Awesome

KRF

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.

You can read more about KRF's implementation in our blog post.

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 (>=2.4, for code generation).

GNU Make is required on all platforms; FreeBSD additionally requires BSD Make.

For systems with apt:

sudo apt install gcc make libelf-dev ruby linux-headers-$(uname -r)

Building

git clone https://github.com/trailofbits/krf && cd krf
make -j$(nproc)
sudo make install # Installs module to /lib/modules and utils to /usr/local/bin
sudo make insmod # Loads module

or, if you're using Vagrant:

git clone https://github.com/trailofbits/krf && cd krf
vagrant up linux && vagrant ssh linux
# inside the VM
cd /vagrant
make -j$(nproc)
sudo make install # Installs module to /lib/modules and utils to /usr/local/bin
sudo make insmod # Loads module

or, for FreeBSD:

git clone https://github.com/trailofbits/krf && cd krf
cd vagrant up freebsd && vagrant ssh freebsd
# inside the VM
cd /vagrant
gmake # NOT make!
gmake install-module # Installs module to /boot/modules/
sudo gmake install-utils # Installs utils to /usr/local/bin
gmake insmod # Loads module

Usage

KRF has three components:

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

To load the kernel module, run make insmod. To unload it, run make rmmod.

For first time use it might be useful to launch sudo krfmesg on a separate terminal to see messages logged from krfx.

KRF begins in a neutral state: no syscalls will be intercepted or faulted until the user specifies some behavior via krfctl:

# no induced faults, even with KRF loaded
ls

# tell krf to fault read(2) and write(2) calls
# note that krfctl requires root privileges
sudo krfctl -F 'read,write'

# tell krf to fault any program started by
# krfexec, meaning a personality of 28
sudo krfctl -T personality=28

# may fault!
krfexec ls

# tell krf to fault with a 1/100 (or 1%) probability
# note that this value is represented as a reciprocal
# so e.g. 1 means all faultable syscalls will fault
# and 500 means that on average every 500 syscalls will fault (1/500 or 0.2%)
sudo krfctl -p 100

# tell krf to fault `io` profile (and so i/o related syscalls)
sudo krfctl -P io

# krfexec will pass options correctly as well
krfexec echo -n 'no newline'

# clear the fault specification
sudo krfctl -c

# clear the targeting specification
sudo krfctl -C

# no induced faults, since no syscalls are being faulted
krfexec firefox

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:

echo "1234" | sudo tee /proc/krf/rng_state
echo "0777" | sudo tee /proc/krf/rng_state
echo "0xFF" | sudo tee /proc/krf/rng_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.

echo "0 28" | sudo tee /proc/krf/targeting

A personality of 28 is hardcoded into krfexec, and must be set in order for things executed by krfexec to be faulted.

`/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.

echo "100000" | sudo tee /proc/krf/probability

`/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.

# replace the syscall in slot 0 (usually SYS_read) with its faulty wrapper
echo "0" | sudo tee /proc/krf/control

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.

# enable fault logging
echo "1" | sudo tee /proc/krf/log_faults
# disable fault logging
echo "0" | sudo tee /proc/krf/log_faults
# read the logging state
cat /proc/krf/log_faults

TODO

Allow users to specify a particular class of faults, e.g. memory pressure (ENOMEM).
- This should be do-able by adding some more bits to the personality(2) value.

Thanks

Many thanks go to Andrew Reiter for the initial port of KRF to FreeBSD. Andrew's work was performed on behalf of the Applied Research Group at Veracode.

Licensing

KRF is licensed under the terms of the GNU GPLv3.

See the LICENSE file for the exact terms.