Detours implementation of overriding CRT

This commit adds an override to Windows for replacing the CRT malloc/free etc routines.

Author: mjp41

.github/workflows/main.yml

@@ -403,7 +403,7 @@ jobs:
         build-type: [ Release, Debug ]
         arch: [ Win32, x64 ]
         toolchain: [ "", "-T ClangCL" ]
-        extra-cmake-flags: [ "" ]
+        extra-cmake-flags: [ "", "-S src/windows" ]
         # Add an extra check for the Windows 8 compatible PAL
         include:
           - os: windows-2022

src/snmalloc/pal/pal_windows.h

@@ -41,8 +41,10 @@
  * be destroyed when the program exits or the DLL is
  * unloaded.
  */
+#  pragma warning(push)
 #  pragma warning(disable : 4075)
 #  pragma init_seg(".CRT$XCB")
+#  pragma warning(pop)
 
 namespace snmalloc
 {

src/windows/CMakeLists.txt

@@ -0,0 +1,46 @@
+cmake_minimum_required(VERSION 3.14)
+include(FetchContent)
+
+FetchContent_Declare(
+  vcpkg
+  GIT_REPOSITORY https://github.com/microsoft/vcpkg
+  GIT_TAG        2025.06.13
+  GIT_SHALLOW    TRUE
+)
+
+FetchContent_MakeAvailable(vcpkg)
+
+set(CMAKE_TOOLCHAIN_FILE
+            ${vcpkg_SOURCE_DIR}/scripts/buildsystems/vcpkg.cmake
+            CACHE STRING "vcpkg toolchain file")
+
+project(snmallocdetours CXX)
+
+add_subdirectory(../.. src/snmalloc EXCLUDE_FROM_ALL)
+
+# On Windows, we need to use the detours library to override the malloc
+# functions.
+find_path(DETOURS_INCLUDE_DIRS "detours/detours.h")
+find_library(DETOURS_LIBRARY detours REQUIRED)
+
+add_library(snmallocdetours OBJECT override/detours.cc)
+
+target_include_directories(snmallocdetours PRIVATE ${DETOURS_INCLUDE_DIRS})
+target_link_libraries(snmallocdetours PUBLIC snmalloc)
+target_link_libraries(snmallocdetours PRIVATE ${DETOURS_LIBRARY})
+
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+add_library(snmallocdetourslib SHARED override/lib.cc)
+target_include_directories(snmallocdetourslib PRIVATE ${DETOURS_INCLUDE_DIRS})
+target_link_libraries(snmallocdetourslib PRIVATE snmallocdetours)
+
+add_executable(snmallocdetoursexample test/main.cc)
+target_link_libraries(snmallocdetoursexample snmallocdetourslib)
+# Check if 32 bit windows.
+if (CMAKE_SIZEOF_VOID_P EQUAL 4)
+  target_link_options(snmallocdetoursexample PRIVATE "/INCLUDE:_is_snmalloc_detour")
+else()
+  target_link_options(snmallocdetoursexample PRIVATE "/INCLUDE:is_snmalloc_detour")
+endif()
+
+add_test(NAME snmallocdetours COMMAND snmallocdetoursexample)

src/windows/README.md

@@ -0,0 +1,81 @@
+# Windows snmalloc override
+
+This directory contains an implementation of the snmalloc allocator that
+overrides the default Windows allocator for the C-runtime (CRT).
+The implementation uses the detours library to intercept calls to the CRT
+and redirects them to snmalloc.
+
+# Building
+
+Build the cmake projects in this directory.  From the root of the repository, do
+
+```bash
+cmake -B build_detours -S src/windows
+cmake --build build_detours
+```
+
+This will install the detours library using vcpkg and then build the snmalloc
+static library, and a DLL that automatically load snmalloc as the CRT allocator.
+
+# Using
+
+## DLL 
+To use the snmalloc allocator in your Windows application, you need to link against the `snmallocdetourslib.dll` that was built in the previous step.
+You can do this by adding the following line to your CMakeLists.txt:
+
+```cmake
+target_link_libraries(your_target_name PRIVATE snmallocdetourslib)
+target_link_options(snmallocdetoursexample PRIVATE "/INCLUDE:is_snmalloc_detour")
+```
+
+The second line is necessary to ensure that the linker does not optimize away the DLL import.
+
+
+## Static Library
+
+To use the snmalloc allocator as a static library, you can link against the `snmallocdetourslib.lib` file that was built in the previous step.
+You can do this by adding the following line to your CMakeLists.txt:
+
+```cmake
+target_link_libraries(your_target_name PRIVATE snmallocdetourslib_static)
+```
+
+Then you need to cause the detours routine to be run by adding the following lines to a C++ source file:
+
+```cpp
+#include "detours.h"
+#pragma warning(push)
+#pragma warning(disable : 4075)
+#pragma init_seg(".CRT$XCV")
+static SnmallocDetour snmalloc_detour;
+#pragma warning(pop)
+```
+
+This causes the detours code to be run early in the program startup.
+
+## Locally scoped detour
+
+Finally, you can use the detour in a locally scoped manner.
+This requires linking against the `snmallocdetourslib.lib`, and then can be used as:
+
+```cpp
+#include "detours.h"
+
+
+void my_function()
+{
+  SnmallocDetour snmalloc_detour;
+  // snmalloc is now the CRT allocator for this function
+  // ...
+} // snmalloc_detour goes out of scope and the CRT allocator is restored
+```
+
+Upon exiting the scope of `snmalloc_detour`, the CRT allocator will be restored to its original state.
+
+
+# Status
+
+This implementation is currently in an alpha state, and is not yet suitable for production use.
+
+Significant testing is required to ensure that it works for a range of applications.
+If you sucessfully use this in your application, please let us know by commenting on #700.

