Merge branch 'dev-steal' into dev

Author: daanx

ide/vs2022/mimalloc.vcxproj

@@ -116,7 +116,7 @@
       <SDLCheck>true</SDLCheck>
       <ConformanceMode>Default</ConformanceMode>
       <AdditionalIncludeDirectories>../../include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>MI_DEBUG=4;MI_GUARDED=1;%(PreprocessorDefinitions);</PreprocessorDefinitions>
+      <PreprocessorDefinitions>MI_DEBUG=3;MI_GUARDED=0;%(PreprocessorDefinitions);</PreprocessorDefinitions>
       <CompileAs>CompileAsCpp</CompileAs>
       <SupportJustMyCode>false</SupportJustMyCode>
       <LanguageStandard>stdcpp20</LanguageStandard>

include/mimalloc.h

@@ -148,6 +148,7 @@ typedef void (mi_cdecl mi_error_fun)(int err, void* arg);
 mi_decl_export void mi_register_error(mi_error_fun* fun, void* arg);
 
 mi_decl_export void mi_collect(bool force)    mi_attr_noexcept;
+mi_decl_export void mi_collect_reduce(size_t target_thread_owned) mi_attr_noexcept;
 mi_decl_export int  mi_version(void)          mi_attr_noexcept;
 mi_decl_export void mi_stats_reset(void)      mi_attr_noexcept;
 mi_decl_export void mi_stats_merge(void)      mi_attr_noexcept;
@@ -377,6 +378,7 @@ typedef enum mi_option_e {
   mi_option_guarded_precise,            // disregard minimal alignment requirement to always place guarded blocks exactly in front of a guard page (=0)
   mi_option_guarded_sample_rate,        // 1 out of N allocations in the min/max range will be guarded (=1000)
   mi_option_guarded_sample_seed,        // can be set to allow for a (more) deterministic re-execution when a guard page is triggered (=0)
+  mi_option_target_segments_per_thread, // experimental (=0)
   _mi_option_last,
   // legacy option names
   mi_option_large_os_pages = mi_option_allow_large_os_pages,

include/mimalloc/internal.h

@@ -178,6 +178,8 @@ void       _mi_page_retire(mi_page_t* page) mi_attr_noexcept;                  /
 void       _mi_page_unfull(mi_page_t* page);
 void       _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force);   // free the page
 void       _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq);            // abandon the page, to be picked up by another thread...
+void       _mi_page_force_abandon(mi_page_t* page);
+
 void       _mi_heap_delayed_free_all(mi_heap_t* heap);
 bool       _mi_heap_delayed_free_partial(mi_heap_t* heap);
 void       _mi_heap_collect_retired(mi_heap_t* heap, bool force);
@@ -625,9 +627,9 @@ static inline bool mi_heap_malloc_use_guarded(mi_heap_t* heap, size_t size) {
   }
   else {
     // failed size criteria, rewind count (but don't write to an empty heap)
-    if (heap->guarded_sample_rate != 0) { heap->guarded_sample_count = 1; } 
+    if (heap->guarded_sample_rate != 0) { heap->guarded_sample_count = 1; }
     return false;
-  }  
+  }
 }
 
 mi_decl_restrict void* _mi_heap_malloc_guarded(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept;

include/mimalloc/types.h

@@ -416,7 +416,8 @@ typedef struct mi_segment_s {
   // segment fields
   struct mi_segment_s* next;             // must be the first (non-constant) segment field  -- see `segment.c:segment_init`
   struct mi_segment_s* prev;
-  bool                 was_reclaimed;    // true if it was reclaimed (used to limit on-free reclamation)
+  bool                 was_reclaimed;    // true if it was reclaimed (used to limit reclaim-on-free reclamation)
+  bool                 dont_free;        // can be temporarily true to ensure the segment is not freed
 
   size_t               abandoned;        // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`)
   size_t               abandoned_visits; // count how often this segment is visited for reclaiming (to force reclaim if it is too long)

src/options.c

@@ -65,6 +65,7 @@ typedef struct mi_option_desc_s {
 #define MI_DEFAULT_ARENA_EAGER_COMMIT 2
 #endif
 
