Enable the GIL while loading C extension modules

Include/internal/pycore_ceval.h

@@ -130,9 +130,43 @@ extern int _PyEval_ThreadsInitialized(void);
 extern void _PyEval_InitGIL(PyThreadState *tstate, int own_gil);
 extern void _PyEval_FiniGIL(PyInterpreterState *interp);
 
-extern void _PyEval_AcquireLock(PyThreadState *tstate);
+// Acquire the GIL and return 1. In free-threaded builds, this function may
+// return 0 to indicate that the GIL was disabled and therefore not acquired.
+extern int _PyEval_AcquireLock(PyThreadState *tstate);
+
 extern void _PyEval_ReleaseLock(PyInterpreterState *, PyThreadState *);
 
+#ifdef Py_GIL_DISABLED
+// Enable or disable the GIL used by the interpreter that owns tstate, which
+// must be the current thread. This may affect other interpreters, if the GIL
+// is shared. All three functions will be no-ops (and return 0) if the
+// interpreter's `enable_gil' config is not _PyConfig_GIL_DEFAULT.
+//
+// Every call to _PyEval_EnableGILTransient() must be paired with exactly one
+// call to either _PyEval_EnableGILPermanent() or
+// _PyEval_DisableGIL(). _PyEval_EnableGILPermanent() and _PyEval_DisableGIL()
+// must only be called while the GIL is enabled from a call to
+// _PyEval_EnableGILTransient().
+//
+// _PyEval_EnableGILTransient() returns 1 if it enabled the GIL, or 0 if the
+// GIL was already enabled, whether transiently or permanently. The caller will
+// hold the GIL upon return.
+//
+// _PyEval_EnableGILPermanent() returns 1 if it permanently enabled the GIL
+// (which must already be enabled), or 0 if it was already permanently
+// enabled. Once _PyEval_EnableGILPermanent() has been called once, all
+// subsequent calls to any of the three functions will be no-ops.
+//
+// _PyEval_DisableGIL() returns 1 if it disabled the GIL, or 0 if the GIL was
+// kept enabled because of another request, whether transient or permanent.
+//
+// All three functions must be called by an attached thread (this implies that
+// if the GIL is enabled, the current thread must hold it).
+extern int _PyEval_EnableGILTransient(PyThreadState *tstate);
+extern int _PyEval_EnableGILPermanent(PyThreadState *tstate);
+extern int _PyEval_DisableGIL(PyThreadState *state);
+#endif
+
 extern void _PyEval_DeactivateOpCache(void);
 
 

Include/internal/pycore_gil.h

@@ -21,8 +21,20 @@ extern "C" {
 
 struct _gil_runtime_state {
 #ifdef Py_GIL_DISABLED
-    /* Whether or not this GIL is being used. Can change from 0 to 1 at runtime
-       if, for example, a module that requires the GIL is loaded. */
+    /* If this GIL is disabled, enabled == 0.
+
+       If this GIL is enabled transiently (most likely to initialize a module
+       of unknown safety), enabled indicates the number of active transient
+       requests.
+
+       If this GIL is enabled permanently, enabled == INT_MAX.
+
+       It must not be modified directly; use _PyEval_EnableGILTransiently(),
+       _PyEval_EnableGILPermanently(), and _PyEval_DisableGIL()
+
+       It is always read and written atomically, but a thread can assume its
+       value will be stable as long as that thread is attached or knows that no
+       other threads are attached (e.g., during a stop-the-world.). */
     int enabled;
 #endif
     /* microseconds (the Python API uses seconds, though) */

Include/internal/pycore_import.h

@@ -83,6 +83,9 @@ struct _import_state {
        This is initialized lazily in PyState_AddModule(), which is also
        where modules get added. */
     PyObject *modules_by_index;
+#ifdef Py_GIL_DISABLED
+    PyObject *module_gil_by_index;
+#endif
     /* importlib module._bootstrap */
     PyObject *importlib;
     /* override for config->use_frozen_modules (for tests)
@@ -206,6 +209,22 @@ extern int _PyImport_CheckSubinterpIncompatibleExtensionAllowed(
 // Export for '_testinternalcapi' shared extension
 PyAPI_FUNC(int) _PyImport_ClearExtension(PyObject *name, PyObject *filename);
 
+#ifdef Py_GIL_DISABLED
+extern int _PyImport_SetModuleGIL(PyObject *module, void *gil);
+extern void *_PyImport_GetModuleDefGIL(PyModuleDef *def);
+
+// Assuming that the GIL is enabled from a call to
+// _PyEval_EnableGILTransient(), resolve the transient request depending on the
+// state of the module argument:
+// - If module is NULL or a PyModuleObject with md_gil == Py_MOD_GIL_NOT_USED,
+//   call _PyEval_DisableGIL().
+// - Otherwise, call _PyEval_EnableGILPermanent(). If the GIL was not already
+//   enabled permanently, issue a warning referencing the module's name.
+//
+// This function may raise an exception.
+extern int _PyImport_CheckGILForModule(PyObject *module, PyObject *module_name);
+#endif
+
 #ifdef __cplusplus
 }
 #endif