src/windows/override/detours.cc

@@ -0,0 +1,156 @@
+/*
+ * This file provides a Windows-specific overriding of the malloc,
+ * free, calloc, realloc, and msize functions using the Detours library.
+ */
+#ifdef _DEBUG
+#  include <crtdbg.h>
+#endif
+
+#include "detours.h"
+
+#include <new>
+#include <stdlib.h>
+
+// Symbols for the original malloc, free, calloc, realloc, and msize functions
+auto* original_malloc = malloc;
+auto* original_calloc = calloc;
+auto* original_realloc = realloc;
+auto* original_free = free;
+auto* original_msize = _msize;
+
+void* (*original_new)(size_t) = operator new;
+void* (*original_new2)(size_t, const std::nothrow_t&) = operator new;
+
+void* (*original_new_array)(size_t) = operator new[];
+void* (*original_new_array2)(size_t, const std::nothrow_t&) = operator new[];
+
+void (*original_delete)(void*) = operator delete;
+void (*original_delete2)(void*, size_t) = operator delete;
+void (*original_delete3)(void*, const std::nothrow_t&) = operator delete;
+
+void (*original_delete_array)(void*) = operator delete[];
+void (*original_delete_array2)(void*, size_t) = operator delete[];
+
+#include <snmalloc/snmalloc_core.h>
+// Provides the global configuration for the snmalloc implementation.
+#include <snmalloc/backend/globalconfig.h>
+
+#define SNMALLOC_PROVIDE_OWN_CONFIG
+
+namespace snmalloc
+{
+  class WindowsHeapAsSecondaryAllocator
+  {
+  public:
+    // This flag is used to turn off checks on fast paths if the secondary
+    // allocator does not own the memory at all.
+    static constexpr inline bool pass_through = false;
+
+    SNMALLOC_FAST_PATH
+    static void initialize() {}
+
+    // We always use snmalloc for allocation.
+    template<class SizeAlign>
+    SNMALLOC_FAST_PATH static void* allocate(SizeAlign&&)
+    {
+      return nullptr;
+    }
+
+    // If the memory was not deallocated by snmalloc, then try the
+    // original free.
+    SNMALLOC_FAST_PATH
+    static void deallocate(void* pointer)
+    {
+      if (pointer == nullptr)
+        return;
+
+      original_free(pointer);
+    }
+
+    SNMALLOC_FAST_PATH
+    static size_t alloc_size(const void* p)
+    {
+      return original_msize(const_cast<void*>(p));
+    }
+  };
+
+  // Root failed deallocations and msize requests to the Windows heap.
+  using Config = snmalloc::StandardConfigClientMeta<
+    NoClientMetaDataProvider,
+    WindowsHeapAsSecondaryAllocator>;
+  using Alloc = snmalloc::Allocator<Config>;
+} // namespace snmalloc
+
+#define SNMALLOC_STATIC_LIBRARY_PREFIX snmalloc_
+#include "detours/detours.h"
+
+#include <snmalloc/override/malloc.cc>
+#include <stdio.h>
+#include <windows.h>
+
+// This name is not provided by malloc.cc above, so we define it here.
+size_t snmalloc_msize(void* ptr)
+{
+  // Call the snmalloc function to get the allocation size.
+  // This is not accurate as it rounds up, whereas the original msize
+  // function returns the exact size of the allocation.
+  return snmalloc::alloc_size(ptr);
+}
+
+SnmallocDetour::SnmallocDetour()
+{
+  // Initilialize snmalloc.
+  snmalloc_free(snmalloc_malloc(1));
+
+  DetourTransactionBegin();
+  DetourAttach(&(PVOID&)original_free, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete2, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete3, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete_array, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete_array2, snmalloc_free);
+  DetourAttach(&(PVOID&)original_malloc, snmalloc_malloc);
+  DetourAttach(&(PVOID&)original_calloc, snmalloc_calloc);
+  DetourAttach(&(PVOID&)original_realloc, snmalloc_realloc);
+  DetourAttach(&(PVOID&)original_msize, snmalloc_msize);
+  DetourAttach(&(PVOID&)original_new, snmalloc_malloc);
+  DetourAttach(&(PVOID&)original_new2, snmalloc_malloc);
+  DetourAttach(&(PVOID&)original_new_array, snmalloc_malloc);
+  DetourAttach(&(PVOID&)original_new_array2, snmalloc_malloc);
+
+  DetourTransactionCommit();
+}
+
+SnmallocDetour::~SnmallocDetour()
+{
+  // Detours performs allocation so during this some data structures will
+  // be allocated with snmalloc.  These cannot be handled by the Windows heap
+  // so leave snmalloc::free in place to handle these allocations.
+
+  DetourTransactionBegin();
+  DetourDetach(&(PVOID&)original_calloc, snmalloc_calloc);
+  DetourDetach(&(PVOID&)original_realloc, snmalloc_realloc);
+  DetourDetach(&(PVOID&)original_malloc, snmalloc_malloc);
+  DetourDetach(&(PVOID&)original_msize, snmalloc_msize);
+  DetourDetach(&(PVOID&)original_new, snmalloc_malloc);
+  DetourDetach(&(PVOID&)original_new2, snmalloc_malloc);
+  DetourDetach(&(PVOID&)original_new_array, snmalloc_malloc);
+  DetourDetach(&(PVOID&)original_new_array2, snmalloc_malloc);
+  DetourTransactionCommit();
+
+  // This transaction's allocation will come from the Windows heap, so it is
+  // safe to use the Windows heap's free during teardown.
+  DetourTransactionBegin();
+  DetourDetach(&(PVOID&)original_free, snmalloc_free);
+  DetourDetach(&(PVOID&)original_delete, snmalloc_free);
+  DetourAttach(&(PVOID&)original_delete2, snmalloc_free);
+  DetourDetach(&(PVOID&)original_delete3, snmalloc_free);
+  DetourDetach(&(PVOID&)original_delete_array, snmalloc_free);
+  DetourDetach(&(PVOID&)original_delete_array2, snmalloc_free);
+  DetourTransactionCommit();
+}
+
+extern "C" bool is_snmalloc_detour(void* ptr)
+{
+  return snmalloc::is_owned(ptr);
+}

