merge from dev

include/mimalloc/atomic.h

@@ -1,5 +1,5 @@
 /* ----------------------------------------------------------------------------
-Copyright (c) 2018-2023 Microsoft Research, Daan Leijen
+Copyright (c) 2018-2024 Microsoft Research, Daan Leijen
 This is free software; you can redistribute it and/or modify it under the
 terms of the MIT license. A copy of the license can be found in the file
 "LICENSE" at the root of this distribution.
@@ -419,28 +419,7 @@ static inline void mi_atomic_yield(void) {
 
 #if defined(_WIN32)
 
-#if 0
+#if 1
-
-#define mi_lock_t  CRITICAL_SECTION
-
-static inline bool mi_lock_try_acquire(mi_lock_t* lock) {
-  return TryEnterCriticalSection(lock);
-}
-static inline void mi_lock_acquire(mi_lock_t* lock) {
-  EnterCriticalSection(lock);
-}
-static inline void mi_lock_release(mi_lock_t* lock) {
-  LeaveCriticalSection(lock);
-}
-static inline void mi_lock_init(mi_lock_t* lock) {
-  InitializeCriticalSection(lock);
-}
-static inline void mi_lock_done(mi_lock_t* lock) {
-  DeleteCriticalSection(lock);
-}
-
-#else
-
 #define mi_lock_t  SRWLOCK   // slim reader-writer lock
 
 static inline bool mi_lock_try_acquire(mi_lock_t* lock) {
@@ -459,10 +438,31 @@ static inline void mi_lock_done(mi_lock_t* lock) {
   (void)(lock);
 }
 
+#else
+#define mi_lock_t  CRITICAL_SECTION
+
+static inline bool mi_lock_try_acquire(mi_lock_t* lock) {
+  return TryEnterCriticalSection(lock);
+}
+static inline void mi_lock_acquire(mi_lock_t* lock) {
+  EnterCriticalSection(lock);
+}
+static inline void mi_lock_release(mi_lock_t* lock) {
+  LeaveCriticalSection(lock);
+}
+static inline void mi_lock_init(mi_lock_t* lock) {
+  InitializeCriticalSection(lock);
+}
+static inline void mi_lock_done(mi_lock_t* lock) {
+  DeleteCriticalSection(lock);
+}
+
 #endif
 
 #elif defined(MI_USE_PTHREADS)
 
+void _mi_error_message(int err, const char* fmt, ...);
+
 #define mi_lock_t  pthread_mutex_t
 
 static inline bool mi_lock_try_acquire(mi_lock_t* lock) {
@@ -471,7 +471,7 @@ static inline bool mi_lock_try_acquire(mi_lock_t* lock) {
 static inline void mi_lock_acquire(mi_lock_t* lock) {
   const int err = pthread_mutex_lock(lock);
   if (err != 0) {
-    mi_error_message(EFAULT, "internal error: lock cannot be acquired\n");
+    _mi_error_message(err, "internal error: lock cannot be acquired\n");
   }
 }
 static inline void mi_lock_release(mi_lock_t* lock) {
@@ -535,6 +535,4 @@ static inline void mi_lock_done(mi_lock_t* lock) {
 #endif
 
 
-
-
 #endif // MI_ATOMIC_H

include/mimalloc/types.h

@@ -545,6 +545,7 @@ typedef struct mi_subproc_s {
   _Atomic(size_t)       arena_count;                    // current count of arena's
   _Atomic(mi_arena_t*)  arenas[MI_MAX_ARENAS];          // arena's of this sub-process
   mi_lock_t             arena_reserve_lock;             // lock to ensure arena's get reserved one at a time
+  _Atomic(int64_t)      purge_expire;                   // expiration is set if any arenas can be purged
 
