cpython/Include/internal/pycore_lock.h

296 lines
9.7 KiB
C
Raw Normal View History

gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Lightweight locks and other synchronization mechanisms.
//
// These implementations are based on WebKit's WTF::Lock. See
// https://webkit.org/blog/6161/locking-in-webkit/ for a description of the
// design.
#ifndef Py_INTERNAL_LOCK_H
#define Py_INTERNAL_LOCK_H
#ifdef __cplusplus
extern "C" {
#endif
#ifndef Py_BUILD_CORE
# error "this header requires Py_BUILD_CORE define"
#endif
// A mutex that occupies one byte. The lock can be zero initialized.
//
// Only the two least significant bits are used. The remaining bits should be
// zero:
// 0b00: unlocked
// 0b01: locked
// 0b10: unlocked and has parked threads
// 0b11: locked and has parked threads
//
// Typical initialization:
// PyMutex m = (PyMutex){0};
//
// Or initialize as global variables:
// static PyMutex m;
//
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Typical usage:
// PyMutex_Lock(&m);
// ...
// PyMutex_Unlock(&m);
// NOTE: In Py_GIL_DISABLED builds, `struct _PyMutex` is defined in Include/object.h.
// The Py_GIL_DISABLED builds need the definition in Include/object.h for the
// `ob_mutex` field in PyObject. For the default (non-free-threaded) build,
// we define the struct here to avoid exposing it in the public API.
#ifndef Py_GIL_DISABLED
struct _PyMutex { uint8_t v; };
#endif
typedef struct _PyMutex PyMutex;
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
#define _Py_UNLOCKED 0
#define _Py_LOCKED 1
#define _Py_HAS_PARKED 2
#define _Py_ONCE_INITIALIZED 4
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// (private) slow path for locking the mutex
PyAPI_FUNC(void) _PyMutex_LockSlow(PyMutex *m);
// (private) slow path for unlocking the mutex
PyAPI_FUNC(void) _PyMutex_UnlockSlow(PyMutex *m);
static inline int
PyMutex_LockFast(uint8_t *lock_bits)
{
uint8_t expected = _Py_UNLOCKED;
return _Py_atomic_compare_exchange_uint8(lock_bits, &expected, _Py_LOCKED);
}
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Locks the mutex.
//
// If the mutex is currently locked, the calling thread will be parked until
// the mutex is unlocked. If the current thread holds the GIL, then the GIL
// will be released while the thread is parked.
static inline void
PyMutex_Lock(PyMutex *m)
{
uint8_t expected = _Py_UNLOCKED;
if (!_Py_atomic_compare_exchange_uint8(&m->v, &expected, _Py_LOCKED)) {
_PyMutex_LockSlow(m);
}
}
// Unlocks the mutex.
static inline void
PyMutex_Unlock(PyMutex *m)
{
uint8_t expected = _Py_LOCKED;
if (!_Py_atomic_compare_exchange_uint8(&m->v, &expected, _Py_UNLOCKED)) {
_PyMutex_UnlockSlow(m);
}
}
// Checks if the mutex is currently locked.
static inline int
PyMutex_IsLocked(PyMutex *m)
{
return (_Py_atomic_load_uint8(&m->v) & _Py_LOCKED) != 0;
}
// Re-initializes the mutex after a fork to the unlocked state.
static inline void
_PyMutex_at_fork_reinit(PyMutex *m)
{
memset(m, 0, sizeof(*m));
}
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
typedef enum _PyLockFlags {
// Do not detach/release the GIL when waiting on the lock.
_Py_LOCK_DONT_DETACH = 0,
// Detach/release the GIL while waiting on the lock.
_PY_LOCK_DETACH = 1,
// Handle signals if interrupted while waiting on the lock.
_PY_LOCK_HANDLE_SIGNALS = 2,
} _PyLockFlags;
// Lock a mutex with an optional timeout and additional options. See
// _PyLockFlags for details.
extern PyLockStatus
_PyMutex_LockTimed(PyMutex *m, PyTime_t timeout_ns, _PyLockFlags flags);
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Lock a mutex with aditional options. See _PyLockFlags for details.
static inline void
PyMutex_LockFlags(PyMutex *m, _PyLockFlags flags)
{
uint8_t expected = _Py_UNLOCKED;
if (!_Py_atomic_compare_exchange_uint8(&m->v, &expected, _Py_LOCKED)) {
_PyMutex_LockTimed(m, -1, flags);
}
}
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Unlock a mutex, returns 0 if the mutex is not locked (used for improved
// error messages).
extern int _PyMutex_TryUnlock(PyMutex *m);
// PyEvent is a one-time event notification
typedef struct {
uint8_t v;
} PyEvent;
// Check if the event is set without blocking. Returns 1 if the event is set or
// 0 otherwise.
