Threading and Tasks in Chrome

Note: See Threading and Tasks FAQ for more examples.


Chromium is a very multithreaded product. We try to keep the UI as responsive as possible, and this means not blocking the UI thread with any blocking I/O or other expensive operations. Our approach is to use message passing as the way of communicating between threads. We discourage locking and threadsafe objects. Instead, objects live on only one thread, we pass messages between threads for communication, and we use callback interfaces (implemented by message passing) for most cross-thread requests.


  • Thread-unsafe: The vast majority of types in Chromium are thread-unsafe by design. Access to such types/methods must be synchronized, typically by sequencing access through a single base::SequencedTaskRunner (this should be enforced by a SEQUENCE_CHECKER) or via low-level synchronization (e.g. locks -- but prefer sequences).
  • Thread-affine: Such types/methods need to be always accessed from the same physical thread (i.e. from the same base::SingleThreadTaskRunner) and should use THREAD_CHECKER to verify that they are. Short of using a third-party API or having a leaf dependency which is thread-affine: there's pretty much no reason for a type to be thread-affine in Chromium. Note that base::SingleThreadTaskRunner is-a base::SequencedTaskRunner so thread-affine is a subset of thread-unsafe. Thread-affine is also sometimes referred to as thread-hostile.
  • Thread-safe: Such types/methods can be safely accessed concurrently.
  • Thread-compatible: Such types provide safe concurrent access to const methods but require synchronization for non-const (or mixed const/non-const access). Chromium doesn't expose reader-writer locks; as such, the only use case for this is objects (typically globals) which are initialized once in a thread-safe manner (either in the single-threaded phase of startup or lazily through a thread-safe static-local-initialization paradigm a la base::NoDestructor) and forever after immutable.
  • Immutable: A subset of thread-compatible types which cannot be modified after construction.
  • Sequence-friendly: Such types/methods are thread-unsafe types which support being invoked from a base::SequencedTaskRunner. Ideally this would be the case for all thread-unsafe types but legacy code sometimes has overzealous checks that enforce thread-affinity in mere thread-unsafe scenarios. See Prefer Sequences to Threads below for more details.


Every Chrome process has

  • a main thread
    • in the browser process: updates the UI
    • in renderer processes: runs most of Blink
  • an IO thread
    • in the browser process: handles IPCs and network requests
    • in renderer processes: handles IPCs
  • a few more special-purpose threads
  • and a pool of general-purpose threads

Most threads have a loop that gets tasks from a queue and runs them (the queue may be shared between multiple threads).


A task is a base::OnceClosure added to a queue for asynchronous execution.

A base::OnceClosure stores a function pointer and arguments. It has a Run() method that invokes the function pointer using the bound arguments. It is created using base::BindOnce. (ref. Callback<> and Bind() documentation).

void TaskA() {}
void TaskB(int v) {}

auto task_a = base::BindOnce(&TaskA);
auto task_b = base::BindOnce(&TaskB, 42);

A group of tasks can be executed in one of the following ways:

  • Parallel: No task execution ordering, possibly all at once on any thread
  • Sequenced: Tasks executed in posting order, one at a time on any thread.
  • Single Threaded: Tasks executed in posting order, one at a time on a single thread.

Prefer Sequences to Threads

Sequenced execution mode is far preferred to Single Threaded in scenarios that require mere thread-safety as it opens up scheduling paradigms that wouldn‘t be possible otherwise (sequences can hop threads instead of being stuck behind unrelated work on a dedicated thread). Ability to hop threads also means the thread count can dynamically adapt to the machine’s true resource availability (increased parallelism on bigger machines, avoids trashing resources on smaller machines).

Many core APIs were recently made sequence-friendly (classes are rarely thread-affine -- i.e. only when using third-party APIs that are thread-affine; even ThreadLocalStorage has a SequenceLocalStorage equivalent). But the codebase has long evolved assuming single-threaded contexts... If your class could run on a sequence but is blocked by an overzealous use of ThreadChecker/ThreadTaskRunnerHandle/SingleThreadTaskRunner in a leaf dependency, consider fixing that dependency for everyone's benefit (or at the very least file a blocking bug against and flag your use of base::CreateSingleThreadTaskRunnerWithTraits() with a TODO against your bug to use base::CreateSequencedTaskRunnerWithTraits() when fixed).

Detailed documentation on how to migrate from single-threaded contexts to sequenced contexts can be found here.

The discussion below covers all of these ways to execute tasks in details.

