Merge branch 'dev-slice-steal' into dev-slice

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>
@@ -213,9 +213,7 @@
     </ClCompile>
     <ClCompile Include="..\..\src\alloc-posix.c" />
     <ClCompile Include="..\..\src\alloc.c" />
-    <ClCompile Include="..\..\src\arena-abandoned.c">
+    <ClCompile Include="..\..\src\arena-abandon.c">
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
     </ClCompile>

ide/vs2022/mimalloc.vcxproj.filters

@@ -61,7 +61,7 @@
     <ClCompile Include="..\..\src\free.c">
       <Filter>Sources</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\arena-abandoned.c">
+    <ClCompile Include="..\..\src\arena-abandon.c">
       <Filter>Sources</Filter>
     </ClCompile>
   </ItemGroup>

include/mimalloc.h

@@ -149,6 +149,7 @@ 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;
@@ -378,6 +379,7 @@ 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

@@ -181,6 +181,8 @@ 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);

include/mimalloc/types.h

@@ -200,7 +200,7 @@ typedef int32_t  mi_ssize_t;
 #define MI_SMALL_OBJ_SIZE_MAX             (MI_SMALL_PAGE_SIZE/4)   // 8KiB on 64-bit
 #define MI_MEDIUM_OBJ_SIZE_MAX            (MI_MEDIUM_PAGE_SIZE/4)  // 128KiB on 64-bit
 #define MI_MEDIUM_OBJ_WSIZE_MAX           (MI_MEDIUM_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
-#define MI_LARGE_OBJ_SIZE_MAX             (MI_SEGMENT_SIZE/2)      // 32MiB on 64-bit
+#define MI_LARGE_OBJ_SIZE_MAX             (MI_SEGMENT_SIZE/2)      // 16MiB on 64-bit
 #define MI_LARGE_OBJ_WSIZE_MAX            (MI_LARGE_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
 
 // Maximum number of size classes. (spaced exponentially in 12.5% increments)
@@ -475,6 +475,7 @@ typedef struct mi_segment_s {
   // from here is zero initialized
   struct mi_segment_s* next;            // the list of freed segments in the cache (must be first field, see `segment.c:mi_segment_init`)
   bool              was_reclaimed;      // true if it was reclaimed (used to limit 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 during abondoned reclamation (to force reclaim if it takes 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
@@ -89,6 +90,14 @@ typedef struct mi_option_desc_s {
 #define MI_DEFAULT_RESERVE_OS_MEMORY 0
 #endif
 
+#ifndef MI_DEFAULT_GUARDED_SAMPLE_RATE
+#if MI_GUARDED
+#define MI_DEFAULT_GUARDED_SAMPLE_RATE 4000
+#else
+#define MI_DEFAULT_GUARDED_SAMPLE_RATE 0
+#endif
+#endif
+
 
 static mi_option_desc_t options[_mi_option_last] =
 {
@@ -145,12 +154,10 @@ static mi_option_desc_t options[_mi_option_last] =
   { 0,   UNINIT, MI_OPTION(guarded_min) },              // only used when building with MI_GUARDED: minimal rounded object size for guarded objects
   { MI_GiB, UNINIT, MI_OPTION(guarded_max) },           // only used when building with MI_GUARDED: maximal rounded object size for guarded objects
   { 0,   UNINIT, MI_OPTION(guarded_precise) },          // disregard minimal alignment requirement to always place guarded blocks exactly in front of a guard page (=0)
-#if MI_GUARDED
+  { MI_DEFAULT_GUARDED_SAMPLE_RATE,
-  { 4000,UNINIT, MI_OPTION(guarded_sample_rate)},       // 1 out of N allocations in the min/max range will be guarded(= 1000)
+         UNINIT, MI_OPTION(guarded_sample_rate)},       // 1 out of N allocations in the min/max range will be guarded (=4000)
-#else
-  { 0,   UNINIT, MI_OPTION(guarded_sample_rate)},
-#endif
   { 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);
@@ -180,7 +187,7 @@ void _mi_options_init(void) {
       _mi_warning_message("option 'allow_large_os_pages' is disabled to allow for guarded objects\n");
     }
   }
-  _mi_verbose_message("guarded build: %s\n", mi_option_get(mi_option_guarded_max) > 0 ? "enabled" : "disabled");
+  _mi_verbose_message("guarded build: %s\n", mi_option_get(mi_option_guarded_sample_rate) != 0 ? "enabled" : "disabled");
   #endif
 }
 

src/page-queue.c

@@ -264,8 +264,16 @@ 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));
@@ -278,6 +286,8 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
                      (mi_page_is_large_or_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;
@@ -288,6 +298,9 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
     mi_heap_queue_first_update(heap, from);
   }
 