PyAPI_FUNC(int) _PyEvent_IsSet(PyEvent *evt);
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// Set the event and notify any waiting threads.
// Export for '_testinternalcapi' shared extension
PyAPI_FUNC(void) _PyEvent_Notify(PyEvent *evt);
// Wait for the event to be set. If the event is already set, then this returns
// immediately.
PyAPI_FUNC(void) PyEvent_Wait(PyEvent *evt);
// Wait for the event to be set, or until the timeout expires. If the event is
// already set, then this returns immediately. Returns 1 if the event was set,
// and 0 if the timeout expired or thread was interrupted.
PyAPI_FUNC(int) PyEvent_WaitTimed(PyEvent *evt, PyTime_t timeout_ns);
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
// _PyRawMutex implements a word-sized mutex that that does not depend on the
// parking lot API, and therefore can be used in the parking lot
// implementation.
//
// The mutex uses a packed representation: the least significant bit is used to
// indicate whether the mutex is locked or not. The remaining bits are either
// zero or a pointer to a `struct raw_mutex_entry` (see lock.c).
typedef struct {
uintptr_t v;
} _PyRawMutex;
// Slow paths for lock/unlock
extern void _PyRawMutex_LockSlow(_PyRawMutex *m);
extern void _PyRawMutex_UnlockSlow(_PyRawMutex *m);
static inline void
_PyRawMutex_Lock(_PyRawMutex *m)
{
uintptr_t unlocked = _Py_UNLOCKED;
if (_Py_atomic_compare_exchange_uintptr(&m->v, &unlocked, _Py_LOCKED)) {
return;
}
_PyRawMutex_LockSlow(m);
}
static inline void
_PyRawMutex_Unlock(_PyRawMutex *m)
{
uintptr_t locked = _Py_LOCKED;
if (_Py_atomic_compare_exchange_uintptr(&m->v, &locked, _Py_UNLOCKED)) {
return;
}
_PyRawMutex_UnlockSlow(m);
}
// A data structure that can be used to run initialization code once in a
// thread-safe manner. The C++11 equivalent is std::call_once.
typedef struct {
uint8_t v;
} _PyOnceFlag;
// Type signature for one-time initialization functions. The function should
// return 0 on success and -1 on failure.
typedef int _Py_once_fn_t(void *arg);
// (private) slow path for one time initialization
PyAPI_FUNC(int)
_PyOnceFlag_CallOnceSlow(_PyOnceFlag *flag, _Py_once_fn_t *fn, void *arg);
// Calls `fn` once using `flag`. The `arg` is passed to the call to `fn`.
//
// Returns 0 on success and -1 on failure.
//
// If `fn` returns 0 (success), then subsequent calls immediately return 0.
// If `fn` returns -1 (failure), then subsequent calls will retry the call.
static inline int
_PyOnceFlag_CallOnce(_PyOnceFlag *flag, _Py_once_fn_t *fn, void *arg)
{
if (_Py_atomic_load_uint8(&flag->v) == _Py_ONCE_INITIALIZED) {
return 0;
}
return _PyOnceFlag_CallOnceSlow(flag, fn, arg);
}
// A readers-writer (RW) lock. The lock supports multiple concurrent readers or
// a single writer. The lock is write-preferring: if a writer is waiting while
// the lock is read-locked then, new readers will be blocked. This avoids
// starvation of writers.
//
// In C++, the equivalent synchronization primitive is std::shared_mutex
// with shared ("read") and exclusive ("write") locking.
//
// The two least significant bits are used to indicate if the lock is
// write-locked and if there are parked threads (either readers or writers)
// waiting to acquire the lock. The remaining bits are used to indicate the
// number of readers holding the lock.
//
// 0b000..00000: unlocked
// 0bnnn..nnn00: nnn..nnn readers holding the lock
// 0bnnn..nnn10: nnn..nnn readers holding the lock and a writer is waiting
// 0b00000..010: unlocked with awoken writer about to acquire lock
// 0b00000..001: write-locked
// 0b00000..011: write-locked and readers or other writers are waiting
//
// Note that reader_count must be zero if the lock is held by a writer, and
// vice versa. The lock can only be held by readers or a writer, but not both.
