add initial support for _mi_prim_reuse and MADV_FREE_REUSABLE on macOS (issue #1097)

azure-pipelines.yml

@@ -6,10 +6,8 @@
 trigger:
   branches:
     include:
-    - master
+    - main
-    - dev3
+    - dev*
-    - dev2
-    - dev
   tags:
     include:
     - v*
@@ -184,35 +182,6 @@ jobs:
 # Other OS versions (just debug mode)
 # ----------------------------------------------------------
 
-- job:
-  displayName: Windows 2019
-  pool:
-    vmImage:
-      windows-2019
-  strategy:
-    matrix:
-      Debug:
-        BuildType: debug
-        cmakeExtraArgs: -DCMAKE_BUILD_TYPE=Debug -DMI_DEBUG_FULL=ON
-        MSBuildConfiguration: Debug
-      Release:
-        BuildType: release
-        cmakeExtraArgs: -DCMAKE_BUILD_TYPE=Release
-        MSBuildConfiguration: Release
-  steps:
-  - task: CMake@1
-    inputs:
-      workingDirectory: $(BuildType)
-      cmakeArgs: .. $(cmakeExtraArgs)
-  - task: MSBuild@1
-    inputs:
-      solution: $(BuildType)/libmimalloc.sln
-      configuration: '$(MSBuildConfiguration)'
-      msbuildArguments: -m
-  - script: ctest --verbose --timeout 240 -C $(MSBuildConfiguration)
-    workingDirectory: $(BuildType)
-    displayName: CTest
-
 - job:
   displayName: Ubuntu 24.04
   pool:

include/mimalloc/internal.h

@@ -8,7 +8,6 @@ terms of the MIT license. A copy of the license can be found in the file
 #ifndef MIMALLOC_INTERNAL_H
 #define MIMALLOC_INTERNAL_H
 
-
 // --------------------------------------------------------------------------
 // This file contains the internal API's of mimalloc and various utility
 // functions and macros.
@@ -17,50 +16,88 @@ terms of the MIT license. A copy of the license can be found in the file
 #include "types.h"
 #include "track.h"
 
+
+// --------------------------------------------------------------------------
+// Compiler defines
+// --------------------------------------------------------------------------
+
 #if (MI_DEBUG>0)
 #define mi_trace_message(...)  _mi_trace_message(__VA_ARGS__)
 #else
 #define mi_trace_message(...)
 #endif
 
-#define MI_CACHE_LINE          64
+#define mi_decl_cache_align     mi_decl_align(64)
+
 #if defined(_MSC_VER)
 #pragma warning(disable:4127)   // suppress constant conditional warning (due to MI_SECURE paths)
 #pragma warning(disable:26812)  // unscoped enum warning
 #define mi_decl_noinline        __declspec(noinline)
 #define mi_decl_thread          __declspec(thread)
-#define mi_decl_cache_align     __declspec(align(MI_CACHE_LINE))
+#define mi_decl_align(a)        __declspec(align(a))
+#define mi_decl_noreturn        __declspec(noreturn)
 #define mi_decl_weak
 #define mi_decl_hidden
+#define mi_decl_cold
 #elif (defined(__GNUC__) && (__GNUC__ >= 3)) || defined(__clang__) // includes clang and icc
 #define mi_decl_noinline        __attribute__((noinline))
 #define mi_decl_thread          __thread
-#define mi_decl_cache_align     __attribute__((aligned(MI_CACHE_LINE)))
+#define mi_decl_align(a)        __attribute__((aligned(a)))
+#define mi_decl_noreturn        __attribute__((noreturn))
 #define mi_decl_weak            __attribute__((weak))
 #define mi_decl_hidden          __attribute__((visibility("hidden")))
+#if (__GNUC__ >= 4) || defined(__clang__)
+#define mi_decl_cold            __attribute__((cold))
+#else
+#define mi_decl_cold
+#endif
 #elif __cplusplus >= 201103L    // c++11
 #define mi_decl_noinline
 #define mi_decl_thread          thread_local
-#define mi_decl_cache_align     alignas(MI_CACHE_LINE)
+#define mi_decl_align(a)        alignas(a)
+#define mi_decl_noreturn        [[noreturn]]
 #define mi_decl_weak
 #define mi_decl_hidden
+#define mi_decl_cold
 #else
 #define mi_decl_noinline
 #define mi_decl_thread          __thread        // hope for the best :-)
-#define mi_decl_cache_align
+#define mi_decl_align(a)
+#define mi_decl_noreturn        
 #define mi_decl_weak
 #define mi_decl_hidden
+#define mi_decl_cold
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#define mi_unlikely(x)     (__builtin_expect(!!(x),false))
+#define mi_likely(x)       (__builtin_expect(!!(x),true))
+#elif (defined(__cplusplus) && (__cplusplus >= 202002L)) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+#define mi_unlikely(x)     (x) [[unlikely]]
+#define mi_likely(x)       (x) [[likely]]
+#else
+#define mi_unlikely(x)     (x)
+#define mi_likely(x)       (x)
+#endif
+
+#ifndef __has_builtin
+#define __has_builtin(x)    0
+#endif
+
+#if defined(__cplusplus)
+#define mi_decl_externc     extern "C"
+#else
+#define mi_decl_externc
 #endif
 
