Internet Key Exchange Protocol version 2

A Security Association (SA) is an agreement between two network nodes on how to process certain traffic between them. This processing may involve encapsulation, authentication, encryption, and compression. The Internet Key Exchange Protocol (IKE) provides a standardized way to establish these associations dynamically.

IKE can be deployed on a network node to negotiate Security Associations for that node. These IKE implementations can only negotiate with other IKE implementations, so IKE must be on each node that is to be an endpoint of an IKE-negotiated Security Association. No other nodes need to be running IKE.

An IKE instance (i.e. an IKE implementation on a particular network node) communicates with other IKE instance using either UDP or TCP packets, so there must be a route between the nodes in each direction.

The negotiation of Security Associations requires a number of choices that involve tradeoffs between security, convenience, trust, and efficiency. These are policy issues and are normally specified to the IKE instance by the system administrator.

IKE deals with two kinds of Security Associations. The first part of a negotiation between IKE instances is to build an IKE SA. An IKE SA is used to establish an authenticated and private communication channel between the two IKE nodes. This communication channel is used to negotiate the actually Child SAs, which are the actual IPsec SAs. The IPsec SAs carry protected IP traffic between the systems.

Negotiating an IKE SA (referred to as the The Initial Exchanges) consists of at least the IKE_SA_INIT exchange to establish a private channel, and the IKE_AUTH exchange to authenticate the channel. Extensions, such as EAP and INTERMEDIATE, may add further exchanges.

Negotiating a Child SA requires a single CREATE_CHILD_SA exchange. As an optimization, the first Child SA's negotiation may be piggybacked on the IKE_AUTH exchange.

IKE negotiation can be initiated by any instance with any other instance. If both can find an agreeable set of characteristics for a Security Association, and both recognize each others authenticity, they can set up a Security Association. The standards do not specify what causes an IKE instance to initiate a negotiation.

In summary, an IKE instance is prepared to automate the management of Security Associations in an IPsec environment, but a number of issues are considered policy and are left in the system administrator's hands.

Libreswan

Libreswan implements the Internet Key Exchange (both versions 1 and 2). It runs as a daemon (called pluto) on a network node. Currently, this network node must be a Linux, NetBSD, FreeBSD, or OpenBSD system.

The policy for acceptable characteristics for Security Associations are specified using the configuration file ipsec.conf (see ipsec.conf(5)).

Libreswan can be configured to authenticate its peer using shared secrets or using public and private keys (see ipsec.conf(5)). Shared secrets are stored in the file ipsec.secrets (see ipsec.secrets(5)). Public keys (for instance, X.509 certificates) are stored in either Libreswan's NSS trust store, obtained from the peer, or using DNS(SEC). Corresponding Private keys are always stored in the Libreswan's NSS trust store.

Libreswan initiates negotiation of a Security Association when:

Libreswan implements IKE SAs itself. After it has negotiated the characteristics for a Child SA (IPsec SA), it directs the kernel to install the association. If necessary, it also invokes a script to adjust any firewall or routing rules (see ipsec-updown(5)).

When Libreswan shuts down, it closes all Security Associations.

The updown command

Whenever libreswan brings a connection up or down, it invokes the updown command. This command is specified either using ipsec.conf configuration file option leftupdown= or the ipsec whack option --updown option. This allows for customized control over routing and firewall manipulation.

The updown script is invoked for five different operations. The operation name, with -host, -client and -v6, appended is passed to the updown script using the environment variable PLUTO_VERB (see below) The operations, in the order they are normally invoked, are as follows:

The script is passed a large number of environment variables to specify what needs to be done:

All output sent by the script to stderr or stdout is logged. The script should return an exit status of 0 if and only if it succeeds.

pluto waits for the script to finish and will not do any other processing while it is waiting. The script may assume that pluto will not change anything while the script runs. The script should avoid doing anything that takes much time and it should not issue any command that requires processing by pluto. Either of these activities could be performed by a background subprocess of the script.

Before Running Pluto

pluto runs as a daemon with userid root. Before running it, a few things must be set up.

pluto requires a working IPsec stack.

pluto supports multiple public networks (that is, networks that are considered insecure and thus need to have their traffic encrypted or authenticated). It discovers the public interfaces to use by looking at all interfaces that are configured (the --interface option can be used to limit the interfaces considered). It does this only when whack tells it to --listen, so the interfaces must be configured by then. The --listen can be used to limit listening on only 1 IP address of a certain interface. ip(8) with the addr option will show the name and status of each network interface.

pluto requires a database of preshared secrets and RSA private keys. This is described in the ipsec.secrets(5). pluto is told of RSA public keys via whack commands. If the connection is Opportunistic, and no RSA public key is known, pluto will attempt to fetch RSA keys using the Domain Name System.

