Hook lua库函数时遇到的问题

最近在为distri.lua实现一个lua调试系统,有一个简单的需求,lua导入一个文件的时候,将这个文件的文件名记录下来,

以方便调试器在设置断点的时候判断是否一个合法的文件.

lua导入文件是通过luaL_loadfilex实现的,一个简单的思路就是修改luaL_loadfilex,在luaL_loadfilex中调用一个外部定义的函数将导入的文件名传给那个外部函数，由它记录下来.

但这种侵入式的方案，除非在逼不得已的情况下不应该使用.

另一个思路是hook luaL_loadfilex,在运行时用另外一个函数替换luaL_loadfilex,由这个替换函数去记录下需要的信息

然后在跳转回原luaL_loadfilex的执行流程上.

与是我从decode中提取除了Hook.h,Hook.c稍加调整以适应Linux系统.

hook的原理很简单:

    * 首先,使用mprotect将luaL_loadfilex所在代码段设置为可读/可写/可执行,以避免在修改代码时出现段访问异常

    * 之后需要把luaL_loadfilex最前面的一段指令替换成一个跳转指令,跳转到替换函数中去执行.为了在替换函数执行
    完之后可以正确的回到luaL_loadfilex的正常执行路径上,需要把luaL_loadfilex中被替换部分的指令保存下来,然后
    在后面再添加一条跳转指令,调到luaL_loadfilex后面的执行路径去.

被替换掉的指令布局和执行流程如下图:

    luaL_loadfilex:
             jmp hook
             -------------
             其余指令  <-------------------------------------------------|
                                                                         |
    hook:                                                                |
            执行必要的记录                                               |
            --保存的代码----                                             |
            luaL_loadfilex中被替换的指令                                 |
            jmp其余指令--------------------------------------------------|

HookFunction实现如下:

    void* HookFunction(void* function, void* hook)
    {

        // Don't allow rehooking of the same function since that screws things up.

        assert(!GetIsHooked(function, hook));

        if (GetIsHooked(function, hook))
        {
            return NULL;
        }

        // Missing from windows.h
        //#define HEAP_CREATE_ENABLE_EXECUTE 0x00040000

        // Jump instruction is 5 bytes.
        const int jumpSize = 5;

        // Compute the instruction boundary so we don't override half an instruction.
        int boundary = GetInstructionBoundary(function, jumpSize);

        size_t pagesize = sysconf(_SC_PAGE_SIZE);
        unsigned char* trampoline = NULL;
        trampoline = (unsigned char*)/*aligned_alloc*/memalign(pagesize,pagesize);
    	if(mprotect(trampoline, pagesize, PROT_WRITE|PROT_READ|PROT_EXEC)){
    		free(trampoline);
    		return NULL;
    	}

        // Copy the original bytes to the trampoline area and append a jump back
        // to the original function (after our jump).

        memcpy(trampoline, function, boundary);
        AdjustRelativeJumps(trampoline, boundary, ((unsigned char*)function) - trampoline);

        WriteJump(trampoline + boundary, ((unsigned char*)function) + boundary);

    	void *ptr = (void*)(((size_t)function/pagesize)*pagesize);
        // Give ourself write access to the region.
        if(!mprotect(ptr, pagesize, PROT_WRITE|PROT_READ|PROT_EXEC))
        {
            // Fill the area with nops and write the jump instruction to our
            // hook function.
            memset(function, 0x90, boundary);
            WriteJump(function, hook);

            // Restore the original protections.
            //VirtualProtect(function, boundary, protection, &protection);
    		mprotect(ptr, pagesize, PROT_READ|PROT_EXEC);
            // Flush the cache so we know that our new code gets executed.
            //FlushInstructionCache(GetCurrentProcess(), NULL, NULL);

            return trampoline;
            //return 0;

        }    

    	free(trampoline);
    	return NULL;

        //return -1;
    }

本以为一切就这样结束了,运行程序的时候,正确的进入了替换函数,但在执行完记录操作要回到luaL_loadfilex后续执行流程的时候程序

挂了,报段访问异常.