 #if defined(__EMSCRIPTEN__) && !defined(__wasi__)
 #define __wasi__
 #endif
 
-#if defined(__cplusplus)
+
-#define mi_decl_externc       extern "C"
+// --------------------------------------------------------------------------
-#else
+// Internal functions
-#define mi_decl_externc
+// --------------------------------------------------------------------------
-#endif
 
 // "libc.c"
 #include    <stdarg.h>
@@ -130,6 +167,7 @@ bool        _mi_os_decommit(void* addr, size_t size);
 bool        _mi_os_unprotect(void* addr, size_t size);
 bool        _mi_os_purge(void* p, size_t size);
 bool        _mi_os_purge_ex(void* p, size_t size, bool allow_reset, size_t stat_size);
+void        _mi_os_reuse(void* p, size_t size);
 mi_decl_nodiscard bool _mi_os_commit(void* p, size_t size, bool* is_zero);
 mi_decl_nodiscard bool _mi_os_commit_ex(void* addr, size_t size, bool* is_zero, size_t stat_size);
 mi_decl_nodiscard bool _mi_os_protect(void* addr, size_t size);
@@ -256,26 +294,6 @@ bool        _mi_page_is_valid(mi_page_t* page);
 #endif
 
 
-// ------------------------------------------------------
-// Branches
-// ------------------------------------------------------
-
-#if defined(__GNUC__) || defined(__clang__)
-#define mi_unlikely(x)     (__builtin_expect(!!(x),false))
-#define mi_likely(x)       (__builtin_expect(!!(x),true))
-#elif (defined(__cplusplus) && (__cplusplus >= 202002L)) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
-#define mi_unlikely(x)     (x) [[unlikely]]
-#define mi_likely(x)       (x) [[likely]]
-#else
-#define mi_unlikely(x)     (x)
-#define mi_likely(x)       (x)
-#endif
-
-#ifndef __has_builtin
-#define __has_builtin(x)  0
-#endif
-
-
 /* -----------------------------------------------------------
   Error codes passed to `_mi_fatal_error`
   All are recoverable but EFAULT is a serious error and aborts by default in secure mode.
@@ -300,6 +318,32 @@ bool        _mi_page_is_valid(mi_page_t* page);
 #endif
 
 
+// ------------------------------------------------------
+// Assertions
+// ------------------------------------------------------
+
+#if (MI_DEBUG)
+// use our own assertion to print without memory allocation
+mi_decl_noreturn mi_decl_cold void _mi_assert_fail(const char* assertion, const char* fname, unsigned int line, const char* func) mi_attr_noexcept;
+#define mi_assert(expr)     ((expr) ? (void)0 : _mi_assert_fail(#expr,__FILE__,__LINE__,__func__))
+#else
+#define mi_assert(x)
+#endif
+
+#if (MI_DEBUG>1)
+#define mi_assert_internal    mi_assert
+#else
+#define mi_assert_internal(x)
+#endif
+
+#if (MI_DEBUG>2)
+#define mi_assert_expensive   mi_assert
+#else
+#define mi_assert_expensive(x)
+#endif
+
+
+
 /* -----------------------------------------------------------
   Inlined definitions
 ----------------------------------------------------------- */

include/mimalloc/prim.h

@@ -63,6 +63,11 @@ int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit);
 // Returns error code or 0 on success.
 int _mi_prim_reset(void* addr, size_t size);
 
+// Reuse memory. This is called for memory that is already committed but
+// may have been reset (`_mi_prim_reset`) or decommitted (`_mi_prim_decommit`) where `needs_recommit` was false.
+// Returns error code or 0 on success. On most platforms this is a no-op.
+int _mi_prim_reuse(void* addr, size_t size);
+
 // Protect memory. Returns error code or 0 on success.
 int _mi_prim_protect(void* addr, size_t size, bool protect);
 

include/mimalloc/types.h

@@ -574,7 +574,6 @@ struct mi_tld_s {
 };
 
