which should look similar to this:

Nov  1 14:09:59 xxx Pluto[yyy]: "L2TP-PSK" #7: responding to Main Mode

Nov  1 14:09:59 xxx Pluto[yyy]: "L2TP-PSK" #7: Peer ID is ID_IPV4_ADDR: ''
Nov  1 14:09:59 xxx Pluto[yyy]: "L2TP-PSK" #7: STATE_MAIN_R3: sent MR3, ISAKMP SA established
Nov  1 14:09:59 xxx Pluto[yyy]: "L2TP-PSK" #8: responding to Quick Mode
Nov  1 14:10:00 xxx Pluto[yyy]: "L2TP-PSK" #8: STATE_QUICK_R2: IPsec SA established

If you see this, congratulations! You've got the IPsec part nailed. Continue with the L2TP part below. If not, check your configuration or go to 'Troubleshooting'.

Back to Contents

13. L2TP overview

The IPsec connection you just configured is to be used for tunnelling the L2TP protocol (L2TP over IPSEC is defined in RFC 3193). L2TP on its turn will tunnel PPP and PPP is to tunnel the actual payload. This means you will need an L2TP server on your Linux system. I am aware of the following open source implementations:

I have mainly used l2tpd and its fork xl2tpd. It was the first L2TP server available, it has been released under the GPL and it is probably the easiest L2TP server to use because of three reasons:
  1. L2tpd has a simple configuration file called l2tpd.conf which is reasonably intuitive to configure.
  2. L2tpd runs in user mode so there is no kernel recompilation needed. Recompiling the kernel is often a lot of trouble.
  3. L2tpd has built-in support for IP pools which means that l2tpd can dynamically assign internal IP addresses from a pool that l2tpd maintains. The other L2TP servers require installation of a RADIUS server to maintain an IP pool.
  4. L2tpd uses pppd. This is a well-known PPP implementation which is very complete and has several authentication options and plugins.
There are also some drawbacks:
  1. L2tpd runs in user mode and uses pppd, so it is slower than kernel mode L2TP servers.
  2. L2tpd is in maintenance mode. No new features are to be expected, but bugfixes and security vulnerabilities will be fixed if reported. There is not a very active user community (older archive).
  3. There are some concerns about the code quality and possibly security issues in l2tpd.
It seems to me that l2tpd is great to get started and easy to use for small setups. However, for a serious deployment with a considerable number of clients, you will probably want to use one of the other L2TP servers: rp-l2tp, l2tpns or possibly OpenL2TP. For practical use you will also need a RADIUS, LDAP or other authentication server. Debian, for instance, now prefers l2tpns over l2tpd.

The authors of l2tpd, rp-l2tp, l2tpns and OpenL2TP use their software in commercial settings. However, I don't know how robust these L2TP implementations are. I have confidence in Openswan, and pppd is used by lots of people all over the world, but Open Source L2TP implementations are fairly new on the scene. Emphasis seems to be on functionality. Security may not have been that much of a concern. This may not be too much of a problem, as long as the L2TP server is not exposed directly to the Internet (see these security considerations). That is because users can only access the L2TP daemon once they have been authenticated through Openswan, which was designed with security in mind.

Dossy Shiobara reports that rp-l2tp works for L2TP/IPsec, which I can confirm. One thing to note is that both l2tpd and rp-l2tp use the same location for their daemon: /usr/sbin/l2tpd. That is very unfortunate but in most cases you will only install one of these two daemons so it should not be that much of a problem. Rp-l2tp seems to have a better code base than l2tpd.

Most of the L2TP daemons (l2tpd is the exception) have the drawback that they cannot assign dynamic internal (virtual) IP addresses by themselves. This is not an issue if you want to assign fixed internal addresses to your users. But this is a problem if you want to assign dynamic IP addresses to users. Three solutions have been proposed: the L2TP servers could be extended so that they hand out IP addresses dynamically. This approach more or less violates the OSI networking layering model but this is how l2tpd does it. Nobody has implemented this solution for the other L2TP daemons (yet). The second solution is to let the PPP server obtain IP addresses from a DHCP server that you already might have on your network. For this to work you need a pppd plugin called ppp-dhcp (for more information read this thread, with additional configuration tips by Ben McKeegan in this thread). The third solution is to use pppd version 2.4.2 or later which supports RADIUS (alternatively, there is a plugin for pppd 2.4.1 by Anton Voronin, see also this post). The RADIUS solution is of course the most flexible but it requires a RADIUS server which adds to the complexity, especially if you have only a few users. Your RADIUS server must support a feature called "IP pools" if you want to use it with PPP (I believe FreeRADIUS does).

Except when you use l2tpns, you will also need a PPP server since L2TP is used to tunnel PPP. Most distributions ship with a PPP server (pppd). No sample PPP configuration files were included with l2tpd, so I made some myself (included with the l2tpd RPM mentioned above). Again, I don't claim that these are the best but they should get you started. Finetuning the l2tpd and pppd configuration and/or switching to a different L2TP implementation may be required for the best results.

According to Wolfgang Hennerbichler, Windows 2000/XP/Vista and Mac OS X also support DHCP in order to retrieve settings such as domain names, static routes etc. from the VPN server. You need a DHCP server that supports "DHCP Informational" messages, such as ISC DHCPD 3.x or higher. Only Mac OS X 10.5 ("Leopard") and Windows 2000/XP/Vista clients support these messages at this time.

Back to Contents

14. L2TP installation and configuration (Linux)

Probably the easiest way to install l2tpd is to get it readily packaged for your distribution.

l2tpd RPMs by me (based on the Mandrake Cooker RPM by Lenny Cartier and Per Øyvind Karlsen):

   Source RPM:          l2tpd-0.69-12jdl.src.rpm  (based on the l2tpd version of August 20, 2002)
   Source RPM:          l2tpd-0.69cvs20051030-1jdl.src.rpm  (based on the l2tpd CVS version of Oct 30, 2005)
   Mandrake 10.1 (x86): l2tpd-0.69-11jdl.i586.rpm
   Mandrake 10.0 (x86): l2tpd-0.69-10jdl.i586.rpm
   Mandrake 9.2  (x86): l2tpd-0.69-10jdl.i586.rpm
   Mandrake 9.1  (x86): l2tpd-0.69-10jdl.i586.rpm (no longer updated by Mandrake)
   Mandrake 9.1  (PPC): l2tpd-0.69-12jdl.ppc.rpm  (no longer updated by Mandrake)
   Fedora Core 1 (x86): l2tpd-0.69-10jdl.i386.rpm
   Fedora Core 2 (x86): l2tpd-0.69-11jdl.i386.rpm
   Fedora Core 3 (x86): l2tpd-0.69-11jdl.i386.rpm
   Fedora Core 4 (x86): l2tpd-0.69-12jdl.i386.rpm
   Red Hat 9.0   (x86): l2tpd-0.69-10jdl.i386.rpm
   Red Hat 8.0   (x86): l2tpd-0.69-10jdl.i386.rpm
   Red Hat 7.3   (x86):
   SuSE 9.1:     (x86)  l2tpd-0.69-10jdl.i586.rpm (thanks to Bernhard Thoni)
   Tar ball (v.10jdl) with software, patches, config files (sig)
   Versions no longer supported: SuSE 8.0,
Mandrake 9.0, Mandrake 8.1

   You may need to hold the 'Shift' key while clicking these links!

rp-l2tp RPMs by me  (based on the ASP Linux RPM by Alexandr D. Kanevskiy):

   Source RPM:          rp-l2tp-0.4-2jdl.src.rpm
Mandrake 10.1 (x86): rp-l2tp-0.4-1jdl.i586.rpm
   Fedora Core 2 (x86): rp-l2tp-0.4-1jdl.i386.rpm
   Fedora Core 4 (ppc): rp-l2tp-0.4-2jdl.i386.rpm (l2tpd does not seem to work on FC4-PPC)

The RPMs have been signed with my PGP key. I have built several binary RPMs for your convenience but I recommend you get the source RPM, inspect the SPEC file, download the original tarball and the patches from their original locations, review all source files for backdoors and other security risks and then build the RPM yourself. It's a lot more of a hassle, but security has its price... :-) (At least with open source software you get the chance to inspect the source!). Note that the RPM spec file defaults to Red Hat/Mandrake/Fedora. You will need to set a boolean variable for SuSE.

