So you've read the Threading and Tasks documentation, surveyed the associated Threading and Tasks FAQ and have implemented a state-of-the-art component. Now you want to test it :). This document will explain how to write matching state-of-the-art tests.
In order to unit test a component which post tasks, you'll need to bring up the task environment in the scope of your test (or test fixture). It will need to outlive the majority of other members to ensure they have access to the task system throughout their lifetime. There are a rare exceptions, like base::test::ScopedFeatureList
, that need to outlive the task environment. For browser tests, see the Browser tests section below.
Task environments come in various forms but share the same fundamental characteristics:
base::test::TaskEnvironment
and support its ValidTraits
and sometimes more.The TaskEnvironment
member is typically exposed in the protected section of the test fixture to allow tests to drive it directly (there's no need to expose public Run*() methods that merely forward to the private member).
Your component uses base::ThreadTaskRunnerHandle::Get()
or base::SequencedTaskRunnerHandle::Get()
to post tasks to the thread it was created on? You'll need at least a base::test::SingleThreadTaskEnvironment
in order for these APIs to be functional and base::RunLoop
to run the posted tasks.
Typically this will look something like this:
foo.h
class Foo { public: Foo() : owning_sequence_(base::SequencedTaskRunnerHandle::Get()) {} DoSomethingAndReply(base::OnceClosure reply) { DCHECK(owning_sequence_->RunsTasksInCurrentSequence()); something_was_done_ = true; owning_sequence_->PostTask(on_done); } bool something_was_done() const { return something_was_done_; } private: bool something_was_done_ = false; scoped_refptr<base::SequencedTaskRunner> owning_sequence_; };
foo_unittest.cc
TEST(FooTest, DoSomething) { base::test::SingleThreadTaskEnvironment task_environment; Foo foo; RunLoop run_loop; foo.DoSomethingAndReply(run_loop.QuitClosure()); run_loop.Run(); EXPECT_TRUE(foo.something_was_done()); }
Note that RunLoop().RunUntilIdle()
could be used instead of a QuitClosure()
above but best practices favor QuitClosure() over RunUntilIdle() as the latter can lead to flaky tests.
If your components depends on base::ThreadPool
(that‘s a good thing!), you’ll need a full base::test::TaskEnvironment
. Don‘t be afraid to use a full TaskEnvironment
when appropriate: think of “SingleThread” as being a readability term like “const”, it documents that ThreadPool isn’t used when it‘s not but you shouldn’t be afraid to lift it.
Task runners are still obtained by the product code through base/task/post_task.h without necessitating a test-only task runner injection seam :).
Typical use case:
foo_service.h
class FooService { public: FooService() : backend_task_runner_( base::CreateSequencedTaskRunnerWithTraits( base::MayBlock(), base::ThreadPool(), base::TaskPriority::BEST_EFFORT))), backend_(new FooBackend, base::OnTaskRunnerDeleter(backend_task_runner_)) {} // Flushes state to disk async and replies. FlushAndReply(base::OnceClosure on_done) { DCHECK(owning_sequence_->RunsTasksInCurrentSequence()); backend_task_runner_->PostTaskAndReply( base::BindOnce(&FooBackend::Flush, Unretained(backend_.get()), std::move(on_done)); } private: scoped_refptr<base::SequencedTaskRunner> backend_task_runner_; // See https://youtu.be/m6Kz6pMaIxc?t=882 for memory management best // practices. std::unique_ptr<FooBackend, base::OnTaskRunnerDeleter> backend_; };
foo_service_unittest.cc
TEST(FooServiceTest, FlushAndReply) { base::test::TaskEnvironment task_environment; FooService foo_service; RunLoop run_loop; foo_service.FlushAndReply(run_loop.QuitClosure()); run_loop.Run(); EXPECT_TRUE(VerifyFooStateOnDisk()); }
This is the same thing as base::test::TaskEnvironment
with the addition of content::BrowserThread
support. You need this if-and-only-if the code under test is using BrowserThread::UI
or BrowserThread::IO
. For determinism, both BrowserThreads will share the main thread and can be driven by RunLoop. By default the main thread will use MainThreadType::UI
but you can override this via the MainThreadType trait to ask for an IO pump.
