rewrite of low-level OS (aligned) allocation to better handle large pages and aligned allocation

src/init.c

@@ -103,7 +103,11 @@ mi_heap_t _mi_heap_main = {
   NULL,
   0,
   0,
-  0xCDCDCDCDCDCDCDL,
+#if MI_INTPTR_SIZE==8   // the cookie of the main heap can be fixed (unlike page cookies that need to be secure!)
+  0xCDCDCDCDCDCDCDCDUL,
+#else
+  0xCDCDCDCDUL,
+#endif
   0,
   false   // can reclaim
 };

src/os.c

@@ -14,12 +14,6 @@ terms of the MIT license. A copy of the license can be found in the file
 #include <string.h>  // memset
 #include <errno.h>
 
-/* -----------------------------------------------------------
-  Initialization.
-  On windows initializes support for aligned allocation and
-  large OS pages (if MIMALLOC_LARGE_OS_PAGES is true).
------------------------------------------------------------ */
-
 #if defined(_WIN32)
 #include <windows.h>
 #else
@@ -27,6 +21,32 @@ terms of the MIT license. A copy of the license can be found in the file
 #include <unistd.h>    // sysconf
 #endif
 
+/* -----------------------------------------------------------
+  Initialization.
+  On windows initializes support for aligned allocation and
+  large OS pages (if MIMALLOC_LARGE_OS_PAGES is true).
+----------------------------------------------------------- */
+bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
+
+uintptr_t _mi_align_up(uintptr_t sz, size_t alignment) {
+  uintptr_t x = (sz / alignment) * alignment;
+  if (x < sz) x += alignment;
+  if (x < sz) return 0; // overflow
+  return x;
+}
+
+static void* mi_align_up_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_up((uintptr_t)p, alignment);
+}
+
+static uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
+  return (sz / alignment) * alignment;
+}
+
+static void* mi_align_down_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_down((uintptr_t)p, alignment);
+}
+
 // page size (initialized properly in `os_init`)
 static size_t os_page_size = 4096;
 
@@ -61,7 +81,7 @@ static size_t mi_os_good_alloc_size(size_t size, size_t alignment) {
 
 #if defined(_WIN32)
 // We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016.
-// So, we need to look it up dynamically to run on older systems.
+// So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility)
 typedef PVOID(__stdcall *VirtualAlloc2Ptr)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MEM_EXTENDED_PARAMETER*, ULONG);
 static VirtualAlloc2Ptr pVirtualAlloc2 = NULL;
 
@@ -128,29 +148,8 @@ void _mi_os_init() {
 
 /* -----------------------------------------------------------
   Raw allocation on Windows (VirtualAlloc) and Unix's (mmap).  
-  Defines a portable `mmap`, `munmap` and `mmap_trim`.
 ----------------------------------------------------------- */
 
-uintptr_t _mi_align_up(uintptr_t sz, size_t alignment) {
-  uintptr_t x = (sz / alignment) * alignment;
-  if (x < sz) x += alignment;
-  if (x < sz) return 0; // overflow
-  return x;
-}
-
-static void* mi_align_up_ptr(void* p, size_t alignment) {
-  return (void*)_mi_align_up((uintptr_t)p, alignment);
-}
-
-static uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
-  return (sz / alignment) * alignment;
-}
-
-static void* mi_align_down_ptr(void* p, size_t alignment) {
-  return (void*)_mi_align_down((uintptr_t)p, alignment);
-}
-
-
 static bool mi_os_mem_free(void* addr, size_t size, mi_stats_t* stats)
 {
   if (addr == NULL || size == 0) return true;
@@ -172,38 +171,53 @@ static bool mi_os_mem_free(void* addr, size_t size, mi_stats_t* stats)
   }
 }
 
