Can we implement c++ thunk in linux?

Question

I want to use class member functions as callbacks, I don't use libsigc, because it's slow. In ATL, we can use member function for C-style callback(http://www.codeproject.com/KB/cpp/SoloGenericCallBack.aspx), so can we implement c++ thunk in linux?

The code below will crash:

#include <assert.h>
#include <stdio.h>
#include <sys/mman.h>

typedef char BYTE;
typedef int DWORD;
typedef int* DWORD_PTR;
typedef int* INT_PTR;
typedef bool BOOL;
typedef unsigned long ULONG;
typedef unsigned long* ULONG_PTR;
#define PtrToUlong( p ) ((ULONG)(ULONG_PTR) (p) )
#define __stdcall __attribute__((__stdcall__))

//#pragma pack( push, 1 )
struct  MemFunToStdCallThunk
{
    BYTE          m_mov;
    DWORD      m_this; 
    BYTE          m_pushEax;
    BYTE          m_jmp;
    DWORD      m_relproc;

    void  Init( DWORD_PTR proc, void* pThis )
    {
        printf("proc=%x\n", proc);
        m_mov = 0xB8;   // mov eax
        m_this = PtrToUlong(pThis);
        m_pushEax = 0xc3;// push eax
        m_jmp = 0xe9;          //jmp
        m_relproc = DWORD((INT_PTR)proc - ((INT_PTR)this+sizeof(MemFunToStdCallThunk)));
        printf("m_relproc = %x\n", m_relproc);
        mprotect(this, sizeof(MemFunToStdCallThunk), PROT_READ|PROT_WRITE|PROT_EXEC);
    }

    void* GetCodeAddress()
    {
        return this;
    }
}__attribute__ ((packed));
//#pragma  pack( pop )

template< typename TDst, typename TSrc >
TDst  UnionCastType( TSrc src )
{
    union
    {
        struct
        {
            int* pfn;  //function,index
            long delta; // offset, 
        }funcPtr;
        TSrc  uSrc;
    }uMedia;
    uMedia.uSrc  = src;
    return uMedia.funcPtr.pfn;
}




typedef  int  ( __stdcall *StdCallFun)(int, int);
class  CTestClass
{
public:
    int  m_nBase;
    MemFunToStdCallThunk  m_thunk;

    int  memFun( int m, int n )
    {
        int  nSun = m_nBase + m + n;
        printf("m=%d,n=%d,nSun=%d\n", m, n, nSun);
        return 1234;
    }

public:
    CTestClass()
    {
        m_nBase  = 10;
    }

    void  Test()
    {
        printf("%x\n", &CTestClass::memFun);
        m_thunk.Init(UnionCastType<DWORD_PTR>(&CTestClass::memFun), this );
        StdCallFun fun = (StdCallFun)m_thunk.GetCodeAddress();
        assert( fun != NULL );

        int ret =   fun( 9, 3 );
        printf("ret = %x\n", ret);


    }
};



int main()
{
    CTestClass test;
    test.Test();
    return 0;
}

EDIT: Thanks to user786653, I get the right answer:

#include <assert.h>
#include <stdio.h>
#include <sys/mman.h>
#include <unistd.h>

typedef char BYTE;
typedef int DWORD;
typedef int* DWORD_PTR;
typedef int* INT_PTR;
typedef bool BOOL;
typedef unsigned long ULONG;
typedef unsigned long* ULONG_PTR;
#define PtrToUlong(p) ((ULONG)(ULONG_PTR) (p) )
#define __stdcall __attribute__((__stdcall__))

struct  MemFunToStdCallThunk
{
    BYTE          m_repairStack[10];
    DWORD      m_mov;
    DWORD      m_this;
    BYTE          m_jmp;
    DWORD      m_relproc;

