Initialization, Finalization, and Threads

Thread State and the Global Interpreter Lock

The Python interpreter is not fully thread safe. In order to support multi-threaded Python programs, there's a global lock, called the :dfn:`global interpreter lock` or :dfn:`GIL`, that must be held by the current thread before it can safely access 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.

Therefore, the rule exists that only the thread that has acquired the global interpreter lock may operate on Python objects or call Python/C API functions. In order to support multi-threaded Python programs, the interpreter regularly releases and reacquires the lock --- by default, every 100 bytecode instructions (this can be changed with :func:`sys.setcheckinterval`). The lock is also released and reacquired around potentially blocking I/O operations like reading or writing a file, so that other threads can run while the thread that requests the I/O is waiting for the I/O operation to complete.

The Python interpreter needs to keep some bookkeeping information separate per thread --- for this it uses a data structure called :ctype:`PyThreadState`. There's one global variable, however: the pointer to the current :ctype:`PyThreadState` structure. Before the addition of :dfn:`thread-local storage` (:dfn:`TLS`) the current thread state had to be manipulated explicitly.

This is easy enough in most cases. Most code manipulating the global interpreter lock has the following simple structure:

Save the thread state in a local variable.
Release the global interpreter lock.
...Do some blocking I/O operation...
Reacquire the global interpreter lock.
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 :cmacro:`Py_BEGIN_ALLOW_THREADS` macro opens a new block and declares a hidden local variable; the :cmacro:`Py_END_ALLOW_THREADS` macro closes the block. Another advantage of using these two macros is that when Python is compiled without thread support, they are defined empty, thus saving the thread state and GIL manipulations.

When thread support is enabled, the block above expands to the following code:

PyThreadState *_save;

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

Using even lower level primitives, we can get roughly the same effect as follows:

PyThreadState *_save;

_save = PyThreadState_Swap(NULL);
PyEval_ReleaseLock();
...Do some blocking I/O operation...
PyEval_AcquireLock();
PyThreadState_Swap(_save);

There are some subtle differences; in particular, :cfunc:`PyEval_RestoreThread` saves and restores the value of the global variable :cdata:`errno`, since the lock manipulation does not guarantee that :cdata:`errno` is left alone. Also, when thread support is disabled, :cfunc:`PyEval_SaveThread` and :cfunc:`PyEval_RestoreThread` don't manipulate the GIL; in this case, :cfunc:`PyEval_ReleaseLock` and :cfunc:`PyEval_AcquireLock` are not available. This is done so that dynamically loaded extensions compiled with thread support enabled can be loaded by an interpreter that was compiled with disabled thread support.

The global interpreter lock is used to protect the pointer to the current thread state. When releasing the lock and saving the thread state, the current thread state pointer must be retrieved before the lock is released (since another thread could immediately acquire the lock and store its own thread state in the global variable). Conversely, when acquiring the lock and restoring the thread state, the lock must be acquired before storing the thread state pointer.

It is important to note that when threads are created from C, they don't have the global interpreter lock, nor is there a thread state data structure for them. Such threads must bootstrap themselves into existence, by first creating a thread state data structure, then acquiring the lock, and finally storing their thread state pointer, before they can start using the Python/C API. When they are done, they should reset the thread state pointer, release the lock, and finally free their thread state data structure.

Threads can take advantage of the :cfunc:`PyGILState_\*` functions to do all of the above automatically. The typical idiom for calling into Python from a C thread is now:

PyGILState_STATE gstate;
gstate = PyGILState_Ensure();

/* Perform Python actions here.  */
result = CallSomeFunction();
/* evaluate result */

/* Release the thread. No Python API allowed beyond this point. */
PyGILState_Release(gstate);

Note that the :cfunc:`PyGILState_\*` functions assume there is only one global interpreter (created automatically by :cfunc:`Py_Initialize`). Python still supports the creation of additional interpreters (using :cfunc:`Py_NewInterpreter`), but mixing multiple interpreters and the :cfunc:`PyGILState_\*` API is unsupported.

Another important thing to note about threads is their behaviour in the face of the C :cfunc:`fork` call. On most systems with :cfunc:`fork`, after a process forks only the thread that issued the fork will exist. That also means any locks held by other threads will never be released. Python solves this for :func:`os.fork` by acquiring the locks it uses internally before the fork, and releasing them afterwards. In addition, it resets any :ref:`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 :cfunc:`posix_atfork` would need to be used to accomplish the same thing. Additionally, when extending or embedding Python, calling :cfunc:`fork` directly rather than through :func:`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. :cfunc:`PyOS_AfterFork` tries to reset the necessary locks, but is not always able to.

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

All of the following functions are only available when thread support is enabled at compile time, and must be called only when the global interpreter lock has been created.

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 argument.

Every check interval, when the global interpreter lock is released and reacquired, Python will also call any such provided functions. This can be used for example by asynchronous IO handlers. The notification can be scheduled from a worker thread and the actual call than made at the earliest convenience by the main thread where it has possession of the global interpreter lock and can perform any Python API calls.

Profiling and Tracing

The Python interpreter provides some low-level support for attaching profiling and execution tracing facilities. These are used for profiling, debugging, and coverage analysis tools.

This C interface allows the profiling or tracing code to avoid the overhead of calling through Python-level callable objects, making a direct C function call instead. The essential attributes of the facility have not changed; the interface allows trace functions to be installed per-thread, and the basic events reported to the trace function are the same as had been reported to the Python-level trace functions in previous versions.

Advanced Debugger Support

These functions are only intended to be used by advanced debugging tools.