-static void* mi_os_mem_alloc(void* addr, size_t size, bool commit, int extra_flags, mi_stats_t* stats) {
+#ifdef _WIN32
-  if (size == 0) return NULL;
+static void* mi_win_virtual_allocx(void* addr, size_t size, size_t try_alignment, DWORD flags) {
+#if defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
+  if (try_alignment > 0 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) {
+    // on modern Windows try use VirtualAlloc2
+    MEM_ADDRESS_REQUIREMENTS reqs = { 0 };
+    reqs.Alignment = try_alignment;
+    MEM_EXTENDED_PARAMETER param = { 0 };
+    param.Type = MemExtendedParameterAddressRequirements;
+    param.Pointer = &reqs;
+    return (*pVirtualAlloc2)(addr, NULL, size, flags, PAGE_READWRITE, &param, 1);
+  }
+#endif
+  return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
+}
+
+static void* mi_win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags) {
   void* p = NULL;
-#if defined(_WIN32)
+  if (use_large_os_page(size, try_alignment)) {
-  int flags = MEM_RESERVE | extra_flags;
+    p = mi_win_virtual_allocx(addr, size, try_alignment, MEM_LARGE_PAGES | flags);
-  if (commit) flags |= MEM_COMMIT;
+    // fall back to non-large page allocation on error (`p == NULL`).
-  if (use_large_os_page(size, 0)) {
-    p = VirtualAlloc(addr, size, MEM_LARGE_PAGES | flags, PAGE_READWRITE);
   }
   if (p == NULL) {
-    p = VirtualAlloc(addr, size, flags, PAGE_READWRITE);
+    p = mi_win_virtual_allocx(addr, size, try_alignment, flags);
   }
+  return p;
+}
+
 #else
+static void* mi_unix_mmap(size_t size, size_t try_alignment, int protect_flags) {
+  void* p = NULL;
   #if !defined(MAP_ANONYMOUS)
   #define MAP_ANONYMOUS  MAP_ANON
   #endif
-  int flags = MAP_PRIVATE | MAP_ANONYMOUS | extra_flags;
+  int flags = MAP_PRIVATE | MAP_ANONYMOUS;
-  if (addr != NULL) {
+  #if defined(MAP_ALIGNED)  // BSD
-    #if defined(MAP_EXCL)
+  if (try_alignment > 0) {
-      flags |= MAP_FIXED | MAP_EXCL;  // BSD
+    size_t n = _mi_bsr(try_alignment);
-    #elif defined(MAP_FIXED_NOREPLACE)
+    if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) {  // alignment is a power of 2 and 4096 <= alignment <= 1GiB
-      flags |= MAP_FIXED_NOREPLACE;   // Linux
+      flags |= MAP_ALIGNED(n);
-    #elif defined(MAP_FIXED)
+    }
-      flags |= MAP_FIXED;
-    #endif
   }
-  int pflags = (commit ? (PROT_READ | PROT_WRITE) : PROT_NONE);
-  #if defined(PROT_MAX)
-  pflags |= PROT_MAX(PROT_READ | PROT_WRITE); // BSD
   #endif
-
+  #if defined(PROT_MAX)
-  if (large_os_page_size > 0 && use_large_os_page(size, 0) && ((uintptr_t)addr % large_os_page_size) == 0) {
+  protect_flags |= PROT_MAX(PROT_READ | PROT_WRITE); // BSD
+  #endif
+  if (large_os_page_size > 0 && use_large_os_page(size, try_alignment)) {
     int lflags = flags;
     #ifdef MAP_ALIGNED_SUPER
     lflags |= MAP_ALIGNED_SUPER;
@@ -216,65 +230,144 @@ static void* mi_os_mem_alloc(void* addr, size_t size, bool commit, int extra_fla
     #endif
     if (lflags != flags) {
       // try large page allocation 
-      p = mmap(addr, size, pflags, lflags, -1, 0);
+      // TODO: if always failing due to permissions or no huge pages, try to avoid repeatedly trying? 
+      // Should we check this in _mi_os_init? (as on Windows)
+      p = mmap(NULL, size, protect_flags, lflags, -1, 0);
       if (p == MAP_FAILED) p = NULL; // fall back to regular mmap if large is exhausted or no permission
     }
   }
   if (p == NULL) {
-    p = mmap(addr, size, pflags, flags, -1, 0);
+    p = mmap(NULL, size, protect_flags, flags, -1, 0);
     if (p == MAP_FAILED) p = NULL;
   }
-  if (addr != NULL && p != addr) {
+  return p;
+}
+#endif
+
+// Primitive allocation from the OS.
+// Note: the `alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
+static void* mi_os_mem_alloc(size_t size, size_t try_alignment, bool commit, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  if (size == 0) return NULL;
+
+  void* p = NULL;
+#if defined(_WIN32)
+  int flags = MEM_RESERVE;
+  if (commit) flags |= MEM_COMMIT;
+  p = mi_win_virtual_alloc(NULL, size, try_alignment, flags);
+#else
+  int protect_flags = (commit ? (PROT_WRITE | PROT_READ) : PROT_NONE);
+  p = mi_unix_mmap(size, try_alignment, protect_flags);
+#endif  
+  _mi_stat_increase(&stats->mmap_calls, 1);
+  if (p != NULL) {
+    _mi_stat_increase(&stats->reserved, size);
+    if (commit) _mi_stat_increase(&stats->committed, size);
+  }
+  return p;
+}
+
+
+// Primitive aligned allocation from the OS.
+// This function guarantees the allocated memory is aligned.
+static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, mi_stats_t* stats) {
