gh-115103: Delay reuse of mimalloc pages that store PyObjects (#115435)

This implements the delayed reuse of mimalloc pages that contain Python
objects in the free-threaded build.

Allocations of the same size class are grouped in data structures called
pages. These are different from operating system pages. For thread-safety, we
want to ensure that memory used to store PyObjects remains valid as long as
there may be concurrent lock-free readers; we want to delay using it for
other size classes, in other heaps, or returning it to the operating system.

When a mimalloc page becomes empty, instead of immediately freeing it, we tag
it with a QSBR goal and insert it into a per-thread state linked list of
pages to be freed. When mimalloc needs a fresh page, we process the queue and
free any still empty pages that are now deemed safe to be freed. Pages
waiting to be freed are still available for allocations of the same size
class and allocating from a page prevent it from being freed. There is
additional logic to handle abandoned pages when threads exit.

Objects/mimalloc/heap.c

@@ -98,7 +98,10 @@ static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t
   if (mi_page_all_free(page)) {
     // no more used blocks, free the page.
     // note: this will free retired pages as well.
-    _mi_page_free(page, pq, collect >= MI_FORCE);
+    bool freed = _PyMem_mi_page_maybe_free(page, pq, collect >= MI_FORCE);
+    if (!freed && collect == MI_ABANDON) {
+      _mi_page_abandon(page, pq);
+    }
   }
   else if (collect == MI_ABANDON) {
     // still used blocks but the thread is done; abandon the page
@@ -153,6 +156,9 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
   // collect retired pages
   _mi_heap_collect_retired(heap, force);
 
+  // free pages that were delayed with QSBR
+  _PyMem_mi_heap_collect_qsbr(heap);
+
   // collect all pages owned by this thread
   mi_heap_visit_pages(heap, &mi_heap_page_collect, &collect, NULL);
   mi_assert_internal( collect != MI_ABANDON || mi_atomic_load_ptr_acquire(mi_block_t,&heap->thread_delayed_free) == NULL );

Objects/mimalloc/page.c

@@ -225,6 +225,9 @@ 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+    // any previous QSBR goals are no longer valid because we reused the page
+    _PyMem_mi_page_clear_qsbr(page);
+
     if mi_likely(page->free == NULL) {
       // usual case
       page->free = page->local_free;
@@ -267,6 +270,7 @@ void _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page) {
   // TODO: push on full queue immediately if it is full?
   mi_page_queue_t* pq = mi_page_queue(heap, mi_page_block_size(page));
   mi_page_queue_push(heap, pq, page);
+  _PyMem_mi_page_reclaimed(page);
   mi_assert_expensive(_mi_page_is_valid(page));
 }
 
@@ -383,6 +387,13 @@ void _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq) {
 
   mi_heap_t* pheap = mi_page_heap(page);
 
+#ifdef Py_GIL_DISABLED
+  if (page->qsbr_node.next != NULL) {
+    // remove from QSBR queue, but keep the goal
+    llist_remove(&page->qsbr_node);
+  }
+#endif
+
   // remove from our page list
   mi_segments_tld_t* segments_tld = &pheap->tld->segments;
   mi_page_queue_remove(pq, page);
@@ -417,6 +428,11 @@ void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
 
   mi_heap_t* heap = mi_page_heap(page);
 
+#ifdef Py_GIL_DISABLED
+  mi_assert_internal(page->qsbr_goal == 0);
+  mi_assert_internal(page->qsbr_node.next == NULL);
+#endif
+
   // remove from the page list
   // (no need to do _mi_heap_delayed_free first as all blocks are already free)
   mi_segments_tld_t* segments_tld = &heap->tld->segments;
@@ -444,6 +460,9 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
 
   mi_page_set_has_aligned(page, false);
 
+  // any previous QSBR goals are no longer valid because we reused the page
+  _PyMem_mi_page_clear_qsbr(page);
+
   // don't retire too often..
   // (or we end up retiring and re-allocating most of the time)
   // NOTE: refine this more: we should not retire if this
@@ -465,7 +484,7 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
       return; // dont't free after all
     }
   }
-  _mi_page_free(page, pq, false);
+  _PyMem_mi_page_maybe_free(page, pq, false);
 }
 
 // free retired pages: we don't need to look at the entire queues
@@ -480,7 +499,10 @@ void _mi_heap_collect_retired(mi_heap_t* heap, bool force) {
       if (mi_page_all_free(page)) {
         page->retire_expire--;
         if (force || page->retire_expire == 0) {
-          _mi_page_free(pq->first, pq, force);
+#ifdef Py_GIL_DISABLED
+          mi_assert_internal(page->qsbr_goal == 0);
+#endif
+          _PyMem_mi_page_maybe_free(page, pq, force);
         }
         else {
           // keep retired, update min/max
@@ -661,6 +683,7 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
   // set fields
   mi_page_set_heap(page, heap);
   page->tag = heap->tag;
+  page->use_qsbr = heap->page_use_qsbr;
   page->debug_offset = heap->debug_offset;
   page->xblock_size = (block_size < MI_HUGE_BLOCK_SIZE ? (uint32_t)block_size : MI_HUGE_BLOCK_SIZE); // initialize before _mi_segment_page_start
   size_t page_size;
@@ -691,6 +714,10 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
   mi_assert_internal(page->xthread_free == 0);
   mi_assert_internal(page->next == NULL);
   mi_assert_internal(page->prev == NULL);
+#ifdef Py_GIL_DISABLED
+  mi_assert_internal(page->qsbr_goal == 0);
+  mi_assert_internal(page->qsbr_node.next == NULL);
+#endif
   mi_assert_internal(page->retire_expire == 0);
   mi_assert_internal(!mi_page_has_aligned(page));
   #if (MI_PADDING || MI_ENCODE_FREELIST)
@@ -750,6 +777,7 @@ 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) {
+    _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);
     if (page == NULL && first_try) {
@@ -760,6 +788,7 @@ static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* p
   else {
     mi_assert(pq->first == page);
     page->retire_expire = 0;
+    _PyMem_mi_page_clear_qsbr(page);
   }
   mi_assert_internal(page == NULL || mi_page_immediate_available(page));
   return page;
@@ -785,6 +814,7 @@ static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
 
     if (mi_page_immediate_available(page)) {
       page->retire_expire = 0;
+      _PyMem_mi_page_clear_qsbr(page);
       return page; // fast path
     }
   }
@@ -878,6 +908,7 @@ static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size, size_t huge_alignme
       return NULL;
     }
     else {
+      _PyMem_mi_heap_collect_qsbr(heap);
       return mi_large_huge_page_alloc(heap,size,huge_alignment);
     }
   }

Objects/mimalloc/segment.c

@@ -982,6 +982,10 @@ static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld
   mi_assert_internal(mi_page_all_free(page));
   mi_segment_t* segment = _mi_ptr_segment(page);
   mi_assert_internal(segment->used > 0);
+#ifdef Py_GIL_DISABLED
+  mi_assert_internal(page->qsbr_goal == 0);
+  mi_assert_internal(page->qsbr_node.next == NULL);
+#endif
 
   size_t inuse = page->capacity * mi_page_block_size(page);
   _mi_stat_decrease(&tld->stats->page_committed, inuse);
@@ -1270,10 +1274,13 @@ static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, s
       // ensure used count is up to date and collect potential concurrent frees
       mi_page_t* const page = mi_slice_to_page(slice);
       _mi_page_free_collect(page, false);
-      if (mi_page_all_free(page)) {
+      if (mi_page_all_free(page) && _PyMem_mi_page_is_safe_to_free(page)) {
         // if this page is all free now, free it without adding to any queues (yet)
         mi_assert_internal(page->next == NULL && page->prev==NULL);
         _mi_stat_decrease(&tld->stats->pages_abandoned, 1);
+#ifdef Py_GIL_DISABLED
+        page->qsbr_goal = 0;
+#endif
         segment->abandoned--;
         slice = mi_segment_page_clear(page, tld); // re-assign slice due to coalesce!
         mi_assert_internal(!mi_slice_is_used(slice));
@@ -1344,15 +1351,18 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
       mi_page_set_heap(page, target_heap);
       _mi_page_use_delayed_free(page, MI_USE_DELAYED_FREE, true); // override never (after heap is set)
       _mi_page_free_collect(page, false); // ensure used count is up to date
-      if (mi_page_all_free(page)) {
+      if (mi_page_all_free(page) && _PyMem_mi_page_is_safe_to_free(page)) {
         // if everything free by now, free the page
+#ifdef Py_GIL_DISABLED
+        page->qsbr_goal = 0;
+#endif
         slice = mi_segment_page_clear(page, tld);   // set slice again due to coalesceing
       }
       else {
         // otherwise reclaim it into the heap
         _mi_page_reclaim(target_heap, page);
         if (requested_block_size == page->xblock_size && mi_page_has_any_available(page) &&
-            heap == target_heap) {
+            requested_block_size <= MI_MEDIUM_OBJ_SIZE_MAX && heap == target_heap) {
           if (right_page_reclaimed != NULL) { *right_page_reclaimed = true; }
         }
       }