Thread states and the global interpreter lock

Unless on a free-threaded build of CPython, the Python interpreter is generally not thread-safe. In order to support multi-threaded Python programs, there's a global lock, called the global interpreter lock or GIL, that must be held by a thread before accessing Python objects. Without the lock, even the simplest operations could cause problems in a multi-threaded program: for example, when two threads simultaneously increment the reference count of the same object, the reference count could end up being incremented only once instead of twice.

As such, only a thread that holds the GIL may operate on Python objects or invoke Python's C API.

In order to emulate concurrency, the interpreter regularly tries to switch threads between bytecode instructions (see sys.setswitchinterval). This is why locks are also necessary for thread-safety in pure-Python code.

Additionally, the global interpreter lock is released around blocking I/O operations, such as reading or writing to a file. From the C API, this is done by detaching the thread state <detaching-thread-state>.

The Python interpreter keeps some thread-local information inside a data structure called PyThreadState, known as a thread state. Each thread has a thread-local pointer to a PyThreadState; a thread state referenced by this pointer is considered to be attached <attached thread state>.

A thread can only have one attached thread state at a time. An attached thread state is typically analogous with holding the GIL, except on free-threaded builds. On builds with the GIL enabled, attaching a thread state will block until the GIL can be acquired. However, even on builds with the GIL disabled, it is still required to have an attached thread state, as the interpreter needs to keep track of which threads may access Python objects.

[!NOTE] Even on the free-threaded build, attaching a thread state may block, as the GIL can be re-enabled or threads might be temporarily suspended (such as during a garbage collection).

Generally, there will always be an attached thread state when using Python's C API, including during embedding and when implementing methods, so it's uncommon to need to set up a thread state on your own. Only in some specific cases, such as in a Py_BEGIN_ALLOW_THREADS block or in a fresh thread, will the thread not have an attached thread state. If uncertain, check if PyThreadState_GetUnchecked returns NULL.

If it turns out that you do need to create a thread state, it is recommended to use PyThreadState_Ensure or PyThreadState_EnsureFromView, which will manage the thread state for you.

Detaching the thread state from extension code

Most extension code manipulating the thread state has the following simple structure:

Save the thread state in a local variable.
... Do some blocking I/O operation ...
Restore the thread state from the local variable.

This is so common that a pair of macros exists to simplify it:

Py_BEGIN_ALLOW_THREADS
... Do some blocking I/O operation ...
Py_END_ALLOW_THREADS

The Py_BEGIN_ALLOW_THREADS macro opens a new block and declares a hidden local variable; the Py_END_ALLOW_THREADS macro closes the block.

The block above expands to the following code:

PyThreadState *_save;

_save = PyEval_SaveThread();
... Do some blocking I/O operation ...
PyEval_RestoreThread(_save);

Here is how these functions work:

The attached thread state implies that the GIL is held for the interpreter. To detach it, PyEval_SaveThread is called and the result is stored in a local variable.

By detaching the thread state, the GIL is released, which allows other threads to attach to the interpreter and execute while the current thread performs blocking I/O. When the I/O operation is complete, the old thread state is reattached by calling PyEval_RestoreThread, which will wait until the GIL can be acquired.

[!NOTE] Performing blocking I/O is the most common use case for detaching the thread state, but it is also useful to call it over long-running native code that doesn't need access to Python objects or Python's C API. For example, the standard zlib and hashlib modules detach the thread state <attached thread state> when compressing or hashing data.

On a free-threaded build, the GIL is usually out of the question, but detaching the thread state is still required, because the interpreter periodically needs to block all threads to get a consistent view of Python objects without the risk of race conditions. For example, CPython currently suspends all threads for a short period of time while running the garbage collector.

[!WARNING] Detaching the thread state can lead to unexpected behavior during interpreter finalization. See cautions-regarding-runtime-finalization for more details.

APIs

The following macros are normally used without a trailing semicolon; look for example usage in the Python source distribution.

[!NOTE] These macros are still necessary on the free-threaded build to prevent deadlocks.

This macro expands to { PyThreadState *_save; _save = PyEval_SaveThread();. Note that it contains an opening brace; it must be matched with a following Py_END_ALLOW_THREADS macro. See above for further discussion of this macro.