   _Atomic(size_t)       abandoned_count[MI_BIN_COUNT];  // total count of abandoned pages for this sub-process
   mi_page_t*            os_abandoned_pages;             // list of pages that OS allocated and not in an arena (only used if `mi_option_visit_abandoned` is on)

src/arena-abandon.c

@@ -0,0 +1,346 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2024, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#if !defined(MI_IN_ARENA_C)
+#error "this file should be included from 'arena.c' (so mi_arena_t is visible)"
+// add includes help an IDE
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "bitmap.h"
+#endif
+
+// Minimal exports for arena-abandoned.
+size_t      mi_arena_id_index(mi_arena_id_t id);
+mi_arena_t* mi_arena_from_index(size_t idx);
+size_t      mi_arena_get_count(void);
+void*       mi_arena_block_start(mi_arena_t* arena, mi_bitmap_index_t bindex);
+bool        mi_arena_memid_indices(mi_memid_t memid, size_t* arena_index, mi_bitmap_index_t* bitmap_index);
+
+/* -----------------------------------------------------------
+  Abandoned blocks/segments:
+
+  _mi_arena_segment_clear_abandoned
+  _mi_arena_segment_mark_abandoned
+
+  This is used to atomically abandon/reclaim segments
+  (and crosses the arena API but it is convenient to have here).
+
+  Abandoned segments still have live blocks; they get reclaimed
+  when a thread frees a block in it, or when a thread needs a fresh
+  segment.
+
+  Abandoned segments are atomically marked in the `block_abandoned`
+  bitmap of arenas. Any segments allocated outside arenas are put
+  in the sub-process `abandoned_os_list`. This list is accessed
+  using locks but this should be uncommon and generally uncontended.
+  Reclaim and visiting either scan through the `block_abandoned`
+  bitmaps of the arena's, or visit the `abandoned_os_list`
+
+  A potentially nicer design is to use arena's for everything
+  and perhaps have virtual arena's to map OS allocated memory
+  but this would lack the "density" of our current arena's. TBC.
+----------------------------------------------------------- */
+
+
+// reclaim a specific OS abandoned segment; `true` on success.
+// sets the thread_id.
+static bool mi_arena_segment_os_clear_abandoned(mi_segment_t* segment, bool take_lock) {
+  mi_assert(segment->memid.memkind != MI_MEM_ARENA);
+  // not in an arena, remove from list of abandoned os segments
+  mi_subproc_t* const subproc = segment->subproc;
+  if (take_lock && !mi_lock_try_acquire(&subproc->abandoned_os_lock)) {
+    return false;  // failed to acquire the lock, we just give up
+  }
+  // remove atomically from the abandoned os list (if possible!)
+  bool reclaimed = false;
+  mi_segment_t* const next = segment->abandoned_os_next;
+  mi_segment_t* const prev = segment->abandoned_os_prev;
+  if (next != NULL || prev != NULL || subproc->abandoned_os_list == segment) {
+    #if MI_DEBUG>3
+    // find ourselves in the abandoned list (and check the count)
+    bool found = false;
+    size_t count = 0;
+    for (mi_segment_t* current = subproc->abandoned_os_list; current != NULL; current = current->abandoned_os_next) {
+      if (current == segment) { found = true; }
+      count++;
+    }
+    mi_assert_internal(found);
+    mi_assert_internal(count == mi_atomic_load_relaxed(&subproc->abandoned_os_list_count));
+    #endif
+    // remove (atomically) from the list and reclaim
+    if (prev != NULL) { prev->abandoned_os_next = next; }
+    else { subproc->abandoned_os_list = next; }
+    if (next != NULL) { next->abandoned_os_prev = prev; }
+    else { subproc->abandoned_os_list_tail = prev; }
+    segment->abandoned_os_next = NULL;
+    segment->abandoned_os_prev = NULL;
+    mi_atomic_decrement_relaxed(&subproc->abandoned_count);
+    mi_atomic_decrement_relaxed(&subproc->abandoned_os_list_count);
+    if (take_lock) { // don't reset the thread_id when iterating
+      mi_atomic_store_release(&segment->thread_id, _mi_thread_id());
+    }
+    reclaimed = true;
+  }
+  if (take_lock) { mi_lock_release(&segment->subproc->abandoned_os_lock); }
+  return reclaimed;
+}
+
+// reclaim a specific abandoned segment; `true` on success.
+// sets the thread_id.
+bool _mi_arena_segment_clear_abandoned(mi_segment_t* segment) {
