pythongh-112532: Use separate mimalloc heaps for GC objects

In `--disable-gil` builds, we now use four separate heaps in
anticipation of using mimalloc to find GC objects when the GIL is
disabled. To support this, we also make a few changes to mimalloc:

* `mi_heap_t` and `mi_tld_t` initialization is split from allocation.
  This allows us to have a `mi_tld_t` per-`PyThreadState`, which is
  important to keep interpreter isolation, since the same OS thread may
  run in multiple interpreters (using different PyThreadStates.)

* Heap abandoning (mi_heap_collect_ex) can now be called from a
  different thread than the one that created the heap. This is necessary
  because we may clear and delete the containing PyThreadStates from a
  different thread during finalization and after fork().

Include/internal/mimalloc/mimalloc/internal.h

@@ -85,6 +85,7 @@ mi_threadid_t _mi_thread_id(void) mi_attr_noexcept;
 mi_heap_t*    _mi_heap_main_get(void);     // statically allocated main backing heap
 void       _mi_thread_done(mi_heap_t* heap);
 void       _mi_thread_data_collect(void);
+void       _mi_tld_init(mi_tld_t* tld, mi_heap_t* bheap);
 
 // os.c
 void       _mi_os_init(void);                                            // called from process init
@@ -170,6 +171,7 @@ size_t     _mi_bin_size(uint8_t bin);           // for stats
 uint8_t    _mi_bin(size_t size);                // for stats
 
 // "heap.c"
+void       _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id);
 void       _mi_heap_destroy_pages(mi_heap_t* heap);
 void       _mi_heap_collect_abandon(mi_heap_t* heap);
 void       _mi_heap_set_default_direct(mi_heap_t* heap);

Include/internal/pycore_mimalloc.h

@@ -9,11 +9,34 @@
 #  error "pycore_mimalloc.h must be included before mimalloc.h"
 #endif
 
+#define _Py_MIMALLOC_HEAP_MEM       0   // PyMem_Malloc() and friends
+#define _Py_MIMALLOC_HEAP_OBJECT    1   // non-GC objects
+#define _Py_MIMALLOC_HEAP_GC        2   // GC objects without pre-header
+#define _Py_MIMALLOC_HEAP_GC_PRE    3   // GC objects with pre-header
+#define _Py_MIMALLOC_HEAP_COUNT     4
+
 #include "pycore_pymem.h"
 #define MI_DEBUG_UNINIT     PYMEM_CLEANBYTE
 #define MI_DEBUG_FREED      PYMEM_DEADBYTE
 #define MI_DEBUG_PADDING    PYMEM_FORBIDDENBYTE
+#ifdef Py_DEBUG
+#  define MI_DEBUG 1
+#else
+#  define MI_DEBUG 0
+#endif
 
 #include "mimalloc.h"
+#include "mimalloc/types.h"
+#include "mimalloc/internal.h"
+
+struct _mimalloc_thread_state {
+#ifdef Py_GIL_DISABLED
+    mi_heap_t *current_object_heap;
+    mi_heap_t heaps[_Py_MIMALLOC_HEAP_COUNT];
+    mi_tld_t tld;
+#else
+    char _unused;  // empty structs are not allowed
+#endif
+};
 
 #endif // Py_INTERNAL_MIMALLOC_H

Include/internal/pycore_pystate.h

@@ -187,6 +187,7 @@ extern PyThreadState * _PyThreadState_New(
     int whence);
 extern void _PyThreadState_Bind(PyThreadState *tstate);
 extern void _PyThreadState_DeleteExcept(PyThreadState *tstate);
+extern void _PyThreadState_ClearMimallocHeaps(PyThreadState *tstate);
 
 // Export for '_testinternalcapi' shared extension
 PyAPI_FUNC(PyObject*) _PyThreadState_GetDict(PyThreadState *tstate);

Include/internal/pycore_tstate.h

@@ -8,6 +8,8 @@ extern "C" {
 #  error "this header requires Py_BUILD_CORE define"
 #endif
 
+#include "pycore_mimalloc.h"      // struct _mimalloc_thread_state
+
 
 // Every PyThreadState is actually allocated as a _PyThreadStateImpl. The
 // PyThreadState fields are exposed as part of the C API, although most fields
@@ -16,7 +18,10 @@ typedef struct _PyThreadStateImpl {
     // semi-public fields are in PyThreadState.
     PyThreadState base;
 
-    // TODO: add private fields here
+#ifdef Py_GIL_DISABLED
+    struct _mimalloc_thread_state mimalloc;
+#endif
+
 } _PyThreadStateImpl;
 
 

Objects/mimalloc/heap.c

@@ -123,14 +123,17 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
   const bool force = collect >= MI_FORCE;
   _mi_deferred_free(heap, force);
 
