Linux IPsec Summit 2018 wishlist

A scratchpad for things we'd like to talk about during the ipsec meetup

INVALID_SPI acquires

When one endpoint crashes and restarts, and does not need to send traffic, it will lead to traffic for because the other end's ESP/AH packets are getting lost without triggering any ACQUIRE.

It would be nice of the kernel could send a rate limited INVALID_SPI message, so the node can see if it has an ondemand/ready tunnel that it should bring up.

larval acquire saying "transport mode" - would be nice to not say mode at all

src 192.0.2.100 dst 192.1.2.23
	proto esp spi 0xSPISPIXX reqid REQID mode transport
	replay-window 0 
	sel src 192.0.2.100/32 dst 192.1.2.23/32 proto icmp type 8 code 0 dev eth0 

add support for Populate-From-Packet flag. Cause acquires for each different policy hit

Support rate limited acquires on a wider policy for individual policy hits, so we can setup individual IPsec SA's for a part of a single SPD policy entry.

some clarification or documentation for IPsec SA flags

FLAG := noecn | decap-dscp | nopmtudisc | wildrecv | icmp | af-unspec | align4 | esn
EXTRA-FLAG-LIST := [ EXTRA-FLAG-LIST ] EXTRA-FLAG
EXTRA-FLAG := dont-encap-dscp

ip xfrm policy help shows:

FLAG := localok | icmp

XFRM-PROTO := esp | ah | comp | route2 | hao
MODE := transport | tunnel | beet | ro | in_trigger
LEVEL := required | use

some clarification or documentation for /proc values

/proc/sys/net/core/xfrm_acq_expires
/proc/sys/net/core/xfrm_aevent_etime
/proc/sys/net/core/xfrm_aevent_rseqth
/proc/sys/net/core/xfrm_larval_drop

fixup for userland using xfrm.h include

Our kernel_netlink.c code contains:

#include "linux/xfrm.h" /* local (if configured) or system copy */
#include "libreswan.h" /* before xfrm.h otherwise break on F22 */

Depending on how new gcc/glibc/userland and/or kernel is we need to swap these two lines :(

Introduce some kind of #ifdef _KERNEL_ that protects xfrm.h from loading too much kernel related defines, so we only get the XFRM_ values we need to have available in userland. Now on older glibc we get:

In file included from /source/programs/pluto/linux-copy/linux/xfrm.h:4:0,
                 from /source/programs/pluto/kernel_netlink.c:54:
/usr/include/netinet/in.h:99:5: error: expected identifier before numeric constant
     IPPROTO_HOPOPTS = 0,   /* IPv6 Hop-by-Hop options.  */
     ^
In file included from /source/linux/include/libreswan.h:76:0,
                 from /source/programs/pluto/kernel_netlink.c:55:
/usr/include/netinet/in.h:209:8: error: redefinition of ‘struct in6_addr’
 struct in6_addr
        ^
In file included from /source/programs/pluto/linux-copy/linux/xfrm.h:4:0,
                 from /source/programs/pluto/kernel_netlink.c:54:
/usr/include/linux/in6.h:32:8: note: originally defined here
 struct in6_addr {
        ^
[more errors left out]

Note that we have linux-copy/linux/xfrm.h because sometimes we need newer XFRM values then the system provided version has, eg if people upgrade kernel but not glibc.

Comply with RFC 7296 NAT-T requirements

The kernel currently marks an IPsec SA as not natted or encaps-udp. It rejects packets based on this. To comply to the RFC, it should:

   When either side is using port 4500, sending ESP with UDP encapsulation is
   not required, but understanding received UDP-encapsulated ESP packets
   is required.  UDP encapsulation MUST NOT be done on port 500.  If
   Network Address Translation Traversal (NAT-T) is supported (that is,
   if NAT_DETECTION_*_IP payloads were exchanged during IKE_SA_INIT),
   all devices MUST be able to receive and process both UDP-encapsulated
   ESP and non-UDP-encapsulated ESP packets at any time.  Either side
   can decide whether or not to use UDP encapsulation for ESP
   irrespective of the choice made by the other side.  However, if a NAT
   is detected, both devices MUST use UDP encapsulation for ESP.