//
// The design is optimized for simplicity of the implementation. The lock is
// not fair: if fairness is desired, use an additional PyMutex to serialize
// writers. The lock is also not reentrant.
typedef struct {
uintptr_t bits;
} _PyRWMutex;
// Read lock (i.e., shared lock)
PyAPI_FUNC(void) _PyRWMutex_RLock(_PyRWMutex *rwmutex);
PyAPI_FUNC(void) _PyRWMutex_RUnlock(_PyRWMutex *rwmutex);
// Write lock (i.e., exclusive lock)
PyAPI_FUNC(void) _PyRWMutex_Lock(_PyRWMutex *rwmutex);
PyAPI_FUNC(void) _PyRWMutex_Unlock(_PyRWMutex *rwmutex);
// Similar to linux seqlock: https://en.wikipedia.org/wiki/Seqlock
// We use a sequence number to lock the writer, an even sequence means we're unlocked, an odd
// sequence means we're locked. Readers will read the sequence before attempting to read the
// underlying data and then read the sequence number again after reading the data. If the
// sequence has not changed the data is valid.
//
// Differs a little bit in that we use CAS on sequence as the lock, instead of a separate spin lock.
// The writer can also detect that the undelering data has not changed and abandon the write
// and restore the previous sequence.
typedef struct {
uint32_t sequence;
} _PySeqLock;
// Lock the sequence lock for the writer
PyAPI_FUNC(void) _PySeqLock_LockWrite(_PySeqLock *seqlock);
// Unlock the sequence lock and move to the next sequence number.
PyAPI_FUNC(void) _PySeqLock_UnlockWrite(_PySeqLock *seqlock);
// Abandon the current update indicating that no mutations have occurred
// and restore the previous sequence value.
PyAPI_FUNC(void) _PySeqLock_AbandonWrite(_PySeqLock *seqlock);
// Begin a read operation and return the current sequence number.
PyAPI_FUNC(uint32_t) _PySeqLock_BeginRead(_PySeqLock *seqlock);
// End the read operation and confirm that the sequence number has not changed.
// Returns 1 if the read was successful or 0 if the read should be re-tried.
PyAPI_FUNC(uint32_t) _PySeqLock_EndRead(_PySeqLock *seqlock, uint32_t previous);
// Check if the lock was held during a fork and clear the lock. Returns 1
// if the lock was held and any associated datat should be cleared.
PyAPI_FUNC(uint32_t) _PySeqLock_AfterFork(_PySeqLock *seqlock);
gh-108724: Add PyMutex and _PyParkingLot APIs (gh-109344) PyMutex is a one byte lock with fast, inlineable lock and unlock functions for the common uncontended case. The design is based on WebKit's WTF::Lock. PyMutex is built using the _PyParkingLot APIs, which provides a cross-platform futex-like API (based on WebKit's WTF::ParkingLot). This internal API will be used for building other synchronization primitives used to implement PEP 703, such as one-time initialization and events. This also includes tests and a mini benchmark in Tools/lockbench/lockbench.py to compare with the existing PyThread_type_lock. Uncontended acquisition + release: * Linux (x86-64): PyMutex: 11 ns, PyThread_type_lock: 44 ns * macOS (arm64): PyMutex: 13 ns, PyThread_type_lock: 18 ns * Windows (x86-64): PyMutex: 13 ns, PyThread_type_lock: 38 ns PR Overview: The primary purpose of this PR is to implement PyMutex, but there are a number of support pieces (described below). * PyMutex: A 1-byte lock that doesn't require memory allocation to initialize and is generally faster than the existing PyThread_type_lock. The API is internal only for now. * _PyParking_Lot: A futex-like API based on the API of the same name in WebKit. Used to implement PyMutex. * _PyRawMutex: A word sized lock used to implement _PyParking_Lot. * PyEvent: A one time event. This was used a bunch in the "nogil" fork and is useful for testing the PyMutex implementation, so I've included it as part of the PR. * pycore_llist.h: Defines common operations on doubly-linked list. Not strictly necessary (could do the list operations manually), but they come up frequently in the "nogil" fork. ( Similar to https://man.freebsd.org/cgi/man.cgi?queue) --------- Co-authored-by: Eric Snow <ericsnowcurrently@gmail.com>
2023-09-19 12:54:29 -03:00
#ifdef __cplusplus
}
#endif
#endif /* !Py_INTERNAL_LOCK_H */