Setting up XFRM for pluto

No special requirements are necessary to use XFRM - it ships with all modern versions of Linux 2.4 and later.

ipsec.secrets file

A pluto daemon and another IKE daemon (for example, another instance of pluto) must convince each other that they are who they are supposed to be before any negotiation can succeed. This authentication is accomplished by using either secrets that have been shared beforehand (manually) or by using RSA signatures. There are other techniques, but they have not been implemented in pluto.

The file /etc/ipsec.secrets is used to keep preshared secret keys and XAUTH passwords. RSA private keys, X.509 certificates, CRLs, OCSP and smartcards are handled via NSS. For debugging, there is an argument to the pluto command to use a different file. This file is described in ipsec.secrets(5).

Running Pluto

To fire up the daemon, just type pluto (be sure to be running as the superuser). The default IKE port number is 500, the UDP port assigned by IANA for IKE Daemons. pluto must be run by the superuser to be able to use the UDP 500 port. If pluto is told to enable NAT-Traversal, then UDP port 4500 is also taken by pluto to listen on.

Pluto supports different IPstacks on different operating systems. This can be configured using one of the options --use-netkey (Linux), --use-bsdkame (BSD). On startup, pluto might also read the protostack= option to select the IPsec stack to use if --config /etc/ipsec.conf is given as argument to pluto. If both --use-XXX and --config /etc/ipsec.conf are specified, the last command line argument specified takes precedence.

Pluto supports RFC 3947 NAT-Traversal. The allowed range behind the NAT routers is submitted using the --virtual-private option. See ipsec.conf(5) for the syntax. The option --force-keepalive forces the sending of the keep-alive packets, which are send to prevent the NAT router from closing its port when there is not enough traffic on the IPsec connection. The --keep-alive sets the delay (in seconds) of these keep-alive packets. The newer NAT-T standards support port floating, and Libreswan enables this per default.

Pluto supports the use of X.509 certificates and sends certificates when needed. Pluto uses NSS for all X.509 related data, including CAcerts, certs, CRLs and private keys. The Certificate Revocation Lists can also be retrieved from an URL. The option --crlcheckinterval sets the time between checking for CRL expiration and issuing new fetch commands. Pluto logs a warning if no valid CRL was loaded or obtained for a connection. If --crl-strict is given, the connection will be rejected until a valid CRL has been loaded.

Pluto can also use helper children to off-load cryptographic operations. This behavior can be fine tuned using the --nhelpers. Pluto will start (n-1) of them, where n is the number of CPU's you have (including hypherthreaded CPU's). A value of 0 forces pluto to do all operations in the main process. A value of -1 tells pluto to perform the above calculation. Any other value forces the number to that amount.

Pluto uses the NSS crypto library as its random source. Some government Three Letter Agency requires that pluto reads 440 bits from /dev/random and feed this into the NSS RNG before drawing random from the NSS library, despite the NSS library itself already seeding its internal state. As this process can block pluto for an extended time, the default is to not perform this redundant seeding. The --seedbits option can be used to specify the number of bits that will be pulled from /dev/random and seeded into the NSS RNG. This can also be accomplished by specifying seedbits in the "config setup" section of ipsec.conf. This option should not be used by most people.

pluto attempts to create a lockfile with the name /run/pluto/pluto.pid. If the lockfile cannot be created, pluto exits - this prevents multiple plutos from competing Any "leftover" lockfile must be removed before pluto will run. pluto writes its PID into this file so that scripts can find it. This lock will not function properly if it is on an NFS volume (but sharing locks on multiple machines doesn't make sense anyway).

pluto then forks and the parent exits. This is the conventional "daemon fork". It can make debugging awkward, so there is an option to suppress this fork. In certain configurations, pluto might also launch helper programs to assist with DNS queries or to offload cryptographic operations.

All logging, including diagnostics, is sent to syslog(3) with facility=authpriv; it decides where to put these messages (possibly in /var/log/secure or /var/log/auth.log). Since this too can make debugging awkward, the option --stderrlog is used to steer logging to stderr.

Alternatively, --logfile can be used to send all logging information to a specific file.

Once pluto is started, it waits for requests from whack.

Pluto's Internal State

To understand how to use pluto, it is helpful to understand a little about its internal state. Furthermore, the terminology is needed to decipher some of the diagnostic messages.

Pluto supports food groups for Opportunistic IPsec. The policies for these are located in /etc/ipsec.d/policies, or another directory as specified by --ipsecdir.

