I have recently been involved in reviewing some large feature patchsets for a project at work, and thought it might be interesting to discuss some of the principles we have been trying to stick to when going about these reviews.
These are just suggestions which will not apply verbatim to every project, but they might provoke some ideas which are relevant to your project. In many cases, a developer might find a checklist is not useful for them — it is too rigid, or slows down the pace of development too much. Checklists are probably best treated as strong suggestions, rather than strict requirements for a review to pass; developers need to use them as reminders for things they should think about when submitting a patch, rather than a box-ticking exercise. Checklists probably work better in larger projects with lots of infrequent or inexperienced contributors, who may not be aware of all of the conventions of the project.
In order to break a set of reviews down into chunks, there a few key principles to stick to:
- Review patches which are as small as possible, but no smaller (see here, here and here)
- Learn from each review so the same review comments do not need to be made more than once
- Use automated tools to eliminate many of the repetitive and time consuming parts of patch review and rework
- Do high-level API review first, ideally before implementing those APIs, to avoid unnecessary rework
(A rationale for each of these points is given in the section below to avoid cluttering this one.)
This is the pre-submission checklist we have been using to check patches against before submission for review:
- All new code follows the coding guidelines, especially the namespacing guidelines, memory management guidelines, pre- and post-condition guidelines, and introspection guidelines — some key points from these are pulled out below, but these are not the only points to pay attention to.
- All new public API must be namespaced correctly.
- All new public API must have a complete and useful documentation comment.
- All new public API documentation comments must have GObject Introspection annotations where appropriate; g-ir-scanner (part of the build process) must emit no warnings when run with --warn-all --warn-error (which should be set by $(WARN_SCANNERFLAGS) from AX_COMPILER_FLAGS).
- All new public methods must have pre- and post-conditions to enforce constraints on the accepted parameter values.
- The code must compile without warnings, after ensuring that AX_COMPILER_FLAGS is used and enabled in configure.ac (if it is correctly enabled, compiling the module should fail if there are any compiler warnings) — remember to add $(WARN_CFLAGS), $(WARN_LDFLAGS) and $(WARN_SCANNERFLAGS) to new Makefile.am targets as appropriate.
- The introduction documentation comment for each new object must give a usage example for each of the main ways that object is intended to be used.
- All new code must be formatted as per the project’s coding guidelines.
- If possible, there should be an example program for each new feature or object, which can be used to manually test that functionality — these examples may be submitted in a separate patch from the object implementation, but must be submitted at the same time as the implementation in order to allow review in parallel. Example programs must be usable when installed or uninstalled, so they can be used during development and on production machines.
- There must be automated tests (using the GTest framework in GLib) for construction of each new object, and for getting and setting each of its properties.
- The code coverage of the automated tests must be checked (using make check-code-coverage and AX_CODE_COVERAGE) before submission, and if it’s possible to add more automated tests (and for them to be reliable) to improve the coverage, this should be done; the final code coverage figure for the object should be mentioned in a comment on the patch review, and it would be helpful to have the lcov reports for the object saved somewhere for analysis as part of the review.
- There must be no definite memory leaks reported by Valgrind when running the automated tests under it (using AX_VALGRIND_CHECK and make check-valgrind).
- All automated tests must be installed as installed-tests so they can be run as integration tests on a production system.
- make distcheck must pass before submission of any patch, especially if it touches the build system.
- All new code has been checked to ensure it doesn’t contradict review comments from previous reviews of other patches (i.e. we want to avoid making the same review comments on every submitted patch).
- Commit messages must explain why they make the changes they do.
- Each coding guideline has its own rationale for why it’s useful, and many of them significantly affect the structure of a patch, so are important to get right early on.
- Namespacing is important for the correct functioning of a lot of the developer tools (for example, GObject Introspection), and to avoid symbol collisions between libraries — checking it is a very mechanical process which it is best to not have to spend review time on.
- Documentation comments are useful to both the reviewer and to end users of the API — for the reviewer, they act as an explanation of why a particular API is necessary, how it is meant to be used, and can provide insight into implementation choices. These are questions which the reviewer would otherwise have to ask in the review, so writing them up lucidly in a documentation comment saves time in the long run.
- Pre- and post-conditions are a form of assertion in the code, which check for programmer errors at runtime. If they are used consistently throughout the code on every API entry point, they can catch programmer errors much nearer their origin than otherwise, which speeds up debugging both during development of the library, and when end users are using the public APIs. They also act as a loose form of documentation of what each API will allow as its inputs and outputs, which helps review (see the comments about documentation above).
- The set of compiler warnings enabled by AX_COMPILER_FLAGS have been chosen to balance false positives against false negatives in detecting bugs in the code. Each compiler warning typically identifies a single bug in the code which would otherwise have to be fixed later in the life of the library — fixing bugs later is always more expensive in terms of debugging time.
- Usage examples are another form of documentation (as discussed above), which specifically make it clearer to a reviewer how a particular feature is intended to be used. In writing usage examples, the author of a patch can often notice awkwardnesses in their API design, which can then be fixed before review — this is faster than them being caught in review and sent back for modification.
- Well formatted code is a lot easier to read and review than poorly formatted code. It allows the reviewer to think about the function of the code they are reviewing, rather than (for example) which function call a given argument actually applies to, or which block of code a statement is actually part of.
- Example programs are a loose form of testing, and also act as usage examples and documentation for the feature (see above). They provide an easy way for the reviewer to test a feature, especially if it affects a UI or has some interactive element; this is very hard to do by simply looking at the code in a patch. Their biggest benefit will be when the feature is modified in future — the example programs can be used to test changes to the feature and ensure that its behaviour changes (or does not) as expected.
- For each unit test for a piece of code, the behaviour checked by that unit test can be guaranteed to be unchanged across modifications to the code in future. This prevents regressions (especially if the unit tests for the project are set up to be run automatically on each commit by a continuous integration system). The value of unit tests when initially implementing a feature is in the way they guide API design to be testable in the first place. It is often the case that an API will be written without unit tests, and later someone will try to add unit tests and find that the API is untestable; typically because it relies on internal state which the test harness cannot affect. By that point, the API is stable and cannot be changed to allow testing.
- Looking at code coverage reports is a good way to check that unit tests are actually checking what they are expected to check about the code. Code coverage provides a simple, coarse-grained metric of code quality — the quality of untested code is unknown.
- Every memory leak is a bug, and hence needs to be fixed at some point. Checking for memory leaks in a code review is a very mechanical, time-consuming process. If memory leaks can be detected automatically, by using valgrind on the unit tests, this reduces the amount of time needed to catch them during review. This is an area where higher code coverage provides immediate benefits. Another way to avoid leaks is to use g_autoptr() to automatically free memory when leaving a control block.
- If all automated tests are available, they can be run as part of system-wide integration tests, to check that the project behaviour doesn’t change when other system libraries (its dependencies) are changed. This is one of the motivations behind installed-tests. This is a one-time setup needed for your project, and once it’s set up, does not need to be done for each commit.
- make distcheck ensures that a tarball can be created successfully from the code, which entails building it, running all the unit tests, and checking that examples compile.
- If each patch is updated to learn from the results of previous patch reviews, the amount of time spent making and explaining repeated patch review comments should be significantly reduced, which saves everyone’s time.
- Commit messages are a form of documentation of the changes being made to a project. They should explain the motivation behind the changes, and clarify any design decisions which the author thinks the reviewer might question. If a commit message is inadequate, the reviewer is going to ask questions in the review which could have been avoided otherwise.