A kernel is a hairy beast, and sometimes changes cause or reveal bugs
in other portions of the code. User-level testing programs --
programs that stress-test features based on their usage in the "real
world" -- for each little slice of the code would, I'm sure, be an
immense help to future kernel developers everywhere.
I have a couple stress testing programs written for my tmpfs project
that I simply cannot live without, and I know I have found at least
two hard-to-spot bugs because of the tests. Adding regression tests
would be a great diagnostic tool, an opportunity perhaps to move some
kernel debug code into separate test programs, etc...
Jeff