BrowserTaskEnvironment::REAL_IO_THREAD
can be also used as a construction trait for rare instances that desire distinct physical BrowserThreads.
This is the //ios equivalent of content::BrowserTaskEnvironment
to simulate web::WebThread
.
All task environments support the following methods to run tasks:
base::RunLoop:Run()
: run the main thread until the QuitClosure()
is invoked (note: other threads also run in parallel by default).base::RunLoop::RunUntilIdle()
: run the main thread until it is idle. This is typically not what you want in multi-threaded environments as it may resume before ThreadPool
is idle.TaskEnvironment::RunUntilIdle()
: Runs everything the TaskEnvironment is aware of. This excludes system events and any threads outside of the main thread and ThreadPool. It should be used with care when such external factors can be involved.TaskEnvironment::FastForward*()
: More on this in the TimeSource section below.By default tests run under TimeSource::SYSTEM_TIME
which means delays are real-time and base::Time::Now()
and base::TimeTicks::Now()
return live system times (context).
Whenever testing code with delays, you should favor TimeSource::MOCK_TIME
as a trait. This makes it such that delayed tasks and base::Time::Now()
+ base::TimeTicks::Now()
use a mock clock.
Under this mode, the mock clock will start at the current system time but will then only advance when explicitly requested by TaskEnvironment::FastForward*()
and TaskEnvironment::AdvanceClock()
methods or when RunLoop::Run()
is running and all managed threads become idle (auto-advances to the soonest delayed task, if any, amongst all managed threads).
TaskEnvironment::FastForwardBy()
repeatedly runs existing immediately executable tasks until idle and then advances the mock clock incrementally to run the next delayed task within the time delta. It may advance time by more than the requested amount if running the tasks causes nested time-advancing-method calls.
This makes it possible to test code with flush intervals, repeating timers, timeouts, etc. without any test-specific seams in the product code, e.g.:
foo_storage.h
class FooStorage { public: static constexpr base::TimeDelta::kFlushInterval = base::TimeDelta::FromSeconds(30); // Sets |key| to |value|. Flushed to disk on the next flush interval. void Set(base::StringPiece key, base::StringPiece value); };
foo_unittest.cc
class FooStorageTest { public: FooStorageTest() = default; // Test helper that returns true if |key| is found in the on disk storage. bool FindKeyInOnDiskStorage(base::StringPiece key); protected: base::test::TaskEnvironment task_environment{ base::test::TaskEnvironment::TimeSource::MOCK_TIME}; FooStorage foo_storage_; }; TEST_F(FooStorageTest, Set) { foo_storage_.Set("mykey", "myvalue"); EXPECT_FALSE(FindKeyInOnDiskStorage("mykey")); task_environment.FastForwardBy(FooStorage::kFlushInterval); EXPECT_TRUE(FindKeyInOnDiskStorage("mykey")); }
In contrast, TaskEnvironment::AdvanceClock()
simply advances the mock time by the requested amount, and does not run tasks. This may be useful in cases where TaskEnvironment::FastForwardBy()
would result in a livelock. For example, if one task is blocked on a WaitableEvent
and there is a delayed task that would signal the event (e.g., a timeout), then TaskEnvironment::FastForwardBy()
will never complete. In this case, you could advance the clock enough that the delayed task becomes runnable, and then TaskEnvironment::RunUntilIdle()
would run the delayed task, signalling the event.
