Bug，解决DllMain DLL_THREAD_ATTACH时间接加载msvcrt导致死锁风险

Sources/ucrt/inc/corecrt_internal.h

@@ -1488,7 +1488,38 @@ __declspec(dllimport) void __cdecl _amsg_exit(
 
 
 __acrt_ptd* __cdecl __acrt_getptd_head(void);
-#ifdef __BuildWithMSVCRT
+#if WindowsTargetPlatformMinVersion < WindowsTargetPlatformWindows10_10240
+__declspec(noinline) __inline bool __fastcall IsPointerInMsvcrtDll(const void* p)
+{
+    if (!p)
+        return false;
+
+    static uintptr_t s_pBegin;
+    static uintptr_t s_pEnd;
+
+    if (s_pBegin == (uintptr_t)0 || s_pEnd == (uintptr_t)0)
+    {
+        MEMORY_BASIC_INFORMATION _BaseInfo;
+        if (VirtualQuery(&_amsg_exit, &_BaseInfo, sizeof(_BaseInfo)) == 0)
+        {
+            return false;
+        }
+        auto _pBegin = (char*)_BaseInfo.AllocationBase;
+        auto _pEnd = (char*)_BaseInfo.AllocationBase + _BaseInfo.RegionSize;
+        for (; VirtualQuery(_pEnd + 1, &_BaseInfo, sizeof(_BaseInfo));)
+        {
+            if (_pBegin != _BaseInfo.AllocationBase)
+                break;
+
+            _pEnd = (char*)_BaseInfo.BaseAddress + _BaseInfo.RegionSize;
+        }
+        s_pBegin = (uintptr_t)_pBegin;
+        s_pEnd = (uintptr_t)_pEnd;
+    }
+
+    return s_pBegin <= (uintptr_t)p && (uintptr_t)p < s_pEnd;
+}
+
 __declspec(noinline) __inline __acrt_ptd* __cdecl __acrt_getptd_noexit(void)
 {
     // ptd->_thandle 一共有3中情况：
@@ -1500,46 +1531,9 @@ __declspec(noinline) __inline __acrt_ptd* __cdecl __acrt_getptd_noexit(void)
     // 这时必须借助GetThreadId，可是它开销是惊人的，不容易做到快速判断。
     // 结合总总情况，我们现在通过判断 _errno() 返回地址是否在msvcrt模块范围来判断 __getptd_noexit 是否发生内存申请失败。
     // 具体Bug请参考：https://github.com/Chuyu-Team/VC-LTL5/issues/49
-    typedef struct _DllAddressInfo
-    {
-        uintptr_t uBaseAddress;
-        uintptr_t uEndAddress;
-    } DllAddressInfo;
-
-    static DllAddressInfo s_DllAddressCache;
-
-    if (s_DllAddressCache.uBaseAddress == 0)
-    {
-        DllAddressInfo _DllAddressInfo = { (uintptr_t)-1, (uintptr_t)-1};
-        MEMORY_BASIC_INFORMATION _BaseInfo;
-        if (VirtualQuery(&_amsg_exit, &_BaseInfo, sizeof(_BaseInfo)))
-        {
-            _DllAddressInfo.uBaseAddress = (uintptr_t)_BaseInfo.AllocationBase;
-            _DllAddressInfo.uEndAddress = (uintptr_t)_BaseInfo.BaseAddress + _BaseInfo.RegionSize;
-
-            for (; VirtualQuery((void*)(_DllAddressInfo.uEndAddress + 1), &_BaseInfo, sizeof(_BaseInfo));)
-            {
-                if (_DllAddressInfo.uBaseAddress != (uintptr_t)_BaseInfo.AllocationBase)
-                    break;
 