I include a minimal l2tpd configuration file with the RPMs as an example (/etc/l2tpd/l2tpd.conf). I will discuss it here:

; Sample l2tpd.conf

; listen-addr =

[lns default]
ip range =
local ip =
require chap = yes
refuse pap = yes
require authentication = yes
name = LinuxVPNserver
ppp debug = yes
pppoptfile = /etc/ppp/options.l2tpd
length bit = yes

The 'listen-addr' parameter has already been discussed above. By default, l2tpd will listen on all interfaces. The parameter 'ip range' specifies a range of IP addresses on the internal LAN that will be allocated to remote users. With 'local ip' you specify what IP address will be used by the Linux server on pppX interfaces created by l2tpd. Please note: 'local ip' must be an IP address on the internal network (i.e. your intranet). In the example above 'local ip' is You cannot use the IP address of the internal interface for 'local ip', nor can you use the IP address of 'listen-addr' (both are in the example above). 'Local ip' should be a free IP address within the same subnet of the internal interface.

CHAP is enabled and PAP is disabled because otherwise the Microsoft clients will complain that the password is not encrypted (which is of course nonsense because the connection is already encrypted by IPsec). I also had to set 'length bit' to yes, because the connection was unstable without this parameter.

14.1 L2TP authentication and client IP restrictions (not essential)

IPsec supports authentication through Preshared Keys and certificates. PPP also supports authentication e.g. through passwords; see below). It turns out that L2TP also supports authentication. The problem is that you cannot specify those passwords anywhere in the Windows/Mac L2TP clients. I guess none of the vendors thought that L2TP authentication was important. And rightly so, because it does not seem very useful anyway. IPsec and PPP authentication should be enough for anyone.

The confusion comes from the 'require authentication' parameter in l2tpd.conf. This parameter has nothing to do with enabling L2TP authentication. It is actually for PPP authentication (i.e. PAP/CHAP). The Windows clients use this by default, so you should enable PPP authentication by including 'require authentication' in your l2tpd configuration file.

L2TP authentication, on the other hand, can be enabled by specifying the parameters 'auth file' and 'challenge'. But as explained above, normally you do not need L2TP authentication.

l2tpd can also do access control based on IP addresses. This is slightly more interesting than L2TP authentication. However, Openswan already does access control on IP addresses. You could use l2tpd's access control as an extra security measure ('belt and suspenders' approach). There's nothing wrong with that but it only works if you know all the IP addresses of the clients in advance (e.g. they all have fixed IP addresses). This rules out "Roadwarriors" with dynamic addresses. Let's say that you want to restrict l2tpd access to a client with the fixed IP address Then your l2tpd.conf should be extended as follows:


access control = yes

[lns default]
lac =

Back to Contents

15. PPP authentication, compression and encryption

15.1 PPP authentication: overview

IPsec is used to tunnel L2TP, which on its turn tunnels PPP. Several authentication methods exist for PPP. The most popular ones are PAP (unencrypted passwords) and CHAP (challenge/response based authentication). You are free to choose either one of these but I recommend CHAP. If you use PAP, the Microsoft clients will complain that the password is not encrypted. This is besides the point because IPsec already does encryption. Both PAP and CHAP are IETF standards but Microsoft has 'embraced and extended' CHAP into a new dialect: MS-CHAP (later fixed as MS-CHAPv2).

Windows 2000/XP/Vista clients also support EAP for PPP authentication. This might require a pppd plugin by Michael Heiart which allows you to use a RADIUS server for EAP.

The easiest way to use PAP or (MS-)CHAP is through a passwords ("secrets") file, see below. These passwords are specified in the file/etc/ppp/chap-secrets or /etc/ppp/pap-secrets which are for PAP and (MS-)CHAP, respectively. In more complex situations with a larger number of users, you may be looking for something more flexible. This is a bit out of the scope of this document (it is a PPP issue after all) but there are several possible solutions: For testing purposes, you could skip PPP authentication altogether by specifying "noauth" in /etc/ppp/options.l2tpd. Then it will work with any version of pppd because it won't ask the client to do PPP authentication (of course the client still has to do IPsec authentication). But this is not recommended for real-world usage. Clients will be able to fake internal IP addresses if PPP authentication is disabled.

15.2 MS-CHAP not always supported on Linux

By default, Microsoft clients will want to use MS-CHAP for PPP authentication (either MS-CHAPv1 or MS-CHAPv2, with the exception of Windows Vista which only supports MS-CHAPv2). Obviously, this means that MS-CHAP support at the Linux server side is highly preferable. There are however two things to take into account with MS-CHAP. First, if you decided to use l2tpns as your L2TP and PPP daemon, then you cannot use MS-CHAP. It is currently not supported by l2tpns. A problem with older Linux distributions is that their PPP daemons do not support MS-CHAP. Examples of such distributions are Red Hat Linux 9 and earlier. A solution is to configure every Windows client to use CHAP. Another solution is to upgrade to pppd version 2.4.2 or later which contains MS-CHAP(v2) support. If you connect to a server without MS-CHAP support and the client is configured to use MS-CHAP, pppd will complain about "auth chap 81" and "peer will not authorize". Once you have installed a pppd with MS-CHAP support, you enable it by adding an extra parameter to /etc/ppp/options.l2tpd. Depending on your pppd version (check man pppd), you have to use:


Less common are:




15.3 MPPE encryption

Microsoft has also made MPPE, an encryption protocol for PPP. It is based on RSA Security's RC4 encryption algorithm and used in PPTP. Normally you do not want to use MPPE in combination with L2TP/IPsec because it means that you will have double encryption: both IPsec's and MPPE's. In some cases the Windows clients want to force MPPE. In that case, 'disable encryption' on those Windows clients. Note that this may confuse the user. He has to disable MPPE encryption, thinking there is no encryption at all. In reality, IPsec already provides the encryption but the user may not be aware of this.
There is a situation where you do want to use double encryption: when the Preshared Key is public knowledge and shared by multiple users, posted on a website somewhere. Some organisation want to avoid using certificates by doing this but it will compromise the IPsec encryption. Instead they rely on the MPPE encryption in the PPP stage.

If the client is configured to require MPPE but the server (pppd) is not, you would see it in the pppd logs as something like:  sent [CCP ConfRej id=0x1 <mppe +H -M +S -L -D -C>] See this explanation of the MPPE bits.

15.4 MPPC compression

Some Windows clients also support compression. Mind you, this is PPP compression. The Windows IPsec implementation does not support IPsec compression (IPCOMP). The compression protocol used by Microsoft (MPPC) is patent encumbered (in those countries which know software patents, mainly the US and Japan). If you want to use MPPC nevertheless, check the documentation for setting up a Linux PPTP server.

Back to Contents

16.1 PPP installation and configuration (Linux)

Once the L2TP connection is up, it hands over control to the PPP daemon. Obviously you will need a PPP server. Almost every distribution has one: pppd. Install a recent version, i.e. ppp 2.4.1 or higher. The same PPP software can be used for something else as well (e.g. an analogue modem for dial-up purposes). Fortunately, with the parameter 'pppoptfile' in l2tpd.conf you can specify a separate PPP options file for the L2TP daemon.

Note: The PPP daemon's passwords are shared by all PPP processes, not just the ones started by l2tpd. However, you can restrict the validity of usernames/passwords to certain IP addresses, such as in this example chap-secrets file:

# Secrets for authentication using CHAP
# client        server  secret                  IP addresses
jacco           *       "mysecret"    
*               jacco   "mysecret"    
sam             *       "rumpelstiltskin"
*               sam     "rumpelstiltskin"

There are two entries for every user since both sides authenticate. In this case, user "jacco" will get an address from l2tpd's IP address pool as specified in /etc/l2tpd/l2tpd.conf (in my example configuration this is which agrees with User "sam" however always gets from the PPP server. This way you can give users fixed virtual IP addresses on your internal LAN (it seems to work, although I am not 100% sure it works in all cases). If you don't want any restrictions on the IP addresses, you can use the wildcard "*" as an IP address instead (I don't recommend this since this will allow the client to determine its own address, which may be a risk).

The Microsoft clients have an option "Use Windows logon". If you enable this, the client will want to authenticate with the username "\\DOMAINNAME\username". You will have to specify this username format in chap-secrets/pap-secrets as well.