Pluto supports X.509 Certificates. All certificate handling is done using the NSS library and all certificate material is stored in an NSS database in /var/lib/ipsec/nss or another directory as specified by --nssdir.

Pluto may core dump. It will normally do so into the current working directory. You can specify the --coredir option for pluto, or specify the dumpdir= option in ipsec.conf.

If you are investigating a potential memory leak in pluto, start pluto with the --leak-detective option. Before the leak causes the system or pluto to die, shut down pluto in the regular way. pluto will display a list of leaks it has detected.

If you are investigating a potential use-after-free or double-free in pluto, first build pluto with USE_EFENCE=true and then start pluto with --efence-protect. See efence(2) and EF_PROTECT_BELOW and EF_PROTECT_FREE.

The (potential) connection database describes attributes of a connection. These include the IP addresses of the hosts and client subnets and the security characteristics desired. pluto requires this information (simply called a connection) before it can respond to a request to build an SA. Each connection is given a name when it is created, and all references are made using this name.

During the IKE exchange to build an SA, the information about the negotiation is represented in a state object. Each state object reflects how far the negotiation has reached. Once the negotiation is complete and the SA established, the state object remains to represent the SA. When the SA is terminated, the state object is discarded. Each State object is given a serial number and this is used to refer to the state objects in logged messages.

Each state object corresponds to a connection and can be thought of as an instantiation of that connection. At any particular time, there may be any number of state objects corresponding to a particular connection. Often there is one representing an ISAKMP SA and another representing an IPsec SA.

XFRM requires no special routing.

Each connection may be routed, and must be while it has an IPsec SA. The connection specifies the characteristics of the route: the interface on this machine, the "gateway" (the nexthop), and the peer's client subnet. Two connections may not be simultaneously routed if they are for the same peer's client subnet but use different interfaces or gateways (pluto's logic does not reflect any advanced routing capabilities).

When pluto needs to install a route for a connection, it must make sure that no conflicting route is in use. If another connection has a conflicting route, that route will be taken down, as long as there is no IPsec SA instantiating that connection. If there is such an IPsec SA, the attempt to install a route will fail.

There is an exception. If pluto, as Responder, needs to install a route to a fixed client subnet for a connection, and there is already a conflicting route, then the SAs using the route are deleted to make room for the new SAs. The rationale is that the new connection is probably more current. The need for this usually is a product of Road Warrior connections (these are explained later; they cannot be used to initiate).

When pluto needs to install an eroute for an IPsec SA (for a state object), first the state object's connection must be routed (if this cannot be done, the eroute and SA will not be installed). If a conflicting eroute is already in place for another connection, the eroute and SA will not be installed (but note that the routing exception mentioned above may have already deleted potentially conflicting SAs). If another IPsec SA for the same connection already has an eroute, all its outgoing traffic is taken over by the new eroute. The incoming traffic will still be processed. This characteristic is exploited during rekeying.

Old Examples

It would be normal to start pluto in one of the system initialization scripts. It needs to be run by the superuser. Generally, no arguments are needed. To run in manually, the superuser can simply type

ipsec pluto

The command will immediately return, but a pluto process will be left running, waiting for requests from whack or a peer.

Using whack, several potential connections would be described:

ipsec whack --name silly --host 127.0.0.1 --to --host 127.0.0.2 --ikelifetime 900 --ipseclifetime 800 --keyingtries 3

Since this silly connection description specifies neither encryption, authentication, nor tunneling, it could only be used to establish an ISAKMP SA.

ipsec whack --name conn_name --host 10.0.0.1 --client 10.0.1.0/24 --to --host 10.0.0.2 --client 10.0.2.0/24 --encrypt

This is something that must be done on both sides. If the other side is pluto, the same whack command could be used on it (the command syntax is designed to not distinguish which end is ours).

Now that the connections are specified, pluto is ready to handle requests and replies via the public interfaces. We must tell it to discover those interfaces and start accepting messages from peers:

ipsec whack --listen

If we don't immediately wish to bring up a secure connection between the two clients, we might wish to prevent insecure traffic. The routing form asks pluto to cause the packets sent from our client to the peer's client to be routed through the ipsec0 device; if there is no SA, they will be discarded:

ipsec whack --route conn_name

Finally, we are ready to get pluto to initiate negotiation for an IPsec SA (and implicitly, an ISAKMP SA):

ipsec whack --initiate --name conn_name

A small log of interesting events will appear on standard output (other logging is sent to syslog).

whack can also be used to terminate pluto cleanly, tearing down all SAs that it has negotiated.

ipsec whack --shutdown

Notification of any IPSEC SA deletion, but not ISAKMP SA deletion is sent to the peer. Unfortunately, such Notification is not reliable. Furthermore, pluto itself ignores Notifications.