+  // gh-112532: we may be called from a thread that is not the owner of the heap
+  bool is_main_thread = _mi_is_main_thread() && heap->thread_id == _mi_thread_id();
+
   // note: never reclaim on collect but leave it to threads that need storage to reclaim
   const bool force_main =
     #ifdef NDEBUG
       collect == MI_FORCE
     #else
       collect >= MI_FORCE
     #endif
-      && _mi_is_main_thread() && mi_heap_is_backing(heap) && !heap->no_reclaim;
+      && is_main_thread && mi_heap_is_backing(heap) && !heap->no_reclaim;
 
   if (force_main) {
     // the main thread is abandoned (end-of-program), try to reclaim all abandoned segments.
@@ -164,7 +167,7 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
   }
 
   // collect regions on program-exit (or shared library unload)
-  if (force && _mi_is_main_thread() && mi_heap_is_backing(heap)) {
+  if (force && is_main_thread && mi_heap_is_backing(heap)) {
     _mi_thread_data_collect();  // collect thread data cache
     _mi_arena_collect(true /* force purge */, &heap->tld->stats);
   }
@@ -206,18 +209,28 @@ mi_heap_t* mi_heap_get_backing(void) {
   return bheap;
 }
 
-mi_decl_nodiscard mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id) {
-  mi_heap_t* bheap = mi_heap_get_backing();
-  mi_heap_t* heap = mi_heap_malloc_tp(bheap, mi_heap_t);  // todo: OS allocate in secure mode?
-  if (heap == NULL) return NULL;
+void _mi_heap_init_ex(mi_heap_t* heap, mi_tld_t* tld, mi_arena_id_t arena_id)
+{
   _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(mi_heap_t));
-  heap->tld = bheap->tld;
+  heap->tld = tld;
   heap->thread_id = _mi_thread_id();
   heap->arena_id = arena_id;
-  _mi_random_split(&bheap->random, &heap->random);
+  if (heap == tld->heap_backing) {
+    _mi_random_init(&heap->random);
+  }
+  else {
+    _mi_random_split(&tld->heap_backing->random, &heap->random);
+  }
   heap->cookie = _mi_heap_random_next(heap) | 1;
   heap->keys[0] = _mi_heap_random_next(heap);
   heap->keys[1] = _mi_heap_random_next(heap);
+}
+
+mi_decl_nodiscard mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id) {
+  mi_heap_t* bheap = mi_heap_get_backing();
+  mi_heap_t* heap = mi_heap_malloc_tp(bheap, mi_heap_t);  // todo: OS allocate in secure mode?
+  if (heap == NULL) return NULL;
+  _mi_heap_init_ex(heap, bheap->tld, arena_id);
   heap->no_reclaim = true;  // don't reclaim abandoned pages or otherwise destroy is unsafe
   // push on the thread local heaps list
   heap->next = heap->tld->heaps;

Objects/mimalloc/init.c

@@ -297,24 +297,20 @@ static bool _mi_heap_init(void) {
     mi_thread_data_t* td = mi_thread_data_zalloc();
     if (td == NULL) return false;
 
-    mi_tld_t*  tld = &td->tld;
-    mi_heap_t* heap = &td->heap;
+    _mi_tld_init(&td->tld, &td->heap);
+    _mi_heap_init_ex(&td->heap, &td->tld, _mi_arena_id_none());
+    _mi_heap_set_default_direct(&td->heap);
+  }
+  return false;
+}
+
+void _mi_tld_init(mi_tld_t* tld, mi_heap_t* bheap) {
     _mi_memcpy_aligned(tld, &tld_empty, sizeof(*tld));
-    _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(*heap));
-    heap->thread_id = _mi_thread_id();
-    _mi_random_init(&heap->random);
-    heap->cookie  = _mi_heap_random_next(heap) | 1;
-    heap->keys[0] = _mi_heap_random_next(heap);
-    heap->keys[1] = _mi_heap_random_next(heap);
-    heap->tld = tld;
-    tld->heap_backing = heap;
-    tld->heaps = heap;
     tld->segments.stats = &tld->stats;
     tld->segments.os = &tld->os;
     tld->os.stats = &tld->stats;
-    _mi_heap_set_default_direct(heap);
-  }
-  return false;
+    tld->heap_backing = bheap;
+    tld->heaps = bheap;
 }
 
 // Free the thread local default heap (called from `mi_thread_done`)

Objects/obmalloc.c

@@ -88,19 +88,37 @@ _PyMem_RawFree(void *Py_UNUSED(ctx), void *ptr)
 void *
 _PyMem_MiMalloc(void *ctx, size_t size)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = &tstate->mimalloc.heaps[_Py_MIMALLOC_HEAP_MEM];
+    return mi_heap_malloc(heap, size);
+#else
     return mi_malloc(size);
+#endif
 }
 
 void *
 _PyMem_MiCalloc(void *ctx, size_t nelem, size_t elsize)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = &tstate->mimalloc.heaps[_Py_MIMALLOC_HEAP_MEM];
+    return mi_heap_calloc(heap, nelem, elsize);
+#else
     return mi_calloc(nelem, elsize);
+#endif
 }
 