This macro expands to PyEval_RestoreThread(_save); }. Note that it contains a closing brace; it must be matched with an earlier Py_BEGIN_ALLOW_THREADS macro. See above for further discussion of this macro.

This macro expands to PyEval_RestoreThread(_save);: it is equivalent to Py_END_ALLOW_THREADS without the closing brace.

This macro expands to _save = PyEval_SaveThread();: it is equivalent to Py_BEGIN_ALLOW_THREADS without the opening brace and variable declaration.

Using the C API from foreign threads

When threads are created using the dedicated Python APIs (such as the threading module), a thread state is automatically associated with them, However, when a thread is created from native code (for example, by a third-party library with its own thread management), it doesn't hold an attached thread state.

If you need to call Python code from these threads (often this will be part of a callback API provided by the aforementioned third-party library), you must first register these threads with the interpreter by creating a new thread state and attaching it.

The easiest way to do this is through PyThreadState_Ensure or PyThreadState_EnsureFromView.

[!NOTE] These functions require an argument pointing to the desired interpreter; such a pointer can be acquired via a call to PyInterpreterGuard_FromCurrent (for PyThreadState_Ensure) or PyInterpreterView_FromCurrent (for PyThreadState_EnsureFromView) from the function that creates the thread. If no pointer is available (such as when the given native thread library doesn't provide a data argument), PyInterpreterView_FromMain can be used to get a view for the main interpreter, but note that this will make the code incompatible with subinterpreters.

For example:

// The return value of PyInterpreterGuard_FromCurrent() from the
// function that created this thread.
PyInterpreterGuard *guard = thread_data->guard;

// Create a new thread state for the interpreter.
PyThreadStateToken *token = PyThreadState_Ensure(guard);
if (token == NULL) {
   PyInterpreterGuard_Close(guard);
   return;
}

// We have a valid thread state -- perform Python actions here.
result = CallSomeFunction();
// Evaluate result or handle exceptions.

// Release the thread state. No calls to the C API are allowed beyond this
// point.
PyThreadState_Release(token);
PyInterpreterGuard_Close(guard);

Keep in mind that calling PyThreadState_Ensure might not always create a new thread state, and calling PyThreadState_Release might not always detach it. These functions may reuse an existing attached thread state, or may re-attach a thread state that was previously attached for the current thread.

Reusing a thread state across repeated calls

Creating and destroying a PyThreadState is not free, and is more expensive on a free-threaded build. A foreign thread that calls into the interpreter many times -- for example, a worker thread in a native thread pool -- should avoid creating a fresh thread state on every entry and destroying it on every exit. Instead, set up one thread state when the thread starts (or lazily on its first call into Python), attach and detach it around each call, and tear it down once when the thread exits.

Manage the thread state explicitly with PyThreadState_New, attaching and detaching it with PyEval_RestoreThread and PyEval_SaveThread. This happens in three distinct phases, at different points in the thread's life.

When the thread starts, create one thread state for it. interp is the target interpreter, captured by the code that created this thread while it held an attached thread state (for example via PyInterpreterState_Get):

PyThreadState *tstate = PyThreadState_New(interp);

Then, on each call into Python -- which may happen many times over the thread's life -- attach the thread state, make the Python C API calls that require it, and detach again so the thread does not hold the GIL while off doing non-Python work:

PyEval_RestoreThread(tstate);
result = CallSomeFunction();  /* your Python C API calls go here */
PyEval_SaveThread();

When the thread is finished calling into Python, destroy the thread state once:

PyEval_RestoreThread(tstate);
PyThreadState_Clear(tstate);
PyThreadState_DeleteCurrent();

The general-purpose entry points for calling in from a foreign thread -- PyThreadState_Ensure and the older PyGILState_Ensure -- do not guarantee a persistent thread state: their thread-state lifetime is deliberately implementation-defined, so a matched acquire/release pair may create and destroy a thread state each time. Use PyThreadState_New, as shown here, whenever you specifically want to reuse one thread state across calls.

The code that created the foreign thread must arrange for the shutdown sequence to run before the thread exits, and before Py_FinalizeEx is called. If interpreter finalization begins first, the shutdown PyEval_RestoreThread call will hang the thread rather than return (see cautions-regarding-runtime-finalization). If the thread exits without running the shutdown sequence, the thread state is leaked for the remainder of the process.

