| 1 | |
| 2 |
|
| 3 | ====================
|
| 4 | Kernel Testing Guide
|
| 5 | ====================
|
| 6 |
|
| 7 |
|
| 8 | There are a number of different tools for testing the Linux kernel, so knowing
|
| 9 | when to use each of them can be a challenge. This document provides a rough
|
| 10 | overview of their differences, and how they fit together.
|
| 11 |
|
| 12 |
|
| 13 | Writing and Running Tests
|
| 14 | =========================
|
| 15 |
|
| 16 | The bulk of kernel tests are written using either the kselftest or KUnit
|
| 17 | frameworks. These both provide infrastructure to help make running tests and
|
| 18 | groups of tests easier, as well as providing helpers to aid in writing new
|
| 19 | tests.
|
| 20 |
|
| 21 | If you're looking to verify the behaviour of the Kernel — particularly specific
|
| 22 | parts of the kernel — then you'll want to use KUnit or kselftest.
|
| 23 |
|
| 24 |
|
| 25 | The Difference Between KUnit and kselftest
|
| 26 | ------------------------------------------
|
| 27 |
|
| 28 | KUnit (Documentation/dev-tools/kunit/index.rst) is an entirely in-kernel system
|
| 29 | for "white box" testing: because test code is part of the kernel, it can access
|
| 30 | internal structures and functions which aren't exposed to userspace.
|
| 31 |
|
| 32 | KUnit tests therefore are best written against small, self-contained parts
|
| 33 | of the kernel, which can be tested in isolation. This aligns well with the
|
| 34 | concept of 'unit' testing.
|
| 35 |
|
| 36 | For example, a KUnit test might test an individual kernel function (or even a
|
| 37 | single codepath through a function, such as an error handling case), rather
|
| 38 | than a feature as a whole.
|
| 39 |
|
| 40 | This also makes KUnit tests very fast to build and run, allowing them to be
|
| 41 | run frequently as part of the development process.
|
| 42 |
|
| 43 | There is a KUnit test style guide which may give further pointers in
|
| 44 | Documentation/dev-tools/kunit/style.rst
|
| 45 |
|
| 46 |
|
| 47 | kselftest (Documentation/dev-tools/kselftest.rst), on the other hand, is
|
| 48 | largely implemented in userspace, and tests are normal userspace scripts or
|
| 49 | programs.
|
| 50 |
|
| 51 | This makes it easier to write more complicated tests, or tests which need to
|
| 52 | manipulate the overall system state more (e.g., spawning processes, etc.).
|
| 53 | However, it's not possible to call kernel functions directly from kselftest.
|
| 54 | This means that only kernel functionality which is exposed to userspace somehow
|
| 55 | (e.g. by a syscall, device, filesystem, etc.) can be tested with kselftest. To
|
| 56 | work around this, some tests include a companion kernel module which exposes
|
| 57 | more information or functionality. If a test runs mostly or entirely within the
|
| 58 | kernel, however, KUnit may be the more appropriate tool.
|
| 59 |
|
| 60 | kselftest is therefore suited well to tests of whole features, as these will
|
| 61 | expose an interface to userspace, which can be tested, but not implementation
|
| 62 | details. This aligns well with 'system' or 'end-to-end' testing.
|
| 63 |
|
| 64 | For example, all new system calls should be accompanied by kselftest tests.
|
| 65 |
|
| 66 | Code Coverage Tools
|
| 67 | ===================
|
| 68 |
|
| 69 | The Linux Kernel supports two different code coverage measurement tools. These
|
| 70 | can be used to verify that a test is executing particular functions or lines
|
| 71 | of code. This is useful for determining how much of the kernel is being tested,
|
| 72 | and for finding corner-cases which are not covered by the appropriate test.
|
| 73 |
|
| 74 | Documentation/dev-tools/gcov.rst is GCC's coverage testing tool, which can be
|
| 75 | used with the kernel to get global or per-module coverage. Unlike KCOV, it
|
| 76 | does not record per-task coverage. Coverage data can be read from debugfs,
|
| 77 | and interpreted using the usual gcov tooling.
|
| 78 |
|
| 79 | Documentation/dev-tools/kcov.rst is a feature which can be built in to the
|
| 80 | kernel to allow capturing coverage on a per-task level. It's therefore useful
|
| 81 | for fuzzing and other situations where information about code executed during,
|
| 82 | for example, a single syscall is useful.
|
| 83 |
|
| 84 |
|
| 85 | Dynamic Analysis Tools
|
| 86 | ======================
|
| 87 |
|
| 88 | The kernel also supports a number of dynamic analysis tools, which attempt to
|
| 89 | detect classes of issues when they occur in a running kernel. These typically
|
| 90 | each look for a different class of bugs, such as invalid memory accesses,
|
| 91 | concurrency issues such as data races, or other undefined behaviour like
|
| 92 | integer overflows.
