The note about instruction timings being random rings true to me. A while back I created a small kernel driver that created random bits by taking the last few bits of a high-resolution timer whenever the timer interrupt went off.
I tried some of the standard toolkits for measuring randomness and they all concluded the data was totally random. The basic conclusion was that you could easily generate hundreds of random bits per second without too much effort. This would be useful for machines currently suffering from loss of randomness because they're headless.
It also seems a much better option than extracting randomness from network traffic, since while an attacker might be able to affect the network card, there's no way they're going to be able to affect the system timer.