Configuring the PPP server is not explained here. There is lots of documentation on this subject, for instance the PPP Howto. Check out my sample configuration file /etc/ppp/options.l2tpd included with the RPMs mentioned above.

For simplicity, let's assume that the external interface of your Linux server is eth0 and the internal interface is eth1. The PPP server is to provide the remote user with an IP address from the internal (protected) network. Once the user is connected interface ppp0 will be up. My sample L2TP configuration file l2tpd.conf has a line 'local ip'. With this parameter you specify a dedicated IP address for the L2TP daemon on the internal subnet.

Note that options.l2tpd contains the parameter proxyarp. This parameter will set a proxy arp entry on the internal interface (eth1 in the example above) for the remote user. If this keyword is not specified, packets from the L2TP/IPsec clients will arrive at internal servers but those servers will not know where to send responses to because nobody replies to its ARP requests. With the proxyarp parameter, the machines on the internal network are fooled into sending packets for the remote Windows clients to the gateway. The gateway has IP forwarding on, so it knows how to send the packets through to the Windows clients.

16.2 Issue: specified DNS is ignored

As you can see from the sample PPP options file (/etc/ppp/options.l2tpd), you can specify DNS and WINS servers. Remote clients will pick these up automatically once the connection has been established. Normally you will want remote clients to use the same DNS/WINS servers that clients on the internal network use. However, it seems that a native Windows 2000/XP/Vista client will only pick up these DNS/WINS servers when its Internet connection is configured with a dynamic IP address (e.g. through DHCP or PPP's IPCP). This puzzles me and I don't have an explanation or a workaround for this. It could be that the DNS entry is not updated (check with IPCONFIG /ALL on the Windows client) but that nevertheless the correct DNS server (as specified through the ms-dns parameter) is used.

16.3 MTU problems

If you see PAYLOAD_MALFORMED error or if you have intermittent problems with your VPN connection, it could be an MTU problem. What you might see is that you can ping machines on the internal network and you can surf or transfer very small files, but you can't copy large files because the connection stalls. Then try decreasing the MTU to 1410 (or perhaps even lower?) by adding the following parameter to /etc/ppp/options.l2tp:

mtu 1400
mru 1400
This problem may be especially apparent if you connect through the VPN to sites with broken 'Path MTU (PMTU) Discovery' (a letter with an executive summary describing the situation can be found here). Perhaps people with ADSL connections that use PPTP or PPPoE have this problem as well. I myself had no problem with a PPTP ADSL connection ('KPN Mxstream') which uses Alcatel equipment. Reducing the PPP packet size does not help if the connection fails before the authentication has succeeded. Try again with a smaller certificate and without a NATed connection, if you can. If you use NETKEY on kernel 2.6, then kernel 2.6.12+ is recommended if you experience PMTU problems. On a 2.6.12+ kernel you could do: echo "0" > /proc/sys/net/ipv4/ip_no_pmtu_disc

There has been some discussion on the FreeS/WAN mailinglist (for instance, by Sam Sgro here and here). The FreeS/WAN team writes: "The MTU can be changed with an overridemtu= parameter in the config setup section of ipsec.conf". Currently this parameter is only supported on KLIPS, not NETKEY. Other suggestions might be reducing the size of your certificates (shorter names, fewer X.509v3 options etc.) or forcing your Ethernet interfaces to use a lower MTU: ifconfig ethX mtu 1400

Another option that you might want to try on Openswan versions before 2.4.5 is: fragicmp=no

Microsoft has a Knowledge Base article on how to change the MTU for VPN connections. A ready to use batch file that changes the MTU parameter can be downloaded from JSI FAQ. Alternatively, there is a registry patch file that can be downloaded from (sets an MTU size of 1200). There is a similar KB article here which explains how to set the MTU for a specific network adapter, among other things. Microsoft also has a troubleshooting guide for "black hole" routing problems (Q314825). Cisco also has some documentation on this problem (here, here and here).

In some cases (especially when certificates are used) fragmentation may lead to a problem with IKE. An extension to IKE has been proposed by Cisco. This extension allows fragmented IKE packets pass through (broken) routers which cause the fragmentation. Openswan does not support this extension but ipsec-tools (racoon) does. Fragmentation is less likely with PSKs. Other options might be to reduce the size of certificates (e.g. by reducing the size of the RSA key, shortening the Distinguished Name fields or removing extended attributes) and eliminating (NAT) devices that don't support fragments.

Another option is to "clamp" the Maximum Segment Size (MSS) of the application protocols that are sent through the VPN. This only works with TCP protocols, not with UDP based protocols.

16.4 Tip for multi-homed DSL users

Most DSL users have a static IP address or a dynamic IP address assigned through DHCP. However, some DSL services use PPP for the Internet connection (e.g. through PPTP or PPPoE). If you have such a DSL service, you often use interface ppp0 for your Internet connection. And when you decide to use an L2TP/IPsec VPN over that Internet connection, the PPP daemon spawned by the L2TP daemon will automatically pick the next free interface name, ppp1.

However, a problem might arise when you have multiple DSL links and one of them goes down for some reason and then comes up again. In the mean time, the PPP interface might have been picked up by an L2TP/IPsec client because pppd detected that it was free. This might screw up several of your settings, for instance, your firewall rules might expect the DSL links on fixed PPP interfaces.

A workaround for this problem is adding the parameter 'unit xxxx' to the PPP options file for your DSL links, where xxxx is a high number, say 1000. Each DSL link will have to have a unique PPP interface (ppp1000, ppp1001 resp.) that does not conflict with the lower PPP unit numbers which are reserved for L2TP/IPsec client (ppp0, ppp1, etc.).

You do need ppp-2.4.2 or higher for this particular 'unit' parameter. The pppd man page says that the parameter only works for outbound connections but it works for incoming connections too (looks like an undocumented feature in pppd?). The 'unit' parameter does not have to be used in /etc/ppp/options.l2tpd, only in the PPP options files of your DSL links. This workaround should work as long as the highest PPP unit number assigned to the DSL links is higher than the total number of clients.

Back to Contents

17. Initiating the L2TP/IPsec connection once again

At this point you should have IPsec, L2TP and PPP configured. Try initiating the VPN connection on the Windows or Macintosh client again. The procedure is the same as described above. But this time the L2TP connection should come up as well. If it does, congratulations, you've got a working L2TP/IPsec setup! If it does not work, check and double check your settings. See also 'Troubleshooting'.

There are two final checks that I highly recommend before you consider rolling out to a live network:

  1. Put a third machine between the client and the server. You should be then use that machine to sniff the communication between the client and the server with a network monitoring program such as tcpdump or Ethereal. The packets should be encrypted. If you see unencrypted packets (e.g. plain text L2TP), there is something wrong in your setup.
  2. Use nmap (or any other good portscanner) on the client and scan for open UDP ports on the server: nmap -sU You should not see the L2TP daemon (UDP port 1701). The only open ports should be UDP 500 (IKE) and optionally UDP 4500 (NAT-T).

Back to Contents

18. Some remarks on L2TP/IPsec

At one time during testing, I forgot to enter a default gateway and a hostname in the TCP/IP settings of a Windows workstation. This was not really a problem since the Linux server was on the local network. However, when I set up a connection from the Windows client to the Linux server, l2tpd complained with a rather cryptic error message in /var/log/messages: "Specify your hostname". At first I thought that perhaps l2tpd could not determine the hostname of the Linux server it was running on. I solved the problem by entering a hostname on the Windows workstation. I simply did not expect that l2tpd relied on it.

Back to Contents

19. NAT-Traversal

19.1 Introduction

If you don't need NAT-Traversal (or if you don't want to deal with the complications at this moment) you can skip this part.

Sometimes Network Address Translation (NAT) is used between the client and the IPsec server, e.g. when the user has networked his home and all his computers are sharing the same Internet connection through a NAT router. Another case where NAT is used is when the VPN server itself is behind a NAT device (UDP ports 500 and 4500 are forwarded to the VPN server). In some situations even both the client and the server may be behind NAT. Some ISPs also use NAT on the route. I have also read reports that many GPRS providers employ NAT, i.e. they assign 10.x.x.x or 192.168.x.x addresses to GPRS phones. 

