[3.14] gh-145566: Skip stop-the-world when reassigning __class__ on newly created objects (gh-145567)

(cherry picked from commit 1d091a3)

Co-authored-by: Sam Gross <colesbury@gmail.com>

Author: colesbury

Misc/NEWS.d/next/Core_and_Builtins/2026-03-05-19-10-56.gh-issue-145566.H4RupyYN.rst

@@ -0,0 +1,2 @@
+In the free threading build, skip the stop-the-world pause when reassigning
+``__class__`` on a newly created object.

Objects/mimalloc/alloc.c

@@ -625,6 +625,10 @@ bool _mi_free_delayed_block(mi_block_t* block) {
   }
 
   // collect all other non-local frees to ensure up-to-date `used` count
+  if (page->qsbr_node.next != NULL && (page->local_free != NULL || mi_page_thread_free(page) != NULL)) {
+    static _Atomic(int) _c; int _n = 1+atomic_fetch_add(&_c,1);
+    if (_n%100==0||_n<=3) printf("QSBR CLEAR from _mi_free_delayed_block page=%p all_free=%d used=%d (%d)\n",(void*)page,(int)mi_page_all_free(page),(int)page->used,_n);
+  }
   _mi_page_free_collect(page, false);
 
   // and free the block (possibly freeing the page as well since used is updated)

Objects/mimalloc/page.c

@@ -226,6 +226,13 @@ void _mi_page_free_collect(mi_page_t* page, bool force) {
   // and the local free list
   if (page->local_free != NULL) {
     // any previous QSBR goals are no longer valid because we reused the page
+    if (page->qsbr_node.next != NULL) {
+      extern _Atomic(int) _debug_qsbr_clear_in_collect;
+      int n = 1 + atomic_fetch_add(&_debug_qsbr_clear_in_collect, 1);
+      if (n%100==0||n<=3) printf("QSBR CLEAR generic page=%p all_free=%d used=%d xfree=%d lfree=%d (%d)\n",
+             (void*)page,(int)mi_page_all_free(page),(int)page->used,
+             (mi_page_thread_free(page)!=NULL),(page->local_free!=NULL),n);
+    }
     _PyMem_mi_page_clear_qsbr(page);
 
     if mi_likely(page->free == NULL) {
@@ -371,6 +378,10 @@ static void mi_page_to_full(mi_page_t* page, mi_page_queue_t* pq) {
 
   if (mi_page_is_in_full(page)) return;
   mi_page_queue_enqueue_from(&mi_page_heap(page)->pages[MI_BIN_FULL], pq, page);
+  if (page->qsbr_node.next != NULL && (page->local_free != NULL || mi_page_thread_free(page) != NULL)) {
+    static _Atomic(int) _c; int _n = 1+atomic_fetch_add(&_c,1);
+    if (_n%100==0||_n<=3) printf("QSBR CLEAR from mi_page_to_full page=%p all_free=%d used=%d (%d)\n",(void*)page,(int)mi_page_all_free(page),(int)page->used,_n);
+  }
   _mi_page_free_collect(page,false);  // try to collect right away in case another thread freed just before MI_USE_DELAYED_FREE was set
 }
 
@@ -752,6 +763,10 @@ static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* p
     #endif
 
     // 0. collect freed blocks by us and other threads
+    if (page->qsbr_node.next != NULL && (page->local_free != NULL || mi_page_thread_free(page) != NULL)) {
+      static _Atomic(int) _c; int _n = 1+atomic_fetch_add(&_c,1);
+      if (_n%100==0||_n<=3) printf("QSBR CLEAR from find_free_ex page=%p all_free=%d used=%d (%d)\n",(void*)page,(int)mi_page_all_free(page),(int)page->used,_n);
+    }
     _mi_page_free_collect(page, false);
 