TEST(FooTest, TimeoutExceeded) { base::test::TaskEnvironment task_environment{ base::test::TaskEnvironment::TimeSource::MOCK_TIME}; base::WaitableEvent event; base::RunLoop run_loop; base::PostTaskAndReply( FROM_HERE, {base::ThreadPool(), base::MayBlock()}, base::BindOnce(&BlocksOnEvent, base::Unretained(&event)), run_loop.QuitClosure()); base::PostDelayedTask( FROM_HERE, {base::ThreadPool()}, base::BindOnce(&WaitableEvent::Signal, base::Unretained(&event)), kTimeout); // Can't use task_environment.FastForwardBy() since BlocksOnEvent blocks // and the task pool will not become idle. // Instead, advance time until the timeout task becomes runnable. task_environment.AdvanceClock(kTimeout); // Now the timeout task is runable. task_environment.RunUntilIdle(); // The reply task should already have been executed, but run the run_loop to // verify. run_loop.Run(); }
The average component only cares about running its tasks and MainThreadType::DEFAULT
is sufficient. Components that care to interact asynchronously with the system will likely need a MainThreadType::UI
to be able to receive system events (e.g,. UI or clipboard events).
Some components will prefer a main thread that handles asynchronous IO events and will use MainThreadType::IO
. Such components are typically the ones living on BrowserThread::IO and being tested with a BrowserTaskEnvironment
initialized with MainThreadType::IO
.
Note: This is strictly about requesting a specific MessagePumpType
for the main thread. It has nothing to do with BrowserThread::UI
or BrowserThread::IO
which are named threads in the //content/browser code.
By default non-delayed tasks posted to base::ThreadPool
may run at any point. Tests that require more determinism can request ThreadPoolExecutionMode::QUEUED
to enforce that tasks posted to base::ThreadPool
only run when TaskEnvironment::RunUntilIdle()
or TaskEnvironment::FastForward*()
are invoked. Note that RunLoop::Run()
does not unblock the ThreadPool in this mode and thus strictly runs only the main thread.
When ThreadPoolExecutionMode::QUEUED
is mixed with TimeSource::MOCK_TIME
, time will auto-advance to the soonest task that is allowed to run when required (i.e. it will ignore delayed tasks in the thread pool while in RunLoop::Run()
). See TaskEnvironmentTest.MultiThreadedMockTimeAndThreadPoolQueuedMode
for an example.
This trait is of course irrelevant under SingleThreadTaskEnvironment
.
Prefer an explicit SingleThreadTaskEnvironment
over using ThreadingMode::MAIN_THREAD_ONLY
. The only reason to use ThreadingMode::MAIN_THREAD_ONLY
explicitly is if the parent class of your test fixture manages the TaskEnvironment
but takes TaskEnvironmentTraits
to let its subclasses customize it and you really need a SingleThreadTaskEnvironment
.
In some cases it makes sense to have the base fixture of an entire section of the codebase be managing the TaskEnvironment
(e.g. ViewsTestBase). It's useful if such base fixture exposes TaskEnvironmentTraits
to their subclasses so that individual tests within that domain can fine-tune their traits as desired.
This typically looks like this (in this case FooTestBase
opts to enforce MainThreadType::UI
and leaves other traits to be specified as desired):
// Constructs a FooTestBase with |traits| being forwarded to its // TaskEnvironment. MainThreadType always defaults to UI and must not be // specified. template <typename... TaskEnvironmentTraits> NOINLINE explicit FooTestBase(TaskEnvironmentTraits&&... traits) : task_environment_(base::test::TaskEnvironment::MainThreadType::UI, std::forward<TaskEnvironmentTraits>(traits)...) {}
Note, if you're not familiar with traits: TaskEnvironment traits use base/traits_bag.h and will automatically complain at compile-time if an enum-based trait is specified more than once (i.e. subclasses will not compile if re-specifying MainThreadType
in the above example).
This is all nice and fancy for unit tests, but what about browser_tests, ui_integration_tests, etc? Tests that subclass content::BrowserTestBase
bring up the entire environment (tasks & more) by default.