NAT changes packets on the fly. The problem is that this kind of 'tampering' is exactly what a VPN tries to prevent! Fortunately, some NAT devices support 'IPsec passthrough' which allows IPsec to be used through the device. However, it seems that IPsec passthrough only works for one user at a time, not multiple concurrent users behind the same NAT device. I have not tested L2TP/IPsec through a NAT device with IPsec passthrough myself, but I received a report by Tim Behrsin about 'UDP: bad checksum' problems. Some NAT devices are broken in the sense that you cannot disable IPsec passthrough (the SMC 2804 comes to mind?). Pure IPsec on the other hand has been reported to work with these devices (i.e. Tunnel Mode, not L2TP/IPsec in Transport Mode).

Instead of modifying the NAT device, you could modify the IPsec standard itself. And this is exactly what has happened. Extensions to IPsec are in development so that you can set up IPsec connections through a NAT device without IPsec passthrough support. This is called "NAT-Traversal" (NAT-T) which was published by the IETF as a Proposed Standard, RFC 3947. Openswan versions 1.0.9+ and 2.3.1+ support this RFC. The NAT-T RFC is based on the NAT-T requirements stipulated in RFC 3715. NAT-Traversal works around the NAT problem by encapsulating IPsec ESP packets within UDP (RFC 3948).

This article by Debra Schinder provides some more background information on NAT-Traversal. And these articles by Stefaan Pouseele provide detailed info on the packet structures and mechanisms behind NAT-T.

NAT-Traversal is considered by Mathieu Lafon to be experimental and unsafe if used with L2TP/IPsec (see section 5 "Security Considerations" of the "IPsec within UDP" IETF document mentioned above). Mathieu is one of the people involved with the implementation of NAT-T on Linux. Linux kernel 2.6 contains a different IPsec implementation (NETKEY) which includes also supports NAT-T. Apparently it has the same security issue as the NAT-T patch for Openswan et al. The NETKEY implementation supports racoon (from ipsec-tools) as well as Openswan/ strongSwan/ FreeS/WAN 2.x. Switching to ipsec-tools won't fix the security issue either.

19.2 Limitations and requirements

NAT-T is supported natively by NETKEY (kernel 2.6.6 or higher required). Recent distributions such as Fedora Core 2+, Debian "(un)stable", Mandrake 10.0+ and SuSE 9.1+ ship with a NETKEY kernel that supports NAT-T. If you use a kernel with KLIPS instead of NETKEY, you will need Mathieu Lafon's NAT-Traversal patch. Note: according to Mathieu Lafon and Openswan team member Ken Bantoft, NAT-T in Transport Mode is (currently) unsafe. Microsoft, SSH and SafeNet do support NAT-T but they have come up with a workaround for the unsafe part.

The NAT-T patch for KLIPS is not included with older Red Hat, Mandrake and SuSE kernels (except Mandrake 9.2). For these versions you will have to recompile the kernel, because the NAT-T patch touches the TCP/IP part of the kernel. If you use FreeS/WAN, this patch will also have to be applied to the FreeS/WAN userland utilities. The userland utilities included with Openswan and strongSwan already contain the patch. Transport Mode NAT-T has to be explicitly enabled on FreeS/WAN by patching the Makefile. Openswan and strongSwan already support Transport Mode NAT-T.

Openswan's NAT-T support is activated on both NETKEY and KLIPS by adding the parameter "nat_traversal=yes" to the configuration file ipsec.conf. Clients not behind NAT are not affected by this line. Everything will still work for these clients. For testing purposes you might be interested to know that recent versions of Openswan can be forced to use NAT-T even if the client and server are not behind NAT. You can do this by adding the line "forceencaps=yes" to the connection section in ipsec.conf.

The NAT-T patch for KLIPS currently does not support connections that authenticate through Preshared Keys (PSKs). If you use a Preshared Key with KLIPS and NAT-T, Openswan will not authorise the connection because the sender's port is not 500 and the NAT-T patch fails to support this. This should be looked into by a programmer. PSKs have some drawbacks so generally you want to use certificates anyway. The NETKEY implementation does support PSKs but you will have to use right=%any and use leftid=someservername and rightid=someclientname. If you do specify an IP address with right=a.b.c.d without leftid=someservername and rightid=someclientname, Openswan will complain about "no connection authorized". If you use Windows Vista with a PSK, and NAT is involved, you are required to use rightsubnet=vhost:%no,%priv. This is probably true for other RFC 3947 compliant clients (such as Mac OS X) as well.

If you have multiple clients behind the same NAT device, only the first client will be able to connect. Openswan logs an error similar to this: Virtual IP is already used by 'C=NL, ST=ST, L=L, O=TESTORG, CN=TESTUSER' (or ... used by '@Pocket_PC' in the case of a PSK). Another limitation of Openswan is that clients cannot share the same NAT-ed (internal) address. This is of course difficult to avoid completely, especially when there is little coordination between remote clients. Many users will be using the same address if they are using a Linksys router, for instance. These are limitations of Openswan. Patrick Naubert, the CEO of Xelerance, has released this statement on the subject. He writes that they have implemented a solution for KLIPS but they need $55,000 to recoup their costs before they can release the source code. Paul Wouters of Xelerance has said that perhaps a solution is in the works. A number of alternative solutions exist. For example, OpenL2TP with ipsec-tools (racoon) does support multiple clients behind the same NAT device. Another alternative is the workaround by the Finnish company Stinghorn which is based on modifications to the Linux kernel and ipsec-tools. Some more details can be found here.

If you have multiple users behind the same NAT device and one client connects to the Openswan server, the other users behind that NAT device will not be able to set up non-IPsec connections (WWW, SSH etc.) to that server. You can solve this by adding an extra 'passthrough' section to ipsec.conf as described by Paul Wouters of the Openswan team.

NAT-T is supported by almost all of the L2TP/IPsec clients. Windows Server 2003 also supports NAT-T (a test report can be found here). Microsoft has released IPsec updates (MS KB Q818043) for Windows XP and Windows 2000 Professional, as promised. This update is also included with XP ServicePack 2. However it turned out that the NAT-T update had issues with some third-party applications which caught a lot of bad press. So Microsoft had to recall the NAT-T update. In August 2003 they re-released it. The patch can be installed through WindowsUpdate. If you use Windows 2000 Professional, you will need to install ServicePack 3 or higher first, otherwise the IPsec update will not show up in WindowsUpdate. For XP you will need ServicePack1 or higher. If the update still does not show up in WindowsUpdate, go to the Windows Update Catalog and search for "818043" using the Advanced Search Options feature. If you prefer to download the NAT-T update to disk and then distribute it to multiple clients, you can use Windows Update's Download Basket or Software Update Services.

According to posts by Microsoft employees, updates for Windows 2000 Server and ISA Server would become available as well. I had some doubts when I read this because Microsoft would rather see you upgrade to Windows Server 2003 instead. It turns out I was right. Quoting Microsoft (KB Q818043): "NAT-T server-side support will not be added to Windows 2000 Routing and Remote Access". Fortunately this decision by Microsoft will not affect you if you use Linux on the server...

All clients except SSH Sentinel (and possibly Forticlient) automatically detect if NAT-T has to be enabled. SSH Sentinel has to be manually configured to use NAT-T: in the "Properties" window of a VPN connection, select the "Advanced" tab and enable "NAT-Traversal" (Sentinel 1.4 screenshot). Sentinel 1.4 also has a feature called "UDP encapsulation to port xxxx" but that variant is not supported by Linux. You will have to use the other Sentinel option called "NAT-T". Sentinel 1.3 has only one NAT-T variant but that is the non-supported type so you can't use it. Next, click "OK". Don't enable this option in Sentinel if the NAT device itself already supports "IPsec-passthrough". The documentation of SSH Sentinel states that NAT-T cannot be used in combination with IPCOMP, the IPsec compression protocol. This is a limitation of Sentinel itself, not a limitation of NAT-T or IPCOMP.

19.3 Status of NAT-T support in L2TP/IPsec clients

Successfully tested:
Probably working:
Not working:
19.4 Preparing for NAT-T

The procedure for enabling NAT-T starts with determining whether the kernel included with your distribution supports either NETKEY or KLIPS. If your kernel support NETKEY, then it already has kernel support for NAT-T. If your distribution's kernel does not contain either NETKEY or KLIPS, then it obviously lacks NAT-T support for IPsec. Even if your kernel supports KLIPS, it probably does not support NAT-T yet (Mandrake 9.2 is an exception).  Distributions that do not ship with any IPsec support at all or with a KLIPS-based kernel without NAT-T patch will probably require installation of a new kernel. If you do need to install and reboot to a new kernel RPM, be sure to take the usual precautions with kernel RPMs, i.e. do not upgrade but install the new kernel in addition to the current kernel and add an entry for the new kernel to lilo.conf. Then rerun lilo (or use GRUB).