+// in KiB
 #ifndef MI_DEFAULT_ARENA_RESERVE
  #if (MI_INTPTR_SIZE>4)
   #define MI_DEFAULT_ARENA_RESERVE 1024L*1024L
@@ -156,6 +157,7 @@ static mi_option_desc_t options[_mi_option_last] =
   { MI_DEFAULT_GUARDED_SAMPLE_RATE,
          UNINIT, MI_OPTION(guarded_sample_rate)},       // 1 out of N allocations in the min/max range will be guarded (=4000)
   { 0,   UNINIT, MI_OPTION(guarded_sample_seed)},
+  { 0,   UNINIT, MI_OPTION(target_segments_per_thread) }, // abandon segments beyond this point, or 0 to disable.
 };
 
 static void mi_option_init(mi_option_desc_t* desc);

src/page-queue.c

@@ -259,8 +259,16 @@ static void mi_page_queue_push(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_
   heap->page_count++;
 }
 
+static void mi_page_queue_move_to_front(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_t* page) {
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_assert_internal(mi_page_queue_contains(queue, page));
+  if (queue->first == page) return;
+  mi_page_queue_remove(queue, page);
+  mi_page_queue_push(heap, queue, page);
+  mi_assert_internal(queue->first == page);
+}
 
-static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* from, mi_page_t* page) {
+static void mi_page_queue_enqueue_from_ex(mi_page_queue_t* to, mi_page_queue_t* from, bool enqueue_at_end, mi_page_t* page) {
   mi_assert_internal(page != NULL);
   mi_assert_expensive(mi_page_queue_contains(from, page));
   mi_assert_expensive(!mi_page_queue_contains(to, page));
@@ -273,6 +281,8 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
                      (mi_page_is_huge(page) && mi_page_queue_is_full(to)));
 
   mi_heap_t* heap = mi_page_heap(page);
+
+  // delete from `from`
   if (page->prev != NULL) page->prev->next = page->next;
   if (page->next != NULL) page->next->prev = page->prev;
   if (page == from->last)  from->last = page->prev;
@@ -283,22 +293,59 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
     mi_heap_queue_first_update(heap, from);
   }
 
-  page->prev = to->last;
-  page->next = NULL;
-  if (to->last != NULL) {
-    mi_assert_internal(heap == mi_page_heap(to->last));
-    to->last->next = page;
-    to->last = page;
+  // insert into `to`
+  if (enqueue_at_end) {
+    // enqueue at the end
+    page->prev = to->last;
+    page->next = NULL;
+    if (to->last != NULL) {
+      mi_assert_internal(heap == mi_page_heap(to->last));
+      to->last->next = page;
+      to->last = page;
+    }
+    else {
+      to->first = page;
+      to->last = page;
+      mi_heap_queue_first_update(heap, to);
+    }
   }
   else {
-    to->first = page;
-    to->last = page;
-    mi_heap_queue_first_update(heap, to);
+    if (to->first != NULL) {
+      // enqueue at 2nd place
+      mi_assert_internal(heap == mi_page_heap(to->first));
+      mi_page_t* next = to->first->next;
+      page->prev = to->first;
+      page->next = next;
+      to->first->next = page;
+      if (next != NULL) { 
+        next->prev = page; 
+      }
+      else {
+        to->last = page;
+      }
+    }
+    else {
+      // enqueue at the head (singleton list)
+      page->prev = NULL;
+      page->next = NULL;
+      to->first = page;
+      to->last = page;
+      mi_heap_queue_first_update(heap, to);
+    }
   }
 
