merge from dev, upstream of python/cpython#113263

include/mimalloc/internal.h

@@ -88,6 +88,7 @@ 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
@@ -186,6 +187,7 @@ 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(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/mimalloc/prim.h

@@ -222,7 +222,13 @@ extern bool _mi_process_is_initialized;             // has mi_process_init been
 static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
 
 // Get a unique id for the current thread.
-#if defined(_WIN32)
+#if defined(MI_PRIM_THREAD_ID)
+
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  return MI_PRIM_THREAD_ID();  // used for example by CPython for a free threaded build (see python/cpython#115488)
+}
+
+#elif defined(_WIN32)
 
 #ifndef WIN32_LEAN_AND_MEAN
 #define WIN32_LEAN_AND_MEAN

src/heap.c

@@ -128,6 +128,9 @@ 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);
 
+  // python/cpython#112532: we may be called from a thread that is not the owner of the heap
+  const 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
@@ -135,7 +138,7 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
     #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)
   _mi_abandoned_collect(heap, collect == MI_FORCE /* force? */, &heap->tld->segments);
   
   // if forced, collect thread data cache 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
   }
   
@@ -208,19 +211,29 @@ mi_heap_t* mi_heap_get_backing(void) {
   return bheap;
 }
 
+void _mi_heap_init(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 = tld;
+  heap->thread_id = _mi_thread_id();
+  heap->arena_id = arena_id;
+  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_memcpy_aligned(heap, &_mi_heap_empty, sizeof(mi_heap_t));
+  _mi_heap_init(heap, bheap->tld, arena_id);
-  heap->tld = bheap->tld;
+  // don't reclaim abandoned pages or otherwise destroy is unsafe
-  heap->thread_id = _mi_thread_id();
+  heap->no_reclaim = true;
-  heap->arena_id = arena_id;
-  _mi_random_split(&bheap->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);
-  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;
   heap->tld->heaps = heap;

src/init.c

@@ -286,7 +286,7 @@ void _mi_thread_data_collect(void) {
 }
 
 // Initialize the thread local default heap, called from `mi_thread_init`
-static bool _mi_heap_init(void) {
+static bool _mi_thread_heap_init(void) {
   if (mi_heap_is_initialized(mi_prim_get_default_heap())) return true;
   if (_mi_is_main_thread()) {
     // mi_assert_internal(_mi_heap_main.thread_id != 0);  // can happen on freeBSD where alloc is called before any initialization
@@ -302,26 +302,25 @@ static bool _mi_heap_init(void) {
 
     mi_tld_t*  tld = &td->tld;
     mi_heap_t* heap = &td->heap;
-    _mi_memcpy_aligned(tld, &tld_empty, sizeof(*tld));
+    _mi_tld_init(tld, heap);  // must be before `_mi_heap_init`
-    _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(*heap));
+    _mi_heap_init(heap, tld, _mi_arena_id_none());
-    heap->thread_id = _mi_thread_id();
+    _mi_heap_set_default_direct(heap);   
-    _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;
 }
 
+// initialize thread local data
+void _mi_tld_init(mi_tld_t* tld, mi_heap_t* bheap) {
+  _mi_memcpy_aligned(tld, &tld_empty, sizeof(mi_tld_t));
+  tld->heap_backing = bheap;
+  tld->heaps = bheap;
+  tld->segments.stats = &tld->stats;
+  tld->segments.os = &tld->os;
+  tld->os.stats = &tld->stats;
+}
+
 // Free the thread local default heap (called from `mi_thread_done`)
-static bool _mi_heap_done(mi_heap_t* heap) {
+static bool _mi_thread_heap_done(mi_heap_t* heap) {
   if (!mi_heap_is_initialized(heap)) return true;
 
   // reset default heap
@@ -418,7 +417,7 @@ void mi_thread_init(void) mi_attr_noexcept
   // initialize the thread local default heap
   // (this will call `_mi_heap_set_default_direct` and thus set the
   //  fiber/pthread key to a non-zero value, ensuring `_mi_thread_done` is called)
-  if (_mi_heap_init()) return;  // returns true if already initialized
+  if (_mi_thread_heap_init()) return;  // returns true if already initialized
 
   _mi_stat_increase(&_mi_stats_main.threads, 1);
   mi_atomic_increment_relaxed(&thread_count);
@@ -450,7 +449,7 @@ void _mi_thread_done(mi_heap_t* heap)
   if (heap->thread_id != _mi_thread_id()) return;
 
   // abandon the thread local heap
-  if (_mi_heap_done(heap)) return;  // returns true if already ran
+  if (_mi_thread_heap_done(heap)) return;  // returns true if already ran
 }
 
 void _mi_heap_set_default_direct(mi_heap_t* heap)  {