Unit testing should be the primary method of testing the bulk of code written, because a unit test can be written once and run many times — manual tests have to be planned once and then manually run each time.
Development of unit tests starts with the architecture and API design of the code to be tested: code should be designed to be easily testable, or will potentially be very difficult to test.
- Write unit tests to be as small as possible, but no smaller.
- Use code coverage tools to write tests to get high code coverage.
- Use installed tests to test integration.
- Run all unit tests under Valgrind to check for leaks and other problems.
- Use appropriate tools to automatically generate unit tests where possible.
- Design code to be testable from the beginning.
Writing unit tests
Unit tests should be written in conjunction with looking at code coverage information gained from running the tests. This typically means writing an initial set of unit tests, running them to get coverage data, then reworking and expanding them to increase the code coverage levels. Coverage should be increased first by ensuring all functions are covered (at least in part), and then by ensuring all lines of code are covered. By covering functions first, API problems which will prevent effective testing can be found quickly. These typically manifest as internal functions which cannot easily be called from unit tests. Overall, coverage levels of over 90% should be aimed for; don’t just test cases covered by project requirements, test everything.
Like git commits, each unit test should be ‘as small as possible, but no smaller’, testing a single specific API or behaviour. Each test case must be able to be run individually, without depending on state from other test cases. This is important to allow debugging of a single failing test, without having to step through all the other test code as well. It also means that a single test failure can easily be traced back to a specific API, rather than a generic “unit tests failed somewhere” message.
GLib has support for unit testing with its
allowing tests to be arranged in groups and hierarchies. This means that groups
of related tests can be run together for enhanced debugging too, by running the
test binary with the
./test-suite-name -p /path/to/test/group
All unit tests should be installed system-wide, following the installed-tests standard.
By installing the unit tests, continuous integration (CI) is made easier, since tests for one project can be re-run after changes to other projects in the CI environment, thus testing the interfaces between modules. That is useful for a highly-coupled set of projects like Apertis.
To add support for installed-tests, add the following to
Once unit tests with high code coverage have been written, they can be run under various dynamic analysis tools, such as Valgrind to check for leaks, threading errors, allocation problems, etc. across the entire code base. The higher the code coverage of the unit tests, the more confidence the Valgrind results can be treated with. See the Tooling page for more information, including build system integration instructions.
Critically, this means that unit tests should not leak memory or other resources themselves, and similarly should not have any threading problems. Any such problems would effectively be false positives in the analysis of the actual project code. (False positives which need to be fixed by fixing the unit tests.)
Certain types of code are quite repetitive, and require a lot of unit tests to gain good coverage; but are appropriate for test data generation, where a tool is used to automatically generate test vectors for the code. This can drastically reduce the time needed for writing unit tests, for code in these specific domains.
One example of a domain amenable to test data generation is parsing, where the data to be parsed is required to follow a strict schema — this is the case for XML and JSON documents. For JSON, a tool such as Walbottle can be used to generate test vectors for all types of valid and invalid input according to the schema.
Every type of JSON document should have a JSON Schema defined for it, which can then be passed to Walbottle to generate test vectors:
These test vectors can then be passed to the code under test in its unit tests.
The JSON instances generated by
--valid-only should be accepted; those from
--invalid-only should be rejected.
Writing testable code
Code should be written with testability in mind from the design stage, as it affects API design and architecture in fundamental ways. A few key principles:
- Do not use global state. Singleton objects are usually a bad idea as they can’t be instantiated separately or controlled in the unit tests.
- Separate out use of external state, such as databases, networking, or the file system. The unit tests can then replace the accesses to external state with mocked objects. A common approach to this is to use dependency injection to pass a file system wrapper object to the code under test. For example, a class should not load a global database (from a fixed location in the file system) because the unit tests would then potentially overwrite the running system’s copy of the database, and could never be executed in parallel. They should be passed an object which provides an interface to the database: in a production system, this would be a thin wrapper around the database API; for testing, it would be a mock object which checks the requests given to it and returns hard-coded responses for various tests.
- Expose utility functions where they might be generally useful.
- Split projects up into collections of small, private libraries which are then linked together with a minimal amount of glue code into the overall executable. Each can be tested separately.
The topic of software testability is covered in the following articles: