Keyboard smashing can be made quite unpredictable if implemented correctly. The randomness should come from time difference between keystrokes, not from the key values.
I have used the Unix signal(3), using SIGINFO, in conjunction with a tight CPU loop to generate random numbers for e.g. WiFi passwords. The technique is something like this:
volatile int signal_info = 0;
void sighdlr(int sig)
{ signal_info = 1; }
int main(...) {...
unsigned count = 0;
signal(SIGINFO, sighdlr);
...
for(;;) {
count++;
if (signal_info) {
signal_info = 0;
...
}
}
Before running this program I use
stty -kerninfo
to avoid seeing CPU statistics.Then every time I press ^T I display a new random number displayed based on the LSBs of the count. E.g. to simulate a die roll I print count%6.
This technique is good for not just a few bits of randomness per ^T, it is good for over 20 bits. That's because when I examine the timing of the inner loop generated I find it taking about 2 or 3 CPU cycles per increment (depending on compiler). So on a 2 GHz machine the counter is incremented at a 1 GHz rate. That's about 30 binary bits, the least significant of which are the most unpredictable. Assuming you press ^T at about a 3 Hz rate, you're probably getting over 20 bits (more or less) of randomness each time you do so.
The above only works if your system supports a decent SIGINFO setup (works fine on OpenBSD and OS X, I don't know about Linux). It probably has serious problems in a virtual machine, I use a real keyboard that has its own independently clocked CPU.