XAUTH

If pluto needs additional authentication, such as defined by the XAUTH specifications, then it may ask whack to prompt the operator for username or passwords. Typically, these will be entered interactively. A GUI that wraps around whack may look for the 041 (username) or 040 (password) prompts, and display them to the user.

For testing purposes, the options --xauthuser user --xauthpass pass may be be given prior to the --initiate to provide responses to the username and password prompts.

Rekeying

When an SA that was initiated by pluto has only a bit of lifetime left, pluto will initiate the creation of a new SA. This applies to ISAKMP and IPsec SAs. The rekeying will be initiated when the SA's remaining lifetime is less than the rekeymargin plus a random percentage, between 0 and rekeyfuzz, of the rekeymargin.

Similarly, when an SA that was initiated by the peer has only a bit of lifetime left, pluto will try to initiate the creation of a replacement. To give preference to the initiator, this rekeying will only be initiated when the SA's remaining lifetime is half of rekeymargin. If rekeying is done by the responder, the roles will be reversed: the responder for the old SA will be the initiator for the replacement. The former initiator might also initiate rekeying, so there may be redundant SAs created. To avoid these complications, make sure that rekeymargin is generous.

One risk of having the former responder initiate is that perhaps none of its proposals is acceptable to the former initiator (they have not been used in a successful negotiation). To reduce the chances of this happening, and to prevent loss of security, the policy settings are taken from the old SA (this is the case even if the former initiator is initiating). These may be stricter than those of the connection.

pluto will not rekey an SA if that SA is not the most recent of its type (IPsec or ISAKMP) for its potential connection. This avoids creating redundant SAs.

The random component in the rekeying time (rekeyfuzz) is intended to make certain pathological patterns of rekeying unstable. If both sides decide to rekey at the same time, twice as many SAs as necessary are created. This could become a stable pattern without the randomness.

Another more important case occurs when a security gateway has SAs with many other security gateways. Each of these connections might need to be rekeyed at the same time. This would cause a high peek requirement for resources (network bandwidth, CPU time, entropy for random numbers). The rekeyfuzz can be used to stagger the rekeying times.

Once a new set of SAs has been negotiated, pluto will never send traffic on a superseded one. Traffic will be accepted on an old SA until it expires.

Selecting a Connection When Responding: Road Warrior Support

When pluto receives an initial Main Mode message, it needs to decide which connection this message is for. It picks based solely on the source and destination IP addresses of the message. There might be several connections with suitable IP addresses, in which case one of them is arbitrarily chosen. (The ISAKMP SA proposal contained in the message could be taken into account, but it is not.)

The ISAKMP SA is negotiated before the parties pass further identifying information, so all ISAKMP SA characteristics specified in the connection description should be the same for every connection with the same two host IP addresses. At the moment, the only characteristic that might differ is authentication method.

Up to this point, all configuring has presumed that the IP addresses are known to all parties ahead of time. This will not work when either end is mobile (or assigned a dynamic IP address for other reasons). We call this situation "Road Warrior". It is fairly tricky and has some important limitations, most of which are features of the IKE protocol.

Only the initiator may be mobile: the initiator may have an IP number unknown to the responder. When the responder doesn't recognize the IP address on the first Main Mode packet, it looks for a connection with itself as one end and %any as the other. If it cannot find one, it refuses to negotiate. If it does find one, it creates a temporary connection that is a duplicate except with the %any replaced by the source IP address from the packet; if there was no identity specified for the peer, the new IP address will be used.

When pluto is using one of these temporary connections and needs to find the preshared secret or RSA private key in ipsec.secrets, and the connection specified no identity for the peer, %any is used as its identity. After all, the real IP address was apparently unknown to the configuration, so it is unreasonable to require that it be used in this table.

Part way into the Phase 1 (Main Mode) negotiation using one of these temporary connection descriptions, pluto will receive an Identity Payload. At this point, pluto checks for a more appropriate connection, one with an identity for the peer that matches the payload and would use the same keys as so far used for authentication. If it finds one, it will switch to using this better connection (or a temporary one derived from this, if it has %any for the peer's IP address). It may even turn out that no connection matches the newly discovered identity, including the current connection; if so, pluto terminates negotiation.

Unfortunately, if preshared secret authentication is being used, the Identity Payload is encrypted using this secret, so the secret must be selected by the responder without knowing this payload. This limits there to being at most one preshared secret for all Road Warrior systems connecting to a host. RSA Signature authentication does not require that the responder knows how to select the initiator's public key until after the initiator's Identity Payload is decoded (using the responder's private key, so that must be preselected).