src/windows/override/detours.h

@@ -0,0 +1,11 @@
+#pragma once
+
+class SnmallocDetour
+{
+public:
+  SnmallocDetour();
+  ~SnmallocDetour();
+};
+
+// Used to check that a pointer is from the snmalloc detour.
+extern "C" __declspec(dllexport) bool is_snmalloc_detour(void* ptr);

src/windows/override/lib.cc

@@ -0,0 +1,11 @@
+#include "detours.h"
+
+#pragma warning(push)
+#pragma warning(disable : 4075)
+// This pragma uses a segment that is alphabetically later than the
+// one used in pal_windows.h.  This is required to ensure that the
+// global function pointers have been initialized before we attempt to
+// detour them.
+#pragma init_seg(".CRT$XCV")
+static SnmallocDetour snmalloc_detour;
+#pragma warning(pop)

src/windows/test/main.cc

@@ -0,0 +1,22 @@
+#include "../override/detours.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+
+// This is done in CMake, but could be done here as well
+// #pragma comment(linker, "/include:is_snmalloc_detour")
+
+int main()
+{
+  auto p_old = malloc(16);
+
+  if (p_old != nullptr && !is_snmalloc_detour(p_old))
+  {
+    printf("Detouring malloc and free failed...\n");
+  }
+
+  free(p_old);
+
+  printf("Test passed: Detouring malloc and free succeeded.\n");
+  return 0;
+}

src/windows/vcpkg-configuration.json

@@ -0,0 +1,14 @@
+{
+  "default-registry": {
+    "kind": "git",
+    "baseline": "1a66c32c6f90c2f646529975c3c076ed3dbdae0c",
+    "repository": "https://github.com/microsoft/vcpkg"
+  },
+  "registries": [
+    {
+      "kind": "artifact",
+      "location": "https://github.com/microsoft/vcpkg-ce-catalog/archive/refs/heads/main.zip",
+      "name": "microsoft"
+    }
+  ]
+}

src/windows/vcpkg.json

@@ -0,0 +1,5 @@
+{
+  "dependencies": [
+    "detours"
+  ]
+}