Python/ceval_gil.c

@@ -205,6 +205,16 @@ static void recreate_gil(struct _gil_runtime_state *gil)
 }
 #endif
 
+static void
+drop_gil_impl(struct _gil_runtime_state *gil)
+{
+    MUTEX_LOCK(gil->mutex);
+    _Py_ANNOTATE_RWLOCK_RELEASED(&gil->locked, /*is_write=*/1);
+    _Py_atomic_store_int_relaxed(&gil->locked, 0);
+    COND_SIGNAL(gil->cond);
+    MUTEX_UNLOCK(gil->mutex);
+}
+
 static void
 drop_gil(PyInterpreterState *interp, PyThreadState *tstate)
 {
@@ -220,7 +230,7 @@ drop_gil(PyInterpreterState *interp, PyThreadState *tstate)
 
     struct _gil_runtime_state *gil = ceval->gil;
 #ifdef Py_GIL_DISABLED
-    if (!gil->enabled) {
+    if (!_Py_atomic_load_int_relaxed(&gil->enabled)) {
         return;
     }
 #endif
@@ -236,11 +246,7 @@ drop_gil(PyInterpreterState *interp, PyThreadState *tstate)
         _Py_atomic_store_ptr_relaxed(&gil->last_holder, tstate);
     }
 
-    MUTEX_LOCK(gil->mutex);
-    _Py_ANNOTATE_RWLOCK_RELEASED(&gil->locked, /*is_write=*/1);
-    _Py_atomic_store_int_relaxed(&gil->locked, 0);
-    COND_SIGNAL(gil->cond);
-    MUTEX_UNLOCK(gil->mutex);
+    drop_gil_impl(gil);
 
 #ifdef FORCE_SWITCHING
     /* We check tstate first in case we might be releasing the GIL for
@@ -275,8 +281,10 @@ drop_gil(PyInterpreterState *interp, PyThreadState *tstate)
 
    The function saves errno at entry and restores its value at exit.
 
-   tstate must be non-NULL. */
-static void
+   tstate must be non-NULL.
+
+   Returns 1 if the GIL was acquired, and 0 if not. */
+static int
 take_gil(PyThreadState *tstate)
 {
     int err = errno;
@@ -300,8 +308,8 @@ take_gil(PyThreadState *tstate)
     PyInterpreterState *interp = tstate->interp;
     struct _gil_runtime_state *gil = interp->ceval.gil;
 #ifdef Py_GIL_DISABLED
-    if (!gil->enabled) {
-        return;
+    if (!_Py_atomic_load_int_relaxed(&gil->enabled)) {
+        return 0;
     }
 #endif
 
@@ -346,6 +354,17 @@ take_gil(PyThreadState *tstate)
         }
     }
 
+#ifdef Py_GIL_DISABLED
+    if (!_Py_atomic_load_int_relaxed(&gil->enabled)) {
+        // Another thread disabled the GIL between our check above and
+        // now. Don't take the GIL, signal any other waiting threads, and
+        // return 0.
+        COND_SIGNAL(gil->cond);
+        MUTEX_UNLOCK(gil->mutex);
+        return 0;
+    }
+#endif
+
 #ifdef FORCE_SWITCHING
     /* This mutex must be taken before modifying gil->last_holder:
        see drop_gil(). */
@@ -387,6 +406,7 @@ take_gil(PyThreadState *tstate)
     MUTEX_UNLOCK(gil->mutex);
 
     errno = err;
+    return 1;
 }
 
 void _PyEval_SetSwitchInterval(unsigned long microseconds)
@@ -451,7 +471,8 @@ init_own_gil(PyInterpreterState *interp, struct _gil_runtime_state *gil)
 {
     assert(!gil_created(gil));
 #ifdef Py_GIL_DISABLED
-    gil->enabled = _PyInterpreterState_GetConfig(interp)->enable_gil == _PyConfig_GIL_ENABLE;
+    const PyConfig *config = _PyInterpreterState_GetConfig(interp);
+    gil->enabled = config->enable_gil == _PyConfig_GIL_ENABLE ? INT_MAX : 0;
 #endif
     create_gil(gil);
     assert(gil_created(gil));
@@ -545,11 +566,11 @@ PyEval_ReleaseLock(void)
     drop_gil(tstate->interp, tstate);
 }
 