19.4.1 Using NAT-T with NETKEY based kernels

The NETKEY implementation supports NAT-T with authentication through certificates but also through PSKs, unlike the NAT-T patch for KLIPS. However, you cannot specify a fixed IP address with the right=a.b.c.d parameter. You will have to specify right=%any and use leftid= / rightid=, which means that the PSK is shared by all Road Warriors. You will also have to use %any in ipsec.secrets: %any: PSK "thisismytopsecretkey"

I could not use the l2tpd parameter listen-addr because NETKEY does not use ipsec0 style interfaces.

19.4.2 Using NAT-T with KLIPS based kernels

Preshared Keys are not supported by the NAT-T patch for KLIPS. Using NAT-T with KLIPS on kernel 2.6 is currently not supported. The Openswan team is working on this. NAT-T with KLIPS on kernel 2.4 does work.

19.5 Distribution-specific information regarding NAT-T

The sections below contain information on how to obtain support for NAT-T on several Linux distributions.

19.5.1 Red Hat Linux, Fedora Core 1

I recommend the RPMs made by Axel Thimm for Red Hat Linux 7-9 and Fedora Core 1. You will need Axel's Openswan userland utilities RPMs and kernel with built-in Openswan support. Do not install Axel's Openswan kernel-module RPM (kernel-module-openswan-*.rpm) because it does not add support for NAT-T. Axel's kernel RPM provides the NAT-T support.

19.5.2 Fedora Core 2+, RHEL

Fedora Core 2+ and Red Hat Enterprise Linux ship with a NETKEY enabled kernel so they support NAT-T out of the box. Reminder: Fedora Core 2+ does not support Legacy PTYs so you need a recent l2tpd RPM version.

19.5.3 Mandrake 10.x, Mandriva 2005+

Mandrake 10.x and Mandriva 2005+ ship with a 2.6 kernel so it supports NAT-T out of the box. Mandrake 10.0 ships with a FreeS/WAN 2.04 userland utilities RPM which does not support NAT-T. The FreeS/WAN 2.06 included with Mandriva 2005+ and Mandrake 10.1 does not support Transport Mode so L2TP/IPsec with or without NAT-T will not work. Mandrake 10.0+ and Mandriva 2005-2006 ship with a SuperFreeS/WAN RPM but it does not support NETKEY so it will not work either. One solution is to use the Openswan or strongSwan userland RPM from the contrib directory or Mandriva Cooker. Reminder: Mandrake 10.1 and Mandriva do not support Legacy PTYs so you need a recent l2tpd RPM version.

19.5.4 Mandrake 9.2

Mandrake 9.2 supports SuperFreeS/WAN out of the box which contains the NAT-T patch for KLIPS. I suggest you do not use any older Mandrake version than 9.2. Unfortunately the FreeS/WAN userland RPM included with Mandrake 9.2 does not work. You will need the updated SuperFreeS/WAN 1.99.8 userland RPMs made available by Mandrake employee Florin Grad. Florin's RPMs do not contain two patches that are required for NAT-T support in Transport Mode. Those two patches need to be copied over to your RPM SOURCES directory. The first patch enables NAT-T support in Transport Mode (this is considered unsafe by Mathieu Lafon). The second patch (version 3) by Mikael Lönnroth works around a problem with the NAT-T implementation in Windows 2000/XP/Vista. The patch doesn't actually add support for this implementation, it is only a hack and probably a security risk. Then you apply this diff against Florin's original SPEC file which ties the two patches into the original SRPM. You can get the whole lot from me with this modified SRPM (super-freeswan-1.99.8-7jdl.src.rpm). Suggested build command: rpmbuild -ba super-freeswan.mdk.spec. Or, if you really want, you could download the binary RPM (super-freeswan-1.99.8-7jdl.i586.rpm) and the documentation RPM (super-freeswan-doc-1.99.8-7jdl.i586.rpm) from my webpage. You may need to hold the 'Shift' key while clicking these links!

19.5.5 SuSE

SuSE 9.x uses kernel 2.6 so it supports NAT-T out of the box. SuSE 9.2 ships with an Openswan 2.2.0 userland utilities RPM. Alternatively, you can download Openswan RPMs for SuSE 9.x from the Openswan website. Alternatively, SuSE employee Kurt Garloff has made Openswan packages and KLIPS enabled kernels for older SuSE versions.

19.5.6 Debian

Debian uses kernel 2.4 with the backported NETKEY implementation. It should work. Openswan packages are available.

19.5.7 Other distributions

See if there are any Openswan packages (kernel and userland RPMs, deb, whatever) available. If none are available, you might have to create them yourself but that may be a lot of work.

19.6 Procedure for enabling NAT-T (server not NATed)

Once you have a kernel that supports NAT-T, you can follow the procedure below to enable NAT-T. If your Openswan server is behind NAT, see the next section. The procedure assumes that your setup looks like this:

                   NAT-        Internet      Openswan
Client  --------- device  =================== Server ...      /     \                      /     \
                /       \                    /                               

The internal subnet behind your Openswan server is Note that l2tpd.conf's local ip (and ip range) must be an IP address (and subrange) within this internal network (in my example above they are and, respectively). The IP address of the client is Often the NAT device assigns this IP address to the client through DHCP. The client is on a subnet behind the NAT device, in this example Openswan needs to know what remote subnets the clients use. You specify these subnet(s) with the virtual_private= parameter in ipsec.conf. Most people prefer to enumerate all network subnets that are defined in RFC 1918 because there can be many clients with many different subnets. But of course you can also specify only the network if there is only one client. You should however always exclude the subnet(s) that are behind the Openswan server. An exclamation mark is used for this purpose (see the example below). In this example we assume that the internal IP address of the NAT device is The external (public) IP address of the NAT device (in this case is specified with the right= parameter. For Road Warriors you would probably want to use right=%any instead of right= The %no option of the rightsubnet= parameter allows the same configuration to be used for other clients that are not behind NAT.

The procedure is as follows:

Note that, the internal network of the NAT device, is covered by the of the virtual_private= parameter. The nat_traversal, virtual_private and vhost parameters listed above are implemented by the NAT-T patch (see this README for a description of the parameters). That means they are supported by all versions of Openswan and strongSwan on both KLIPS and NETKEY. They are also supported on FreeS/WAN but only if you apply the NAT-T patch. I am not sure why the rightsubnet= parameter is required for NAT-T. Transport Mode uses host-to-host connections so in reality there are no subnets(!). Sam Sgro of the FreeS/WAN team discussed this on the mailinglist and provided an explanation.

19.7 Procedure for enabling NAT-T (server NATed)

If your Openswan server itself is behind NAT, the procedure is slightly different. You can use almost the same setup as described in the previous section but it now looks like this:

                   NAT-        Internet        NAT-
Client  --------- device  =================== device -------------+-------- ...      /     \                      /     \             |
                /       \                    /   Openswan         Server

The Openswan server is now only connected to the internal network behind the NAT device ( The diagram shows a NATed client but the procedure is also to be used for non-NATed clients. L2tpd.conf's local ip (and ip range) are still required to be an IP address (and subrange) within this internal network (in my example above they are and, respectively). The main difference with the previous setup is that you need to add a leftnexthop= parameter and that the left= parameter is the IP address of the Openswan server's network interface, not the public IP address of the NAT device.

The procedure is now as follows:

With older Openswan versions there is a problem when the server itself is behind NAT (or when both the client and the server are behind NAT). The Openswan team was notified of this problem by Bernd Galonska. It has been fixed in Openswan 2.4.5 and higher. For older versions you can download Bernd's experimental patch from this webpage which I have modified slightly so that it applies cleanly to Openswan 2.3.0 and 2.3.1--2.4.4 respectively.

Back to Contents

20. Windows Networking (WINS etc.)

In many cases VPNs are used to tunnel NetBIOS/SMB/CIFS network traffic ("Windows Networking"). Unfortunately, Windows Networking is a terrible protocol (ask the Samba guys) and it often results in all kinds of problems. This is not a tutorial on Windows Networking so I can't help you with that. But I can provide some hints.

