Revisiting "too small to fail"

It sounds like we need a fault injection test framework that takes ideas from fuzzers like AFL and continues to try variants until it exercises all of the code paths.

Even fancier, if the test framework could examine branch conditions and work backward to create the conditions to exercise untested branches. Or if it can't work it out, log it for developer attention so a custom test rule can be created.

Although, it'd also need a lot of verification code written. Sure, the memory allocation failure during bad block handling during a btrfs scrub, at the same time the SAS link got hot-removed got handled, but is the filesystem still correct? Someone has to write that.

Or possibly a simpler idea, take one of the existing automatic unit test generators and modify it for kernel code, although there's always the problem of knowing if the hardware is being simulated correctly.