The downside is that you don't have fine-grained control over it like you would with all the TaskEnvironment
methods.
The favored paradigm is RunLoop::Run()
+ QuitClosure()
. The asynchronous nature of Chromium code makes this the most reliable way to wait for an event.
There are fancy versions of this to perform common actions, e.g. content/public/test/browser_test_utils.h content/public/test/content_browser_test_utils.h which will let you navigate, execute scripts, simulate UI interactions, etc.
But the idea is always the same :
RunLoop run_loop;
run_loop.QuitClosure()
run_loop.Run()
until the QuitClosure()
is called.So you fell in love with TimeSource::MOCK_TIME
but now you're in a browser test... yeah, sorry about that...
The eventual goal is to make it possible to set up TaskEnvironmentTraits from your test fixture just like you can override command-line, etc. but we're not there yet...
In the mean time you can still use the old base::ScopedMockTimeMessageLoopTaskRunner
to mock delayed tasks on the main thread (you're out of luck on other threads for now). And you can use base::subtle::ScopedTimeClockOverrides
if you want to override Now()
.
You think that‘s a mess? Just think that it used to be this way in unit tests too and you’ll be happy again :).
Here are some paradigms you might see throughout the code base and some insight on what to do about them (hint: copying them is not one!). Migration help is welcome crbug.com/984323!
This is the ancestor of SingleThreadTaskEnvironment
+ TimeSource::MOCK_TIME
. It's sort of equivalent but prefer task environments for consistency.
The only case where base::TestMockTimeTaskRunner
is still the only option is when writing regression tests that simulate a specific task execution order across multiple sequences. To do so, use two base::TestMockTimeTaskRunner
and have the racing components post their tasks to separate ones. You can then explicitly run tasks posted to each one from the main test thread in a way that deterministically exercises the race resolution under test. This only applies to task execution order races, data races still require parallel execution and this is the main reason TaskEnvironment
doesn't multiplex the ThreadPool
tasks onto the main thread (i.e. exercise data races, especially in the scope of TSAN).
Prefer using SingleThreadTaskEnvironment
over base::TestSimpleTaskRunner
. TestSimpleTaskRunner
isn‘t as “simple” as it seems specifically because it runs tasks in a surprising order (delays aren’t respected and nesting doesn't behave as usual). Should you prefer to flush all tasks regardless of delays, TimeSource::MOCK_TIME
and TaskEnvironment::FastForwardUntilNoTasksRemain()
have you covered.
Prefer SingleThreadTaskEnvironment
or TaskEnvironment
with ThreadPoolExecutionMode::QUEUED
over base::NullTaskRunner
. A NullTaskRunner
might seem appealing, but not posting tasks is under-testing the potential side-effects of the code under tests. All tests should be okay if tasks born from their actions are run or deleted at a later point.
This is the ancestor of base::TestMockTimeTaskRunner
which is itself mostly deprecated. As mentioned above in the TimeSource trait section: This should never be used anymore except to mock time when there's already a task system in place, e.g. in browser_tests.
Prior to TaskEnvironment::TimeSource::MOCK_TIME
, many components had SetClockForTesting()
in their product code. And before modern Threading and Tasks, some components had SetTaskRunnerForTesting(). Neither of these test-only seams are required anymore now that task environments can mock those from under-the-hood. Cleanup in favor of modern TaskEnvironment paradigms is always appreciated (crbug.com/984323).
Different parts of the codebase have their own helper task runners. Please migrate away from them or document them above. Ultimately the goal is for TaskEnvironment
and its subclasses to rule them all and to have a consistent task testing API surface once and for all.
It is still okay for specific areas to have a base fixture that configures a default TaskEnvironment
appropriate for that area and use the TaskEnvironmentTraits
paradigm outlined in the Base Fixture managed TaskEnvironment section above to let individual tests provide additional traits.