improved accounting of committed bytes (issue #1035)

include/mimalloc/internal.h

@@ -127,6 +127,7 @@ bool        _mi_os_has_virtual_reserve(void);
 
 bool        _mi_os_reset(void* addr, size_t size);
 bool        _mi_os_commit(void* p, size_t size, bool* is_zero);
+bool        _mi_os_commit_ex(void* addr, size_t size, bool* is_zero, size_t stat_size);
 bool        _mi_os_decommit(void* addr, size_t size);
 bool        _mi_os_protect(void* addr, size_t size);
 bool        _mi_os_unprotect(void* addr, size_t size);
@@ -947,6 +948,21 @@ static inline size_t mi_bsr(size_t x) {
   return (x==0 ? MI_SIZE_BITS : MI_SIZE_BITS - 1 - mi_clz(x));
 }
 
+size_t _mi_popcount_generic(size_t x);
+
+static inline size_t mi_popcount(size_t x) {
+  if (x<=1) return x;
+  if (x==SIZE_MAX) return MI_SIZE_BITS;
+  #if defined(__GNUC__)
+    #if (SIZE_MAX == ULONG_MAX)
+      return __builtin_popcountl(x);
+    #else
+      return __builtin_popcountll(x);
+    #endif
+  #else
+    return _mi_popcount_generic(x);
+  #endif
+}
 
 // ---------------------------------------------------------------------------------
 // Provide our own `_mi_memcpy` for potential performance optimizations.

src/arena.c

@@ -255,7 +255,7 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
 
   // set the dirty bits (todo: no need for an atomic op here?)
   if (arena->memid.initially_zero && arena->blocks_dirty != NULL) {
-    memid->initially_zero = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
+    memid->initially_zero = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL, NULL);
   }
 
   // set commit state
@@ -267,10 +267,14 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
     // commit requested, but the range may not be committed as a whole: ensure it is committed now
     memid->initially_committed = true;
     bool any_uncommitted;
-    _mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted);
+    size_t already_committed = 0;
+    _mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted, &already_committed);
     if (any_uncommitted) {
+      mi_assert_internal(already_committed < needed_bcount);
+      const size_t commit_size = mi_arena_block_size(needed_bcount);
+      const size_t stat_commit_size = commit_size - mi_arena_block_size(already_committed);
       bool commit_zero = false;
-      if (!_mi_os_commit(p, mi_arena_block_size(needed_bcount), &commit_zero)) {
+      if (!_mi_os_commit_ex(p, commit_size, &commit_zero, stat_commit_size)) {
         memid->initially_committed = false;
       }
       else {
@@ -280,7 +284,14 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
   }
   else {
     // no need to commit, but check if already fully committed
-    memid->initially_committed = _mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
+    size_t already_committed = 0;
+    memid->initially_committed = _mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &already_committed);
+    if (!memid->initially_committed && already_committed > 0) {
+      // partially committed: as it will be committed at some time, adjust the stats and pretend the range is fully uncommitted.
+      mi_assert_internal(already_committed < needed_bcount);
+      _mi_stat_decrease(&_mi_stats_main.committed, mi_arena_block_size(already_committed));
+      _mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
+    }
   }
 
   return p;
@@ -464,17 +475,19 @@ static void mi_arena_purge(mi_arena_t* arena, size_t bitmap_idx, size_t blocks)
   const size_t size = mi_arena_block_size(blocks);
   void* const p = mi_arena_block_start(arena, bitmap_idx);
   bool needs_recommit;
-  if (_mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx)) {
+  size_t already_committed = 0;
+  if (_mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx, &already_committed)) {
     // all blocks are committed, we can purge freely
+    mi_assert_internal(already_committed == blocks);
     needs_recommit = _mi_os_purge(p, size);
   }
   else {
     // some blocks are not committed -- this can happen when a partially committed block is freed
     // in `_mi_arena_free` and it is conservatively marked as uncommitted but still scheduled for a purge
-    // we need to ensure we do not try to reset (as that may be invalid for uncommitted memory),
+    // we need to ensure we do not try to reset (as that may be invalid for uncommitted memory).
-    // and also undo the decommit stats (as it was already adjusted)
+    mi_assert_internal(already_committed < blocks);
     mi_assert_internal(mi_option_is_enabled(mi_option_purge_decommits));
-    needs_recommit = _mi_os_purge_ex(p, size, false /* allow reset? */, 0);    
+    needs_recommit = _mi_os_purge_ex(p, size, false /* allow reset? */, mi_arena_block_size(already_committed));    
   }
 
   // clear the purged blocks
@@ -508,7 +521,7 @@ static void mi_arena_schedule_purge(mi_arena_t* arena, size_t bitmap_idx, size_t
     else {
       // already an expiration was set
     }
-    _mi_bitmap_claim_across(arena->blocks_purge, arena->field_count, blocks, bitmap_idx, NULL);
+    _mi_bitmap_claim_across(arena->blocks_purge, arena->field_count, blocks, bitmap_idx, NULL, NULL);
   }
 }
 
@@ -648,7 +661,7 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
   if (p==NULL) return;
   if (size==0) return;
   const bool all_committed = (committed_size == size);
-  const bool decommitted_size = (committed_size <= size ? size - committed_size : 0);
+  const size_t decommitted_size = (committed_size <= size ? size - committed_size : 0);
 
   // need to set all memory to undefined as some parts may still be marked as no_access (like padding etc.)
   mi_track_mem_undefined(p,size);
@@ -691,14 +704,14 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
       mi_assert_internal(arena->blocks_purge != NULL);
 
       if (!all_committed) {
-        // mark the entire range as no longer committed (so we recommit the full range when re-using)
+        // mark the entire range as no longer committed (so we will recommit the full range when re-using)
         _mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
         mi_track_mem_noaccess(p,size);
-        if (committed_size > 0) {
+        //if (committed_size > 0) {
           // if partially committed, adjust the committed stats (is it will be recommitted when re-using)
           // in the delayed purge, we do no longer decrease the commit if the range is not marked entirely as committed.
           _mi_stat_decrease(&_mi_stats_main.committed, committed_size);
-        }
+        //}
         // note: if not all committed, it may be that the purge will reset/decommit the entire range
         // that contains already decommitted parts. Since purge consistently uses reset or decommit that
         // works (as we should never reset decommitted parts).

src/bitmap.c

@@ -351,7 +351,7 @@ bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t
 
 // Set `count` bits at `bitmap_idx` to 1 atomically
 // Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
-bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero, size_t* already_set) {
   size_t idx = mi_bitmap_index_field(bitmap_idx);
   size_t pre_mask;
   size_t mid_mask;
@@ -359,28 +359,31 @@ bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t co
   size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
   bool all_zero = true;
   bool any_zero = false;
+  size_t one_count = 0;
   _Atomic(size_t)*field = &bitmap[idx];
   size_t prev = mi_atomic_or_acq_rel(field++, pre_mask);
-  if ((prev & pre_mask) != 0) all_zero = false;
+  if ((prev & pre_mask) != 0) { all_zero = false; one_count += mi_popcount(prev & pre_mask); }
   if ((prev & pre_mask) != pre_mask) any_zero = true;
   while (mid_count-- > 0) {
     prev = mi_atomic_or_acq_rel(field++, mid_mask);
-    if ((prev & mid_mask) != 0) all_zero = false;
+    if ((prev & mid_mask) != 0) { all_zero = false; one_count += mi_popcount(prev & mid_mask); }
     if ((prev & mid_mask) != mid_mask) any_zero = true;
   }
   if (post_mask!=0) {
     prev = mi_atomic_or_acq_rel(field, post_mask);
-    if ((prev & post_mask) != 0) all_zero = false;
+    if ((prev & post_mask) != 0) { all_zero = false; one_count += mi_popcount(prev & post_mask); }
     if ((prev & post_mask) != post_mask) any_zero = true;
   }
   if (pany_zero != NULL) { *pany_zero = any_zero; }
+  if (already_set != NULL) { *already_set = one_count; };
+  mi_assert_internal(all_zero ? one_count == 0 : one_count <= count);
   return all_zero;
 }
 
 
 // Returns `true` if all `count` bits were 1.
 // `any_ones` is `true` if there was at least one bit set to one.
-static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
+static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones, size_t* already_set) {
   size_t idx = mi_bitmap_index_field(bitmap_idx);
   size_t pre_mask;
   size_t mid_mask;
@@ -388,30 +391,33 @@ static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_field
   size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
   bool all_ones = true;
   bool any_ones = false;
+  size_t one_count = 0;
   mi_bitmap_field_t* field = &bitmap[idx];
   size_t prev = mi_atomic_load_relaxed(field++);
   if ((prev & pre_mask) != pre_mask) all_ones = false;
-  if ((prev & pre_mask) != 0) any_ones = true;
+  if ((prev & pre_mask) != 0) { any_ones = true; one_count += mi_popcount(prev & pre_mask); }
   while (mid_count-- > 0) {
     prev = mi_atomic_load_relaxed(field++);
     if ((prev & mid_mask) != mid_mask) all_ones = false;
-    if ((prev & mid_mask) != 0) any_ones = true;
+    if ((prev & mid_mask) != 0) { any_ones = true; one_count += mi_popcount(prev & mid_mask); }
   }
   if (post_mask!=0) {
     prev = mi_atomic_load_relaxed(field);
     if ((prev & post_mask) != post_mask) all_ones = false;
-    if ((prev & post_mask) != 0) any_ones = true;
+    if ((prev & post_mask) != 0) { any_ones = true; one_count += mi_popcount(prev & post_mask); }
   }
   if (pany_ones != NULL) { *pany_ones = any_ones; }
+  if (already_set != NULL) { *already_set = one_count; }
+  mi_assert_internal(all_ones ? one_count == count : one_count < count);
   return all_ones;
 }
 
-bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, size_t* already_set) {
-  return mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, NULL);
+  return mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, NULL, already_set);
 }
 
 bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   bool any_ones;
-  mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
+  mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, &any_ones, NULL);
   return any_ones;
 }

src/bitmap.h

@@ -102,9 +102,9 @@ bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t
 
 // Set `count` bits at `bitmap_idx` to 1 atomically
 // Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
-bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero);
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero, size_t* already_set);
 
-bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, size_t* already_set);
 bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
 
 #endif

src/libc.c

@@ -275,3 +275,60 @@ int _mi_snprintf(char* buf, size_t buflen, const char* fmt, ...) {
   va_end(args);
   return written;
 }
+
+
+#if MI_SIZE_SIZE == 4
+#define mi_mask_even_bits32      (0x55555555)
+#define mi_mask_even_pairs32     (0x33333333)
+#define mi_mask_even_nibbles32   (0x0F0F0F0F)
+
+// sum of all the bytes in `x` if it is guaranteed that the sum < 256!
+static size_t mi_byte_sum32(uint32_t x) {
+  // perform `x * 0x01010101`: the highest byte contains the sum of all bytes.
+  x += (x << 8);
+  x += (x << 16);
+  return (size_t)(x >> 24);
+}
+
+static size_t mi_popcount_generic32(uint32_t x) {
+  // first count each 2-bit group `a`, where: a==0b00 -> 00, a==0b01 -> 01, a==0b10 -> 01, a==0b11 -> 10
+  // in other words, `a - (a>>1)`; to do this in parallel, we need to mask to prevent spilling a bit pair
+  // into the lower bit-pair:
+  x = x - ((x >> 1) & mi_mask_even_bits32);
+  // add the 2-bit pair results
+  x = (x & mi_mask_even_pairs32) + ((x >> 2) & mi_mask_even_pairs32);
+  // add the 4-bit nibble results
+  x = (x + (x >> 4)) & mi_mask_even_nibbles32;
+  // each byte now has a count of its bits, we can sum them now:
+  return mi_byte_sum32(x);
+}
+
+mi_decl_noinline size_t _mi_popcount_generic(size_t x) {
+  return mi_popcount_generic32(x);
+}
+
+#else
+#define mi_mask_even_bits64      (0x5555555555555555)
+#define mi_mask_even_pairs64     (0x3333333333333333)
+#define mi_mask_even_nibbles64   (0x0F0F0F0F0F0F0F0F)
+
+// sum of all the bytes in `x` if it is guaranteed that the sum < 256!
+static size_t mi_byte_sum64(uint64_t x) {
+  x += (x << 8);
+  x += (x << 16);
+  x += (x << 32);
+  return (size_t)(x >> 56);
+}
+
+static size_t mi_popcount_generic64(uint64_t x) {
+  x = x - ((x >> 1) & mi_mask_even_bits64);
+  x = (x & mi_mask_even_pairs64) + ((x >> 2) & mi_mask_even_pairs64);
+  x = (x + (x >> 4)) & mi_mask_even_nibbles64;
+  return mi_byte_sum64(x);
+}
+
+mi_decl_noinline size_t _mi_popcount_generic(size_t x) {
+  return mi_popcount_generic64(x);
+}
+#endif
+

src/stats.c

@@ -30,6 +30,7 @@ static void mi_stat_update(mi_stat_count_t* stat, int64_t amount) {
   {
     // add atomically (for abandoned pages)
     int64_t current = mi_atomic_addi64_relaxed(&stat->current, amount);
+    // if (stat == &_mi_stats_main.committed) { mi_assert_internal(current + amount >= 0); };
     mi_atomic_maxi64_relaxed(&stat->peak, current + amount);
     if (amount > 0) {
       mi_atomic_addi64_relaxed(&stat->total,amount);