+  // insert into `to`
+  if (enqueue_at_end) {
+    // enqueue at the end
     page->prev = to->last;
     page->next = NULL;
     if (to->last != NULL) {
@@ -300,10 +313,44 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
       to->last = page;
       mi_heap_queue_first_update(heap, to);
     }
+  }
+  else {
+    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

@@ -358,7 +358,7 @@ 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) {
@@ -404,6 +404,28 @@ 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) {
@@ -451,6 +473,7 @@ 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?
@@ -466,6 +489,7 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
       return; // don't free after all
     }
   }
+  #endif
   _mi_page_free(page, pq, false);
 }
 
@@ -712,6 +736,17 @@ 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)
 {
@@ -719,39 +754,76 @@ static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* p
   #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_MAX_CANDIDATE_SEARCH > 1
-    if (mi_page_immediate_available(page)) {
+    // search up to N pages for a best candidate
+
+    // 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
     }
 
-    // 2. Try to extend
+    // If the page is completely full, move it to the `mi_pages_full`
-    if (page->capacity < page->reserved) {
-      mi_page_extend_free(heap, page, heap->tld);
-      mi_assert_internal(mi_page_immediate_available(page));
-      break;
-    }
-
-    // 3. 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);
@@ -761,10 +833,14 @@ 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;
 }
 
@@ -773,6 +849,8 @@ static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* p
 // 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);
+
+  // 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
@@ -791,6 +869,7 @@ 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);
 }
 

src/segment.c

@@ -693,6 +693,8 @@ static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_
       // free previous slice -- remove it from free and merge
       mi_assert_internal(prev->slice_count > 0 && prev->slice_offset==0);
       slice_count += prev->slice_count;
+      slice->slice_count = 0;
+      slice->slice_offset = (uint32_t)((uint8_t*)slice - (uint8_t*)prev); // set the slice offset for `segment_force_abandon` (in case the previous free block is very large).
       if (!is_abandoned) { mi_segment_span_remove_from_queue(prev, tld); }
       slice = prev;
     }
@@ -957,6 +959,9 @@ static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t
   mi_assert_internal(segment->next == NULL);
   mi_assert_internal(segment->used == 0);
   
+  // in `mi_segment_force_abandon` we set this to true to ensure the segment's memory stays valid
+  if (segment->dont_free) return;
+
   // Remove the free pages
   mi_slice_t* slice = &segment->slices[0];
   const mi_slice_t* end = mi_segment_slices_end(segment);
@@ -1259,6 +1264,9 @@ bool _mi_segment_attempt_reclaim(mi_heap_t* heap, mi_segment_t* segment) {
   if (mi_atomic_load_relaxed(&segment->thread_id) != 0) return false;  // it is not abandoned
   if (segment->subproc != heap->tld->segments.subproc)  return false;  // only reclaim within the same subprocess
   if (!_mi_heap_memid_is_suitable(heap,segment->memid)) return false;  // don't reclaim between exclusive and non-exclusive arena's
+  const long target = _mi_option_get_fast(mi_option_target_segments_per_thread);
+  if (target > 0 && (size_t)target <= heap->tld->segments.count) return false; // don't reclaim if going above the target count
+
   // don't reclaim more from a `free` call than half the current segments
   // this is to prevent a pure free-ing thread to start owning too many segments
   // (but not for out-of-arena segments as that is the main way to be reclaimed for those)
@@ -1283,6 +1291,13 @@ void _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld) {
   _mi_arena_field_cursor_done(&current);
 }
 
+
+static bool segment_count_is_within_target(mi_segments_tld_t* tld, size_t* ptarget) {
+  const size_t target = (size_t)mi_option_get_clamp(mi_option_target_segments_per_thread, 0, 1024);
+  if (ptarget != NULL) { *ptarget = target; }
+  return (target == 0 || tld->count < target);
+}
+
 static long mi_segment_get_reclaim_tries(mi_segments_tld_t* tld) {
   // limit the tries to 10% (default) of the abandoned segments with at least 8 and at most 1024 tries.
   const size_t perc = (size_t)mi_option_get_clamp(mi_option_max_segment_reclaim, 0, 100);
@@ -1305,7 +1320,7 @@ static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slice
   mi_segment_t* segment = NULL;
   mi_arena_field_cursor_t current;
   _mi_arena_field_cursor_init(heap, tld->subproc, false /* non-blocking */, &current);
-  while ((max_tries-- > 0) && ((segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL))
+  while (segment_count_is_within_target(tld,NULL) && (max_tries-- > 0) && ((segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL))
   {
     mi_assert(segment->subproc == heap->tld->segments.subproc); // cursor only visits segments in our sub-process
     segment->abandoned_visits++;
@@ -1330,7 +1345,7 @@ static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slice
       break;
     }
     else if (segment->abandoned_visits > 3 && is_suitable) {
-      // always reclaim on 3rd visit to limit the abandoned queue length.
+      // always reclaim on 3rd visit to limit the abandoned segment count.
       mi_segment_reclaim(segment, heap, 0, NULL, tld);
     }
     else {
@@ -1343,7 +1358,7 @@ static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slice
   return result;
 }
 
