merge from dev-reset

include/mimalloc/internal.h

@@ -88,8 +88,10 @@ void       _mi_thread_data_collect(void);
 
 // os.c
 void       _mi_os_init(void);                                            // called from process init
-void*      _mi_os_alloc(size_t size, bool* is_zero, mi_stats_t* stats);  // to allocate thread local data
+void*      _mi_os_alloc(size_t size, bool* is_zero, mi_stats_t* stats);  
-void       _mi_os_free(void* p, size_t size, mi_stats_t* stats);         // to free thread local data
+void       _mi_os_free(void* p, size_t size, mi_stats_t* stats);
+void       _mi_os_free_ex(void* p, size_t size, bool is_committed, mi_stats_t* stats);
+
 size_t     _mi_os_page_size(void);
 size_t     _mi_os_good_alloc_size(size_t size);
 bool       _mi_os_has_overcommit(void);
@@ -104,16 +106,16 @@ bool       _mi_os_unprotect(void* addr, size_t size);
 bool       _mi_os_purge(void* p, size_t size, mi_stats_t* stats);
 bool       _mi_os_purge_ex(void* p, size_t size, bool allow_reset, mi_stats_t* stats);
 
-void*      _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, bool* is_zero, mi_stats_t* stats);
+void*      _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* stats);
-void*      _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t align_offset, bool commit, bool* large, bool* is_zero, mi_stats_t* tld_stats);
+void*      _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t align_offset, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* tld_stats);
-void       _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats);
+void       _mi_os_free_aligned_at_offset(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats);
+
 void*      _mi_os_get_aligned_hint(size_t try_alignment, size_t size);
 bool       _mi_os_use_large_page(size_t size, size_t alignment);
 size_t     _mi_os_large_page_size(void);
 
-void       _mi_os_free_ex(void* p, size_t size, bool was_committed, mi_stats_t* stats);
 void*      _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_secs, size_t* pages_reserved, size_t* psize, bool* is_zero);
-void       _mi_os_free_huge_pages(void* p, size_t size, mi_stats_t* stats);
+void       _mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats);
 
 // arena.c
 mi_arena_id_t _mi_arena_id_none(void);
@@ -918,6 +920,8 @@ static inline void _mi_memzero(void* dst, size_t n) {
 }
 #endif
 
+// initialize a local variable to zero; use memset as compilers optimize constant sized memset's
+#define _mi_memzero_var(x)  memset(&x,0,sizeof(x))
 
 // -------------------------------------------------------------------------------
 // The `_mi_memcpy_aligned` can be used if the pointers are machine-word aligned

include/mimalloc/prim.h

@@ -26,8 +26,8 @@ typedef struct mi_os_mem_config_s {
   size_t  large_page_size;      // 2MiB
   size_t  alloc_granularity;    // smallest allocation size (on Windows 64KiB)
   bool    has_overcommit;       // can we reserve more memory than can be actually committed?
-  bool    must_free_whole;      // must allocated blocks free as a whole (false for mmap, true for VirtualAlloc)
+  bool    must_free_whole;      // must allocated blocks be freed as a whole (false for mmap, true for VirtualAlloc)
-  bool    has_virtual_reserve;  // has virtual reserve? (if true we can reserve virtual address space without using commit or physical memory)
+  bool    has_virtual_reserve;  // supports virtual address space reservation? (if true we can reserve virtual address space without using commit or physical memory)
 } mi_os_mem_config_t;
 
 // Initialize

include/mimalloc/types.h

@@ -325,6 +325,10 @@ typedef struct mi_page_s {
 
 
 
+// ------------------------------------------------------
+// Mimalloc segments contain mimalloc pages
+// ------------------------------------------------------
+
 typedef enum mi_page_kind_e {
   MI_PAGE_SMALL,    // small blocks go into 64KiB pages inside a segment
   MI_PAGE_MEDIUM,   // medium blocks go into medium pages inside a segment
@@ -393,7 +397,6 @@ typedef struct mi_memid_s {
     mi_memid_arena_info_t arena;// only used for MI_MEM_ARENA
   } mem;
   bool          is_pinned;      // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large OS pages)
-  bool          is_large;       // `true` if the memory is in OS large (2MiB) or huge (1GiB) pages. (`is_pinned` will be true)
   bool          was_committed;  // `true` if the memory was originally allocated as committed
   bool          was_zero;       // `true` if the memory was originally zero initialized
   mi_memkind_t  memkind;