     // 1. if the page contains free blocks, we are done
@@ -777,6 +792,15 @@ static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* p
   mi_heap_stat_counter_increase(heap, searches, count);
 
   if (page == NULL) {
+    {
+      static _Atomic(int) null_page_count;
+      int n = 1 + atomic_fetch_add(&null_page_count, 1);
+      if (n % 50 == 0 || n <= 5) {
+        printf("find_free_ex: page==NULL tid=%zu heap_tid=%zu use_qsbr=%d (call #%d)\n",
+               (size_t)_mi_thread_id(), (size_t)heap->thread_id,
+               heap->page_use_qsbr, n);
+      }
+    }
     _PyMem_mi_heap_collect_qsbr(heap); // some pages might be safe to free now
     _mi_heap_collect_retired(heap, false); // perhaps make a page available?
     page = mi_page_fresh(heap, pq);
@@ -809,6 +833,10 @@ static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
     else
    #endif
     {
+      if (page->qsbr_node.next != NULL && (page->local_free != NULL || mi_page_thread_free(page) != NULL)) {
+        static _Atomic(int) _c; int _n = 1+atomic_fetch_add(&_c,1);
+        if (_n%100==0||_n<=3) printf("QSBR CLEAR from mi_page_fresh_alloc page=%p all_free=%d used=%d (%d)\n",(void*)page,(int)mi_page_all_free(page),(int)page->used,_n);
+      }
       _mi_page_free_collect(page,false);
     }
 

Objects/obmalloc.c

@@ -151,6 +151,8 @@ should_advance_qsbr_for_page(struct _qsbr_thread_state *qsbr, mi_page_t *page)
 }
 #endif
 
+_Atomic(int) _debug_qsbr_clear_in_collect;
+
 static bool
 _PyMem_mi_page_maybe_free(mi_page_t *page, mi_page_queue_t *pq, bool force)
 {
@@ -174,7 +176,19 @@ _PyMem_mi_page_maybe_free(mi_page_t *page, mi_page_queue_t *pq, bool force)
             page->qsbr_goal = _Py_qsbr_shared_next(tstate->qsbr->shared);
         }
 
-        llist_insert_tail(&tstate->mimalloc.page_list, &page->qsbr_node);
+        mi_heap_t *page_heap = mi_page_heap(page);
+        _PyThreadStateImpl *heap_tstate = _Py_CONTAINER_OF(page_heap->tld, _PyThreadStateImpl, mimalloc.tld);
+        if (page_heap->thread_id != _mi_thread_id()) {
+            static _Atomic(int) cross_thread_qsbr_count;
+            int n = 1 + atomic_fetch_add(&cross_thread_qsbr_count, 1);
+            if (n % 100 == 0) {
+                _PyThreadStateImpl *cur_tstate = (_PyThreadStateImpl *)PyThreadState_GET();
+                printf("cross-thread QSBR page count: %d (page_tid=%zu cur_tid=%zu heap_tstate=%p cur_tstate=%p)\n",
+                       n, (size_t)page_heap->thread_id, (size_t)_mi_thread_id(),
+                       (void*)heap_tstate, (void*)cur_tstate);
+            }
+        }
+        llist_insert_tail(&heap_tstate->mimalloc.page_list, &page->qsbr_node);
         return false;
     }
 #endif
@@ -212,25 +226,61 @@ _PyMem_mi_heap_collect_qsbr(mi_heap_t *heap)
     _PyThreadStateImpl *tstate = (_PyThreadStateImpl *)_PyThreadState_GET();
     struct llist_node *head = &tstate->mimalloc.page_list;
     if (llist_empty(head)) {
+        static _Atomic(int) empty_qsbr_count;
+        int n = 1 + atomic_fetch_add(&empty_qsbr_count, 1);
+        if (n % 50 == 0 || n <= 5) {
+            _PyThreadStateImpl *heap_ts = _Py_CONTAINER_OF(heap->tld, _PyThreadStateImpl, mimalloc.tld);
+            printf("qsbr_collect EMPTY tid=%zu heap_tid=%zu tstate=%p heap_tstate=%p (call #%d)\n",
+                   (size_t)_mi_thread_id(), (size_t)heap->thread_id,
+                   (void*)tstate, (void*)heap_ts, n);
+        }
         return;
     }
 