+  if mi_unlikely(segment->memid.memkind != MI_MEM_ARENA) {
+    return mi_arena_segment_os_clear_abandoned(segment, true /* take lock */);
+  }
+  // arena segment: use the blocks_abandoned bitmap.
+  size_t arena_idx;
+  size_t bitmap_idx;
+  mi_arena_memid_indices(segment->memid, &arena_idx, &bitmap_idx);
+  mi_arena_t* arena = mi_arena_from_index(arena_idx);
+  mi_assert_internal(arena != NULL);
+  // reclaim atomically
+  bool was_marked = _mi_bitmap_unclaim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx);
+  if (was_marked) {
+    mi_assert_internal(mi_atomic_load_acquire(&segment->thread_id) == 0);
+    mi_atomic_decrement_relaxed(&segment->subproc->abandoned_count);
+    mi_atomic_store_release(&segment->thread_id, _mi_thread_id());
+  }
+  // mi_assert_internal(was_marked);
+  mi_assert_internal(!was_marked || _mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
+  //mi_assert_internal(arena->blocks_committed == NULL || _mi_bitmap_is_claimed(arena->blocks_committed, arena->field_count, 1, bitmap_idx));
+  return was_marked;
+}
+
+
+// mark a specific OS segment as abandoned
+static void mi_arena_segment_os_mark_abandoned(mi_segment_t* segment) {
+  mi_assert(segment->memid.memkind != MI_MEM_ARENA);
+  // not in an arena; we use a list of abandoned segments
+  mi_subproc_t* const subproc = segment->subproc;
+  mi_lock(&subproc->abandoned_os_lock) {
+    // push on the tail of the list (important for the visitor)
+    mi_segment_t* prev = subproc->abandoned_os_list_tail;
+    mi_assert_internal(prev == NULL || prev->abandoned_os_next == NULL);
+    mi_assert_internal(segment->abandoned_os_prev == NULL);
+    mi_assert_internal(segment->abandoned_os_next == NULL);
+    if (prev != NULL) { prev->abandoned_os_next = segment; }
+    else { subproc->abandoned_os_list = segment; }
+    subproc->abandoned_os_list_tail = segment;
+    segment->abandoned_os_prev = prev;
+    segment->abandoned_os_next = NULL;
+    mi_atomic_increment_relaxed(&subproc->abandoned_os_list_count);
+    mi_atomic_increment_relaxed(&subproc->abandoned_count);
+    // and release the lock
+  }
+  return;
+}
+
+// mark a specific segment as abandoned
+// clears the thread_id.
+void _mi_arena_segment_mark_abandoned(mi_segment_t* segment)
+{
+  mi_assert_internal(segment->used == segment->abandoned);
+  mi_atomic_store_release(&segment->thread_id, (uintptr_t)0);  // mark as abandoned for multi-thread free's
+  if mi_unlikely(segment->memid.memkind != MI_MEM_ARENA) {
+    mi_arena_segment_os_mark_abandoned(segment);
+    return;
+  }
+  // segment is in an arena, mark it in the arena `blocks_abandoned` bitmap
+  size_t arena_idx;
+  size_t bitmap_idx;
+  mi_arena_memid_indices(segment->memid, &arena_idx, &bitmap_idx);
+  mi_arena_t* arena = mi_arena_from_index(arena_idx);
+  mi_assert_internal(arena != NULL);
+  // set abandonment atomically
+  mi_subproc_t* const subproc = segment->subproc; // don't access the segment after setting it abandoned
+  const bool was_unmarked = _mi_bitmap_claim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx, NULL);
+  if (was_unmarked) { mi_atomic_increment_relaxed(&subproc->abandoned_count); }
+  mi_assert_internal(was_unmarked);
+  mi_assert_internal(_mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
+}
+
+
+/* -----------------------------------------------------------
+  Iterate through the abandoned blocks/segments using a cursor.
+  This is used for reclaiming and abandoned block visiting.
+----------------------------------------------------------- */
+
+// start a cursor at a randomized arena
+void _mi_arena_field_cursor_init(mi_heap_t* heap, mi_subproc_t* subproc, bool visit_all, mi_arena_field_cursor_t* current) {
+  mi_assert_internal(heap == NULL || heap->tld->segments.subproc == subproc);
+  current->bitmap_idx = 0;
+  current->subproc = subproc;
+  current->visit_all = visit_all;
+  current->hold_visit_lock = false;
+  const size_t abandoned_count = mi_atomic_load_relaxed(&subproc->abandoned_count);
+  const size_t abandoned_list_count = mi_atomic_load_relaxed(&subproc->abandoned_os_list_count);
+  const size_t max_arena = mi_arena_get_count();
+  if (heap != NULL && heap->arena_id != _mi_arena_id_none()) {
