gh-126096: Use a thread-local freelist for thread states

Misc/NEWS.d/next/Core_and_Builtins/2024-10-28-19-13-52.gh-issue-126096.3x4spF.rst

@@ -0,0 +1 @@
+Optimized switching between interpreters via a freelist.

Python/pystate.c

@@ -1390,10 +1390,26 @@ allocate_chunk(int size_in_bytes, _PyStackChunk* previous)
     return res;
 }
 
+#ifdef HAVE_THREAD_LOCAL
+_Py_thread_local static const struct _Py_freelist _Py_tss_tstates_freelist = { 0 };
+#define _PyThreadState_FREELIST() ((struct _Py_freelist *) &_Py_tss_tstates_freelist)
+#endif
+
 static _PyThreadStateImpl *
 alloc_threadstate(void)
 {
+#ifdef HAVE_THREAD_LOCAL
+    _PyThreadStateImpl *impl = _PyFreeList_PopMem(_PyThreadState_FREELIST());
+    if (impl == NULL)
+    {
+        return PyMem_RawCalloc(1, sizeof(_PyThreadStateImpl));
+    }
+    // Zero-out the allocation again.
+    memset(impl, 0, sizeof(_PyThreadStateImpl));
+    return impl;
+#else
     return PyMem_RawCalloc(1, sizeof(_PyThreadStateImpl));
+#endif
 }
 
 static void
@@ -1408,7 +1424,12 @@ free_threadstate(_PyThreadStateImpl *tstate)
                sizeof(*tstate));
     }
     else {
+#ifdef HAVE_THREAD_LOCAL
+        // Maximum of 100 thread states in the freelist
+        _PyFreeList_Free(_PyThreadState_FREELIST(), tstate, 100, PyMem_RawFree);
+#else
         PyMem_RawFree(tstate);
+#endif
     }
 }
 

Tools/c-analyzer/cpython/ignored.tsv

@@ -185,6 +185,7 @@ Python/pyfpe.c	-	PyFPE_counter	-
 
 Python/import.c	-	pkgcontext	-
 Python/pystate.c	-	_Py_tss_tstate	-
+Python/pystate.c	-	_Py_tss_tstates_freelist	-
 
 ##-----------------------
 ## should be const