When pluto is responding to a Quick Mode negotiation via one of these temporary connection descriptions, it may well find that the subnets specified by the initiator don't match those in the temporary connection description. If so, it will look for a connection with matching subnets, its own host address, a peer address of %any and matching identities. If it finds one, a new temporary connection is derived from this one and used for the Quick Mode negotiation of IPsec SAs. If it does not find one, pluto terminates negotiation.

Be sure to specify an appropriate nexthop for the responder to send a message to the initiator: pluto has no way of guessing it (if forwarding isn't required, use an explicit %direct as the nexthop and the IP address of the initiator will be filled in; the obsolete notation 0.0.0.0 is still accepted).

pluto has no special provision for the initiator side. The current (possibly dynamic) IP address and nexthop must be used in defining connections. These must be properly configured each time the initiator's IP address changes. pluto has no mechanism to do this automatically.

Although we call this Road Warrior Support, it could also be used to support encrypted connections with anonymous initiators. The responder's organization could announce the preshared secret that would be used with unrecognized initiators and let anyone connect. Of course the initiator's identity would not be authenticated.

If any Road Warrior connections are supported, pluto cannot reject an exchange initiated by an unknown host until it has determined that the secret is not shared or the signature is invalid. This must await the third Main Mode message from the initiator. If no Road Warrior connection is supported, the first message from an unknown source would be rejected. This has implications for ease of debugging configurations and for denial of service attacks.

Although a Road Warrior connection must be initiated by the mobile side, the other side can and will rekey using the temporary connection it has created. If the Road Warrior wishes to be able to disconnect, it is probably wise to set --keyingtries to 1 in the connection on the non-mobile side to prevent it trying to rekey the connection. Unfortunately, there is no mechanism to unroute the connection automatically.

Debugging

pluto accepts several optional arguments, useful mostly for debugging. Except for --interface, each should appear at most once.

pluto is willing to produce a prodigious amount of debugging information. There are several classes of debugging output, and pluto may be directed to produce a selection of them. All lines of debugging output are prefixed with "|" to distinguish them from normal diagnostic messages.

When pluto is invoked, it may be given arguments to specify which debug classes to output. The current options are:

The debug form of the whack command will change the selection in a running pluto. If a connection name is specified, the flags are added whenever pluto has identified that it is dealing with that connection. Unfortunately, this is often part way into the operation being observed.

For example, to start pluto with both base and cpu-usage debug-logging enabled:

ipsec pluto --debug base --debug cpu-usage

To later change this pluto to disable base debug-logging use either:

ipsec whack --no-debug base

or:

ipsec whack --debug none --debug cpu-usage

Impairing

pluto and whack accept several optional arguments that alter (impair) correct behaviour.

These options are solely intended for use by developers when testing pluto.

Pluto's Behaviour When Things Go Wrong

When pluto doesn't understand or accept a message, it just ignores the message. It is not yet capable of communicating the problem to the other IKE daemon (in the future it might use Notifications to accomplish this in many cases). It does log a diagnostic.

When pluto gets no response from a message, it resends the same message (a message will be sent at most three times). This is appropriate: UDP is unreliable.

When pluto gets a message that it has already seen, there are many cases when it notices and discards it. This too is appropriate for UDP.

Combine these three rules, and you can explain many apparently mysterious behaviours. In a pluto log, retrying isn't usually the interesting event. The critical thing is either earlier (pluto got a message that it didn't like and so ignored, so it was still awaiting an acceptable message and got impatient) or on the other system (pluto didn't send a reply because it wasn't happy with the previous message).

Notes

Each IPsec SA is assigned an SPI, a 32-bit number used to refer to the SA. The IKE protocol lets the destination of the SA choose the SPI. The range 0 to 0xFF is reserved for IANA. Pluto also avoids choosing an SPI in the range 0x100 to 0xFFF, leaving these SPIs free for manual keying. Remember that the peer, if not pluto, may well chose SPIs in this range.

Policies

This catalogue of policies may be of use when trying to configure pluto and another IKE implementation to interoperate.

In Phase 1, only Main Mode is supported. We are not sure that Aggressive Mode is secure. For one thing, it does not support identity protection. It may allow more severe Denial Of Service attacks.

No Informational Exchanges are supported. These are optional and since their delivery is not assured, they must not matter. It is the case that some IKE implementations won't interoperate without Informational Exchanges, but we feel they are broken.

No Informational Payloads are supported. These are optional, but useful. It is of concern that these payloads are not authenticated in Phase 1, nor in those Phase 2 messages authenticated with HASH(3).