Attaching/detaching thread states

Ensure that the thread has an attached thread state for the interpreter protected by guard, and thus can safely invoke that interpreter.

It is OK to call this function if the thread already has an attached thread state, as long as there is a subsequent call to PyThreadState_Release that matches this one (meaning that "nested" calls to this function are permitted).

The function's effect (if any) will be reversed by the matching call to PyThreadState_Release.

On error, this function returns NULL without an exception set. Do not call !PyThreadState_Release in this case.

On success, this function returns a pointer value that must be passed to the matching call to !PyThreadState_Release.

The conditions in which this function creates a new thread state are considered unstable and implementation-dependent. If you need to control the exact lifetime of a thread state, consider using PyThreadState_New. However, do not avoid this function solely on the basis that the lifetime of the thread state may be inconsistent across versions; changes to this function will be done with caution and in a backwards-compatible manner. In particular, the saving of thread-local variables and similar state will be retained across Python versions.

The exact behavior of whether this function creates a new thread state is described below, but be aware that this may change in the future.

First, this function checks if an attached thread state is present. If there is, this function then checks if the interpreter of that thread state matches the interpreter guarded by guard. If that is the case, this function simply marks the thread state as being used by a PyThreadState_Ensure call and returns.

If there is no attached thread state, then this function checks if any thread state has been used by the current OS thread. (This is returned by PyGILState_GetThisThreadState.) If there was, then this function checks if that thread state's interpreter matches guard. If it does, it is re-attached and marked as used.

Otherwise, if both of the above cases fail, a new thread state is created for guard. It is then attached and marked as owned by PyThreadState_Ensure.

3.15

Get an attached thread state for the interpreter referenced by view.

The behavior and return value are the same as for PyThreadState_Ensure; additionally, if the function succeeds, the interpreter referenced by view will be implicitly guarded. The guard will be released upon the corresponding PyThreadState_Release call.

3.15

Undo a PyThreadState_Ensure or PyThreadState_EnsureFromView call.

This must be called exactly once for each successful Ensure call, with token set to that call's return value.

The state that was attached before the corresponding Ensure call (if any) will be attached when PyThreadState_Release returns.

The exact behavior of whether this function deletes a thread state is considered unstable and implementation-dependent.

Currently, this function will decrement an internal counter on the attached thread state. If this counter ever reaches below zero, this function emits a fatal error (via Py_FatalError).

If the attached thread state is owned by PyThreadState_Ensure, then the attached thread state will be deallocated and deleted upon the internal counter reaching zero. Otherwise, nothing happens when the counter reaches zero.

3.15

An opaque token retrieved from a PyThreadState_Ensure call and passed to a corresponding PyThreadState_Release call.

GIL-state APIs

The following APIs are generally not compatible with subinterpreters and will hang the process during interpreter finalization (see cautions-regarding-runtime-finalization). As such, these APIs were soft deprecated in Python 3.15 in favor of the new APIs <c-api-foreign-threads>.

The type of the value returned by PyGILState_Ensure and passed to PyGILState_Release.

The GIL was already held when PyGILState_Ensure was called.

The GIL was not held when PyGILState_Ensure was called.

Ensure that the current thread is ready to call the Python C API regardless of the current state of Python, or of the attached thread state. This may be called as many times as desired by a thread as long as each call is matched with a call to PyGILState_Release. In general, other thread-related APIs may be used between PyGILState_Ensure and PyGILState_Release calls as long as the thread state is restored to its previous state before the Release(). For example, normal usage of the Py_BEGIN_ALLOW_THREADS and Py_END_ALLOW_THREADS macros is acceptable.

The return value is an opaque "handle" to the attached thread state when PyGILState_Ensure was called, and must be passed to PyGILState_Release to ensure Python is left in the same state. Even though recursive calls are allowed, these handles cannot be shared - each unique call to PyGILState_Ensure must save the handle for its call to PyGILState_Release.

When the function returns, there will be an attached thread state and the thread will be able to call arbitrary Python code.

This function has no way to return an error. As such, errors are either fatal (that is, they send SIGABRT and crash the process; see Py_FatalError), or the thread will be permanently blocked (such as during interpreter finalization).