   mi_page_set_in_full(page, mi_page_queue_is_full(to));
 }
 
+static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* from, mi_page_t* page) {
+  mi_page_queue_enqueue_from_ex(to, from, true /* enqueue at the end */, page);
+}
+
+static void mi_page_queue_enqueue_from_full(mi_page_queue_t* to, mi_page_queue_t* from, mi_page_t* page) {
+  // note: we could insert at the front to increase reuse, but it slows down certain benchmarks (like `alloc-test`)
+  mi_page_queue_enqueue_from_ex(to, from, false /* enqueue at the end of the `to` queue? */, page);
+}
+
 // Only called from `mi_heap_absorb`.
 size_t _mi_page_queue_append(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_queue_t* append) {
   mi_assert_internal(mi_heap_contains_queue(heap,pq));

src/page.c

@@ -357,7 +357,7 @@ void _mi_page_unfull(mi_page_t* page) {
   mi_page_set_in_full(page, false); // to get the right queue
   mi_page_queue_t* pq = mi_heap_page_queue_of(heap, page);
   mi_page_set_in_full(page, true);
-  mi_page_queue_enqueue_from(pq, pqfull, page);
+  mi_page_queue_enqueue_from_full(pq, pqfull, page);
 }
 
 static void mi_page_to_full(mi_page_t* page, mi_page_queue_t* pq) {
@@ -403,6 +403,27 @@ void _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq) {
   _mi_segment_page_abandon(page,segments_tld);
 }
 
+// force abandon a page
+void _mi_page_force_abandon(mi_page_t* page) {
+  mi_heap_t* heap = mi_page_heap(page);
+  // mark page as not using delayed free
+  _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
+
+  // ensure this page is no longer in the heap delayed free list
+  _mi_heap_delayed_free_all(heap);
+  // We can still access the page meta-info even if it is freed as we ensure 
+  // in `mi_segment_force_abandon` that the segment is not freed (yet)
+  if (page->capacity == 0) return; // it may have been freed now
+
+  // and now unlink it from the page queue and abandon (or free)
+  mi_page_queue_t* pq = mi_heap_page_queue_of(heap, page);
+  if (mi_page_all_free(page)) {
+    _mi_page_free(page, pq, false);
+  }
+  else {
+    _mi_page_abandon(page, pq);
+  }
+}
 
 // Free a page with no more free blocks
 void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
@@ -448,6 +469,7 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
   // how to check this efficiently though...
   // for now, we don't retire if it is the only page left of this size class.
   mi_page_queue_t* pq = mi_page_queue_of(page);
+  #if MI_RETIRE_CYCLES > 0
   const size_t bsize = mi_page_block_size(page);
   if mi_likely( /* bsize < MI_MAX_RETIRE_SIZE && */ !mi_page_queue_is_special(pq)) {  // not full or huge queue?
     if (pq->last==page && pq->first==page) { // the only page in the queue?
@@ -463,7 +485,7 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
       return; // don't free after all
     }
   }
-
+  #endif
   _mi_page_free(page, pq, false);
 }
 
@@ -709,46 +731,94 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
   Find pages with free blocks
 -------------------------------------------------------------*/
 
+// search for a best next page to use for at most N pages (often cut short if immediate blocks are available)
+#define MI_MAX_CANDIDATE_SEARCH  (8)
+
+// is the page not yet used up to its reserved space?
+static bool mi_page_is_expandable(const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_internal(page->capacity <= page->reserved);
+  return (page->capacity < page->reserved);
+}
+
+
 // Find a page with free blocks of `page->block_size`.
 static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* pq, bool first_try)
 {
   // search through the pages in "next fit" order
   #if MI_STAT
   size_t count = 0;
   #endif
+  size_t candidate_count = 0;        // we reset this on the first candidate to limit the search
+  mi_page_t* page_candidate = NULL;  // a page with free space
   mi_page_t* page = pq->first;
+
   while (page != NULL)
   {
     mi_page_t* next = page->next; // remember next
     #if MI_STAT
     count++;
     #endif
+    candidate_count++;
 