+  mi_assert_internal(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0));
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  if (!(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0))) return NULL;
+  size = _mi_align_up(size, _mi_os_page_size());
+  
+  // try first with a hint (this will be aligned directly on Win 10+ or BSD)
+  void* p = mi_os_mem_alloc(size, alignment, commit, stats);
+  if (p == NULL) return NULL;
+
+  // if not aligned, free it, overallocate, and unmap around it
+  if (((uintptr_t)p % alignment != 0)) {
+    mi_os_mem_free(p, size, stats);
+    if (size >= (SIZE_MAX - alignment)) return NULL; // overflow
+    size_t over_size = size + alignment;
+
+#if _WIN32
+    // over-allocate and than re-allocate exactly at an aligned address in there.
+    // this may fail due to threads allocating at the same time so we
+    // retry this at most 3 times before giving up. 
+    // (we can not decommit around the overallocation on Windows, because we can only
+    //  free the original pointer, not one pointing inside the area)
+    int flags = MEM_RESERVE;
+    if (commit) flags |= MEM_COMMIT;
+    for (int tries = 0; tries < 3; tries++) {
+      // over-allocate to determine a virtual memory range
+      p = mi_os_mem_alloc(over_size, alignment, commit, stats);
+      if (p == NULL) return NULL; // error
+      if (((uintptr_t)p % alignment) == 0) {
+        // if p happens to be aligned, just decommit the left-over area
+        _mi_os_decommit((uint8_t*)p + size, over_size - size, stats);
+        break;
+      }
+      else {
+        // otherwise free and allocate at an aligned address in there
+        mi_os_mem_free(p, over_size, stats);
+        void* aligned_p = mi_align_up_ptr(p, alignment);
+        p = mi_win_virtual_alloc(aligned_p, size, alignment, flags);
+        if (p == aligned_p) break; // success!
+        if (p != NULL) { // should not happen?
           mi_os_mem_free(p, size, stats);  
           p = NULL;
         }
-#endif
-  mi_assert(p == NULL || (addr == NULL && p != addr) || (addr != NULL && p == addr));
-  if (p != NULL) {
-    _mi_stat_increase(&stats->mmap_calls, 1);
-    _mi_stat_increase(&stats->reserved, size);
-    if (commit) _mi_stat_increase(&stats->committed, size);
       }
-  return p;
-}
-
-static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, mi_stats_t* stats) {
-  if (alignment < _mi_os_page_size() || ((alignment & (~alignment + 1)) != alignment)) return NULL;
-  void* p = NULL;
-  #if defined(_WIN32) && defined(MEM_EXTENDED_PARAMETER_TYPE_BITS)
-  if (pVirtualAlloc2 != NULL) {
-    // on modern Windows try use VirtualAlloc2
-    MEM_ADDRESS_REQUIREMENTS reqs = {0};
-    reqs.Alignment = alignment;
-    MEM_EXTENDED_PARAMETER param = { 0 };
-    param.Type = MemExtendedParameterAddressRequirements;
-    param.Pointer = &reqs;
-    DWORD flags = MEM_RESERVE;
-    if (commit) flags |= MEM_COMMIT;
-    if (use_large_os_page(size, alignment)) flags |= MEM_LARGE_PAGES;
-    p = (*pVirtualAlloc2)(NULL, NULL, size, flags, PAGE_READWRITE, &param, 1);
-  }
-  #elif defined(MAP_ALIGNED)
-  // on BSD, use the aligned mmap api
-  size_t n = _mi_bsr(alignment);
-  if (((size_t)1 << n) == alignment && n >= 12) {  // alignment is a power of 2 and >= 4096
-    p = mi_os_mem_alloc(suggest, size, commit, MAP_ALIGNED(n), tld->stats);     // use the NetBSD/freeBSD aligned flags
     }
 #else
-  UNUSED(size);
+    // overallocate...
-  UNUSED(alignment);
+    p = mi_os_mem_alloc(over_size, alignment, commit, stats);
+    if (p == NULL) return NULL;
+    // and selectively unmap parts around the over-allocated area.
+    void* aligned_p = mi_align_up_ptr(p, alignment);
+    size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p;
+    size_t mid_size = _mi_align_up(size, _mi_os_page_size());
+    size_t post_size = over_size - pre_size - mid_size;
+    mi_assert_internal(pre_size < over_size && post_size < over_size && mid_size >= size);
+    if (pre_size > 0)  mi_os_mem_free(p, pre_size, stats);
+    if (post_size > 0) mi_os_mem_free((uint8_t*)aligned_p + mid_size, post_size, stats);
+    // we can return the aligned pointer on `mmap` systems
+    p = aligned_p;
 #endif
-  mi_assert(p == NULL || (uintptr_t)p % alignment == 0);
-  if (p != NULL) {
-    _mi_stat_increase(&stats->mmap_calls, 1);
-    _mi_stat_increase(&stats->reserved, size);
-    if (commit) _mi_stat_increase(&stats->committed, size);
   }
+
+  mi_assert_internal(p == NULL || (p != NULL && ((uintptr_t)p % alignment) == 0));
   return p;
 }
 