|
| 93 |
|
| 94 | Some of these tools are listed below:
|
| 95 |
|
| 96 | * kmemleak detects possible memory leaks. See
|
| 97 | Documentation/dev-tools/kmemleak.rst
|
| 98 | * KASAN detects invalid memory accesses such as out-of-bounds and
|
| 99 | use-after-free errors. See Documentation/dev-tools/kasan.rst
|
| 100 | * UBSAN detects behaviour that is undefined by the C standard, like integer
|
| 101 | overflows. See Documentation/dev-tools/ubsan.rst
|
| 102 | * KCSAN detects data races. See Documentation/dev-tools/kcsan.rst
|
| 103 | * KFENCE is a low-overhead detector of memory issues, which is much faster than
|
| 104 | KASAN and can be used in production. See Documentation/dev-tools/kfence.rst
|
| 105 | * lockdep is a locking correctness validator. See
|
| 106 | Documentation/locking/lockdep-design.rst
|
| 107 | * Runtime Verification (RV) supports checking specific behaviours for a given
|
| 108 | subsystem. See Documentation/trace/rv/runtime-verification.rst
|
| 109 | * There are several other pieces of debug instrumentation in the kernel, many
|
| 110 | of which can be found in lib/Kconfig.debug
|
| 111 |
|
| 112 | These tools tend to test the kernel as a whole, and do not "pass" like
|
| 113 | kselftest or KUnit tests. They can be combined with KUnit or kselftest by
|
| 114 | running tests on a kernel with these tools enabled: you can then be sure
|
| 115 | that none of these errors are occurring during the test.
|
| 116 |
|
| 117 | Some of these tools integrate with KUnit or kselftest and will
|
| 118 | automatically fail tests if an issue is detected.
|
| 119 |
|
| 120 | Static Analysis Tools
|
| 121 | =====================
|
| 122 |
|
| 123 | In addition to testing a running kernel, one can also analyze kernel source code
|
| 124 | directly (**at compile time**) using **static analysis** tools. The tools
|
| 125 | commonly used in the kernel allow one to inspect the whole source tree or just
|
| 126 | specific files within it. They make it easier to detect and fix problems during
|
| 127 | the development process.
|
| 128 |
|
| 129 | Sparse can help test the kernel by performing type-checking, lock checking,
|
| 130 | value range checking, in addition to reporting various errors and warnings while
|
| 131 | examining the code. See the Documentation/dev-tools/sparse.rst documentation
|
| 132 | page for details on how to use it.
|
| 133 |
|
| 134 | Smatch extends Sparse and provides additional checks for programming logic
|
| 135 | mistakes such as missing breaks in switch statements, unused return values on
|
| 136 | error checking, forgetting to set an error code in the return of an error path,
|
| 137 | etc. Smatch also has tests against more serious issues such as integer
|
| 138 | overflows, null pointer dereferences, and memory leaks. See the project page at
|
| 139 | http://smatch.sourceforge.net/.
|
| 140 |
|
| 141 | Coccinelle is another static analyzer at our disposal. Coccinelle is often used
|
| 142 | to aid refactoring and collateral evolution of source code, but it can also help
|
| 143 | to avoid certain bugs that occur in common code patterns. The types of tests
|
| 144 | available include API tests, tests for correct usage of kernel iterators, checks
|
| 145 | for the soundness of free operations, analysis of locking behavior, and further
|
| 146 | tests known to help keep consistent kernel usage. See the
|
| 147 | Documentation/dev-tools/coccinelle.rst documentation page for details.
|
| 148 |
|
| 149 | Beware, though, that static analysis tools suffer from **false positives**.
|
| 150 | Errors and warns need to be evaluated carefully before attempting to fix them.
|
| 151 |
|
| 152 | When to use Sparse and Smatch
|
| 153 | -----------------------------
|
| 154 |
|
| 155 | Sparse does type checking, such as verifying that annotated variables do not
|
| 156 | cause endianness bugs, detecting places that use ``__user`` pointers improperly,
|
| 157 | and analyzing the compatibility of symbol initializers.
|
| 158 |
|
| 159 | Smatch does flow analysis and, if allowed to build the function database, it
|
| 160 | also does cross function analysis. Smatch tries to answer questions like where
|
| 161 | is this buffer allocated? How big is it? Can this index be controlled by the
|
| 162 | user? Is this variable larger than that variable?
|
| 163 |
|
| 164 | It's generally easier to write checks in Smatch than it is to write checks in
|
| 165 | Sparse. Nevertheless, there are some overlaps between Sparse and Smatch checks.
|
| 166 |
|
| 167 | Strong points of Smatch and Coccinelle
|
| 168 | --------------------------------------
|
| 169 |
|
| 170 | Coccinelle is probably the easiest for writing checks. It works before the
|
| 171 | pre-processor so it's easier to check for bugs in macros using Coccinelle.
|
| 172 | Coccinelle also creates patches for you, which no other tool does.
|
| 173 |
|
| 174 | For example, with Coccinelle you can do a mass conversion from
|
| 175 | ``kmalloc(x * size, GFP_KERNEL)`` to ``kmalloc_array(x, size, GFP_KERNEL)``, and
|
| 176 | that's really useful. If you just created a Smatch warning and try to push the
|
| 177 | work of converting on to the maintainers they would be annoyed. You'd have to
|
| 178 | argue about each warning if can really overflow or not.
|
| 179 |
|
| 180 | Coccinelle does no analysis of variable values, which is the strong point of
|
| 181 | Smatch. On the other hand, Coccinelle allows you to do simple things in a simple
|
| 182 | way.
|