-    // 0. collect freed blocks by us and other threads
+    // collect freed blocks by us and other threads
     _mi_page_free_collect(page, false);
 
-    // 1. if the page contains free blocks, we are done
-    if (mi_page_immediate_available(page)) {
-      break;  // pick this one
-    }
+  #if MI_MAX_CANDIDATE_SEARCH > 1
+    // search up to N pages for a best candidate
 
-    // 2. Try to extend
-    if (page->capacity < page->reserved) {
-      mi_page_extend_free(heap, page, heap->tld);
-      mi_assert_internal(mi_page_immediate_available(page));
-      break;
+    // is the local free list non-empty?
+    const bool immediate_available = mi_page_immediate_available(page);
+
+    // if the page is completely full, move it to the `mi_pages_full`
+    // queue so we don't visit long-lived pages too often.
+    if (!immediate_available && !mi_page_is_expandable(page)) {
+      mi_assert_internal(!mi_page_is_in_full(page) && !mi_page_immediate_available(page));
+      mi_page_to_full(page, pq);
+    }
+    else {
+      // the page has free space, make it a candidate
+      // we prefer non-expandable pages with high usage as candidates (to reduce commit, and increase chances of free-ing up pages)
+      if (page_candidate == NULL) {
+        page_candidate = page;
+        candidate_count = 0;
+      }
+      else if (/* !mi_page_is_expandable(page) && */ page->used >= page_candidate->used) {
+        page_candidate = page;
+      }
+      // if we find a non-expandable candidate, or searched for N pages, return with the best candidate
+      if (immediate_available || candidate_count > MI_MAX_CANDIDATE_SEARCH) {
+        mi_assert_internal(page_candidate!=NULL);
+        break;
+      }
+    }
+  #else
+    // first-fit algorithm
+    // If the page contains free blocks, we are done
+    if (mi_page_immediate_available(page) || mi_page_is_expandable(page)) {
+      break;  // pick this one
     }
 
-    // 3. If the page is completely full, move it to the `mi_pages_full`
+    // If the page is completely full, move it to the `mi_pages_full`
     // queue so we don't visit long-lived pages too often.
     mi_assert_internal(!mi_page_is_in_full(page) && !mi_page_immediate_available(page));
     mi_page_to_full(page, pq);
+  #endif
 
     page = next;
   } // for each page
 
   mi_heap_stat_counter_increase(heap, searches, count);
 
+  // set the page to the best candidate
+  if (page_candidate != NULL) {
+    page = page_candidate;
+  }
+  if (page != NULL && !mi_page_immediate_available(page)) {
+    mi_assert_internal(mi_page_is_expandable(page));
+    mi_page_extend_free(heap, page, heap->tld);
+  }
+
   if (page == NULL) {
     _mi_heap_collect_retired(heap, false); // perhaps make a page available
     page = mi_page_fresh(heap, pq);
@@ -758,18 +828,24 @@ 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);
+    // move the page to the front of the queue
+    mi_page_queue_move_to_front(heap, pq, page);
     page->retire_expire = 0;
+    // _mi_heap_collect_retired(heap, false); // update retire counts; note: increases rss on MemoryLoad bench so don't do this
   }
   mi_assert_internal(page == NULL || mi_page_immediate_available(page));
+
+
   return page;
 }
 
 
 
 // Find a page with free blocks of `size`.
 static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
-  mi_page_queue_t* pq = mi_page_queue(heap,size);
+  mi_page_queue_t* pq = mi_page_queue(heap, size);
+
+  // check the first page: we even do this with candidate search or otherwise we re-search every time
   mi_page_t* page = pq->first;
   if (page != NULL) {
    #if (MI_SECURE>=3) // in secure mode, we extend half the time to increase randomness
@@ -788,6 +864,7 @@ static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
       return page; // fast path
     }
   }
+
   return mi_page_queue_find_free_ex(heap, pq, true);
 }