Posting a Parallel Task

Direct Posting to the Thread Pool

A task that can run on any thread and doesn’t have ordering or mutual exclusion requirements with other tasks should be posted using one of the base::PostTask*() functions defined in base/task/post_task.h.

base::PostTask(FROM_HERE, base::BindOnce(&Task));

This posts tasks with default traits.

The base::PostTask*WithTraits() functions allow the caller to provide additional details about the task via TaskTraits (ref. Annotating Tasks with TaskTraits).

    FROM_HERE, {base::TaskPriority::BEST_EFFORT, MayBlock()},

Posting via a TaskRunner

A parallel TaskRunner is an alternative to calling base::PostTask*() directly. This is mainly useful when it isn’t known in advance whether tasks will be posted in parallel, in sequence, or to a single-thread (ref. Posting a Sequenced Task, Posting Multiple Tasks to the Same Thread). Since TaskRunner is the base class of SequencedTaskRunner and SingleThreadTaskRunner, a scoped_refptr<TaskRunner> member can hold a TaskRunner, a SequencedTaskRunner or a SingleThreadTaskRunner.

class A {
  A() = default;

  void set_task_runner_for_testing(
      scoped_refptr<base::TaskRunner> task_runner) {
    task_runner_ = std::move(task_runner);

  void DoSomething() {
    // In production, A is always posted in parallel. In test, it is posted to
    // the TaskRunner provided via set_task_runner_for_testing().
    task_runner_->PostTask(FROM_HERE, base::BindOnce(&A));

  scoped_refptr<base::TaskRunner> task_runner_ =

Unless a test needs to control precisely how tasks are executed, it is preferred to call base::PostTask*() directly (ref. Testing for less invasive ways of controlling tasks in tests).

Posting a Sequenced Task

A sequence is a set of tasks that run one at a time in posting order (not necessarily on the same thread). To post tasks as part of a sequence, use a SequencedTaskRunner.

Posting to a New Sequence

A SequencedTaskRunner can be created by base::CreateSequencedTaskRunnerWithTraits().

scoped_refptr<SequencedTaskRunner> sequenced_task_runner =

// TaskB runs after TaskA completes.
sequenced_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskA));
sequenced_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskB));

Posting to the Current Sequence

The SequencedTaskRunner to which the current task was posted can be obtained via SequencedTaskRunnerHandle::Get().

NOTE: it is invalid to call SequencedTaskRunnerHandle::Get() from a parallel task, but it is valid from a single-threaded task (a SingleThreadTaskRunner is a SequencedTaskRunner).
// The task will run after any task that has already been posted
// to the SequencedTaskRunner to which the current task was posted
// (in particular, it will run after the current task completes).
// It is also guaranteed that it won’t run concurrently with any
// task posted to that SequencedTaskRunner.
    PostTask(FROM_HERE, base::BindOnce(&Task));

Using Sequences Instead of Locks

Usage of locks is discouraged in Chrome. Sequences inherently provide thread-safety. Prefer classes that are always accessed from the same sequence to managing your own thread-safety with locks.

Thread-safe but not thread-affine; how so? Tasks posted to the same sequence will run in sequential order. After a sequenced task completes, the next task may be picked up by a different worker thread, but that task is guaranteed to see any side-effects caused by the previous one(s) on its sequence.

class A {
  A() {
    // Do not require accesses to be on the creation sequence.

  void AddValue(int v) {
    // Check that all accesses are on the same sequence.


  // No lock required, because all accesses are on the
  // same sequence.
  std::vector<int> values_;

A a;
scoped_refptr<SequencedTaskRunner> task_runner_for_a = ...;
                      base::BindOnce(&A::AddValue, base::Unretained(&a), 42));
                      base::BindOnce(&A::AddValue, base::Unretained(&a), 27));

// Access from a different sequence causes a DCHECK failure.
scoped_refptr<SequencedTaskRunner> other_task_runner = ...;
                            base::BindOnce(&A::AddValue, base::Unretained(&a), 1));

Locks should only be used to swap in a shared data structure that can be accessed on multiple threads. If one thread updates it based on expensive computation or through disk access, then that slow work should be done without holding on to the lock. Only when the result is available should the lock be used to swap in the new data. An example of this is in PluginList::LoadPlugins (content/browser/ If you must use locks, here are some best practices and pitfalls to avoid.

In order to write non-blocking code, many APIs in Chromium are asynchronous. Usually this means that they either need to be executed on a particular thread/sequence and will return results via a custom delegate interface, or they take a base::Callback<> object that is called when the requested operation is completed. Executing work on a specific thread/sequence is covered in the PostTask sections above.