+    int freed = 0, not_free = 0, not_reached = 0;
     struct llist_node *node;
     llist_for_each_safe(node, head) {
         mi_page_t *page = llist_data(node, mi_page_t, qsbr_node);
         if (!mi_page_all_free(page)) {
             // We allocated from this page some point after the delayed free
+            not_free++;
             _PyMem_mi_page_clear_qsbr(page);
             continue;
         }
 
         if (!_Py_qsbr_poll(tstate->qsbr, page->qsbr_goal)) {
-            return;
+            not_reached++;
+            // On first failure, log the details
+            if (not_reached == 1) {
+                struct _qsbr_shared *shared = tstate->qsbr->shared;
+                printf("  qsbr FAIL: goal=%llu rd_seq=%llu wr_seq=%llu my_seq=%llu\n",
+                       (unsigned long long)page->qsbr_goal,
+                       (unsigned long long)_Py_atomic_load_uint64(&shared->rd_seq),
+                       (unsigned long long)_Py_atomic_load_uint64(&shared->wr_seq),
+                       (unsigned long long)_Py_atomic_load_uint64(&tstate->qsbr->seq));
+                // scan threads to find the blocker
+                struct _qsbr_pad *array = shared->array;
+                for (Py_ssize_t ii = 0; ii < shared->size; ii++) {
+                    uint64_t s = _Py_atomic_load_uint64(&array[ii].qsbr.seq);
+                    if (s != QSBR_OFFLINE && s < page->qsbr_goal) {
+                        printf("    blocker slot %zd: seq=%llu\n", ii, (unsigned long long)s);
+                    }
+                }
+            }
+            // count remaining
+            while (node->next != head) { not_reached++; node = node->next; }
+            break;
         }
 
+        freed++;
         _PyMem_mi_page_clear_qsbr(page);
         _mi_page_free(page, mi_page_queue_of(page), false);
     }
+    if (freed || not_free || not_reached) {
+        printf("qsbr_collect tid=%zu: freed=%d not_free=%d not_reached=%d heap_tid=%zu\n",
+               (size_t)_mi_thread_id(), freed, not_free, not_reached,
+               (size_t)heap->thread_id);
+    }
 #endif
 }
 

Objects/typeobject.c

@@ -7120,7 +7120,11 @@ object_set_class_world_stopped(PyObject *self, PyTypeObject *newto)
 
             assert(_PyObject_GetManagedDict(self) == dict);
 
-            if (_PyDict_DetachFromObject(dict, self) < 0) {
+            int err;
+            Py_BEGIN_CRITICAL_SECTION(dict);
+            err = _PyDict_DetachFromObject(dict, self);
+            Py_END_CRITICAL_SECTION();
+            if (err < 0) {
                 return -1;
             }
 
@@ -7161,13 +7165,18 @@ object_set_class(PyObject *self, PyObject *value, void *closure)
     }
 
 #ifdef Py_GIL_DISABLED
+    int unique = _PyObject_IsUniquelyReferenced(self);
     PyInterpreterState *interp = _PyInterpreterState_GET();
-    _PyEval_StopTheWorld(interp);
+    if (!unique) {
+        _PyEval_StopTheWorld(interp);
+    }
 #endif
     PyTypeObject *oldto = Py_TYPE(self);
     int res = object_set_class_world_stopped(self, newto);
 #ifdef Py_GIL_DISABLED
-    _PyEval_StartTheWorld(interp);
+    if (!unique) {
+        _PyEval_StartTheWorld(interp);
+    }
 #endif
     if (res == 0) {
         if (oldto->tp_flags & Py_TPFLAGS_HEAPTYPE) {