[!WARNING] Calling this function when the interpreter is finalizing will infinitely hang the thread, which may cause deadlocks. cautions-regarding-runtime-finalization for more details.

In addition, this function generally does not work with subinterpreters when used from foreign threads, because this function has no way of knowing which interpreter created the thread (and as such, will implicitly pick the main interpreter).

3.14 Hangs the current thread, rather than terminating it, if called while the interpreter is finalizing.

3.15 Use PyThreadState_Ensure or PyThreadState_EnsureFromView instead.

Release any resources previously acquired. After this call, Python's state will be the same as it was prior to the corresponding PyGILState_Ensure call (but generally this state will be unknown to the caller, hence the use of the GIL-state API).

Every call to PyGILState_Ensure must be matched by a call to PyGILState_Release on the same thread.

3.15 Use PyThreadState_Release instead.

Get the thread state that was most recently attached <attached thread state> for this thread. (If the most recent thread state has been deleted, this returns NULL.)

If the caller has an attached thread state, it is returned.

In other terms, this function returns the thread state that will be used by PyGILState_Ensure. If this returns NULL, then PyGILState_Ensure will create a new thread state.

This function cannot fail.

3.15 Use PyThreadState_Get or PyThreadState_GetUnchecked instead.

Return 1 if the current thread has an attached thread state that matches the thread state returned by PyGILState_GetThisThreadState. If the caller has no attached thread state or it otherwise doesn't match, then this returns 0.

If the current Python process has ever created a subinterpreter, this function will always return 1.

This is mainly a helper/diagnostic function.

3.4

3.15 Use PyThreadState_GetUnchecked() != NULL instead.

Cautions about fork()

Another important thing to note about threads is their behaviour in the face of the C fork call. On most systems with fork, after a process forks only the thread that issued the fork will exist. This has a concrete impact both on how locks must be handled and on all stored state in CPython's runtime.

The fact that only the "current" thread remains means any locks held by other threads will never be released. Python solves this for os.fork by acquiring the locks it uses internally before the fork, and releasing them afterwards. In addition, it resets any lock-objects in the child. When extending or embedding Python, there is no way to inform Python of additional (non-Python) locks that need to be acquired before or reset after a fork. OS facilities such as !pthread_atfork would need to be used to accomplish the same thing. Additionally, when extending or embedding Python, calling fork directly rather than through os.fork (and returning to or calling into Python) may result in a deadlock by one of Python's internal locks being held by a thread that is defunct after the fork. PyOS_AfterFork_Child tries to reset the necessary locks, but is not always able to.

The fact that all other threads go away also means that CPython's runtime state there must be cleaned up properly, which os.fork does. This means finalizing all other PyThreadState objects belonging to the current interpreter and all other PyInterpreterState objects. Due to this and the special nature of the "main" interpreter <sub-interpreter-support>, fork should only be called in that interpreter's "main" thread, where the CPython global runtime was originally initialized. The only exception is if exec will be called immediately after.

High-level APIs

These are the most commonly used types and functions when writing multi-threaded C extensions.

This data structure represents the state of a single thread. The only public data member is:

This thread's interpreter state.

Detach the attached thread state and return it. The thread will have no thread state upon returning.

Set the attached thread state to tstate. The passed thread state should not be attached <attached thread state>, otherwise deadlock ensues. tstate will be attached upon returning.

[!NOTE] Calling this function from a thread when the runtime is finalizing will hang the thread until the program exits, even if the thread was not created by Python. Refer to cautions-regarding-runtime-finalization for more details.

3.14 Hangs the current thread, rather than terminating it, if called while the interpreter is finalizing.

Return the attached thread state. If the thread has no attached thread state, (such as when inside of Py_BEGIN_ALLOW_THREADS block), then this issues a fatal error (so that the caller needn't check for NULL).

See also PyThreadState_GetUnchecked.

Similar to PyThreadState_Get, but don't kill the process with a fatal error if it is NULL. The caller is responsible to check if the result is NULL.

3.13 In Python 3.5 to 3.12, the function was private and known as _PyThreadState_UncheckedGet().

Set the attached thread state to tstate, and return the thread state that was attached prior to calling.