Posting Multiple Tasks to the Same Thread

If multiple tasks need to run on the same thread, post them to a SingleThreadTaskRunner. All tasks posted to the same SingleThreadTaskRunner run on the same thread in posting order.

Posting to the Main Thread or to the IO Thread in the Browser Process

To post tasks to the main thread or to the IO thread, use base::PostTaskWithTraits() or get the appropriate SingleThreadTaskRunner using base::CreateSingleThreadTaskRunnerWithTraits, supplying a BrowserThread::ID as trait. For this, you'll also need to include content/public/browser/browser_task_traits.h.

base::PostTaskWithTraits(FROM_HERE, {content::BrowserThread::UI}, ...);

    ->PostTask(FROM_HERE, ...);

The main thread and the IO thread are already super busy. Therefore, prefer posting to a general purpose thread when possible (ref. Posting a Parallel Task, Posting a Sequenced task). Good reasons to post to the main thread are to update the UI or access objects that are bound to it (e.g. Profile). A good reason to post to the IO thread is to access the internals of components that are bound to it (e.g. IPCs, network). Note: It is not necessary to have an explicit post task to the IO thread to send/receive an IPC or send/receive data on the network.

Posting to the Main Thread in a Renderer Process


Posting to a Custom SingleThreadTaskRunner

If multiple tasks need to run on the same thread and that thread doesn’t have to be the main thread or the IO thread, post them to a SingleThreadTaskRunner created by base::CreateSingleThreadTaskRunnerWithTraits.

scoped_refptr<SequencedTaskRunner> single_thread_task_runner =

// TaskB runs after TaskA completes. Both tasks run on the same thread.
single_thread_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskA));
single_thread_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskB));
IMPORTANT: You should rarely need this, most classes in Chromium require thread-safety (which sequences provide) not thread-affinity. If an API you’re using is incorrectly thread-affine (i.e. using base::ThreadChecker when it’s merely thread-unsafe and should use base::SequenceChecker), please consider fixing it instead of making things worse by also making your API thread-affine.

Posting to the Current Thread

IMPORTANT: To post a task that needs mutual exclusion with the current sequence of tasks but doesn’t absolutely need to run on the current thread, use SequencedTaskRunnerHandle::Get() instead of ThreadTaskRunnerHandle::Get() (ref. Posting to the Current Sequence). That will better document the requirements of the posted task. In a single-thread task, SequencedTaskRunnerHandle::Get() is equivalent to ThreadTaskRunnerHandle::Get().

To post a task to the current thread, use ThreadTaskRunnerHandle.

// The task will run on the current thread in the future.
    FROM_HERE, base::BindOnce(&Task));
NOTE: It is invalid to call ThreadTaskRunnerHandle::Get() from a parallel or a sequenced task.

Posting Tasks to a COM Single-Thread Apartment (STA) Thread (Windows)

Tasks that need to run on a COM Single-Thread Apartment (STA) thread must be posted to a SingleThreadTaskRunner returned by CreateCOMSTATaskRunnerWithTraits(). As mentioned in Posting Multiple Tasks to the Same Thread, all tasks posted to the same SingleThreadTaskRunner run on the same thread in posting order.

// Task(A|B|C)UsingCOMSTA will run on the same COM STA thread.

void TaskAUsingCOMSTA() {
  // [ This runs on a COM STA thread. ]

  // Make COM STA calls.
  // ...

  // Post another task to the current COM STA thread.
      FROM_HERE, base::BindOnce(&TaskCUsingCOMSTA));
void TaskBUsingCOMSTA() { }
void TaskCUsingCOMSTA() { }

auto com_sta_task_runner = base::CreateCOMSTATaskRunnerWithTraits(...);
com_sta_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskAUsingCOMSTA));
com_sta_task_runner->PostTask(FROM_HERE, base::BindOnce(&TaskBUsingCOMSTA));

Annotating Tasks with TaskTraits

TaskTraits encapsulate information about a task that helps the thread pool make better scheduling decisions.

All PostTask*() functions in base/task/post_task.h have an overload that takes TaskTraits as argument and one that doesn’t. The overload that doesn’t take TaskTraits as argument is appropriate for tasks that:

  • Don’t block (ref. MayBlock and WithBaseSyncPrimitives).
  • Prefer inheriting the current priority to specifying their own.
  • Can either block shutdown or be skipped on shutdown (thread pool is free to choose a fitting default). Tasks that don’t match this description must be posted with explicit TaskTraits.

base/task/task_traits.h provides exhaustive documentation of available traits. The content layer also provides additional traits in content/public/browser/browser_task_traits.h to facilitate posting a task onto a BrowserThread.

