Kernel Testing Guide

There are a number of different tools for testing the Linux kernel, so knowing when to use each of them can be a challenge. This document provides a rough overview of their differences, and how they fit together.

Writing and Running Tests

The bulk of kernel tests are written using either the kselftest or KUnit frameworks. These both provide infrastructure to help make running tests and groups of tests easier, as well as providing helpers to aid in writing new tests.

If you're looking to verify the behaviour of the Kernel — particularly specific parts of the kernel — then you'll want to use KUnit or kselftest.

The Difference Between KUnit and kselftest

KUnit (KUnit - Linux Kernel Unit Testing) is an entirely in-kernel system for "white box" testing: because test code is part of the kernel, it can access internal structures and functions which aren't exposed to userspace.

KUnit tests therefore are best written against small, self-contained parts of the kernel, which can be tested in isolation. This aligns well with the concept of 'unit' testing.

For example, a KUnit test might test an individual kernel function (or even a single codepath through a function, such as an error handling case), rather than a feature as a whole.

This also makes KUnit tests very fast to build and run, allowing them to be run frequently as part of the development process.

There is a KUnit test style guide which may give further pointers in Test Style and Nomenclature

kselftest (Linux Kernel Selftests), on the other hand, is largely implemented in userspace, and tests are normal userspace scripts or programs.

This makes it easier to write more complicated tests, or tests which need to manipulate the overall system state more (e.g., spawning processes, etc.). However, it's not possible to call kernel functions directly from kselftest. This means that only kernel functionality which is exposed to userspace somehow (e.g. by a syscall, device, filesystem, etc.) can be tested with kselftest. To work around this, some tests include a companion kernel module which exposes more information or functionality. If a test runs mostly or entirely within the kernel, however, KUnit may be the more appropriate tool.

kselftest is therefore suited well to tests of whole features, as these will expose an interface to userspace, which can be tested, but not implementation details. This aligns well with 'system' or 'end-to-end' testing.

For example, all new system calls should be accompanied by kselftest tests.

Code Coverage Tools

The Linux Kernel supports two different code coverage measurement tools. These can be used to verify that a test is executing particular functions or lines of code. This is useful for determining how much of the kernel is being tested, and for finding corner-cases which are not covered by the appropriate test.

Using gcov with the Linux kernel is GCC's coverage testing tool, which can be used with the kernel to get global or per-module coverage. Unlike KCOV, it does not record per-task coverage. Coverage data can be read from debugfs, and interpreted using the usual gcov tooling.

KCOV: code coverage for fuzzing is a feature which can be built in to the kernel to allow capturing coverage on a per-task level. It's therefore useful for fuzzing and other situations where information about code executed during, for example, a single syscall is useful.

Dynamic Analysis Tools

The kernel also supports a number of dynamic analysis tools, which attempt to detect classes of issues when they occur in a running kernel. These typically each look for a different class of bugs, such as invalid memory accesses, concurrency issues such as data races, or other undefined behaviour like integer overflows.

Some of these tools are listed below:

These tools tend to test the kernel as a whole, and do not "pass" like kselftest or KUnit tests. They can be combined with KUnit or kselftest by running tests on a kernel with these tools enabled: you can then be sure that none of these errors are occurring during the test.

Some of these tools integrate with KUnit or kselftest and will automatically fail tests if an issue is detected.

Static Analysis Tools

In addition to testing a running kernel, one can also analyze kernel source code directly (at compile time) using static analysis tools. The tools commonly used in the kernel allow one to inspect the whole source tree or just specific files within it. They make it easier to detect and fix problems during the development process.

Sparse can help test the kernel by performing type-checking, lock checking, value range checking, in addition to reporting various errors and warnings while examining the code. See the Sparse documentation page for details on how to use it.

Smatch extends Sparse and provides additional checks for programming logic mistakes such as missing breaks in switch statements, unused return values on error checking, forgetting to set an error code in the return of an error path, etc. Smatch also has tests against more serious issues such as integer overflows, null pointer dereferences, and memory leaks. See the project page at http://smatch.sourceforge.net/.

Coccinelle is another static analyzer at our disposal. Coccinelle is often used to aid refactoring and collateral evolution of source code, but it can also help to avoid certain bugs that occur in common code patterns. The types of tests available include API tests, tests for correct usage of kernel iterators, checks for the soundness of free operations, analysis of locking behavior, and further tests known to help keep consistent kernel usage. See the Coccinelle documentation page for details.

Beware, though, that static analysis tools suffer from false positives. Errors and warns need to be evaluated carefully before attempting to fix them.

When to use Sparse and Smatch

Sparse does type checking, such as verifying that annotated variables do not cause endianness bugs, detecting places that use __user pointers improperly, and analyzing the compatibility of symbol initializers.

Smatch does flow analysis and, if allowed to build the function database, it also does cross function analysis. Smatch tries to answer questions like where is this buffer allocated? How big is it? Can this index be controlled by the user? Is this variable larger than that variable?

It's generally easier to write checks in Smatch than it is to write checks in Sparse. Nevertheless, there are some overlaps between Sparse and Smatch checks.

Strong points of Smatch and Coccinelle

Coccinelle is probably the easiest for writing checks. It works before the pre-processor so it's easier to check for bugs in macros using Coccinelle. Coccinelle also creates patches for you, which no other tool does.

For example, with Coccinelle you can do a mass conversion from kmalloc(x * size, GFP_KERNEL) to kmalloc_array(x, size, GFP_KERNEL), and that's really useful. If you just created a Smatch warning and try to push the work of converting on to the maintainers they would be annoyed. You'd have to argue about each warning if can really overflow or not.

Coccinelle does no analysis of variable values, which is the strong point of Smatch. On the other hand, Coccinelle allows you to do simple things in a simple way.