 
-
 // ------------------------------------------------------
 // Debug
 // ------------------------------------------------------
@@ -589,26 +588,6 @@ struct mi_tld_s {
 #define MI_DEBUG_PADDING    (0xDE)
 #endif
 
-#if (MI_DEBUG)
-// use our own assertion to print without memory allocation
-void _mi_assert_fail(const char* assertion, const char* fname, unsigned int line, const char* func );
-#define mi_assert(expr)     ((expr) ? (void)0 : _mi_assert_fail(#expr,__FILE__,__LINE__,__func__))
-#else
-#define mi_assert(x)
-#endif
-
-#if (MI_DEBUG>1)
-#define mi_assert_internal    mi_assert
-#else
-#define mi_assert_internal(x)
-#endif
-
-#if (MI_DEBUG>2)
-#define mi_assert_expensive   mi_assert
-#else
-#define mi_assert_expensive(x)
-#endif
-
 
 // ------------------------------------------------------
 // Statistics

readme.md

@@ -72,15 +72,14 @@ Enjoy!
 
 ### Branches
 
-* `master`: latest stable release (still based on `dev2`).
+* `main`: latest stable release (still based on `dev2`).
 * `dev`:  development branch for mimalloc v1. Use this branch for submitting PR's.
 * `dev2`: development branch for mimalloc v2. This branch is downstream of `dev` 
           (and is essentially equal to `dev` except for `src/segment.c`). Uses larger sliced segments to manage
           mimalloc pages that can reduce fragmentation.
 * `dev3`: development branch for mimalloc v3-beta. This branch is downstream of `dev`. This version 
           simplifies the lock-free ownership of previous versions, has no thread-local segments any more. 
-          This improves sharing of memory between threads, and on certain large workloads may use less memory 
+          This improves sharing of memory between threads, and on certain large workloads may use (much) less memory.
-          with less fragmentation.
 
 ### Releases
 

src/arena.c

@@ -266,12 +266,12 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
   else if (commit) {
     // commit requested, but the range may not be committed as a whole: ensure it is committed now
     memid->initially_committed = true;
+    const size_t commit_size = mi_arena_block_size(needed_bcount);      
     bool 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_ex(p, commit_size, &commit_zero, stat_commit_size)) {
@@ -281,6 +281,10 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
         if (commit_zero) { memid->initially_zero = true; }
       }
     }
+    else {
+      // all are already committed: signal that we are reusing memory in case it was purged before
+      _mi_os_reuse( p, commit_size );
+    }
   }
   else {
     // no need to commit, but check if already fully committed

src/options.c

@@ -525,7 +525,7 @@ void _mi_warning_message(const char* fmt, ...) {
 
 
 #if MI_DEBUG
-void _mi_assert_fail(const char* assertion, const char* fname, unsigned line, const char* func ) {
+mi_decl_noreturn mi_decl_cold void _mi_assert_fail(const char* assertion, const char* fname, unsigned line, const char* func ) mi_attr_noexcept {
   _mi_fprintf(NULL, NULL, "mimalloc: assertion failed: at \"%s\":%u, %s\n  assertion: \"%s\"\n", fname, line, (func==NULL?"":func), assertion);
   abort();
 }

src/os.c

@@ -512,6 +512,17 @@ bool _mi_os_reset(void* addr, size_t size) {
 }
 
 
+void _mi_os_reuse( void* addr, size_t size ) {
+  // page align conservatively within the range
+  size_t csize = 0;
+  void* const start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return;
+  const int err = _mi_prim_reuse(start, csize);
+  if (err != 0) {
+    _mi_warning_message("cannot reuse OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize);
+  }
+}
+
 // either resets or decommits memory, returns true if the memory needs
 // to be recommitted if it is to be re-used later on.
 bool _mi_os_purge_ex(void* p, size_t size, bool allow_reset, size_t stat_size)

src/prim/emscripten/prim.c

@@ -114,6 +114,11 @@ int _mi_prim_reset(void* addr, size_t size) {
   return 0;
 }
 