-void
+int
 _PyEval_AcquireLock(PyThreadState *tstate)
 {
     _Py_EnsureTstateNotNULL(tstate);
-    take_gil(tstate);
+    return take_gil(tstate);
 }
 
 void
@@ -1011,6 +1032,117 @@ _PyEval_InitState(PyInterpreterState *interp)
     _gil_initialize(&interp->_gil);
 }
 
+#ifdef Py_GIL_DISABLED
+int
+_PyEval_EnableGILTransient(PyThreadState *tstate)
+{
+    if (_PyInterpreterState_GetConfig(tstate->interp)->enable_gil !=
+        _PyConfig_GIL_DEFAULT) {
+        return 0;
+    }
+    struct _gil_runtime_state *gil = tstate->interp->ceval.gil;
+
+    int enabled = _Py_atomic_load_int_relaxed(&gil->enabled);
+    if (enabled == INT_MAX) {
+        // The GIL is already enabled permanently.
+        return 0;
+    }
+    if (enabled == INT_MAX - 1) {
+        Py_FatalError("Too many transient requests to enable the GIL");
+    }
+    if (enabled > 0) {
+        // If enabled is nonzero, we know we hold the GIL. This means that no
+        // other threads are attached, and nobody else can be concurrently
+        // mutating it.
+        _Py_atomic_store_int_relaxed(&gil->enabled, enabled + 1);
+        return 0;
+    }
+
+    // Enabling the GIL changes what it means to be an "attached" thread. To
+    // safely make this transition, we:
+    // 1. Detach the current thread.
+    // 2. Stop the world to detach (and suspend) all other threads.
+    // 3. Enable the GIL, if nobody else did between our check above and when
+    //    our stop-the-world begins.
+    // 4. Start the world.
+    // 5. Attach the current thread. Other threads may attach and hold the GIL
+    //    before this thread, which is harmless.
+    _PyThreadState_Detach(tstate);
+
+    // This could be an interpreter-local stop-the-world in situations where we
+    // know that this interpreter's GIL is not shared, and that it won't become
+    // shared before the stop-the-world begins. For now, we always stop all
+    // interpreters for simplicity.
+    _PyEval_StopTheWorldAll(&_PyRuntime);
+
+    enabled = _Py_atomic_load_int_relaxed(&gil->enabled);
+    int this_thread_enabled = enabled == 0;
+    _Py_atomic_store_int_relaxed(&gil->enabled, enabled + 1);
+
+    _PyEval_StartTheWorldAll(&_PyRuntime);
+    _PyThreadState_Attach(tstate);
+
+    return this_thread_enabled;
+}
+
+int
+_PyEval_EnableGILPermanent(PyThreadState *tstate)
+{
+    if (_PyInterpreterState_GetConfig(tstate->interp)->enable_gil !=
+        _PyConfig_GIL_DEFAULT) {
+        return 0;
+    }
+
+    struct _gil_runtime_state *gil = tstate->interp->ceval.gil;
+    assert(current_thread_holds_gil(gil, tstate));
+
+    int enabled = _Py_atomic_load_int_relaxed(&gil->enabled);
+    if (enabled == INT_MAX) {
+        return 0;
+    }
+
+    _Py_atomic_store_int_relaxed(&gil->enabled, INT_MAX);
+    return 1;
+}
+
+int
+_PyEval_DisableGIL(PyThreadState *tstate)
+{
+    if (_PyInterpreterState_GetConfig(tstate->interp)->enable_gil !=
+        _PyConfig_GIL_DEFAULT) {
+        return 0;
+    }
+
+    struct _gil_runtime_state *gil = tstate->interp->ceval.gil;
+    assert(current_thread_holds_gil(gil, tstate));
+
+    int enabled = _Py_atomic_load_int_relaxed(&gil->enabled);
+    if (enabled == INT_MAX) {
+        return 0;
+    }
+
+    assert(enabled >= 1);
+    enabled--;
+
+    // Disabling the GIL is much simpler than enabling it, since we know we are
+    // the only attached thread. Other threads may start free-threading as soon
+    // as this store is complete, if it sets gil->enabled to 0.
+    _Py_atomic_store_int_relaxed(&gil->enabled, enabled);
+
+    if (enabled == 0) {
+        // We're attached, so we know the GIL will remain disabled until at
+        // least the next time we detach, which must be after this function
+        // returns.
+        //
+        // Drop the GIL, which will wake up any threads waiting in take_gil()
+        // and let them resume execution without the GIL.
+        drop_gil_impl(gil);
+        return 1;
+    }
+    return 0;
+}
+#endif
+
 
 /* Do periodic things, like check for signals and async I/0.
 * We need to do reasonably frequently, but not too frequently.