If you want to browse the Network Neighbourhood across subnets (this includes WANs and VPNs), a WINS server is highly recommended. If you don't have a Windows NT/2000/2003 server available or if you don't want to buy one, you can download Samba and configure it as a WINS server. The important part to remember is that all computers should be configured to use this WINS server. Otherwise some computers might have trouble seeing other computers.

I also noticed that for the best results all clients should be configured to use the same workgroup/domain name, namely that of the office network. Out of the box, many clients have a default workgroup/domain name such as WORKGROUP or MSHOME which is different from the office network's workgroup/domain name. It is better to change these to a common name. Similar tips can be found on the ISA website.

The Microsoft clients have a stetting "Log on to network". Enable this if you want to log on to a domain server.

Note that Windows XP Home cannot join a domain. This is also the case for some versions of Windows Vista. You should be able to access resources in the domain, however.

I have not really tried to get logon scripts (batch files) to work. In theory it should work, since L2TP/IPsec is a Microsoft "approved" protocol. Perhaps the following procedure works. At the end of the New Connection Wizard, you are asked if this connection is for "Me only" or for "Anyone who uses this computer". Select anyone. You might prefer to not enter the password. The next time you logon, click the "Options >>" button. You will notice that a checkbox "Logon through Dial-up Networking" appears. Tick off this checkbox. Logon with your Windows Networking username and password. You will be presented with a window from which you can select the VPN connection that you just created. Select it and (hopefully) you will logon over the VPN connection, and the logon script will start.

Back to Contents

21. Split tunnelling

If you have a VPN connection to the office, normally all packets will be sent through the VPN tunnel. This includes not only traffic to internal servers but also Internet traffic, for instance when you surf to an Internet website. This has the disadvantage that the Internet traffic goes through the office's Internet link twice: once from your location to the office site and a second time from the office site to the Internet.

If you use 'split tunnelling' on the other hand, all internal traffic will be tunnelled through the VPN but all Internet traffic will be sent to the Internet directly, i.e. not through the office Internet connection. The problem is that this is less secure. Client users will then have two connections at the same time: one to the office and one to the Internet. Hackers could in theory break into the user's home machine and access the office from there. With split tunnelling disabled, this is more difficult to do. See also this Cable Guy article on the Microsoft website for more information about split tunnelling.

Here is how to enable split tunnelling:

Generally, I would recommend against enabling split tunnelling. You gain a bit of extra bandwidth but you also introduce a security problem. If you are worried that users might secretly enable split tunnelling without your permission, you could consider assigning "off-subnet" virtual IP addresses to the VPN clients. This is not a water tight solution but it would keep away most users. With off-subnet I mean that the virtual IP addresses (specified with l2tpd's "ip range" parameter) are not within the internal subnet ( in my example configuration files). You will have to do some extra routing on the VPN server so that the client on the virtual IP subnet can still reach resources on the internal subnet. Let's assume you have configured off-subnet virtual IP addresses. So what happens when a user connects? When the user has disabled split tunnelling, the default route will be to the VPN server. The user can access internal resources because of the extra routing on the VPN server. Internet sites are also routed through the VPN server and thus accessible. When the user has enabled split tunnelling, however, things are a bit different. The VPN client's default route will be to the Internet. So Internet sites will be accessible. But resources on the internal subnet are not accessible. The client has an off-subnet virtual IP address so packets to the internal subnet will be sent to the default route (the Internet), not to the VPN server. If you use private non-routable IP addresses on your internal subnet (following RFC 1918, e.g., then the ISP will drop those packets. In other words, the VPN client will not be able to access resources on the internal subnet when split tunnelling is enabled, which was the whole objective. Unfortunately this method can be circumvented by users who define a static route to the internal subnet on their workstation. There is no cure for negligent users, except not using a VPN at all...

Back to Contents

22.1 Troubleshooting

The very nature of a VPN makes it difficult to troubleshoot. VPN servers do not want to give away much information to a potential 'attacker' in case of a problem. Packets may be silently dropped by the server. Error messages sent to the client might not be very helpful. That said, here are some tips in case of problems with your L2TP/IPsec connection.

Use the command ipsec verify. Not every "[Failed]" error that it reports is a real problem but it can be helpful in certain circumstances. If you don't use Opportunistic Encryption you should ignore the error messages about missing reverse DNS entries, for instance.

Sometimes the IPsec packets get blocked. For instance when there is a too restrictive firewall on the route or when an ISP actively blocks VPN protocols. Verify with tcpdump that the IPsec packets (UDP ports 500/4500 and IP protocol 50) actually arrive at your Linux server. You can also use Openswan's ipsec ikeping command for that. Problems can be debugged by tracing packets layer by layer:

tcpdump -n -i eth0 not port 22

tcpdump -n -i ipsec0 -s 1500
tcpdump -n -i ppp0 -s 1500
tcpdump -n -i eth1

Where eth0 is the hostile network, ipsec0 is the IPsec connection running over the hostile network (only if you use KLIPS; if you use NETKEY there is no ipsec0), ppp0 is the L2TP/PPP connection through the IPsec tunnel, and eth1 is the internal network. If your network devices are named or numbered differently, change accordingly.

If you use KLIPS you should not see any unencrypted packets (e.g. pure L2TP) on the eth0 interface. If you do see unencrypted packets with KLIPS, you don't have a secure VPN (make sure that the automatic L2TP/IPsec policy has been activated on the Windows client). But if you use NETKEY there is no ipsec0 interface. If you use NETKEY and do a tcpdump on the external interface eth0, you might see unencrypted packets in one direction. The best (and authoritative) solution is to check with a third system between the client and the server. If you sniff the communication between these two with run tcpdump /Ethereal/etc. you should see encrypted packets all around.

Use ping to test the flow of packets:

ping <external IP>

ping <internal IP >
ping <local ip in l2tpd.conf>
ping <any IP address on internal network>

22.2 IPsec logging on the Linux server

Openswan logs its error messages to /var/log/secure. There can be various reasons why you don't get a working IPsec connection:

This Openswan troubleshooting guide by Michael Richardson may be of help, although it assumes that the Openswan system is the initiator of the connection (in most L2TP/IPsec cases it is the responder because a Windows or Mac client is the initiator).

You might be tempted to enable full Pluto debugging by adding plutodebug=all to ipsec.conf. This will force Pluto (the IKE daemon) to log in much more detail. In most cases this should not be necessary because your problem is likely to be caused by some configuration error. Only when there are serious issues you may need to enable Pluto debugging. For example, when a brand new version of Openswan or the Linux kernel has been released and a problem was introduced in this version. Or when you want to interoperate with an IPsec product that has not yet been tested with Openswan. When you have such advanced implementation or interoperability problems, you may even have to enable KLIPS (kernel) debugging with klipsdebug=all. This will also show packets silently dropped by the kernel. To activate these changes Pluto will have to be restarted. More information can be find in the FreeS/WAN Troubleshooting Guide. Do not use these settings in production setups because they make it very easy for the bad guys to perform a Denial Of Service.

If you have got IPsec working but still you cannot log in, you might want to have a look at the L2TP and PPP error messages. If the L2TP and/or PPP phase fails, the IPsec connection will be terminated, eventually. The PPP and L2TP error messages are typically logged to /var/log/messages. But not all debug messages are logged there. Mandrake logs verbose debug messages to /var/log/daemons/*. For Red Hat you might have to increase the debug level by editing /etc/syslogd.conf and adding *.debug; to the line containing /var/log/messages. Then you'll have to restart syslogd with service syslog restart. On Debian/Ubuntu etc. you have to look in /var/log/debug for L2TP messages and /var/log/auth.log for IPsec messages.

22.3 Gateway not reachable

You are using KLIPS and you get the following error:

route-host output: /usr/lib/ipsec/_updown: doroute `ip route add x.x.x.x/32 via x.x.x.x dev ipsec0 ' failed
(RTNETLINK answers: Network is unreachable)

This means that the packets are not routed to the proper ipsecN interface so they are not processed
by KLIPS. This may be a bug which should be solved in Openswan 2.3.1+. You can work around this problem by adding an extra parameter to the Openswan configuration file, e.g.:

conn L2TP-CERT

There are no ipsecN interfaces if you are using NETKEY instead of KLIPS. So in theory the leftnexthop parameter would not work but I noticed that I did have to use this leftnexthop= parameter when the Openswan server was behind NAT.