-                _DllAddressInfo.uEndAddress = (uintptr_t)_BaseInfo.BaseAddress + _BaseInfo.RegionSize;
-            }
-        }
-
-#if defined(_X86_) || defined(_ARM_)
-        static_assert(sizeof(s_DllAddressCache) == sizeof(LONGLONG), "");
-        InterlockedCompareExchange64((volatile LONGLONG*)&s_DllAddressCache, *(LONGLONG*)&_DllAddressInfo, 0);
-#elif defined(_IA64_) || defined(_AMD64_) || defined(_ARM64_)
-        // AMD 早期不支持 InterlockedCompareExchange128，所以不用……
-        InterlockedCompareExchange64((volatile LONGLONG*)&s_DllAddressCache.uEndAddress, _DllAddressInfo.uEndAddress, 0);
-        InterlockedCompareExchange64((volatile LONGLONG*)&s_DllAddressCache.uBaseAddress, _DllAddressInfo.uBaseAddress, 0);
-#else
-#error unsrpport!
-#endif
-    }
-
-    uintptr_t _p_errno_value = (uintptr_t)_errno();
-    if (_p_errno_value == 0 || (s_DllAddressCache.uBaseAddress <= _p_errno_value && _p_errno_value < s_DllAddressCache.uEndAddress))
+    auto _p_errno_value = (unsigned char*)_errno();
+    if (_p_errno_value == 0 || IsPointerInMsvcrtDll(_p_errno_value))
     {
         return NULL;
     }
@@ -1551,7 +1545,7 @@ __declspec(noinline) __inline __acrt_ptd* __cdecl __acrt_getptd_noexit(void)
 __acrt_ptd* __cdecl __acrt_getptd_noexit(void);
 #endif
 
-#ifdef __BuildWithMSVCRT
+#if WindowsTargetPlatformMinVersion < WindowsTargetPlatformWindows10_10240
 __declspec(noinline) __inline __acrt_ptd* __cdecl __acrt_getptd(void)
 {
     __acrt_ptd* ptd = __acrt_getptd_noexit();
@@ -2785,4 +2779,4 @@ windowing_model_policy __cdecl __acrt_get_windowing_model_policy(void);
 
 _CRT_END_C_HEADER
 
-//#include <corecrt_internal_state_isolation.h>
+//#include <corecrt_internal_state_isolation.h>

Sources/ucrt/internal/initialization.cpp

@@ -1,4 +1,4 @@
-//
+//
 // initialization.cpp
 //
 //      Copyright (c) Microsoft Corporation.  All rights reserved.
@@ -13,6 +13,10 @@
 #include <stdlib.h>
 #include <stdio.h>
 
+#if WindowsTargetPlatformMinVersion < __MakeVersion(10, 0, 10240)
+bool __cdecl __LTL_ThunksInit();
+#endif
+
 extern "C" {
 
 
@@ -192,7 +196,7 @@ static bool __cdecl uninitialize_allocated_io_buffers(bool const /* terminating
     _free_crt(__acrt_stderr_buffer);
     __acrt_stderr_buffer = nullptr;
 
-#if 0 //msvcrt.dl»áÊÍ·ÅËûÃÇ
+#if 0 //msvcrt.dl会释放他们
     _free_crt(__argv);
     __argv = nullptr;
 
@@ -229,6 +233,9 @@ static __acrt_initializer const __acrt_initializers[] =
     // Global pointers are stored in encoded form; they must be dynamically
     // initialized to the encoded nullptr value before they are used by the CRT.
     { initialize_pointers,                     nullptr                                  },
+#if WindowsTargetPlatformMinVersion < __MakeVersion(10, 0, 10240)
+    { __LTL_ThunksInit,                        nullptr                                  },
+#endif
     // Enclaves only require initializers for supported features.
 #ifndef _UCRT_ENCLAVE_BUILD
     { __acrt_initialize_winapi_thunks,         __acrt_uninitialize_winapi_thunks        },

ucrtbase.msvcrt/fputwc_thunks.cpp

@@ -6,44 +6,42 @@
 
 extern "C" __declspec(dllimport) extern _iobuf _iob[20];
 
+__EXPAND_MSVCRT_FUN(fputc);
 
 static int __cdecl fputc_msvcrt(int const c, FILE* const stream)
 {
-	__EXPAND_MSVCRT_FUN(fputc);
-
-	if (auto pfputc_msvcrt = __Get_MSVCRT_FUN(fputc))
+	if (auto pfputc_msvcrt = try_get_fputc())
 	{
 		return pfputc_msvcrt(c, stream);
 	}
 
 	return EOF;
 }
 
+__EXPAND_MSVCRT_FUN(fputwc);
 
 static wint_t __cdecl fputwc_msvcrt(
 	wchar_t _Character,
 	FILE* _Stream)
 {
-	__EXPAND_MSVCRT_FUN(fputwc);
-
-	if (auto pfputwc_msvcrt = __Get_MSVCRT_FUN(fputwc))
+	if (auto pfputwc_msvcrt = try_get_fputwc())
 	{
 		return pfputwc_msvcrt(_Character, _Stream);
 	}
 