为啥呢,我们看下WriteJump的实现:

/**
 * Writes a relative jmp instruction.
 */
void WriteJump(void* dst, void* address)
{

    unsigned char* jump = (unsigned char*)(dst);

    // Setup a jump instruction.
    jump[0] = 0xE9;
    *((unsigned long*)(&jump[1])) = (unsigned long)(address) - (unsigned long)(dst) - 5;

}

使用的是E9跳转指令+4字节的立即数做相对rip计数器的跳转.这在32位程序下这是没问题的,因为trampoline和luaL_loadfilex的位移差必定

在4字节的范围内.但我程序的运行环境是64位的,这个时候程序就出问题了, trampoline和luaL_loadfilex的位移差已经超过4个字节.这就导致

[jmp其余指令]跳转到到错误的地址上了.

如何解决这个问题:

*   用FF指令做跳转,但这个方案要修改的地方就多了,除了` WriteJump`,还有`AdjustRelativeJumps`并且还会导致指令长度变长.

*   用static数据区保存luaL_loadfilex中被替换的指令,使得位移差被控制在4字节以内.

针对我的需求，我选择了方案2,下面是修改过的HookFunction和WriteJump以及使用示例:

    /**
     * Writes a relative jmp instruction.
     */
    void WriteJump(void* dst, void* address)
    {

        unsigned char* jump = (unsigned char*)(dst);

        // Setup a jump instruction.
        jump[0] = 0xE9;
        *((unsigned int*)(&jump[1])) = (unsigned int)((unsigned long)(address) - (unsigned long)(dst) - 5);

    }

    void* HookFunction(void* function, void* hook,void *saveaddr,size_t saveaddr_size)
    {

        // Don't allow rehooking of the same function since that screws things up.

        assert(!GetIsHooked(function, hook));

        if (GetIsHooked(function, hook))
        {
            return NULL;
        }

        // Missing from windows.h
        //#define HEAP_CREATE_ENABLE_EXECUTE 0x00040000

        // Jump instruction is 5 bytes.
        const int jumpSize = 5;

        // Compute the instruction boundary so we don't override half an instruction.
        int boundary = GetInstructionBoundary(function, jumpSize);

        if(saveaddr_size < (size_t)boundary) return NULL;

        size_t pagesize = sysconf(_SC_PAGE_SIZE);
    	if(mprotect(saveaddr, boundary, PROT_WRITE|PROT_READ|PROT_EXEC)){
    		//free(trampoline);
    		return NULL;
    	}

        // Copy the original bytes to the trampoline area and append a jump back
        // to the original function (after our jump).

        memcpy(saveaddr, function, boundary);
        AdjustRelativeJumps(saveaddr, boundary, ((unsigned char*)function) - (unsigned char*)saveaddr);

        WriteJump(saveaddr + boundary, ((unsigned char*)function) + boundary);

    	void *ptr = (void*)(((size_t)function/pagesize)*pagesize);
        // Give ourself write access to the region.
        if(!mprotect(ptr, pagesize, PROT_WRITE|PROT_READ|PROT_EXEC))
        {
            // Fill the area with nops and write the jump instruction to our
            // hook function.
            memset(function, 0x90, boundary);
            WriteJump(function, hook);

            // Restore the original protections.
            //VirtualProtect(function, boundary, protection, &protection);
    		mprotect(ptr, pagesize, PROT_READ|PROT_EXEC);
            // Flush the cache so we know that our new code gets executed.
            //FlushInstructionCache(GetCurrentProcess(), NULL, NULL);

            return saveaddr;
            //return 0;

        }    

    	return NULL;

        //return -1;

    }

////使用示例

    int (*ori_luaL_loadfilex)(lua_State *L, const char *filename,const char *mode) = NULL;

    int my_luaL_loadfilex(lua_State *L, const char *filename,const char *mode){
    	printf("%s\n",filename);//记录导入的lua文件，供调试器使用
    	return ori_luaL_loadfilex(L,filename,mode);
    }

    static char luaL_loadfilex_buf[4096] __attribute__((aligned(4096)));

    int debug_init(){
    	ori_luaL_loadfilex = HookFunction(luaL_loadfilex,my_luaL_loadfilex,luaL_loadfilex_buf,4096);
    	if(!ori_luaL_loadfilex){
    		return -1;
    	}
    	return 0;
    }