-// OS page align within a given area,
+/* -----------------------------------------------------------
-// either conservative (pages inside the area only),
+  OS API: alloc, free, alloc_aligned
+----------------------------------------------------------- */
+
+void* _mi_os_alloc(size_t size, mi_stats_t* stats) {
+  if (size == 0) return NULL;
+  size = mi_os_good_alloc_size(size, 0);
+  return mi_os_mem_alloc(size, 0, true, stats);
+}
+
+void  _mi_os_free(void* p, size_t size, mi_stats_t* stats) {
+  if (size == 0 || p == NULL) return;
+  size = mi_os_good_alloc_size(size, 0);
+  mi_os_mem_free(p, size, stats);
+}
+
+void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, mi_os_tld_t* tld)
+{
+  if (size == 0) return NULL;
+  size = mi_os_good_alloc_size(size, alignment);
+  alignment = _mi_align_up(alignment, _mi_os_page_size());
+  return mi_os_mem_alloc_aligned(size, alignment, commit, tld->stats);
+}
+
+
+
+/* -----------------------------------------------------------
+  OS memory API: reset, commit, decommit, protect, unprotect.
+----------------------------------------------------------- */
+
+
+// OS page align within a given area, either conservative (pages inside the area only),
 // or not (straddling pages outside the area is possible)
 static void* mi_os_page_align_areax(bool conservative, void* addr, size_t size, size_t* newsize) {
   mi_assert(addr != NULL && size > 0);
@@ -377,8 +470,7 @@ bool _mi_os_unprotect(void* addr, size_t size) {
   return mi_os_protectx(addr, size, false);
 }
 