 	return WEOF;
 }
 
+__EXPAND_MSVCRT_FUN(fwrite);
+
 static size_t __cdecl fwrite_msvcrt(
     void const* const buffer,
     size_t      const size,
     size_t      const count,
     FILE*       const stream
     )
 {
-	__EXPAND_MSVCRT_FUN(fwrite);
-
-	if (auto pfwrite_msvcrt = __Get_MSVCRT_FUN(fwrite))
+	if (auto pfwrite_msvcrt = try_get_fwrite())
 	{
 		return pfwrite_msvcrt(buffer, size, count, stream);
 	}

ucrtbase.msvcrt/localeconv.cpp

@@ -67,10 +67,11 @@ static wchar_t* __fastcall MultiByteToWideCharHelper(UINT CodePage, LPCCH lpMult
 
 //
 #if WindowsTargetPlatformMinVersion < WindowsTargetPlatformWindows7
+__EXPAND_MSVCRT_FUN(localeconv);
+
 extern "C" lconv* __cdecl localeconv(void)
 {
-    __EXPAND_MSVCRT_FUN(localeconv);
-
+    auto pMSVCRT_localeconv = try_get_localeconv();
     if (pMSVCRT_localeconv == nullptr)
     {
         //正常不应该为空！！
@@ -171,4 +172,4 @@ extern "C" lconv* __cdecl localeconv(void)
 }
 
 _LCRT_DEFINE_IAT_SYMBOL(localeconv);
-#endif
+#endif

ucrtbase.msvcrt/matherr.cpp

@@ -7,15 +7,19 @@ typedef int(__cdecl* _HANDLE_MATH_ERROR)(struct _exception*);
 
 static _HANDLE_MATH_ERROR user_matherr;
 
+extern "C" void __cdecl __setusermatherr(
+	_HANDLE_MATH_ERROR pf
+	);
+
+__EXPAND_MSVCRT_FUN(__setusermatherr);
+
 extern "C" void __cdecl __setusermatherr(
 	_HANDLE_MATH_ERROR pf
 	)
 {
 	//继续调用msvcrt版本是为了能接管 msvcrt的user_matherr，但是有一个问题，如果其他人调用了 __setusermatherr，这会覆盖msvcrt的user_matherr
 	//二者会发生不一致……这类问题无法一种健全的解决方案，但是可以暂不处理
-	__EXPAND_MSVCRT_FUN(__setusermatherr);
-
-	if (auto p_setusermatherr_msvcrt = __Get_MSVCRT_FUN(__setusermatherr))
+	if (auto p_setusermatherr_msvcrt = try_get___setusermatherr())
 	{
 		p_setusermatherr_msvcrt(pf);
 	}
@@ -45,4 +49,4 @@ extern "C" int __cdecl __acrt_invoke_user_matherr(struct _exception* ex)
 
 
 	return 0;
-}
+}

ucrtbase.msvcrt/msvcrt_loadhelper.cpp

@@ -1,7 +1,15 @@
 #include <Windows.h>
 
+#include <internal_shared.h>
 
-HMODULE __fastcall __LTL_GetMSVCRTModule()
+#include "msvcrt_loadhelper.h"
+
+#pragma comment(linker, "/merge:.YYLT1=.data")
+#pragma comment(linker, "/merge:.YYLT2=.rdata")
+
+typedef void* (__cdecl* InitFunType)();
+
+static HMODULE __fastcall __LTL_GetMSVCRTModule()
 {
 	static HMODULE MSVCRTModuleCache;
 
@@ -21,4 +29,60 @@ HMODULE __fastcall __LTL_GetMSVCRTModule()
 	MSVCRTModuleCache = hMSVCRT ? hMSVCRT : (HMODULE)INVALID_HANDLE_VALUE;
 
