[glib: 1/2] docs: Add a testing policy

[Philip Withnall <pwithnall src gnome org>]

commit 7674fd1504d2156ceb9d3ca18f1432725ca0bb8c
Author: Philip Withnall <pwithnall endlessos org>
Date:   Tue Jun 14 12:55:15 2022 +0100

    docs: Add a testing policy
    
    This documents our existing testing best practices, and tries to explain
    some of the rationale around them.
    
    There’s some new policy around performance tests, specifying that they
    are good to keep around if they are high quality (so they can be used
    again, the effort put into them not thrown away, and the numbers
    compared over time) and that they should be runnable automatically to
    avoid bitrot.
    
    Everything else just documents existing practice.
    
    Signed-off-by: Philip Withnall <pwithnall endlessos org>

 CONTRIBUTING.md |  3 ++-
 docs/testing.md | 71 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 73 insertions(+), 1 deletion(-)
---
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index ba8039d4c3..9f2cdc9842 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -95,7 +95,8 @@ they are imported into GLib and marked as stable.
 
 Each feature should also come fully documented, and with tests which approach
 full branch coverage of the new code. GLib’s CI system generates code coverage
-reports which are viewable for each merge request.
+reports which are viewable for each merge request. See
+[the testing policy](./docs/testing.md) for more details.
 
 If proposing a large feature or change, it’s better to discuss it (on the
 `#gtk` IRC channel or on [Discourse](https://discourse.gnome.org) before
diff --git a/docs/testing.md b/docs/testing.md
new file mode 100644
index 0000000000..1d35964a09
--- /dev/null
+++ b/docs/testing.md
@@ -0,0 +1,71 @@
+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](https://testanything.org/) 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](https://github.com/google/oss-fuzz/) and are very effective at
+   catching exploitable security issues. See the
+   [fuzzing README](../fuzzing/README.md) 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.