+int _mi_prim_reuse(void* addr, size_t size) {
+  MI_UNUSED(addr); MI_UNUSED(size);
+  return 0;
+}
+
 int _mi_prim_protect(void* addr, size_t size, bool protect) {
   MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(protect);
   return 0;

src/prim/unix/prim.c

@@ -433,13 +433,27 @@ int _mi_prim_commit(void* start, size_t size, bool* is_zero) {
   return err;
 }
 
+int _mi_prim_reuse(void* start, size_t size) {
+  #if defined(__APPLE__) && defined(MADV_FREE_REUSE)
+  return unix_madvise(start, size, MADV_FREE_REUSE);
+  #endif
+  return 0;
+}
+
 int _mi_prim_decommit(void* start, size_t size, bool* needs_recommit) {
   int err = 0;
-  // decommit: use MADV_DONTNEED as it decreases rss immediately (unlike MADV_FREE)
-  err = unix_madvise(start, size, MADV_DONTNEED);
   #if !MI_DEBUG && MI_SECURE<=2
     *needs_recommit = false;
+    #if defined(__APPLE__) && defined(MADV_FREE_REUSABLE)
+    // decommit on macOS: use MADV_FREE_REUSABLE as it does immediate rss accounting (issue #1097)
+    err = unix_madvise(start, size, MADV_FREE_REUSABLE);
+    #else
+    // decommit: use MADV_DONTNEED as it decreases rss immediately (unlike MADV_FREE)
+    err = unix_madvise(start, size, MADV_DONTNEED);
+    #endif  
   #else
+    // note: don't use MADV_FREE_REUSABLE as the range may contain protected areas
+    err = unix_madvise(start, size, MADV_DONTNEED);
     *needs_recommit = true;
     mprotect(start, size, PROT_NONE);
   #endif
@@ -454,14 +468,21 @@ int _mi_prim_decommit(void* start, size_t size, bool* needs_recommit) {
 }
 
 int _mi_prim_reset(void* start, size_t size) {
-  // We try to use `MADV_FREE` as that is the fastest. A drawback though is that it
+  int err = 0;
+  #if defined(__APPLE__) && defined(MADV_FREE_REUSABLE)
+  // on macOS we try to use MADV_FREE_REUSABLE as it seems the fastest
+  err = unix_madvise(start, size, MADV_FREE_REUSABLE);  
+  if (err == 0) return 0;
+  // fall through
+  #endif
+
+  #if defined(MADV_FREE)
+  // Otherwise, we try to use `MADV_FREE` as that is the fastest. A drawback though is that it
   // will not reduce the `rss` stats in tools like `top` even though the memory is available
   // to other processes. With the default `MIMALLOC_PURGE_DECOMMITS=1` we ensure that by
   // default `MADV_DONTNEED` is used though.
-  #if defined(MADV_FREE)
   static _Atomic(size_t) advice = MI_ATOMIC_VAR_INIT(MADV_FREE);
   int oadvice = (int)mi_atomic_load_relaxed(&advice);
-  int err;
   while ((err = unix_madvise(start, size, oadvice)) != 0 && errno == EAGAIN) { errno = 0;  };
   if (err != 0 && errno == EINVAL && oadvice == MADV_FREE) {
     // if MADV_FREE is not supported, fall back to MADV_DONTNEED from now on
@@ -469,7 +490,7 @@ int _mi_prim_reset(void* start, size_t size) {
     err = unix_madvise(start, size, MADV_DONTNEED);
   }
   #else
-  int err = unix_madvise(start, size, MADV_DONTNEED);
+  err = unix_madvise(start, size, MADV_DONTNEED);
   #endif
   return err;
 }

src/prim/wasi/prim.c

@@ -149,6 +149,11 @@ int _mi_prim_reset(void* addr, size_t size) {
   return 0;
 }
 
+int _mi_prim_reuse(void* addr, size_t size) {
+  MI_UNUSED(addr); MI_UNUSED(size);
+  return 0;
+}
+
 int _mi_prim_protect(void* addr, size_t size, bool protect) {
   MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(protect);
   return 0;

src/prim/windows/prim.c

@@ -352,6 +352,11 @@ int _mi_prim_reset(void* addr, size_t size) {
   return (p != NULL ? 0 : (int)GetLastError());
 }
 
+int _mi_prim_reuse(void* addr, size_t size) {
+  MI_UNUSED(addr); MI_UNUSED(size);
+  return 0;
+}
+
 int _mi_prim_protect(void* addr, size_t size, bool protect) {
   DWORD oldprotect = 0;
   BOOL ok = VirtualProtect(addr, size, protect ? PAGE_NOACCESS : PAGE_READWRITE, &oldprotect);