 void *
 _PyMem_MiRealloc(void *ctx, void *ptr, size_t size)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = &tstate->mimalloc.heaps[_Py_MIMALLOC_HEAP_MEM];
+    return mi_heap_realloc(heap, ptr, size);
+#else
     return mi_realloc(ptr, size);
+#endif
 }
 
 void
@@ -112,20 +130,38 @@ _PyMem_MiFree(void *ctx, void *ptr)
 void *
 _PyObject_MiMalloc(void *ctx, size_t nbytes)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = tstate->mimalloc.current_object_heap;
+    return mi_heap_malloc(heap, nbytes);
+#else
     return mi_malloc(nbytes);
+#endif
 }
 
 void *
 _PyObject_MiCalloc(void *ctx, size_t nelem, size_t elsize)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = tstate->mimalloc.current_object_heap;
+    return mi_heap_calloc(heap, nelem, elsize);
+#else
     return mi_calloc(nelem, elsize);
+#endif
 }
 
 
 void *
 _PyObject_MiRealloc(void *ctx, void *ptr, size_t nbytes)
 {
+#ifdef Py_GIL_DISABLED
+    _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
+    mi_heap_t *heap = tstate->mimalloc.current_object_heap;
+    return mi_heap_realloc(heap, ptr, nbytes);
+#else
     return mi_realloc(ptr, nbytes);
+#endif
 }
 
 void

Python/pylifecycle.c

@@ -1794,6 +1794,10 @@ finalize_interp_clear(PyThreadState *tstate)
     }
 
     finalize_interp_types(tstate->interp);
+
+    /* finalize_interp_types may allocate Python objects so we may need to
+       abandon mimalloc segments again */
+    _PyThreadState_ClearMimallocHeaps(tstate);
 }
 
 

Python/pystate.c

@@ -236,6 +236,8 @@ tstate_is_bound(PyThreadState *tstate)
 static void bind_gilstate_tstate(PyThreadState *);
 static void unbind_gilstate_tstate(PyThreadState *);
 
+static void tstate_mimalloc_bind(PyThreadState *);
+
 static void
 bind_tstate(PyThreadState *tstate)
 {
@@ -256,6 +258,9 @@ bind_tstate(PyThreadState *tstate)
     tstate->native_thread_id = PyThread_get_thread_native_id();
 #endif
 
+    // mimalloc state needs to be initialized from the active thread.
+    tstate_mimalloc_bind(tstate);
+
     tstate->_status.bound = 1;
 }
 
@@ -1533,6 +1538,8 @@ PyThreadState_Clear(PyThreadState *tstate)
         tstate->on_delete(tstate->on_delete_data);
     }
 
+    _PyThreadState_ClearMimallocHeaps(tstate);
+
     tstate->_status.cleared = 1;
 
     // XXX Call _PyThreadStateSwap(runtime, NULL) here if "current".
@@ -2495,3 +2502,51 @@ _PyThreadState_MustExit(PyThreadState *tstate)
     }
     return 1;
 }
+
+/********************/
+/* mimalloc support */
+/********************/
+
+static void
+tstate_mimalloc_bind(PyThreadState *tstate)
+{
+#ifdef Py_GIL_DISABLED
+    struct _mimalloc_thread_state *mts = &((_PyThreadStateImpl*)tstate)->mimalloc;
+
+    // Initialize the mimalloc thread state. This must be called from the
+    // same thread that will use the thread state. The "mem" heap doubles as
+    // the "backing" heap.
+    mi_tld_t *tld = &mts->tld;
+    _mi_tld_init(tld, &mts->heaps[_Py_MIMALLOC_HEAP_MEM]);
+
+    // Initialize each heap
+    for (Py_ssize_t i = 0; i < _Py_MIMALLOC_HEAP_COUNT; i++) {
+        _mi_heap_init_ex(&mts->heaps[i], tld, _mi_arena_id_none());
+    }
+
+    // By default, object allocations use _Py_MIMALLOC_HEAP_OBJECT.
+    // _PyObject_GC_New() and similar functions temporarily override this to
+    // use one of the GC heaps.
+    mts->current_object_heap = &mts->heaps[_Py_MIMALLOC_HEAP_OBJECT];
+#endif
+}
+
+void
+_PyThreadState_ClearMimallocHeaps(PyThreadState *tstate)
+{
+#ifdef Py_GIL_DISABLED
+    if (!tstate->_status.bound) {
+        // The mimalloc heaps are only initialized when the thread is bound.
+        return;
+    }
+
+    _PyThreadStateImpl *tstate_impl = (_PyThreadStateImpl *)tstate;
+    for (Py_ssize_t i = 0; i < _Py_MIMALLOC_HEAP_COUNT; i++) {
+        // Abandon all segments in use by this thread. This pushes them to
+        // a shared pool to later be reclaimed by other threads. It's important
+        // to do this before the thread state is destroyed so that objects
+        // remain visible to the GC.
+        _mi_heap_collect_abandon(&tstate_impl->mimalloc.heaps[i]);
+    }
+#endif
+}