Dan Kaminsky pointed out that using partial MD4 and using that to generate a sequence number, of which only 24-bits are truly unguessable, seriously undermine the goals of random sequence number generation.

In particular, with only 24-bits being truly unguessable, packet injection into a session using even something like brute force is a real potential possibility.

We only use 24-bits because we regenerate the random number every 5 minutes "just in case."  But what does is trade a "we don't know" kind of theoretical issue for a provably real one (brute force attack).

Therefore [Dave Miller] moving us more in line with RFC1948 (as well as OpenBSD and Solaris), to use MD5 and a full 32-bit result in the generated sequence number.

MD5 was selected as a compromise between performance loss and theoretical ability to be compromised.  Willy Tarreau did extensive testing and SHA1 was found to harm performance too much to be considered seriously at this time.

We may later add a sysctl for various modes (ie. a "super secure" mode that uses SHA1 if people want that, and an "insecure" mode that doesn't use cryptographic hashing at all for people in protected environments where that might be safe to do).

[Dave Miller] also moved the sequence number generators out of random.c (they never really belonged there, and are only there due to historical artifacts), and fixed a bug in DCCP sequence number generation (on ipv6 the 43-bit sequence number was truncated to 32-bits).

Acknowledgements:

Red Hat would like to thank Dan Kaminsky for reporting this issue.