-
+// collect abandoned segments
 void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
 {
   mi_segment_t* segment;
@@ -1367,6 +1382,103 @@ void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
   _mi_arena_field_cursor_done(&current);
 }
 
+/* -----------------------------------------------------------
+   Force abandon a segment that is in use by our thread
+----------------------------------------------------------- */
+
+// force abandon a segment
+static void mi_segment_force_abandon(mi_segment_t* segment, mi_segments_tld_t* tld)
+{
+  mi_assert_internal(!mi_segment_is_abandoned(segment));
+  mi_assert_internal(!segment->dont_free);
+
+  // ensure the segment does not get free'd underneath us (so we can check if a page has been freed in `mi_page_force_abandon`)
+  segment->dont_free = true;
+
+  // for all slices
+  const mi_slice_t* end;
+  mi_slice_t* slice = mi_slices_start_iterate(segment, &end);
+  while (slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    if (mi_slice_is_used(slice)) {
+      // 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);
+      {
+        // abandon the page if it is still in-use (this will free it if possible as well)
+        mi_assert_internal(segment->used > 0);
+        if (segment->used == segment->abandoned+1) {
+          // the last page.. abandon and return as the segment will be abandoned after this
+          // and we should no longer access it.
+          segment->dont_free = false;
+          _mi_page_force_abandon(page);
+          return;
+        }
+        else {
+          // abandon and continue
+          _mi_page_force_abandon(page);
+          // it might be freed, reset the slice (note: relies on coalesce setting the slice_offset)
+          slice = mi_slice_first(slice);
+        }
+      }
+    }
+    slice = slice + slice->slice_count;
+  }
+  segment->dont_free = false;
+  mi_assert(segment->used == segment->abandoned);
+  mi_assert(segment->used == 0);
+  if (segment->used == 0) {  // paranoia
+    // all free now
+    mi_segment_free(segment, false, tld);
+  }
+  else {
+    // perform delayed purges
+    mi_segment_try_purge(segment, false /* force? */, tld->stats);
+  }
+}
+
+
+// try abandon segments.
+// this should be called from `reclaim_or_alloc` so we know all segments are (about) fully in use.
+static void mi_segments_try_abandon_to_target(mi_heap_t* heap, size_t target, mi_segments_tld_t* tld) {
+  if (target <= 1) return;
+  const size_t min_target = (target > 4 ? (target*3)/4 : target);  // 75%
+  // todo: we should maintain a list of segments per thread; for now, only consider segments from the heap full pages
+  for (int i = 0; i < 64 && tld->count >= min_target; i++) {
+    mi_page_t* page = heap->pages[MI_BIN_FULL].first;
+    while (page != NULL && mi_page_block_size(page) > MI_LARGE_OBJ_SIZE_MAX) {
+      page = page->next;
+    }
+    if (page==NULL) {
+      break;
+    }
+    mi_segment_t* segment = _mi_page_segment(page);
+    mi_segment_force_abandon(segment, tld);
+    mi_assert_internal(page != heap->pages[MI_BIN_FULL].first); // as it is just abandoned
+  }
+}
+
+// try abandon segments.
+// this should be called from `reclaim_or_alloc` so we know all segments are (about) fully in use.
+static void mi_segments_try_abandon(mi_heap_t* heap, mi_segments_tld_t* tld) {
+  // we call this when we are about to add a fresh segment so we should be under our target segment count.
+  size_t target = 0;
+  if (segment_count_is_within_target(tld, &target)) return;
+  mi_segments_try_abandon_to_target(heap, target, tld);
+}
+
+void mi_collect_reduce(size_t target_size) mi_attr_noexcept {
+  mi_collect(true);
+  mi_heap_t* heap = mi_heap_get_default();
+  mi_segments_tld_t* tld = &heap->tld->segments;
+  size_t target = target_size / MI_SEGMENT_SIZE;
+  if (target == 0) {
+    target = (size_t)mi_option_get_clamp(mi_option_target_segments_per_thread, 1, 1024);
+  }
+  mi_segments_try_abandon_to_target(heap, target, tld);
+}
+
 /* -----------------------------------------------------------
    Reclaim or allocate
 ----------------------------------------------------------- */
@@ -1375,6 +1487,9 @@ static mi_segment_t* mi_segment_reclaim_or_alloc(mi_heap_t* heap, size_t needed_
 {
   mi_assert_internal(block_size <= MI_LARGE_OBJ_SIZE_MAX);
 
+  // try to abandon some segments to increase reuse between threads
+  mi_segments_try_abandon(heap,tld);
+
   // 1. try to reclaim an abandoned segment
   bool reclaimed;
   mi_segment_t* segment = mi_segment_try_reclaim(heap, needed_slices, block_size, &reclaimed, tld);