-// Commit/Decommit memory.
+// Commit/Decommit memory. Commit is aligned liberal, while decommit is aligned conservative.
-// We page align to a conservative area inside the range to reset.
 static bool mi_os_commitx(void* addr, size_t size, bool commit, mi_stats_t* stats) {
   // page align in the range, commit liberally, decommit conservative
   size_t csize;
@@ -440,106 +532,3 @@ bool _mi_os_shrink(void* p, size_t oldsize, size_t newsize, mi_stats_t* stats) {
 #endif
 }
 
-/* -----------------------------------------------------------
-  OS allocation using mmap/munmap
------------------------------------------------------------ */
-
-void* _mi_os_alloc(size_t size, mi_stats_t* stats) {
-  if (size == 0) return NULL;
-  size = mi_os_good_alloc_size(size, 0);
-  void* p = mi_os_mem_alloc(NULL, size, true, 0, stats);
-  mi_assert(p!=NULL);
-  return p;
-}
-
-void  _mi_os_free(void* p, size_t size, mi_stats_t* stats) {
-  if (size==0) return;
-  size = mi_os_good_alloc_size(size, 0);
-  mi_os_mem_free(p, size, stats);
-}
-
-// Slow but guaranteed way to allocated aligned memory
-// by over-allocating and then reallocating at a fixed aligned
-// address that should be available then.
-static void* mi_os_alloc_aligned_ensured(size_t size, size_t alignment, bool commit, size_t trie, mi_stats_t* stats)
-{
-  if (trie >= 3) return NULL; // stop recursion (only on Windows)
-  if (size > SIZE_MAX - alignment) return NULL; // overflow
-  size_t alloc_size = size + alignment; // no need for -1 as we need to be page aligned anyways
-  
-  // allocate a chunk that includes the alignment
-  void* p = mi_os_mem_alloc(NULL, alloc_size, commit, 0, stats);
-  if (p == NULL) return NULL;
-  // create an aligned pointer in the allocated area
-  void* aligned_p = mi_align_up_ptr(p, alignment);
-  mi_assert(aligned_p != NULL);
-#if _WIN32
-  // free it and try to allocate `size` at exactly `aligned_p`
-  // note: this may fail in case another thread happens to allocate
-  // concurrently at that spot. We try up to 3 times to mitigate this.
-  mi_os_mem_free(p, alloc_size, stats);
-  p = mi_os_mem_alloc(aligned_p, size, commit, 0, stats);
-  if (p != aligned_p) {
-    if (p != NULL) mi_os_mem_free(p, size, stats);
-    return mi_os_alloc_aligned_ensured(size, alignment, commit, trie+1, stats);
-  }
-#else  
-  // we selectively unmap parts around the over-allocated area.
-  size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p;
-  size_t mid_size = _mi_align_up(size, _mi_os_page_size());
-  size_t post_size = alloc_size - pre_size - mid_size;
-  if (pre_size > 0)  mi_os_mem_free(p, pre_size, stats);
-  if (post_size > 0) mi_os_mem_free((uint8_t*)aligned_p + mid_size, post_size, stats);
-#endif
-
-  mi_assert(((uintptr_t)aligned_p) % alignment == 0);
-  return aligned_p;
-}
-
-// Allocate an aligned block.
-// Since `mi_mmap` is relatively slow we try to allocate directly at first and
-// hope to get an aligned address; only when that fails we fall back
-// to a guaranteed method by overallocating at first and adjusting.
-void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, mi_os_tld_t* tld)
-{
-  if (size == 0) return NULL;
-  size = mi_os_good_alloc_size(size,alignment);
-  if (alignment < 1024) return mi_os_mem_alloc(NULL, size, commit, 0, tld->stats);
-
-  // try direct OS aligned allocation; only supported on BSD and Windows 10+
-  void* suggest = NULL;
-  void* p = mi_os_mem_alloc_aligned(size,alignment,commit,tld->stats);
-
-  // Fall back
-  if (p==NULL && (tld->mmap_next_probable % alignment) == 0) {
-    // if the next probable address is aligned,
-    // then try to just allocate `size` and hope it is aligned...
-    p = mi_os_mem_alloc(suggest, size, commit, 0, tld->stats);
-    if (p == NULL) return NULL;
-    if (((uintptr_t)p % alignment) == 0) _mi_stat_increase(&tld->stats->mmap_right_align, 1);
-  }
-  //fprintf(stderr, "segment address guess: %s, p=%lxu, guess:%lxu\n", (p != NULL && (uintptr_t)p % alignment ==0 ? "correct" : "incorrect"), (uintptr_t)p, next_probable);
-
-  if (p==NULL || ((uintptr_t)p % alignment) != 0) {
-    // if `p` is not yet aligned after all, free the block and use a slower
-    // but guaranteed way to allocate an aligned block
-    if (p != NULL) mi_os_mem_free(p, size, tld->stats);
-    _mi_stat_increase( &tld->stats->mmap_ensure_aligned, 1);
-    //fprintf(stderr, "mimalloc: slow mmap 0x%lx\n", _mi_thread_id());
-    p = mi_os_alloc_aligned_ensured(size, alignment,commit,0,tld->stats);
-  }
-  if (p != NULL) {
-    // next probable address is the page-aligned address just after the newly allocated area.
-    size_t probable_size = MI_SEGMENT_SIZE;
-    if (tld->mmap_previous > p) {
-      // Linux tends to allocate downward
-      tld->mmap_next_probable = _mi_align_down((uintptr_t)p - probable_size, os_alloc_granularity); // ((uintptr_t)previous - (uintptr_t)p);
-    }
-    else {
-      // Otherwise, guess the next address is page aligned `size` from current pointer
-      tld->mmap_next_probable = _mi_align_up((uintptr_t)p + probable_size, os_alloc_granularity);
-    }
-    tld->mmap_previous = p;
-  }
-  return p;
-}