src/arena.c

@@ -50,9 +50,9 @@ typedef struct mi_arena_s {
   mi_memid_t meta_memid;                  // memid of the arena structure itself (OS or static allocation)
   int      numa_node;                     // associated NUMA node
   bool     is_zero_init;                  // is the arena zero initialized?
-  bool     allow_decommit;                // is decommit allowed? if true, is_large should be false and blocks_committed != NULL
   bool     is_large;                      // large- or huge OS pages (always committed)
   bool     is_huge_alloc;                 // huge OS pages allocated by `_mi_os_alloc_huge_pages`
+  bool     allow_decommit;                // is decommit allowed? if true, is_large should be false and blocks_committed != NULL
   _Atomic(size_t) search_idx;             // optimization to start the search for free blocks
   _Atomic(mi_msecs_t) purge_expire;       // expiration time when blocks should be decommitted from `blocks_decommit`.  
   mi_bitmap_field_t* blocks_dirty;        // are the blocks potentially non-zero?
@@ -96,22 +96,19 @@ static bool mi_arena_id_is_suitable(mi_arena_id_t arena_id, bool arena_is_exclus
   memory id's
 ----------------------------------------------------------- */
 
-static mi_memid_t mi_memid_none(void) {
-  mi_memid_t memid;
-  _mi_memzero(&memid, sizeof(memid));
-  memid.memkind = MI_MEM_NONE;
-  return memid;
-}
-
 static mi_memid_t mi_memid_create(mi_memkind_t memkind) {
-  mi_memid_t memid = mi_memid_none();
+  mi_memid_t memid;
+  _mi_memzero_var(memid);
   memid.memkind = memkind;
   return memid;
 }
 
+static mi_memid_t mi_memid_none(void) {
+  return mi_memid_create(MI_MEM_NONE);
+}
+
 static mi_memid_t mi_memid_create_os(bool committed) {
-  mi_memid_t memid = mi_memid_none();
+  mi_memid_t memid = mi_memid_create(MI_MEM_OS);
-  memid.memkind = MI_MEM_OS;
   memid.was_committed = committed;
   return memid;
 }
@@ -166,11 +163,10 @@ static bool mi_arena_memid_indices(mi_memid_t memid, size_t* arena_index, mi_bit
 
 /* -----------------------------------------------------------
   Special static area for mimalloc internal structures
-  to avoid OS calls (for example, for the arena and thread
+  to avoid OS calls (for example, for the arena metadata)
-  metadata)
 ----------------------------------------------------------- */
 
-#define MI_ARENA_STATIC_MAX  (MI_INTPTR_SIZE*8*MI_KiB)  // 64 KiB on 64-bit
+#define MI_ARENA_STATIC_MAX  (MI_INTPTR_SIZE*MI_KiB)  // 8 KiB on 64-bit
 
 static uint8_t mi_arena_static[MI_ARENA_STATIC_MAX];
 static _Atomic(size_t) mi_arena_static_top;
@@ -212,7 +208,7 @@ static void* mi_arena_meta_zalloc(size_t size, mi_memid_t* memid, mi_stats_t* st
   p = _mi_os_alloc(size, &is_zero, stats);
   if (p != NULL) {
     *memid = mi_memid_create_os(true);
-    if (!is_zero) { _mi_memzero(p, size); }
+    if (!is_zero) { _mi_memzero_aligned(p, size); }
     return p;
   }
 
@@ -259,7 +255,6 @@ static mi_decl_noinline void* mi_arena_alloc_at(mi_arena_t* arena, size_t arena_
   // claimed it! 
   void* p = arena->start + mi_arena_block_size(mi_bitmap_index_bit(bitmap_index));
   *memid = mi_memid_create_arena(arena->id, arena->exclusive, bitmap_index);
-  memid->is_large = arena->is_large;
   memid->is_pinned = (arena->is_large || !arena->allow_decommit);
 
   // none of the claimed blocks should be scheduled for a decommit
@@ -437,13 +432,12 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset
   }
    
