Software testing is a common practice that helps teams continuously deliver quality code. Tests exercise invariants on the software's behavior, catch and prevent regressions in functionality or other desired properties, and help scale engineering processes.
Measuring test coverage in terms of source code line coverage helps engineers identify gaps in their testing solutions. Using test coverage as a metric promotes higher-quality software and safer development practices. Measuring test coverage continuously helps engineers maintain a high bar of quality.
Test coverage does not guarantee bug-free code. Testing should be used along with other tools such as fuzz testing, static and dynamic analysis, etc.
Absolute test coverage is a measure of all lines of source code that are covered by the complete set of tests. Fuchsia's Continuous Integration (CI) infrastructure produces absolute coverage reports, and continuously refreshes them. Coverage reports are typically at most a couple of hours stale.
The latest absolute coverage report is available here. This dashboard shows a tree layout of all code that was found to have been covered by all tests that were exercised, as a subset of the source tree. You can navigate the tree by the directory structure and view total coverage metrics for directories or individual line coverage information for files.
In addition, coverage information is also available as a layer in Google's internal code search.
Incremental test coverage is shown in the context of a change on the Gerrit code review web UI. Incremental coverage shows, specifically in the context of a given change, which modified lines are covered by tests and which modified lines are not.
Incremental test coverage is collected by Fuchsia's Commit Queue (CQ) infrastructure. When sending a change to CQ (Commit-Queue+1), you can click “show experimental tryjobs” to reveal a tryjob named fuchsia-coverage. When this tryjob is complete, your patch set should have absolute coverage (|Cov.|) and incremental coverage (ΔCov.)
Maintaining high incremental test coverage for changes affecting a project helps keep test coverage high continuously. Particularly, it prevents introducing new untested code into a project. Change authors can review incremental coverage information on their changes to ensure that their test coverage is sufficient. Code reviewers can review incremental test coverage information on changes and ask authors to close any testing gaps that they identify as important.
Only unit tests and hermetic integration tests are considered reliable sources of test coverage. No test coverage is collected or presented for E2E tests.
E2E tests are large system tests that exercise the product as a whole and don‘t necessarily cover a well-defined portion of the source code. For example, it’s common for E2E tests on Fuchsia to boot the system in an emulator, interact with it, and expect certain behaviors.
Because E2E tests exercise the system as a whole:
For top-level buildbot bundles like core{:.external}, a counterpart core_no_e2e{:.external} is provided, so bots that collect coverage can use the no_e2e bundle to avoid building and running any E2E tests.
Currently, there are no reliable ways to identify all E2E tests in-tree. As a proxy, no_e2e bundles maintain the invariant that they don‘t have any known e2e test libraries in their recursive dependencies, through GN’s assert_no_deps{:.external}. The list of E2E test libraries is manually curated and maintained, with the assumption that it changes very infrequently:
{% includecode gerrit_repo="fuchsia/fuchsia" gerrit_path="build/config/BUILDCONFIG.gn" region_tag="e2e_test_libs" adjust_indentation="auto" %}
Currently, test coverage is collected only if:
core.x64 or core.arm64 configuration on qemu. Tests that only run in other configurations, such as on hardware targets, are not supported at this time.On that last note, e2e tests exercise a lot of code throughout the system, but they do so in a manner that's inconsistent between runs (or “flaky”). To achieve higher test coverage for code, it is possible and in fact recommended to do so using unit tests and integration tests.
Support for the following additional use cases is currently under development:
core, for instance bringup or workstation.If you are not seeing absolute or incremental coverage information for your code, first review the limitations and ensure that your code is expected to receive coverage support in the first place.
Absolute coverage reports are generated after the code is merged and may take a few hours to fully compile. The dashboard shows the commit hash for the generated report. If you're not seeing expected results on the dashboard, ensure that the data was generated past any recent changes that would have affected coverage. If the data appears stale, come back later and refresh the page.
Incremental coverage reports are generated by CQ. Ensure that you are looking at a patch set that was sent to CQ. You can click “show experimental tryjobs” to reveal a tryjob named fuchsia-coverage. If the tryjob is still running, come back later and refresh the page.
If your code is missing coverage that you expect to see, then pick a test that should have covered your code and ensure that it ran on the coverage tryjob.
fuchsia-coverage run on the CI dashboard.If your test didn't run on any coverage tryjob as expected then one reason might simply be that it only runs in configurations not currently covered by CI/CQ. Another is that the test is explicitly opted out in coverage variants. For instance a BUILD.gn file referencing your test might look as follows:
group("tests") { deps = [ ":foo_test", ":bar_test", ] # TODO(fxbug.dev/12345): This test is intentionally disabled on coverage. if (!is_coverage) { deps += [ ":qux_test" ] } }
Look for context as to why your test is disabled on coverage and investigate.
Related to the above, tests are more likely to be flaky under coverage example. The extra overhead from collecting coverage at runtime slows down performance, which in turn affects timing, which is often the cause for additional flakiness.
An immediate fix would be to disable your test under coverage (see the GN snippet above), at the immediate cost of not collecting coverage information from your test. As a best practice you should treat flakes on coverage the same as you would treat flakes elsewhere, mainly fix the flakiness.
See also: flaky test policy.
Fuchsia's code coverage build, test runtime support, and processing tools use LLVM source-based code coverage{:.external}. The Fuchsia platform is supported by compiler-rt profile runtime.
Profile instrumentation is activated when the "coverage" build variant is selected. The compiler will then generate counters that each correspond to a branch in the code's control flow, and emit instructions on branch entry to increment the associated counters. In addition, the profile instrumentation runtime library is linked into the executable.
For more implementation details see LLVM Code Coverage Mapping Format{:.external}.
Note that the instrumentation leads to increased binary size, increased memory usage, and slower test execution time. Some steps were taken to offset this:
The profile runtime library on Fuchsia stores the profile data in a VMO, and publishes a handle to the VMO using the fuchsia.debug.DebugData protocol. This protocol is made available to tests at runtime using the Component Framework and is hosted by the Test Runner Framework's on-device controller, Test Manager.
The profiles are collected after the test realm terminates, along with any components hosted in it. The profiles are then processed into a single summary for the test. This is an important optimization that significantly reduces the total profile size. The optimized profile is then sent to the host-side test controller.
The host uses the covargs tool, which itself uses the llvm-profdata{:.external} and llvm-cov{:.external} tools, to convert raw profiles to a summary format and to generate test coverage reports. In addition, covargs converts the data to protobuf format which is used as an interchange format with various dashboards.
Ongoing work:
Areas for future work: