pythongh-112532: Improve mimalloc page visiting (python#114133)

colesbury · Glyphack · commit 8c906f4b6d67 · 2024-01-27T10:36:34.000+01:00
This adds support for visiting abandoned pages in mimalloc and improves
the performance of the page visiting code. Abandoned pages contain
memory blocks from threads that have exited. At some point, they may be
later reclaimed by other threads. We still need to visit those pages in
the free-threaded GC because they contain live objects.

This also reduces the overhead of visiting mimalloc pages:

 * Special cases for full, empty, and pages containing only a single
   block.
 * Fix free_map to use one bit instead of one byte per block.
 * Use fast integer division by a constant algorithm when computing
   block offset from block size and index.
diff --git a/Include/internal/mimalloc/mimalloc/internal.h b/Include/internal/mimalloc/mimalloc/internal.h
@@ -120,6 +120,8 @@ void       _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t*
 void       _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
 bool       _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segments_tld_t* tld);
 void       _mi_segment_thread_collect(mi_segments_tld_t* tld);
+bool       _mi_abandoned_pool_visit_blocks(mi_abandoned_pool_t* pool, uint8_t page_tag, bool visit_blocks, mi_block_visit_fun* visitor, void* arg);
+
 
 #if MI_HUGE_PAGE_ABANDON
 void       _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
@@ -161,6 +163,8 @@ void       _mi_heap_collect_abandon(mi_heap_t* heap);
 void       _mi_heap_set_default_direct(mi_heap_t* heap);
 bool       _mi_heap_memid_is_suitable(mi_heap_t* heap, mi_memid_t memid);
 void       _mi_heap_unsafe_destroy_all(void);
+void       _mi_heap_area_init(mi_heap_area_t* area, mi_page_t* page);
+bool       _mi_heap_area_visit_blocks(const mi_heap_area_t* area, mi_page_t *page, mi_block_visit_fun* visitor, void* arg);
 
 // "stats.c"
 void       _mi_stats_done(mi_stats_t* stats);
diff --git a/Objects/mimalloc/heap.c b/Objects/mimalloc/heap.c
@@ -26,7 +26,7 @@ typedef bool (heap_page_visitor_fun)(mi_heap_t* heap, mi_page_queue_t* pq, mi_pa
 // Visit all pages in a heap; returns `false` if break was called.
 static bool mi_heap_visit_pages(mi_heap_t* heap, heap_page_visitor_fun* fn, void* arg1, void* arg2)
 {
-  if (heap==NULL || heap->page_count==0) return 0;
+  if (heap==NULL || heap->page_count==0) return true;
 
   // visit all pages
   #if MI_DEBUG>1
@@ -521,11 +521,20 @@ typedef struct mi_heap_area_ex_s {
   mi_page_t*     page;
 } mi_heap_area_ex_t;
 
-static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_visit_fun* visitor, void* arg) {
-  mi_assert(xarea != NULL);
-  if (xarea==NULL) return true;
-  const mi_heap_area_t* area = &xarea->area;
-  mi_page_t* page = xarea->page;
+static void mi_fast_divisor(size_t divisor, size_t* magic, size_t* shift) {
+  mi_assert_internal(divisor > 0 && divisor <= UINT32_MAX);
+  *shift = MI_INTPTR_BITS - mi_clz(divisor - 1);
+  *magic = (size_t)(((1ULL << 32) * ((1ULL << *shift) - divisor)) / divisor + 1);
+}
+
+static size_t mi_fast_divide(size_t n, size_t magic, size_t shift) {
+  mi_assert_internal(n <= UINT32_MAX);
+  return ((((uint64_t) n * magic) >> 32) + n) >> shift;
+}
+
+bool _mi_heap_area_visit_blocks(const mi_heap_area_t* area, mi_page_t *page, mi_block_visit_fun* visitor, void* arg) {
+  mi_assert(area != NULL);
+  if (area==NULL) return true;
   mi_assert(page != NULL);
   if (page == NULL) return true;
 
@@ -537,17 +546,39 @@ static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_v
   const size_t ubsize = mi_page_usable_block_size(page); // without padding
   size_t   psize;
   uint8_t* pstart = _mi_page_start(_mi_page_segment(page), page, &psize);
+  mi_heap_t* heap = mi_page_heap(page);
 
   if (page->capacity == 1) {
     // optimize page with one block
     mi_assert_internal(page->used == 1 && page->free == NULL);
-    return visitor(mi_page_heap(page), area, pstart, ubsize, arg);
+    return visitor(heap, area, pstart, ubsize, arg);
+  }
+
+  if (page->used == page->capacity) {
+    // optimize full pages
+    uint8_t* block = pstart;
+    for (size_t i = 0; i < page->capacity; i++) {
+        if (!visitor(heap, area, block, ubsize, arg)) return false;
+        block += bsize;
+    }
+    return true;
   }
 
   // create a bitmap of free blocks.
   #define MI_MAX_BLOCKS   (MI_SMALL_PAGE_SIZE / sizeof(void*))
-  uintptr_t free_map[MI_MAX_BLOCKS / sizeof(uintptr_t)];
-  memset(free_map, 0, sizeof(free_map));
+  uintptr_t free_map[MI_MAX_BLOCKS / MI_INTPTR_BITS];
+  size_t bmapsize = (page->capacity + MI_INTPTR_BITS - 1) / MI_INTPTR_BITS;
+  memset(free_map, 0, bmapsize * sizeof(uintptr_t));
+
+  if (page->capacity % MI_INTPTR_BITS != 0) {
+    size_t shift = (page->capacity % MI_INTPTR_BITS);
+    uintptr_t mask = (UINTPTR_MAX << shift);
+    free_map[bmapsize-1] = mask;
+  }
+
+  // fast repeated division by the block size
+  size_t magic, shift;
+  mi_fast_divisor(bsize, &magic, &shift);
 
   #if MI_DEBUG>1
   size_t free_count = 0;
@@ -559,10 +590,11 @@ static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_v
     mi_assert_internal((uint8_t*)block >= pstart && (uint8_t*)block < (pstart + psize));
     size_t offset = (uint8_t*)block - pstart;
     mi_assert_internal(offset % bsize == 0);
-    size_t blockidx = offset / bsize;  // Todo: avoid division?
-    mi_assert_internal( blockidx < MI_MAX_BLOCKS);
-    size_t bitidx = (blockidx / sizeof(uintptr_t));
-    size_t bit = blockidx - (bitidx * sizeof(uintptr_t));
+    size_t blockidx = mi_fast_divide(offset, magic, shift);
+    mi_assert_internal(blockidx == offset / bsize);
+    mi_assert_internal(blockidx < MI_MAX_BLOCKS);
+    size_t bitidx = (blockidx / MI_INTPTR_BITS);
+    size_t bit = blockidx - (bitidx * MI_INTPTR_BITS);
     free_map[bitidx] |= ((uintptr_t)1 << bit);
   }
   mi_assert_internal(page->capacity == (free_count + page->used));
@@ -571,19 +603,29 @@ static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_v
   #if MI_DEBUG>1
   size_t used_count = 0;
   #endif
-  for (size_t i = 0; i < page->capacity; i++) {
-    size_t bitidx = (i / sizeof(uintptr_t));
-    size_t bit = i - (bitidx * sizeof(uintptr_t));
-    uintptr_t m = free_map[bitidx];
-    if (bit == 0 && m == UINTPTR_MAX) {
-      i += (sizeof(uintptr_t) - 1); // skip a run of free blocks
+  uint8_t* block = pstart;
+  for (size_t i = 0; i < bmapsize; i++) {
+    if (free_map[i] == 0) {
+      // every block is in use
+      for (size_t j = 0; j < MI_INTPTR_BITS; j++) {
+        #if MI_DEBUG>1
+        used_count++;
+        #endif
+        if (!visitor(heap, area, block, ubsize, arg)) return false;
+        block += bsize;
+      }
     }
-    else if ((m & ((uintptr_t)1 << bit)) == 0) {
-      #if MI_DEBUG>1
-      used_count++;
-      #endif
-      uint8_t* block = pstart + (i * bsize);
-      if (!visitor(mi_page_heap(page), area, block, ubsize, arg)) return false;
+    else {
+      uintptr_t m = ~free_map[i];
+      while (m) {
+        #if MI_DEBUG>1
+        used_count++;
+        #endif
+        size_t bitidx = mi_ctz(m);
+        if (!visitor(heap, area, block + (bitidx * bsize), ubsize, arg)) return false;
+        m &= m - 1;
+      }
+      block += bsize * MI_INTPTR_BITS;
     }
   }
   mi_assert_internal(page->used == used_count);
@@ -592,21 +634,24 @@ static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_v
 
 typedef bool (mi_heap_area_visit_fun)(const mi_heap_t* heap, const mi_heap_area_ex_t* area, void* arg);
 
+void _mi_heap_area_init(mi_heap_area_t* area, mi_page_t* page) {
+  const size_t bsize = mi_page_block_size(page);
+  const size_t ubsize = mi_page_usable_block_size(page);
+  area->reserved = page->reserved * bsize;
+  area->committed = page->capacity * bsize;
+  area->blocks = _mi_page_start(_mi_page_segment(page), page, NULL);
+  area->used = page->used;   // number of blocks in use (#553)
+  area->block_size = ubsize;
+  area->full_block_size = bsize;
+}
 
 static bool mi_heap_visit_areas_page(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* vfun, void* arg) {
   MI_UNUSED(heap);
   MI_UNUSED(pq);
   mi_heap_area_visit_fun* fun = (mi_heap_area_visit_fun*)vfun;
   mi_heap_area_ex_t xarea;
-  const size_t bsize = mi_page_block_size(page);
-  const size_t ubsize = mi_page_usable_block_size(page);
   xarea.page = page;
-  xarea.area.reserved = page->reserved * bsize;
-  xarea.area.committed = page->capacity * bsize;
-  xarea.area.blocks = _mi_page_start(_mi_page_segment(page), page, NULL);
-  xarea.area.used = page->used;   // number of blocks in use (#553)
-  xarea.area.block_size = ubsize;
-  xarea.area.full_block_size = bsize;
+  _mi_heap_area_init(&xarea.area, page);
   return fun(heap, &xarea, arg);
 }
 
@@ -627,7 +672,7 @@ static bool mi_heap_area_visitor(const mi_heap_t* heap, const mi_heap_area_ex_t*
   mi_visit_blocks_args_t* args = (mi_visit_blocks_args_t*)arg;
   if (!args->visitor(heap, &xarea->area, NULL, xarea->area.block_size, args->arg)) return false;
   if (args->visit_blocks) {
-    return mi_heap_area_visit_blocks(xarea, args->visitor, args->arg);
+    return _mi_heap_area_visit_blocks(&xarea->area, xarea->page, args->visitor, args->arg);
   }
   else {
     return true;
@@ -637,5 +682,6 @@ static bool mi_heap_area_visitor(const mi_heap_t* heap, const mi_heap_area_ex_t*
 // Visit all blocks in a heap
 bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
   mi_visit_blocks_args_t args = { visit_blocks, visitor, arg };
+  _mi_heap_delayed_free_partial((mi_heap_t *)heap);
   return mi_heap_visit_areas(heap, &mi_heap_area_visitor, &args);
 }
diff --git a/Objects/mimalloc/segment.c b/Objects/mimalloc/segment.c
@@ -1614,3 +1614,53 @@ mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t pag
   mi_assert_expensive(page == NULL || mi_segment_is_valid(_mi_page_segment(page),tld));
   return page;
 }
+
+/* -----------------------------------------------------------
+   Visit blocks in abandoned segments
+----------------------------------------------------------- */
+
+static bool mi_segment_visit_page(mi_segment_t* segment, mi_page_t* page, bool visit_blocks, mi_block_visit_fun* visitor, void* arg)
+{
+  mi_heap_area_t area;
+  _mi_heap_area_init(&area, page);
+  if (!visitor(NULL, &area, NULL, area.block_size, arg)) return false;
+  if (visit_blocks) {
+    return _mi_heap_area_visit_blocks(&area, page, visitor, arg);
+  }
+  else {
+    return true;
+  }
+}
+
+static bool mi_segment_visit_pages(mi_segment_t* segment, uint8_t page_tag, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
+  const mi_slice_t* end;
+  mi_slice_t* slice = mi_slices_start_iterate(segment, &end);
+  while (slice < end) {
+    if (mi_slice_is_used(slice)) {
+      mi_page_t* const page = mi_slice_to_page(slice);
+      if (page->tag == page_tag) {
+        if (!mi_segment_visit_page(segment, page, visit_blocks, visitor, arg)) return false;
+      }
+    }
+    slice = slice + slice->slice_count;
+  }
+  return true;
+}
+
+// Visit all blocks in a abandoned segments
+bool _mi_abandoned_pool_visit_blocks(mi_abandoned_pool_t* pool, uint8_t page_tag, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
+  // Note: this is not safe in any other thread is abandoning or claiming segments from the pool
+  mi_segment_t* segment = mi_tagged_segment_ptr(pool->abandoned);
+  while (segment != NULL) {
+    if (!mi_segment_visit_pages(segment, page_tag, visit_blocks, visitor, arg)) return false;
+    segment = segment->abandoned_next;
+  }
+
+  segment = pool->abandoned_visited;
+  while (segment != NULL) {
+    if (!mi_segment_visit_pages(segment, page_tag, visit_blocks, visitor, arg)) return false;
+    segment = segment->abandoned_next;
+  }
+
+  return true;
+}