   // finally, fall back to the OS
-  bool os_large = allow_large;
+  bool os_is_large = false;
   bool os_is_zero = false;
-  void* p = _mi_os_alloc_aligned_offset(size, alignment, align_offset, commit, &os_large, &os_is_zero, tld->stats);
+  void* p = _mi_os_alloc_aligned_at_offset(size, alignment, align_offset, commit, allow_large, &os_is_large, &os_is_zero, tld->stats);
   if (p != NULL) { 
     *memid = mi_memid_create_os(commit);
-    memid->is_large = os_large;
+    memid->is_pinned = os_is_large;
-    memid->is_pinned = os_large;
     memid->was_zero = os_is_zero;
     memid->mem.os.alignment = alignment;
     memid->mem.os.align_offset = align_offset;
@@ -647,7 +641,7 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
       _mi_stat_decrease(&stats->committed, committed_size);
     }
     if (memid.mem.os.align_offset != 0) {
-      _mi_os_free_aligned(p, size, memid.mem.os.alignment, memid.mem.os.align_offset, all_committed, stats);
+      _mi_os_free_aligned_at_offset(p, size, memid.mem.os.alignment, memid.mem.os.align_offset, all_committed, stats);
     }
     else {
       _mi_os_free(p, size, stats);
@@ -729,7 +723,7 @@ static void mi_arenas_unsafe_destroy(void) {
       if (arena->owned && arena->start != NULL) {
         mi_atomic_store_ptr_release(mi_arena_t, &mi_arenas[i], NULL);
         if (arena->is_huge_alloc) { 
-          _mi_os_free_huge_pages(arena->start, mi_arena_size(arena), &_mi_stats_main); 
+          _mi_os_free_huge_os_pages(arena->start, mi_arena_size(arena), &_mi_stats_main); 
         }
         else {
           _mi_os_free(arena->start, mi_arena_size(arena), &_mi_stats_main); 
@@ -860,16 +854,16 @@ static int mi_reserve_os_memory_ex2(size_t size, bool commit, bool allow_large,
 {
   if (arena_id != NULL) *arena_id = _mi_arena_id_none();
   size = _mi_align_up(size, MI_ARENA_BLOCK_SIZE); // at least one block
-  bool large = allow_large;
+  bool is_large = false;
-  bool is_zero;
+  bool is_zero = false;
-  void* start = _mi_os_alloc_aligned(size, MI_SEGMENT_ALIGN, commit, &large, &is_zero, &_mi_stats_main);
+  void* start = _mi_os_alloc_aligned(size, MI_SEGMENT_ALIGN, commit, allow_large, &is_large, &is_zero, &_mi_stats_main);
   if (start==NULL) return ENOMEM;
-  if (!mi_manage_os_memory_ex2(start, size, (large || commit), large, false, is_zero, -1, exclusive, owned, arena_id)) {
+  if (!mi_manage_os_memory_ex2(start, size, (is_large || commit), is_large, false, is_zero, -1, exclusive, owned, arena_id)) {
     _mi_os_free_ex(start, size, commit, &_mi_stats_main);
     _mi_verbose_message("failed to reserve %zu k memory\n", _mi_divide_up(size,1024));
     return ENOMEM;
   }
-  _mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size,1024), large ? " (in large os pages)" : "");
+  _mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size,1024), is_large ? " (in large os pages)" : "");
   return 0;
 }
 
@@ -943,7 +937,7 @@ int mi_reserve_huge_os_pages_at_ex(size_t pages, int numa_node, size_t timeout_m
   _mi_verbose_message("numa node %i: reserved %zu GiB huge pages (of the %zu GiB requested)\n", numa_node, pages_reserved, pages);
 
   if (!mi_manage_os_memory_ex2(p, hsize, true, true, true, is_zero, numa_node, exclusive, true /* owned */, arena_id)) {
-    _mi_os_free_huge_pages(p, hsize, &_mi_stats_main);
+    _mi_os_free_huge_os_pages(p, hsize, &_mi_stats_main);
     return ENOMEM;
   }
   return 0;

src/init.c

@@ -243,7 +243,7 @@ static mi_thread_data_t* mi_thread_data_zalloc(void) {
   }
   
   if (td != NULL && !is_zero) {
-    _mi_memzero(td, sizeof(*td));
+    _mi_memzero_aligned(td, sizeof(*td));
   }
   return td;
 }

src/os.c

@@ -264,23 +264,19 @@ void* _mi_os_alloc(size_t size, bool* is_zero, mi_stats_t* tld_stats) {
   return p;
 }
 
