KRF is a Kernelspace Randomized Faulter.
It currently supports the Linux and FreeBSD kernels.

What?
Fault injection is a software testing technique that involves inducing failures ("faults") in the functions called by a program. If the callee has failed to perform proper error checking and handling, these faults can result in unreliable application behavior or exploitable vulnerabilities.

Unlike the many userspace fault injection systems out there, KRF runs in kernelspace via a loaded module. This has several advantages:

• It works on static binaries, as it does not rely on LD_PRELOAD for injection.
• Because it intercepts raw syscalls and not their libc wrappers, it can inject faults into calls made by syscall(3) or inline assembly.
• It's probably faster and less error-prone than futzing with dlsym.

There are also several disadvantages:

• You'll probably need to build it yourself.
• It probably only works on x86(_64), since it twiddles cr0 manually. There is probably an architecture-independent way to do that in Linux, somewhere.
• It's essentially a rootkit. You should definitely never, ever run it on a non-testing system.
• It probably doesn't cover everything that the Linux kernel expects of syscalls, and may destabilize its host in weird and difficult to reproduce ways.

How does it work?
KRF rewrites the Linux or FreeBSD system call table: when configured via krfctl, KRF replaces faultable syscalls with thin wrappers.
Each wrapper then performs a check to see whether the call should be faulted using a configurable targeting system capable of targeting a specific personality(2), PID, UID, and/or GID. If the process shouldn't be faulted, the original syscall is invoked.
Finally, the targeted call is faulted via a random failure function. For example, a read(2) call might receive one of EBADF, EINTR, EIO, and so on.

Setup

Compatibility
NOTE: If you have Vagrant, just use the Vagrantfile and jump to the build steps.
KRF should work on any recent-ish (4.15+) Linux kernel with CONFIG_KALLSYMS=1.
This includes the default kernel on Ubuntu 18.04 and probably many other recent distros.

Dependencies
NOTE: Ignore this if you're using Vagrant.
Apart from a C toolchain (GCC is probably necessary for Linux), KRF's only dependencies should be libelf, the kernel headers, and Ruby (for code generation).
GNU Make is required on all platforms; FreeBSD additionally requires BSD Make.
For systems with apt:

Building
or, if you're using Vagrant:
or, for FreeBSD:

Usage
KRF has three components:

• A kernel module (krfx)
• An execution utility (krfexec)
• A control utility (krfctl)

To load the kernel module, run make insmod. To unload it, run make rmmod.
KRF begins in a neutral state: no syscalls will be intercepted or faulted until the user specifies some behavior via krfctl:
On FreeBSD, krfexec requires root privileges. By default, it will attempt to use SUDO_UID and the username returned by getlogin_r to return to a non-root user before executing the target. To force a particular UID, export REAL_UID, e.g.:

Configuration
NOTE: Most users should use krfctl instead of manipulating these files by hand. In FreeBSD, these same values are accessible through sysctl krf.whatever instead of procfs.

/proc/krf/rng_state
This file allows a user to read and modify the internal state of KRF's PRNG.
For example, each of the following will correctly update the state:
The state is a 32-bit unsigned integer; attempting to change it beyond that will fail.

/proc/krf/targeting
This file allows a user set the values used by KRF for syscall targeting.
NOTE: KRF uses a default personality not currently used by the Linux kernel by default. If you change this, you should be careful to avoid making it something that Linux cares about. man 2 personality has the details.
A personality of 28 is hardcoded into krfexec.

/proc/krf/probability
This file allows a user to read and write the probability of inducing fault for a given (faultable) syscall.
The probability is represented as a reciprocal, e.g. 1000 means that, on average, 0.1% of faultable syscalls will be faulted.

/proc/krf/control
This file controls the syscalls that KRF faults.
NOTE: Most users should use krfctl instead of interacting with this file directly — the former will perform syscall name-to-number translation automatically and will provide clearer error messages when things go wrong.
Passing any number greater than KRF_NR_SYSCALLS will cause KRF to flush the entire syscall table, returning it to the neutral state. Since KRF_NR_SYSCALLS isn't necessarily predictable for arbitrary versions of the Linux kernel, choosing a large number (like 65535) is fine.
Passing a valid syscall number that lacks a fault injection wrapper will cause the write(2) to the file to fail with EOPNOTSUPP.

/proc/krf/log_faults
This file controls whether or not KRF emits kernel logs on faulty syscalls. By default, no logging messages are emitted.
NOTE: Most users should use krfctl instead of interacting with this file directly.