+    // for a heap that is bound to one arena, only visit that arena
+    current->start = mi_arena_id_index(heap->arena_id);
+    current->end = current->start + 1;
+    current->os_list_count = 0;
+  }
+  else {
+    // otherwise visit all starting at a random location
+    if (abandoned_count > abandoned_list_count && max_arena > 0) {
+      current->start = (heap == NULL || max_arena == 0 ? 0 : (mi_arena_id_t)(_mi_heap_random_next(heap) % max_arena));
+      current->end = current->start + max_arena;
+    }
+    else {
+      current->start = 0;
+      current->end = 0;
+    }
+    current->os_list_count = abandoned_list_count; // max entries to visit in the os abandoned list
+  }
+  mi_assert_internal(current->start <= max_arena);
+}
+
+void _mi_arena_field_cursor_done(mi_arena_field_cursor_t* current) {
+  if (current->hold_visit_lock) {
+    mi_lock_release(&current->subproc->abandoned_os_visit_lock);
+    current->hold_visit_lock = false;
+  }
+}
+
+static mi_segment_t* mi_arena_segment_clear_abandoned_at(mi_arena_t* arena, mi_subproc_t* subproc, mi_bitmap_index_t bitmap_idx) {
+  // try to reclaim an abandoned segment in the arena atomically
+  if (!_mi_bitmap_unclaim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx)) return NULL;
+  mi_assert_internal(_mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
+  mi_segment_t* segment = (mi_segment_t*)mi_arena_block_start(arena, bitmap_idx);
+  mi_assert_internal(mi_atomic_load_relaxed(&segment->thread_id) == 0);
+  // check that the segment belongs to our sub-process
+  // note: this is the reason we need the `abandoned_visit` lock in the case abandoned visiting is enabled.
+  //  without the lock an abandoned visit may otherwise fail to visit all abandoned segments in the sub-process.
+  //  for regular reclaim it is fine to miss one sometimes so without abandoned visiting we don't need the `abandoned_visit` lock.
+  if (segment->subproc != subproc) {
+    // it is from another sub-process, re-mark it and continue searching
+    const bool was_zero = _mi_bitmap_claim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx, NULL);
+    mi_assert_internal(was_zero); MI_UNUSED(was_zero);
+    return NULL;
+  }
+  else {
+    // success, we unabandoned a segment in our sub-process
+    mi_atomic_decrement_relaxed(&subproc->abandoned_count);
+    return segment;
+  }
+}
+
+static mi_segment_t* mi_arena_segment_clear_abandoned_next_field(mi_arena_field_cursor_t* previous) {
+  const size_t max_arena = mi_arena_get_count();
+  size_t field_idx = mi_bitmap_index_field(previous->bitmap_idx);
+  size_t bit_idx = mi_bitmap_index_bit_in_field(previous->bitmap_idx);
+  // visit arena's (from the previous cursor)
+  for (; previous->start < previous->end; previous->start++, field_idx = 0, bit_idx = 0) {
+    // index wraps around
+    size_t arena_idx = (previous->start >= max_arena ? previous->start % max_arena : previous->start);
+    mi_arena_t* arena = mi_arena_from_index(arena_idx);
+    if (arena != NULL) {
+      bool has_lock = false;
+      // visit the abandoned fields (starting at previous_idx)
+      for (; field_idx < arena->field_count; field_idx++, bit_idx = 0) {
+        size_t field = mi_atomic_load_relaxed(&arena->blocks_abandoned[field_idx]);
+        if mi_unlikely(field != 0) { // skip zero fields quickly
+          // we only take the arena lock if there are actually abandoned segments present
+          if (!has_lock && mi_option_is_enabled(mi_option_visit_abandoned)) {
+            has_lock = (previous->visit_all ? (mi_lock_acquire(&arena->abandoned_visit_lock),true) : mi_lock_try_acquire(&arena->abandoned_visit_lock));
+            if (!has_lock) {
+              if (previous->visit_all) {
+                _mi_error_message(EFAULT, "internal error: failed to visit all abandoned segments due to failure to acquire the visitor lock");
+              }
+              // skip to next arena
+              break;
+            }
+          }
+          mi_assert_internal(has_lock || !mi_option_is_enabled(mi_option_visit_abandoned));
+          // visit each set bit in the field  (todo: maybe use `ctz` here?)
+          for (; bit_idx < MI_BITMAP_FIELD_BITS; bit_idx++) {
+            // pre-check if the bit is set
+            size_t mask = ((size_t)1 << bit_idx);
+            if mi_unlikely((field & mask) == mask) {
+              mi_bitmap_index_t bitmap_idx = mi_bitmap_index_create(field_idx, bit_idx);
+              mi_segment_t* const segment = mi_arena_segment_clear_abandoned_at(arena, previous->subproc, bitmap_idx);
+              if (segment != NULL) {
+                //mi_assert_internal(arena->blocks_committed == NULL || _mi_bitmap_is_claimed(arena->blocks_committed, arena->field_count, 1, bitmap_idx));
+                if (has_lock) { mi_lock_release(&arena->abandoned_visit_lock); }
+                previous->bitmap_idx = mi_bitmap_index_create_ex(field_idx, bit_idx + 1); // start at next one for the next iteration
+                return segment;
+              }
+            }
+          }
+        }
+      }
+      if (has_lock) { mi_lock_release(&arena->abandoned_visit_lock); }
+    }
+  }
+  return NULL;
+}
+
+static mi_segment_t* mi_arena_segment_clear_abandoned_next_list(mi_arena_field_cursor_t* previous) {
+  // go through the abandoned_os_list
+  // we only allow one thread per sub-process to do to visit guarded by the `abandoned_os_visit_lock`.
+  // The lock is released when the cursor is released.
+  if (!previous->hold_visit_lock) {
+    previous->hold_visit_lock = (previous->visit_all ? (mi_lock_acquire(&previous->subproc->abandoned_os_visit_lock),true)
+                                                     : mi_lock_try_acquire(&previous->subproc->abandoned_os_visit_lock));
+    if (!previous->hold_visit_lock) {
+      if (previous->visit_all) {
+        _mi_error_message(EFAULT, "internal error: failed to visit all abandoned segments due to failure to acquire the OS visitor lock");
+      }
+      return NULL; // we cannot get the lock, give up
+    }
+  }
+  // One list entry at a time
+  while (previous->os_list_count > 0) {
+    previous->os_list_count--;
+    mi_lock_acquire(&previous->subproc->abandoned_os_lock); // this could contend with concurrent OS block abandonment and reclaim from `free`
+    mi_segment_t* segment = previous->subproc->abandoned_os_list;
+    // pop from head of the list, a subsequent mark will push at the end (and thus we iterate through os_list_count entries)
+    if (segment == NULL || mi_arena_segment_os_clear_abandoned(segment, false /* we already have the lock */)) {
+      mi_lock_release(&previous->subproc->abandoned_os_lock);
+      return segment;
+    }
+    // already abandoned, try again
+    mi_lock_release(&previous->subproc->abandoned_os_lock);
+  }
+  // done
+  mi_assert_internal(previous->os_list_count == 0);
+  return NULL;
+}
+
+
+// reclaim abandoned segments
+// this does not set the thread id (so it appears as still abandoned)
+mi_segment_t* _mi_arena_segment_clear_abandoned_next(mi_arena_field_cursor_t* previous) {
+  if (previous->start < previous->end) {
+    // walk the arena
+    mi_segment_t* segment = mi_arena_segment_clear_abandoned_next_field(previous);
+    if (segment != NULL) { return segment; }
+  }
+  // no entries in the arena's anymore, walk the abandoned OS list
+  mi_assert_internal(previous->start == previous->end);
+  return mi_arena_segment_clear_abandoned_next_list(previous);
+}
+
+
+bool mi_abandoned_visit_blocks(mi_subproc_id_t subproc_id, int heap_tag, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
+  // (unfortunately) the visit_abandoned option must be enabled from the start.
+  // This is to avoid taking locks if abandoned list visiting is not required (as for most programs)
+  if (!mi_option_is_enabled(mi_option_visit_abandoned)) {
+    _mi_error_message(EFAULT, "internal error: can only visit abandoned blocks when MIMALLOC_VISIT_ABANDONED=ON");
+    return false;
+  }
+  mi_arena_field_cursor_t current;
+  _mi_arena_field_cursor_init(NULL, _mi_subproc_from_id(subproc_id), true /* visit all (blocking) */, &current);
+  mi_segment_t* segment;
+  bool ok = true;
+  while (ok && (segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL) {
+    ok = _mi_segment_visit_blocks(segment, heap_tag, visit_blocks, visitor, arg);
+    _mi_arena_segment_mark_abandoned(segment);
+  }
+  _mi_arena_field_cursor_done(&current);
+  return ok;
+}