Below are some examples of how to specify TaskTraits.

// This task has no explicit TaskTraits. It cannot block. Its priority
// is inherited from the calling context (e.g. if it is posted from
// a BEST_EFFORT task, it will have a BEST_EFFORT priority). It will either
// block shutdown or be skipped on shutdown.
base::PostTask(FROM_HERE, base::BindOnce(...));

// This task has the highest priority. The thread pool will try to
// run it before USER_VISIBLE and BEST_EFFORT tasks.
    FROM_HERE, {base::TaskPriority::USER_BLOCKING},

// This task has the lowest priority and is allowed to block (e.g. it
// can read a file from disk).
    FROM_HERE, {base::TaskPriority::BEST_EFFORT, base::MayBlock()},

// This task blocks shutdown. The process won't exit before its
// execution is complete.
    FROM_HERE, {base::TaskShutdownBehavior::BLOCK_SHUTDOWN},

// This task will run on the Browser UI thread.
    FROM_HERE, {content::BrowserThread::UI},

Keeping the Browser Responsive

Do not perform expensive work on the main thread, the IO thread or any sequence that is expected to run tasks with a low latency. Instead, perform expensive work asynchronously using base::PostTaskAndReply*() or SequencedTaskRunner::PostTaskAndReply(). Note that asynchronous/overlapped I/O on the IO thread are fine.

Example: Running the code below on the main thread will prevent the browser from responding to user input for a long time.

// GetHistoryItemsFromDisk() may block for a long time.
// AddHistoryItemsToOmniboxDropDown() updates the UI and therefore must
// be called on the main thread.

The code below solves the problem by scheduling a call to GetHistoryItemsFromDisk() in a thread pool followed by a call to AddHistoryItemsToOmniboxDropdown() on the origin sequence (the main thread in this case). The return value of the first call is automatically provided as argument to the second call.

    FROM_HERE, {base::MayBlock()},
    base::BindOnce(&GetHistoryItemsFromDisk, "keyword"),

Posting a Task with a Delay

Posting a One-Off Task with a Delay

To post a task that must run once after a delay expires, use base::PostDelayedTask*() or TaskRunner::PostDelayedTask().

  FROM_HERE, {base::TaskPriority::BEST_EFFORT}, base::BindOnce(&Task),

scoped_refptr<base::SequencedTaskRunner> task_runner =
    FROM_HERE, base::BindOnce(&Task), base::TimeDelta::FromHours(1));
NOTE: A task that has a 1-hour delay probably doesn’t have to run right away when its delay expires. Specify base::TaskPriority::BEST_EFFORT to prevent it from slowing down the browser when its delay expires.

Posting a Repeating Task with a Delay

To post a task that must run at regular intervals, use base::RepeatingTimer.

class A {
  ~A() {
    // The timer is stopped automatically when it is deleted.
  void StartDoingStuff() {
    timer_.Start(FROM_HERE, TimeDelta::FromSeconds(1),
                 this, &MyClass::DoStuff);
  void StopDoingStuff() {
  void DoStuff() {
    // This method is called every second on the sequence that invoked
    // StartDoingStuff().
  base::RepeatingTimer timer_;

Cancelling a Task

Using base::WeakPtr

base::WeakPtr can be used to ensure that any callback bound to an object is canceled when that object is destroyed.

int Compute() {  }

class A {
  A() : weak_ptr_factory_(this) {}

  void ComputeAndStore() {
    // Schedule a call to Compute() in a thread pool followed by
    // a call to A::Store() on the current sequence. The call to
    // A::Store() is canceled when |weak_ptr_factory_| is destroyed.
    // (guarantees that |this| will not be used-after-free).
        FROM_HERE, base::BindOnce(&Compute),
        base::BindOnce(&A::Store, weak_ptr_factory_.GetWeakPtr()));

  void Store(int value) { value_ = value; }

  int value_;
  base::WeakPtrFactory<A> weak_ptr_factory_;

Note: WeakPtr is not thread-safe: GetWeakPtr(), ~WeakPtrFactory(), and Compute() (bound to a WeakPtr) must all run on the same sequence.

Using base::CancelableTaskTracker

base::CancelableTaskTracker allows cancellation to happen on a different sequence than the one on which tasks run. Keep in mind that CancelableTaskTracker cannot cancel tasks that have already started to run.

auto task_runner = base::CreateTaskRunnerWithTraits(base::TaskTraits());
base::CancelableTaskTracker cancelable_task_tracker;
cancelable_task_tracker.PostTask(task_runner.get(), FROM_HERE,
// Cancels Task(), only if it hasn't already started running.