22.4 Mars Attacks!

If a tcpdump shows that the Linux server does not seem to respond to packets from the Windows/Mac client, then you might not have disabled Reverse Path filtering. If you see lines like this in /var/log/messages:

martian destination from, dev ipsec0
martian source from, on dev eth0
ll header: 00:00:00:00:00:00:00:40:33:2c:70:c8:08:00

... then you need to clear rp_filter. If you scroll back in /var/log/messages, you will see that FreeS/WAN had complained about it (Openswan by default clears this parameter):

ipsec_setup: (/proc/sys/net/ipv4/conf/eth0/rp_filter = `1', should be 0)

A quick echo 0 > /proc/sys/net/ipv4/conf/eth0/rp_filter fixes the problem, if eth0 is the device. Alternatively, you could add net.ipv4.conf.default.rp_filter = 0 to /etc/sysctl.conf. Openswan supports the parameter rp_filter in the config setup section of ipsec.conf. Possible values are %unchanged (to leave it alone) or 0, 1, 2. It defaults to 0, i.e. disable for all interfaces used by Openswan.

22.5 Troubleshooting on the client (Windows/Mac)

Additionally, you can examine the IPsec and L2TP/PPP logs on the client side. Note that I'm not terribly impressed with the error messages in Windows. I find the debug messages on the Linux side (Openswan, l2tpd, pppd) much more informative. DNS problems are pretty common on VPNs. Here is a webpage with tips on resolving such issues.

22.6 "This binary does not support kernel L2TP"

When you use l2tpd you will see the message "This binary does not support kernel L2TP" in /var/log/messages. This does not indicate an error. It is just an informational message. The original creators of l2tpd intended to implement support for both user mode and kernel mode L2TP. But they never got around to do kernel mode L2TP.

22.7 PPP troubleshooting on Linux

This is a bit outside the scope of this document. You can find some great tips on this PPTP troubleshooting page.

Back to Contents

23.1 Certificates

X.509 certificates can be used as an alternative to Preshared Keys. As mentioned above, certificates require Strongsec's certificate patch. You will need to set up some kind of Public Key Infrastructure (PKI) for creating the private keys and X.509 certificates for the Openswan host and the Windows IPsec clients.

Don't forget to disable the PSK configuration files once you start using certificates. Openswan will get confused if there are PSK and certificate configuration files for the same client / IP address. The PSK configuration will get the upper hand. If you do need certificates and PSKs working at the same time for Road Warriors, check out the instructions by Sverre Gunnersen.

OpenSSL is a popular choice for generating certificates because it is Open Source and freely available. This page is not a tutorial on using OpenSSL or certificates in general. Excellent documentation on generating certificates is available elsewhere:

The thing to remember is that the X.509 support in Openswan and strongSwan is very flexible. There Is More Than One Way To Do certificates. For instance, you can use one configuration file per client where each file has a different rightcert= line. Or you could define only one configuration file which is used for all clients where you specify the name of the CA with rightca= (don't forget to issue CRLs for those certificates that are stolen or lost). Wildcards are supported, Check out the Strongsec documentation for all options.

It is probably wise to avoid special characters in the certificates, as this example shows. Software such as OpenCA, IDX-PKI and TinyCA (included in roCA, a secure off-line CA based on Knoppix) could be useful when setting up a Certificate Authority. For more Open Source CAs, see this Nyetwork Wiki. Alternatively, you could buy software which generates certificates for you. Windows NT/2000/2003 ships with Certificate Authority software. SSH sells Certifier, an "PKI platform product for issuing and managing digital certificates in a managed service provider and enterprise environment". There are probably lots of other vendors, including Verisign and Baltimore.

Either way, you will have to end up with certificates in PKCS#12 form. PKCS#12 is a standard for distributing keys and certificates. A PKCS#12 file with extension .p12. or .pfx contains the client's private key, its corresponding certificate and the root CA's certificate (possibly even a chain of CA certificates). Since it contains the client's private key, the PKCS#12 file is encrypted with a password. Note that client certificates have to be signed by the same Certificate Authority (CA) as Openswan's certificate!

You import certificates for L2TP/IPsec into Windows 2000/XP/Vista using the Microsoft Management Console (see these instructions). Windows 95/98/Me do not ship with MMC. Instead, you import the PKCS#12 file into Internet Explorer,  or to be more precise, the certificate store found under "Tools -> Internet Options" (see these instructions).

The CA included with Windows 2000/2003 Server supports Extended Key Usages properties (in ASN.1 Object Identifier notation) for certificates. These EKUs restrict certificates so that they can only be used for IPsec connections. Other uses (say, for a webserver) are not allowed. This adds a little bit of extra security to the server's PKI. If you use OpenSSL, you can probably create certificates with these EKU properties too, but I haven't looked into it.

There has been a report to the Bugtraq mailinglist by Thor Lancelot Simon about a potential Man-in-the-Middle vulnerability (mirror) with certain IPsec implementations. If one client certificate is compromised (which is not too difficult: a laptop can get stolen, for instance), it could be used by an attacker to masquerade as a server certificate to a second client. After all, the compromised certificate of the first client is a valid certificate issued by the CA. The issue is only true for Windows clients if no EKUs are used. If you do use EKUs (as described above) then the issue is mitigated. Windows Vista and Mac clients are also not vulnerable to this issue because they require the server's hostname or IP address in the certificate's ID (can be disabled on Vista as an option).

23.2 Troubleshooting: Windows 2000/XP/Vista picks the wrong certificate

Some people reported the following problem. They have a Windows 2000, XP or Vista client that should connect to 2 or more Openswan L2TP/IPsec servers that each have their own CA. Different certificates were installed on the client. But when the client connects to a server, it always presents the same certificate to the server. So the client can connect correctly to one server but not to the other VPN server(s). This is not what you want. The client should be able to connect to all servers.

The short story is that adding rightca=%same to the configuration should solve the problem. The long story is described here.

23.3 Certificates on smartcards, two-factor authentication

Certificates can be stored on smartcards (including USB tokens) for extra security. I have tested this with the MSL2TP client and it works. When I started the connection, it asked for the PIN number of my IKey 2032 token and only then the connection was set up. SSH Sentinel should work too (it comes with a separate smartcard application called Accession Lite) but I had a bit of a fight with the IKey driver.

Using smartcards with the built-in IPsec client of Windows 2000/XP/Vista may be a problem. Microsoft distinguishes between 'machine certificates' and 'user certificates'. A machine certificate is only used for IPsec and cannot be installed on a smartcard or token (as mentioned above). A user certificate on the other hand is used in the PPP phase when EAP authentication is selected. I could not find a way to store the 'machine certificate' (also called 'local computer certificate') on my USB token. Please correct me if I am wrong! According to the Microsoft documentation, it is possible to specify a different CSP (Cryptographic Service Provider), such as the one supplied with your smartcard/USB token, but only when you apply for the certificate through the web interface ("web enrolment"). Not when you import your certificate through a file in PKCS#12 format.

So if you want to users to authenticate with certificates using the built-in Windows L2TP/IPsec client, you will need a Linux PPP server with EAP support (or a PPP server with a RADIUS plugin so that it can authenticate against a RADIUS server through EAP-TLS). This is outside the scope of this webpage, but FreeRADIUS should be able to do this. Check out the EAP-TLS support in FreeRADIUS. An extension for the pppd RADIUS plugin by Michael Heiart may also be needed. An alternative is to switch to a third-party Windows IPsec client with smartcard support such as SSH Sentinel.

Back to Contents

24. Protecting wireless connections

Here are some thoughts on using L2TP/IPsec over wireless connections.

The WEP encryption algorithm, which is part of the wireless 802.11 standard (WiFi), is completely broken. Any scriptkiddie can crack WEP passwords in a couple of minutes with tools such as Backtrack or Airsnort. Another problem is that distributing WEP keys does not scale well if you have more than only a couple of users. Most wireless access points have additional security measure (such as SSID authentication, disabling SSID broadcast and access controls based on MAC addresses) but these are also weak. They are easily circumvented with tools such as Airsnort or Kismet. WPA and WPA2 are alternatives but some people have suggested to use VPN protocol such as IPsec or PPTP instead. These protocols have been around for a while and their security is well understood. It is clear that you should not rely on WEP alone.