src/arena.c

@@ -43,7 +43,6 @@ typedef struct mi_arena_s {
   bool                is_exclusive;         // only allow allocations if specifically for this arena
   bool                is_large;             // memory area consists of large- or huge OS pages (always committed)
   _Atomic(mi_msecs_t) purge_expire;         // expiration time when slices can be purged from `slices_purge`.
-  _Atomic(mi_msecs_t) purge_expire_extend;  // the purge expiration may be extended by a bit
 
   mi_bitmap_t*        slices_free;          // is the slice free?
   mi_bitmap_t*        slices_committed;     // is the slice committed? (i.e. accessible)
@@ -56,14 +55,6 @@ typedef struct mi_arena_s {
   // note: when adding bitmaps revise `mi_arena_info_slices_needed`
 } mi_arena_t;
 
-// Every "page" in `pages_purge` points to purge info 
-// (since we use it for any free'd range and not just for pages)
-typedef struct mi_purge_info_s {
-  _Atomic(mi_msecs_t)  expire;
-  _Atomic(size_t)      slice_count;
-} mi_purge_info_t;
-
-
 
 /* -----------------------------------------------------------
   Arena id's
@@ -1140,7 +1131,6 @@ static bool mi_manage_os_memory_ex2(mi_subproc_t* subproc, void* start, size_t s
   arena->numa_node    = numa_node; // TODO: or get the current numa node if -1? (now it allows anyone to allocate on -1)
   arena->is_large     = is_large;
   arena->purge_expire = 0;
-  arena->purge_expire_extend = 0;
   // mi_lock_init(&arena->abandoned_visit_lock);
 
