glib/docs/testing.md

Testing policy

Aims

• Maintainers should be able to make a release of GLib at any time, confident that it will not contain regressions or obvious bugs with new functionality
• Speed up review of submitted changes by deferring some of the review effort to automated testing
• Allow fast detection of bugs in new or changed code, particularly if they are only present on platforms not regularly used by the maintainers
• Allow easy dynamic and static analysis of a significant proportion of the GLib code
• Statistics on tests (such as pass/failure) should be easily and mechanically collectable to allow analysis and highlight problems
• Code for tests and code for production should be easily separable so that statistics on them can be grouped separately
• Performance measurement tools for GLib should be reusable over time to allow comparable measurements to be collected and to discourage use of lower quality and throwaway tests when prototyping improvements to GLib

Policy

• Tests must be written for all new code, and any existing code which is being non-trivially modified (for example to fix a bug), to give confidence to the author and reviewer of the changes that they are correct for all platforms that GLib runs CI on.
• Tests live in the {glib,gobject,gio}/tests directories. This allows their code to be counted separately when analysing statistics such as code coverage.
  • Performance tests live in {glib,gobject,gio}/tests/performance, as they are executed and results interpreted differently due to giving a result on a continuous scale rather than a pass/fail result.
• All tests must use the GTest framework, as it supports structured output which exposes test results to the test runner for analysis.
  • Use g_test_bug() and g_test_summary() in unit tests to link them to contextual information in bug reports, and to provide a summary of what each test checks and how it goes about doing those checks. Sometimes a test’s behaviour can be quite complex, and needs to be explained so that future developers can understand and build on such tests in future.
  • Use the g_assert_*() functions inside unit tests, and do not use g_assert(). The latter is compiled out when GLib is built with G_DISABLE_ASSERT, and the former are not. The g_assert_*() functions also give more helpful error messages on test failure.
• Performance tests must be able to be run unattended. In this mode they must choose default argument values which check that the performance test functions (i.e. without crashing) and doesn’t take too long to complete. This is used to automatically verify that performance tests still work, as they are typically used infrequently and are subject to bitrot.
• Code coverage reports must be used to demonstrate that unit tests reach all newly submitted or significantly modified code, reaching all lines of code and a significant majority of branches. If this is not enforced, code ends up never being tested.
• Code should be structured to be testable, which is typically only possible by writing tests at the same time as the code. Otherwise it is easy to design APIs which cannot easily be unit tested, and once those APIs are stable it is hard to retrofit tests to them.
• Parsers, network-facing code or code which handles untrusted user input must have fuzz tests added, in the fuzzing directory. These are run by oss-fuzz and are very effective at catching exploitable security issues. See the fuzzing README for more details.
• When fixing bugs in existing code, regression tests must be added when it is straightforward to do so. If it’s difficult to do so (such as if the code needs to be significantly restructured or APIs need to be changed), adding the regression tests can be deferred to a follow-up issue so as not to slow down bug fixing. In that case, the bug fix must be carefully manually tested before being merged.