 	return hMSVCRT;
-}
+}
+
+__declspec(allocate(".YYLT1$DAA")) static void* s_pfnReserve[] = { nullptr };
+#define __LTL_THUNKS_FUN_START (s_pfnReserve + 1) //指针缓存开始位置
+__declspec(allocate(".YYLT1$DAC")) static void* __LTL_THUNKS_FUN_END[] = { nullptr };   //指针缓存结束位置
+
+__declspec(allocate(".YYLT2$IFA")) static void* const s_pfnInitReserve[] = { nullptr };
+#define __LTL_THUNKS_INIT_FUN_START (s_pfnInitReserve + 1) //函数初始化开始位置
+__declspec(allocate(".YYLT2$IFC")) static void* const __LTL_THUNKS_INIT_FUN_END[] = { nullptr }; //函数初始化开始位置
+
+bool __cdecl __LTL_ThunksInit()
+{
+    for (auto p = __LTL_THUNKS_FUN_START; p != __LTL_THUNKS_FUN_END; ++p)
+    {
+        *p = __crt_fast_encode_pointer(nullptr);
+    }
+
+    for (auto p = (InitFunType*)__LTL_THUNKS_INIT_FUN_START; p != (InitFunType*)__LTL_THUNKS_INIT_FUN_END; ++p)
+    {
+        if (*p == nullptr)
+            continue;
+
+        (*p)();
+    }
+
+    return true;
+}
+
+namespace YY
+{
+    namespace VC_LTL
+    {
+        void* __fastcall try_get_function(
+            void** _ppFunAddress,
+            char      const* const _szName)
+        {
+            auto _p = __crt_fast_decode_pointer(*_ppFunAddress);
+            if (_p)
+            {
+                return _p == INVALID_HANDLE_VALUE ? nullptr : _p;
+            }
+
+            do
+            {
+                auto _hModule = __LTL_GetMSVCRTModule();
+                if (!_hModule)
+                    break;
+
+                _p = GetProcAddress(_hModule, _szName);
+
+                InterlockedExchange((uintptr_t*)_ppFunAddress, (uintptr_t)__crt_fast_encode_pointer(_p ? _p : INVALID_HANDLE_VALUE));
+            } while (false);
+
+            return _p;
+        }
+    }
+}

ucrtbase.msvcrt/msvcrt_loadhelper.h

@@ -1,16 +1,30 @@
 #pragma once
 #include <Windows.h>
 
-HMODULE __fastcall __LTL_GetMSVCRTModule();
-
-#define __EXPAND_MSVCRT_FUN(name) \
-static decltype(name)* _CRT_CONCATENATE(pMSVCRT_, name); \
-if(_CRT_CONCATENATE(pMSVCRT_, name) == NULL)\
-{ \
-    auto fun = GetProcAddress(__LTL_GetMSVCRTModule(), #name);\
-    _CRT_CONCATENATE(pMSVCRT_, name) = fun ? (decltype(name)*)fun : (decltype(name)*)INVALID_HANDLE_VALUE;\
-}
+#pragma section(".YYLT1$DAA",    long, read, write)
+#pragma section(".YYLT1$DAB",    long, read, write) //LTL函数缓存节点
+#pragma section(".YYLT1$DAC",    long, read, write) //保留，暂时用于边界结束
 
+#pragma section(".YYLT2$IFA",    long, read)
+#pragma section(".YYLT2$IFB",    long, read)        //LTL函数缓存初始化函数
+#pragma section(".YYLT2$IFC",    long, read)        //LTL函数缓存初始化函数结束
 
-#define __Get_MSVCRT_FUN(name) (_CRT_CONCATENATE(pMSVCRT_, name) != (decltype(name)*)INVALID_HANDLE_VALUE ? _CRT_CONCATENATE(pMSVCRT_, name) : (decltype(name)*)NULL)
+namespace YY
+{
+    namespace VC_LTL
+    {
+        void* __fastcall try_get_function(
+            void** _ppFunAddress,
+            char      const* const _szName);
+    }
+}
 
+#define __EXPAND_MSVCRT_FUN(name)                                                                           \
+static decltype(name)* _CRT_CONCATENATE(try_get_, name)()                                                   \
+{                                                                                                           \
+    __declspec(allocate(".YYLT2$IFB")) static void* const s_pfnInitFun = &_CRT_CONCATENATE(try_get_, name); \
+    __pragma(warning(suppress:6031))                                                                        \
+	_bittest((LONG*)(&s_pfnInitFun), 0);                                                                    \
+    __declspec(allocate(".YYLT1$DAB")) static void* _CRT_CONCATENATE(s_pfn_, name);                         \
+    return (decltype(name)*)YY::VC_LTL::try_get_function(&_CRT_CONCATENATE(s_pfn_, name), #name);           \
+}