   // init bitmaps
@@ -1455,14 +1445,15 @@ static void mi_arena_schedule_purge(mi_arena_t* arena, size_t slice_index, size_
   }
   else {
     // schedule purge
+    const mi_msecs_t expire = _mi_clock_now() + delay;
     mi_msecs_t expire0 = 0;
-    if (mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire0, _mi_clock_now() + delay)) {
+    if (mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire0, expire)) {
       // expiration was not yet set
-      mi_atomic_storei64_release(&arena->purge_expire_extend, 0);      
+      // maybe set the global arenas expire as well (if it wasn't set already)
+      mi_atomic_casi64_strong_acq_rel(&arena->subproc->purge_expire, &expire0, expire);
     }
-    else if (mi_atomic_loadi64_acquire(&arena->purge_expire_extend) < 10*delay) {     // limit max extension time
+    else {
       // already an expiration was set
-      mi_atomic_addi64_acq_rel(&arena->purge_expire_extend, (mi_msecs_t)(delay/10));  // add smallish extra delay
     }
     mi_bitmap_setN(arena->slices_purge, slice_index, slice_count, NULL);
   }
@@ -1515,18 +1506,13 @@ static bool mi_arena_try_purge(mi_arena_t* arena, mi_msecs_t now, bool force)
 {
   // check pre-conditions
   if (arena->memid.is_pinned) return false;
-  mi_msecs_t expire_base = mi_atomic_loadi64_relaxed(&arena->purge_expire);
-  mi_msecs_t expire_extend = mi_atomic_loadi64_relaxed(&arena->purge_expire_extend);
-  const mi_msecs_t expire = expire_base + expire_extend;
-  if (expire == 0) return false;
 
   // expired yet?
-  if (!force && expire > now) return false;
+  mi_msecs_t expire = mi_atomic_loadi64_relaxed(&arena->purge_expire);
+  if (!force && (expire == 0 || expire > now)) return false;
 
   // reset expire (if not already set concurrently)
-  if (mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire_base, (mi_msecs_t)0)) {
+  mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire, (mi_msecs_t)0);
-    mi_atomic_storei64_release(&arena->purge_expire_extend, (mi_msecs_t)0); // and also reset the extend
-  }
   mi_subproc_stat_counter_increase(arena->subproc, arena_purges, 1);
 
   // go through all purge info's  (with max MI_BFIELD_BITS ranges at a time)
@@ -1543,8 +1529,13 @@ static void mi_arenas_try_purge(bool force, bool visit_all)
 {
   if (_mi_preloading() || mi_arena_purge_delay() <= 0) return;  // nothing will be scheduled
 
+  // check if any arena needs purging?
   mi_tld_t* tld = _mi_tld();
   mi_subproc_t* subproc = tld->subproc;
+  const mi_msecs_t now = _mi_clock_now();
+  mi_msecs_t arenas_expire = mi_atomic_load_acquire(&subproc->purge_expire);
+  if (!force && (arenas_expire == 0 || arenas_expire < now)) return;
+
   const size_t max_arena = mi_arenas_get_count(subproc);
   if (max_arena == 0) return;
 
@@ -1552,20 +1543,28 @@ static void mi_arenas_try_purge(bool force, bool visit_all)
   static mi_atomic_guard_t purge_guard;
   mi_atomic_guard(&purge_guard)
   {
-    const mi_msecs_t now = _mi_clock_now();
+    // increase global expire: at most one purge per delay cycle
+    mi_atomic_store_release(&subproc->purge_expire, now + mi_arena_purge_delay());
     const size_t arena_start = tld->thread_seq % max_arena;
-    size_t max_purge_count = (visit_all ? max_arena : 1);
+    size_t max_purge_count = (visit_all ? max_arena : 2);
+    bool all_visited = true;
     for (size_t _i = 0; _i < max_arena; _i++) {
       size_t i = _i + arena_start;
       if (i >= max_arena) { i -= max_arena; }
       mi_arena_t* arena = mi_arena_from_index(subproc,i);
       if (arena != NULL) {
         if (mi_arena_try_purge(arena, now, force)) {
-          if (max_purge_count <= 1) break;
+          if (max_purge_count <= 1) {
+            all_visited = false;
+            break;
+          }
           max_purge_count--;
         }
       }
     }
+    if (all_visited) {
+      mi_atomic_store_release(&subproc->purge_expire, (mi_msecs_t)0);
+    }
   }
 }