This is also important for TCP support.

crypto module loading problems

The problem of not autoloading crypto modules means we have to manually load modules. We don't want the daemon to be able to load kernel modules, so we have a helper script that loads everything that could possibly be needed before the daemon starts. The problem is that sometimes loading a crypto module is bad. Sometimes there is something new we don't load.

A new recent issue is with containers, we cannot even run our helper to load kernel modules anymore.

the kernel should really be able to auto load all crypto and IPsec related modules on demand.

also, how to detect af_key is the old PF_KEY api is removed? we detect af_key using /proc/net/pfkey as /proc/net/xfrm_stat is not available everywhere

VTI devices do not work for /32 routes

For Opportunistc IPsec, we want all unauthenticated (anonymous) IPsec to go into a single VTI device, so iptables rules that are installed the machine's plaintext can be applied to the VTI device as well. Currently (last tested early 4.x) only VTI policies with networks, not single hosts, worked properly.

RFC 8229 ESPinTCP support

AFAIK, Herbert Xu is working on this

RFC 8229 ESPinTLS support

AFAIK, no one is working on this. Could it use ktls ?

FIPS mode private key censoring

In FIPS mode, or actually maybe in normal mode too, it would be good if the private keys are not visible in "ip xfrm state" unless there is some debug flag passed/set into the kernel. The debug flag would not be settable in FIPS mode.

Proposal: per default, show a CENSORED string instead of the actual private key

named sockets

getsockopt() and setsockopt() for named sockets. This way a client that gets a DNS name, and resolves it to an IP, can set the socket name. This can then be send along with the ACQUIRE so that the IKE daemon can use that to possibly pull up some authentication mechanism based on this FQDN (eg from DNS, confirm the CERT/ID payload matches this)

Also, this prevents an attack where someone who controls routing (coffeeshop, hotel, etc) puts up an evil hostname to steal IP. Eg evil.nohats.ca IN A 8.8.8.8 with an IPSECKEY XXXX. Then the client talks to the rogue server and sends all encrypted 8.8.8.8 to the attacker that can decrypt it.

The idea is, if you mark a connection to expect it to be "dns.google.com", then if it comes with the mark "evil.nohats.ca" you know there is a MITM.

named IPsec policies

We would like to be able to "get" the name of an IPsec policy from userland. Imagine this would be available via netlink or socket options. This allows us to trigger an Opportunistic IPsec connection based on IP address, then authenticate this via IKE, which gives us an identity (eg FQDN) which we can then "add" to the SPD entry. Userland can then confirm that a connect() or bind() it did based on a DNS name got a policy matching the identity.

a New 'Encryption Required' socket option

We'd like to be able to open a connection only if it is protected by IPsec encryption. If the encryption is terminated, we want the kernel to close the connection with some error.

currently, userland can ask the kernel about the encryption status, and refuse to send data if there is no matching encryption, but obviously this is comletely unsafe as the userland will not be notified if the encryption status changes.

Client Address Translation

For anonymous IPsec, where the client authenticates the server, but the server does not authenticate the client, we have an IP clashing problem of clients behind NAT. Either the server hands out addresses, in which case a client connecting to multiple servers can get the same address resulting in a conflict. Or the client can somehow use its internal address, in which case the server has a clashing address problem with multiple clients behind different NAT routers using the same pre-NAT IP.

We have a solution that currently installs an additional IPsec SA policy using a special IP address. Then we tie that in with iptables rules. This has the extra advantage that the system never needs to configure the IP addresses given my the server side.

This works but it might clash with the administrator's firewall changes. There is no reason this could not just be contained within the IPsec subsystem, and this "CAT" IP address is passed to the kernel to install the additional rule (and reverse rule so the iptables rule is not needed)

We have a draft written for this which we will share soon)

XFRM_MIGRATE support in ip xfrm monitor

Last I checked, ip xfrm monitor still threw errors when seeing XFRM_MIGRATE messages