This function is safe to call without an attached thread state; it will simply return NULL indicating that there was no prior thread state.

PyEval_ReleaseThread

[!NOTE] Similar to PyGILState_Ensure, this function will hang the thread if the runtime is finalizing.

Low-level APIs

Create a new thread state object belonging to the given interpreter object. An attached thread state is not needed.

Reset all information in a thread state object. tstate must be attached <attached thread state>

3.9 This function now calls the !PyThreadState.on_delete callback. Previously, that happened in PyThreadState_Delete.

3.13 The !PyThreadState.on_delete callback was removed.

Destroy a thread state object. tstate should not be attached <attached thread state> to any thread. tstate must have been reset with a previous call to PyThreadState_Clear.

Detach the attached thread state (which must have been reset with a previous call to PyThreadState_Clear) and then destroy it.

No thread state will be attached <attached thread state> upon returning.

Get the current frame of the Python thread state tstate.

Return a strong reference. Return NULL if no frame is currently executing.

See also PyEval_GetFrame.

tstate must not be NULL, and must be attached <attached thread state>.

3.9

Get the unique thread state identifier of the Python thread state tstate.

tstate must not be NULL, and must be attached <attached thread state>.

3.9

Get the interpreter of the Python thread state tstate.

tstate must not be NULL, and must be attached <attached thread state>.

3.9

Suspend tracing and profiling in the Python thread state tstate.

Resume them using the PyThreadState_LeaveTracing function.

3.11

Resume tracing and profiling in the Python thread state tstate suspended by the PyThreadState_EnterTracing function.

See also PyEval_SetTrace and PyEval_SetProfile functions.

3.11

Set the stack protection start address and stack protection size of a Python thread state.

On success, return 0. On failure, set an exception and return -1.

CPython implements recursion control <recursion> for C code by raising RecursionError when it notices that the machine execution stack is close to overflow. See for example the Py_EnterRecursiveCall function. For this, it needs to know the location of the current thread's stack, which it normally gets from the operating system. When the stack is changed, for example using context switching techniques like the Boost library's boost::context, you must call ~PyUnstable_ThreadState_SetStackProtection to inform CPython of the change.

Call ~PyUnstable_ThreadState_SetStackProtection either before or after changing the stack. Do not call any other Python C API between the call and the stack change.

See PyUnstable_ThreadState_ResetStackProtection for undoing this operation.

3.15

Reset the stack protection start address and stack protection size of a Python thread state to the operating system defaults.

See PyUnstable_ThreadState_SetStackProtection for an explanation.

3.15

Return a dictionary in which extensions can store thread-specific state information. Each extension should use a unique key to store a state in the dictionary. It is okay to call this function when no thread state is attached <attached thread state>. If this function returns NULL and no exception has been raised, then the caller should assume no thread state is attached.

Attach <attached thread state> tstate to the current thread, which must not be NULL or already attached <attached thread state>.

The calling thread must not already have an attached thread state.

[!NOTE] Calling this function from a thread when the runtime is finalizing will hang the thread until the program exits, even if the thread was not created by Python. Refer to cautions-regarding-runtime-finalization for more details.

3.8 Updated to be consistent with PyEval_RestoreThread, Py_END_ALLOW_THREADS, and PyGILState_Ensure, and terminate the current thread if called while the interpreter is finalizing.

3.14 Hangs the current thread, rather than terminating it, if called while the interpreter is finalizing.

PyEval_RestoreThread is a higher-level function which is always available (even when threads have not been initialized).

Detach the attached thread state. The tstate argument, which must not be NULL, is only used to check that it represents the attached thread state --- if it isn't, a fatal error is reported.

PyEval_SaveThread is a higher-level function which is always available (even when threads have not been initialized).

Asynchronous notifications

A mechanism is provided to make asynchronous notifications to the main interpreter thread. These notifications take the form of a function pointer and a void pointer argument.

Schedule a function to be called from the main interpreter thread. On success, 0 is returned and func is queued for being called in the main thread. On failure, -1 is returned without setting any exception.