The problem with running a VPN over wireless links is that the bad guys are not prevented from having access to your wireless network. They can still do nasty things such as attacking your access points or your wireless clients. If a user workstation has not been secured, it could get exploited by a hacker or a virus through the wireless link. For instance, the workstation might have open ports, File and Print Sharing has not been disabled, no personal firewall has been installed or security patches are missing. WEP and its stronger counterparts such as WPA, 802.1x, EAP/PEAP and WAP2 / 802.11x prevent people from joining your wireless network if they do not have the proper credentials. Microsoft is actively pushing PEAP for wireless networks in Windows XP/Vista/2003. So if you are not using a good wireless security protocol (WEP doesn't count), it is probably wise to at least install a personal firewall and possibly anti-virus software on the client workstations.

IPsec supports both Preshared Keys and certificates. PSKs are easier to use if the number of users is small but they have the disadvantage that users must have fixed (static) IP addresses. Fortunately, fixed addresses from the RFC 1918 ranges (192.168.x.x, 10.x.x.x etc) can often be assigned to wireless users. With one exception: it will not work under Windows 2000 when the 'New Connection Wizard' is used, because this Wizard does not support PSKs.

Another thing to watch out for is that if the VPN connection is down, clients should obviously not be able to access your protected network. This sounds silly but if you have IP forwarding set to on, packets from the wireless net will be happily forwarded to your internal network. Be sure to firewall the wireless network, or to configure IP forwarding for only the internal addresses.

Roaming from one access point to another access point is currently a problem with wireless networks based on the 802.11 standard. If I'm not mistaken the wireless link gets disconnected. This means that the IPsec connection will get disconnected as well, because the new access point may hand out a new IP address. There are however provisions in the L2TP standard that should help to reconnect fast, but I don't know if it works with Openswan and l2tpd.

An initial test on a wireless network did not show more MTU size problems than usual (i.e. with Ethernet). A complication might be that the built-in L2TP/IPsec client might have trouble recovering from lost packets if the link quality is bad (I have not tested this yet).

Students from Stuttgart, Germany are using L2TP/IPsec on a wireless network in their dorm using a Linksys WRT54g router and a central FreeS/WAN server authenticating through FreeRADIUS: the WH-Funk project.

Back to Contents

25. Other topics

25.1 Linux as an L2TP/IPsec client

This section has been moved to a separate page.

25.2 Dead Pear Detection

Dead Peer Detection (DPD) is an IPsec feature described in RFC 3706. If client and server agree on the use of DPD they send keep-alive packets through the IPsec connection. This allows them to determine whether the remote peer is still responding or not. Connections that are determined to be dead (DPD speeds this up) will be deleted, allowing resources to be reclaimed.

DPD is mainly supported by Cisco, Openswan, Strongswan and Mac OS X 10.5 "Leopard" and higher. None of the L2TP/IPsec clients for Windows support DPD, nor do Mac OS X 10.3 and 10.4. However, clients such as Windows that do not support DPD will simple ignore any DPD packets, so there is no downside in enabling DPD on the VPN server. If you want to support Mac OS 10.5 and Linux clients it is highly recommended to enable DPD. On the Linux server you add the following parameters to your 'conn' section:
conn L2TP-PSK

The values 40 and 130 are provided as an example. On Linux clients you add the following parameters (i.e. both peers must have them and it is recommended to use slightly different values, as to avoid starting DPD negotiations from both sides at the same time):

Both L2TP and PPP also have some form of keep-alive mechanism. L2TP sends "Hello" control messages to keep the link alive (these show up as "check_control: control, cid = 0, Ns = 4, Nr = 17" in the l2tpd logs). The Link Control Protocol in PPP also has a keep-alive mechanism (LCP EchoReq and EchoRep). This option is enabled by default in pppd (see "man pppd"). I suspect that Windows and Mac clients use these keep-alive mechanism to their benefit but the Linux side does not, because there is no interaction between the IPsec daemon and the L2TP/PPP daemons.

Back to Contents

26. This is purely speculation, but...

Could Microsoft also be trying to commoditise the IPsec protocol, while extending/embracing the (Microsoft/Cisco co-developed) L2TP protocol? Microsoft likes to write about L2TP as if it were a VPN protocol. For instance here: "Q. Do I want to use PPTP or L2TP? A. L2TP is considered more secured than PPTP and will allow you interoperability with 3rd parties". What they mean with 'L2TP' is actually 'L2TP/IPsec'. It is IPsec that provides the confidentiality and authentication in L2TP/IPsec connections. They seem to be shifting public attention from IPsec (several implementations by multiple vendors) to L2TP (not that many products available). Another thing is that not everyone is happy with L2TP. It results in 'double tunnelling' (extra overhead) and might very well be only a vehicle for doing extra authentication (of course, preferably against an NT SAM user database, which means buying more Windows client access licences).

I am a bit puzzled why Microsoft released the MSL2TP client. They have recently stopped support for 'old' Windows versions, and here they are releasing software which runs even on Windows 95. Interesting also is that SafeNet is eating into its own third-party Windows IPsec client market, to some extent, by releasing the free MSL2TP client. Why would they do this? An offer by Microsoft "they could not not refuse"?

I would also like to make the following observation. The MSL2TP client was developed for Microsoft by SafeNet. The IKE/IPsec support in Windows 2000/XP and Pocket PC was "jointly developed with Cisco". It appears that IPsec is not a high enough priority for Microsoft to do all development in-house...

Back to Contents

27. Related links
Back to Contents

28. Acknowledgements and disclaimer

Here are some messages by my crack team of lawyers. Openswan is a trademark by Xelerance. The Openswan logo was designed by Nana Manojlovic. Windows is a registered trademark of Microsoft Corporation in the United States and other countries. Apple Macintosh is a trademark of Apple Computer, Inc., registered in the U.S. and other countries. All other trademarks are the property of their respective owners. These pages do not express or imply affiliation, sponsorship, endorsement, certification, or approval by any company.

Back to Contents

29. Revision history

Apr 7, 2008: Openswan L2TP/IPsec setup available on OpenL2TP website.
Mar 12, 2008:
Added remarks about DHCP Inform.
Jun 27, 2007: Apple iPhone ships with L2TP/IPsec client.
Jun 2, 2007: Stinghorn no longer available?
Dec 27, 2006: Added info on OpenL2TP, with thanks to James Chapman.
Nov 22, 2006: Moved Linux as an L2TP/IPsec client to a separate page.
Nov 9, 2006: Vista NAT-T issue resolved.
Nov 5, 2006: Added page for Windows Vista.
Jan 7, 2006: Added link to Xelerance's xl2tpd.
Dec 6, 2005: Updated info on Linux as an L2TP/IPsec client.
Nov 15, 2005: Updated sample Openswan configuration files (rekey=no etc.)
Nov 15, 2005: Openswan 2.4.2 partially supports NATed Mac clients.
Jul 6, 2005:
EU Parliament rejects software patents! (For now, at least).
Jun 25, 2005: Uploaded l2tpd-0.69-12jdl (S)RPM that compiles and works on for FC4. Added ppp 'unit' info.
Jun 14, 2005: Added reference to the Securepoint Personal Firewall & VPN Client.
Jun 2, 2005: I consider removing support for ancient versions such as Mandrake 9.1 and older.
May 31, 2005: l2tpd is now in maintenance mode. The domain has been hijacked. Website has moved.
May 3, 2005: Mac OS X v10.4 ("Tiger") Client and Server support certificates for L2TP/IPsec but I have not read reports of anyone using this method successfully.
Mar 17, 2005: Uploaded l2tpd-0.69-11jdl RPMS for FC2/3 and MDK10.1. These contain Unix98 pty support.
Mar 17, 2005: Uploaded experimental patch to fix server-side NAT-OA bug.
Mar 16, 2005: Found out about another L2TP server, l2tpns.
Jan 21, 2005: Nate Carlson has made an L2TP/IPsec Howto.
Jan 8, 2005: NAT-T standard approved by the IETF as RFC 3947.
[Old revision history]
July 20, 2002: Added report of preliminary l2tpd success.
July 15, 2002: Changed <h4> headings since Opera does not display them. Added PSK remarks.

Jacco de Leeuw