-void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, bool* is_zero, mi_stats_t* tld_stats)
+void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* tld_stats)
 {
   MI_UNUSED(&_mi_os_get_aligned_hint); // suppress unused warnings
   MI_UNUSED(tld_stats);
   if (size == 0) return NULL;
   size = _mi_os_good_alloc_size(size);
   alignment = _mi_align_up(alignment, _mi_os_page_size());
-  bool allow_large = false;
+  
-  if (large != NULL) {
+  bool os_is_large = false;
-    allow_large = *large;
+  bool os_is_zero  = false;
-    *large = false;
+  void* p = mi_os_mem_alloc_aligned(size, alignment, commit, allow_large, &os_is_large, &os_is_zero, &_mi_stats_main /*tld->stats*/ );
-  }
+  if (is_large != NULL) { *is_large = os_is_large; }
-  bool is_largex = false;
+  if (is_zero != NULL)  { *is_zero = os_is_zero; }
-  bool is_zerox = false;
-  void* p = mi_os_mem_alloc_aligned(size, alignment, commit, allow_large, &is_largex, &is_zerox, &_mi_stats_main /*tld->stats*/ );
-  if (large != NULL) { *large = is_largex; }
-  if (is_zero != NULL) { *is_zero = is_zerox; }
   return p;
 }
 
@@ -292,20 +288,20 @@ void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* lar
   to use the actual start of the memory region.
 ----------------------------------------------------------- */
 
-void* _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t offset, bool commit, bool* large, bool* is_zero, mi_stats_t* tld_stats) {
+void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offset, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* tld_stats) {
   mi_assert(offset <= MI_SEGMENT_SIZE);
   mi_assert(offset <= size);
   mi_assert((alignment % _mi_os_page_size()) == 0);
   if (offset > MI_SEGMENT_SIZE) return NULL;
   if (offset == 0) {
     // regular aligned allocation
-    return _mi_os_alloc_aligned(size, alignment, commit, large, is_zero, tld_stats);
+    return _mi_os_alloc_aligned(size, alignment, commit, allow_large, is_large, is_zero, tld_stats);
   }
   else {
     // overallocate to align at an offset
     const size_t extra = _mi_align_up(offset, alignment) - offset;
     const size_t oversize = size + extra;
-    void* start = _mi_os_alloc_aligned(oversize, alignment, commit, large, is_zero, tld_stats);
+    void* start = _mi_os_alloc_aligned(oversize, alignment, commit, allow_large, is_large, is_zero, tld_stats);
     if (start == NULL) return NULL;
     void* p = (uint8_t*)start + extra;
     mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment));
@@ -317,7 +313,7 @@ void* _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t offset,
   }
 }
 
-void _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats) {
+void _mi_os_free_aligned_at_offset(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats) {
   mi_assert(align_offset <= MI_SEGMENT_SIZE);
   const size_t extra = _mi_align_up(align_offset, alignment) - align_offset;
   void* start = (uint8_t*)p - extra;
@@ -594,7 +590,7 @@ void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_mse
 
 // free every huge page in a range individually (as we allocated per page)
 // note: needed with VirtualAlloc but could potentially be done in one go on mmap'd systems.
-void _mi_os_free_huge_pages(void* p, size_t size, mi_stats_t* stats) {
+void _mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats) {
   if (p==NULL || size==0) return;
   uint8_t* base = (uint8_t*)p;
   while (size >= MI_HUGE_OS_PAGE_SIZE) {

src/segment.c

@@ -842,7 +842,7 @@ static mi_segment_t* mi_segment_os_alloc( size_t required, size_t page_alignment
 
   mi_track_mem_undefined(segment, (*pinfo_slices) * MI_SEGMENT_SLICE_SIZE); // todo: should not be necessary?
   segment->memid = memid;
-  segment->allow_decommit = !memid.is_pinned && !memid.is_large;
+  segment->allow_decommit = !memid.is_pinned;
   segment->allow_purge = segment->allow_decommit && mi_option_is_enabled(mi_option_allow_purge);
   segment->segment_size = segment_size;
   segment->commit_mask = commit_mask;

src/stats.c

@@ -444,7 +444,7 @@ mi_msecs_t _mi_clock_end(mi_msecs_t start) {
 mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_msecs, size_t* current_rss, size_t* peak_rss, size_t* current_commit, size_t* peak_commit, size_t* page_faults) mi_attr_noexcept
 {
   mi_process_info_t pinfo;
-  _mi_memzero(&pinfo,sizeof(pinfo));
+  _mi_memzero_var(pinfo);
   pinfo.elapsed        = _mi_clock_end(mi_process_start);
   pinfo.current_commit = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.current));
   pinfo.peak_commit    = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.peak));