When successfully queued, func will be eventually called from the main interpreter thread with the argument arg. It will be called asynchronously with respect to normally running Python code, but with both these conditions met:

• on a bytecode boundary;
• with the main thread holding an attached thread state (func can therefore use the full C API).

func must return 0 on success, or -1 on failure with an exception set. func won't be interrupted to perform another asynchronous notification recursively, but it can still be interrupted to switch threads if the thread state <attached thread state> is detached.

This function doesn't need an attached thread state. However, to call this function in a subinterpreter, the caller must have an attached thread state. Otherwise, the function func can be scheduled to be called from the wrong interpreter.

[!WARNING] This is a low-level function, only useful for very special cases. There is no guarantee that func will be called as quick as possible. If the main thread is busy executing a system call, func won't be called before the system call returns. This function is generally not suitable for calling Python code from arbitrary C threads. Instead, use PyThreadState_EnsureFromView.

3.1

3.9 If this function is called in a subinterpreter, the function func is now scheduled to be called from the subinterpreter, rather than being called from the main interpreter. Each subinterpreter now has its own list of scheduled calls.

3.12 This function now always schedules func to be run in the main interpreter.

Execute all pending calls. This is usually executed automatically by the interpreter.

This function returns 0 on success, and returns -1 with an exception set on failure.

If this is not called in the main thread of the main interpreter, this function does nothing and returns 0. The caller must hold an attached thread state.

3.1

3.12 This function only runs pending calls in the main interpreter.

Schedule an exception to be raised asynchronously in a thread. If the thread has a previously scheduled exception, it is overwritten.

The id argument is the thread id of the target thread, as returned by PyThread_get_thread_ident. exc is the class of the exception to be raised, or NULL to clear the pending exception (if any).

Return the number of affected thread states. This is normally 1 if id is found, even when no change was made (the given exc was already pending, or exc is NULL but no exception is pending). If the thread id isn't found, return 0. This raises no exceptions.

To prevent naive misuse, you must write your own C extension to call this. This function must be called with an attached thread state. This function does not steal any references to exc. This function does not necessarily interrupt system calls such as ~time.sleep.

3.7 The type of the id parameter changed from long to unsigned long.

Operating system thread APIs

Sentinel value for an invalid thread ID.

This is currently equivalent to (unsigned long)-1.

Start function func in a new thread with argument arg. The resulting thread is not intended to be joined.

func must not be NULL, but arg may be NULL.

On success, this function returns the identifier of the new thread; on failure, this returns PYTHREAD_INVALID_THREAD_ID.

The caller does not need to hold an attached thread state.

Return the identifier of the current thread, which will never be zero.

This function cannot fail, and the caller does not need to hold an attached thread state.

threading.get_ident and threading.Thread.ident expose this identifier to Python.

Get general information about the current thread in the form of a struct sequence <struct-sequence-objects> object. This information is accessible as sys.thread_info in Python.

On success, this returns a new strong reference to the thread information; on failure, this returns NULL with an exception set.

The caller must hold an attached thread state.

This macro is defined when the system supports native thread IDs.

Get the native identifier of the current thread as it was assigned by the operating system's kernel, which will never be less than zero.

This function is only available when PY_HAVE_THREAD_NATIVE_ID is defined.

This function cannot fail, and the caller does not need to hold an attached thread state.

threading.get_native_id

Terminate the current thread. This function is generally considered unsafe and should be avoided. It is kept solely for backwards compatibility.

This function is only safe to call if all functions in the full call stack are written to safely allow it.

[!WARNING] If the current system uses POSIX threads (also known as "pthreads"), this calls pthread_exit(3), which attempts to unwind the stack and call C++ destructors on some libc implementations. However, if a noexcept function is reached, it may terminate the process. Other systems, such as macOS, do unwinding.

On Windows, this function calls _endthreadex(), which kills the thread without calling C++ destructors.

In any case, there is a risk of corruption on the thread's stack.

3.14

Initialize PyThread* APIs. Python executes this function automatically, so there's little need to call it from an extension module.

Set the stack size of the current thread to size bytes.

This function returns 0 on success, -1 if size is invalid, or -2 if the system does not support changing the stack size. This function does not set exceptions.

The caller does not need to hold an attached thread state.

Return the stack size of the current thread in bytes, or 0 if the system's default stack size is in use.

The caller does not need to hold an attached thread state.