To test code that uses base::ThreadTaskRunnerHandle, base::SequencedTaskRunnerHandle or a function in base/task/post_task.h, instantiate a base::test::ScopedTaskEnvironment for the scope of the test.

Tests can run the ScopedTaskEnvironment's message pump using a RunLoop, which can be made to run until Quit, or to execute ready-to-run tasks and immediately return.

ScopedTaskEnvironment configures RunLoop::Run() to LOG(FATAL) if it hasn't been explicitly quit after TestTimeouts::action_timeout(). This is preferable to having the test hang if the code under test fails to trigger the RunLoop to quit. The timeout can be overridden with ScopedRunTimeoutForTest.

class MyTest : public testing::Test {
  // ...
   base::test::ScopedTaskEnvironment scoped_task_environment_;

TEST(MyTest, MyTest) {
  base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, base::BindOnce(&A));
      FROM_HERE, base::BindOnce(&C), base::TimeDelta::Max());

  // This runs the (Thread|Sequenced)TaskRunnerHandle queue until it is empty.
  // Delayed tasks are not added to the queue until they are ripe for execution.
  // A and B have been executed. C is not ripe for execution yet.

  base::RunLoop run_loop;
  base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, base::BindOnce(&D));
  base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, run_loop.QuitClosure());
  base::ThreadTaskRunnerHandle::Get()->PostTask(FROM_HERE, base::BindOnce(&E));

  // This runs the (Thread|Sequenced)TaskRunnerHandle queue until QuitClosure is
  // invoked.
  // D and run_loop.QuitClosure() have been executed. E is still in the queue.

  // Tasks posted to thread pool run asynchronously as they are posted.
  base::PostTaskWithTraits(FROM_HERE, base::TaskTraits(), base::BindOnce(&F));
  auto task_runner =
  task_runner->PostTask(FROM_HERE, base::BindOnce(&G));

  // To block until all tasks posted to thread pool are done running:
  // F and G have been executed.

      FROM_HERE, base::TaskTrait(),
      base::BindOnce(&H), base::BindOnce(&I));

  // This runs the (Thread|Sequenced)TaskRunnerHandle queue until both the
  // (Thread|Sequenced)TaskRunnerHandle queue and the TaskSchedule queue are
  // empty:
  // E, H, I have been executed.

Using ThreadPool in a New Process

ThreadPool needs to be initialized in a process before the functions in base/task/post_task.h can be used. Initialization of ThreadPool in the Chrome browser process and child processes (renderer, GPU, utility) has already been taken care of. To use ThreadPool in another process, initialize ThreadPool early in the main function:

// This initializes and starts ThreadPool with default params.
// The base/task/post_task.h API can now be used. Tasks will be // scheduled as
// they are posted.

// This initializes ThreadPool.
// The base/task/post_task.h API can now be used. No threads // will be created
// and no tasks will be scheduled until after Start() is called.
// ThreadPool can now create threads and schedule tasks.

And shutdown ThreadPool late in the main function:

// Tasks posted with TaskShutdownBehavior::BLOCK_SHUTDOWN and
// tasks posted with TaskShutdownBehavior::SKIP_ON_SHUTDOWN that
// have started to run before the Shutdown() call have now completed their
// execution. Tasks posted with
// TaskShutdownBehavior::CONTINUE_ON_SHUTDOWN may still be
// running.

TaskRunner ownership (encourage no dependency injection)

TaskRunners shouldn't be passed through several components. Instead, the components that uses a TaskRunner should be the one that creates it.

See this example of a refactoring where a TaskRunner was passed through a lot of components only to be used in an eventual leaf. The leaf can and should now obtain its TaskRunner directly from base/task/post_task.h.

Dependency injection of TaskRunners can still seldomly be useful to unit test a component when triggering a specific race in a specific way is essential to the test. For such cases the preferred approach is the following:

class FooWithCustomizableTaskRunnerForTesting {

  void SetBackgroundTaskRunnerForTesting(
      scoped_refptr<base::SequencedTaskRunner> background_task_runner);

  scoped_refptr<base::SequencedTaskRunner> background_task_runner_ =
          {base::MayBlock(), base::TaskPriority::BEST_EFFORT});

Note that this still allows removing all layers of plumbing between //chrome and that component since unit tests will use the leaf layer directly.


See Threading and Tasks FAQ for more examples.