gh-115103: Implement delayed free mechanism for free-threaded builds (#115367)

This adds `_PyMem_FreeDelayed()` and supporting functions. The
`_PyMem_FreeDelayed()` function frees memory with the same allocator as
`PyMem_Free()`, but after some delay to ensure that concurrent lock-free
readers have finished.

2 files changed, 9 insertions, 0 deletions

diff --git a/Python/pylifecycle.c b/Python/pylifecycle.c
index 656d821..0448734 100644
--- a/Python/pylifecycle.c
+++ b/Python/pylifecycle.c
@@ -1837,6 +1837,9 @@ finalize_interp_clear(PyThreadState *tstate)
 
     finalize_interp_types(tstate->interp);
 
+    /* Free any delayed free requests immediately */
+    _PyMem_FiniDelayed(tstate->interp);
+
     /* finalize_interp_types may allocate Python objects so we may need to
        abandon mimalloc segments again */
     _PyThreadState_ClearMimallocHeaps(tstate);
diff --git a/Python/pystate.c b/Python/pystate.c
index 1d8c096..bb8e24c 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -617,6 +617,7 @@ init_interpreter(PyInterpreterState *interp,
 #ifdef Py_GIL_DISABLED
     _Py_brc_init_state(interp);
 #endif
+    llist_init(&interp->mem_free_queue.head);
     for (int i = 0; i < _PY_MONITORING_UNGROUPED_EVENTS; i++) {
         interp->monitors.tools[i] = 0;
     }
@@ -1353,6 +1354,7 @@ init_threadstate(_PyThreadStateImpl *_tstate,
     // Initialize biased reference counting inter-thread queue
     _Py_brc_init_thread(tstate);
 #endif
+    llist_init(&_tstate->mem_free_queue);
 
     if (interp->stoptheworld.requested || _PyRuntime.stoptheworld.requested) {
         // Start in the suspended state if there is an ongoing stop-the-world.
@@ -1574,6 +1576,7 @@ PyThreadState_Clear(PyThreadState *tstate)
         // don't call _PyInterpreterState_SetNotRunningMain() yet.
         tstate->on_delete(tstate->on_delete_data);
     }
+
 #ifdef Py_GIL_DISABLED
     // Each thread should clear own freelists in free-threading builds.
     struct _Py_object_freelists *freelists = _Py_object_freelists_GET();
@@ -1583,6 +1586,9 @@ PyThreadState_Clear(PyThreadState *tstate)
     _Py_brc_remove_thread(tstate);
 #endif
 
+    // Merge our queue of pointers to be freed into the interpreter queue.
+    _PyMem_AbandonDelayed(tstate);
+
     _PyThreadState_ClearMimallocHeaps(tstate);
 
     tstate->_status.cleared = 1;