    void  Init( DWORD_PTR proc, void* pThis )
    {
        printf("proc=%p\n", proc);
        m_repairStack[0] = 0x83; //sub esp, 0x4
        m_repairStack[1] = 0xec;
        m_repairStack[2] = 0x04;
        m_repairStack[3] = 0x8b; //mov eax,[esp + 0x4]
        m_repairStack[4] = 0x44;
        m_repairStack[5] = 0x24;
        m_repairStack[6] = 0x04;
        m_repairStack[7] = 0x89;//mov [esp], eax
        m_repairStack[8] = 0x04;
        m_repairStack[9] = 0x24;
        m_mov = 0x042444C7;   // mov   dword   ptr   [esp+0x4], 
        m_this = PtrToUlong(pThis);
        m_jmp = 0xe9;          //jmp
        m_relproc = (DWORD)proc - ((DWORD)this+sizeof(MemFunToStdCallThunk));
        printf("m_relproc = %d\n", m_relproc);
        //long page_size = sysconf(_SC_PAGE_SIZE);
        //mprotect((void*)(PtrToUlong(this) & -page_size), 2*page_size, PROT_READ|PROT_WRITE|PROT_EXEC);
    }

    void* GetCodeAddress()
    {
        return this;
    }
}__attribute__ ((packed));


template< typename TDst, typename TSrc >
TDst  UnionCastType( TSrc src )
{
    union
    {
        struct
        {
            int* pfn;  //function or index
            long delta; // offset
        }funcPtr;
        TSrc  uSrc;
    }uMedia;
    uMedia.uSrc  = src;
    return uMedia.funcPtr.pfn;
}




typedef  int  ( __stdcall *StdCallFun)(int, int);
class  CTestClass
{
public:
    int  m_nBase;
    MemFunToStdCallThunk  m_thunk;

    int  memFun( int m, int n )
    {
    printf("this=%p\n", this);
        int  nSun = m_nBase + m + n;
        printf("m=%d,n=%d,nSun=%d\n", m, n, nSun);
        return nSun;
    }

public:
    CTestClass()
    {
        m_nBase  = 10;
    }

    void  Test()
    {
        int (CTestClass::*abc)(int, int);
        printf("sizeof(MemFunToStdCallThunk)=%d,sizeof(abc)=%d\n", sizeof(MemFunToStdCallThunk), sizeof(abc));
        printf("memFun=%p\n", &CTestClass::memFun);
        m_thunk.Init(UnionCastType<DWORD_PTR>(&CTestClass::memFun), this );
        StdCallFun fun = (StdCallFun)m_thunk.GetCodeAddress();
        assert( fun != NULL );

        int ret = memFun(2, 3);
        printf("ret 1= %d\n", ret);
        ret =  fun( 9, 3 );
        printf("ret 2= %d\n", ret);
    }
};


int main()
{
    CTestClass test;
    test.Test();
    return 0;
}

Answer 1

Yes, but I wouldn't recommend it. It will (obviously) make your code a lot less portable and you're potentially opening a security hole if you're not careful.

You will need to make the code executable with mprotect(2). Something like mprotect(&thunk_struct, sizeof(struct _CallBackProcThunk), PROT_READ|PROT_WRITE|PROT_EXEC).

Also the normal GCC syntax for structure packing is struct S { /* ... */ } __attribute__ ((packed)) though newer versions might support the #pragma pack syntax.

You will probably also want to substitute DWORD with uint32_t from stdint.h and BYTE with uint8_t (or just stick a typedef in there).

EDIT:

From the man page on mprotect "[..]addr must be aligned to a page boundary". You should check the return value. Try doing something like this instead:

long page_size = sysconf(_SC_PAGE_SIZE);
uintptr_t addr = ((uintptr_t)this) & -page_size;
if (mprotect((void*)addr, 2*page_size, PROT_READ|PROT_WRITE|PROT_EXEC)) {
    perror("mprotect"); 
    /* handle error */
}

The following calculation is wrong:

DWORD((INT_PTR)proc - ((INT_PTR)this+sizeof(MemFunToStdCallThunk)))

It's doing its calculations on int*'s.

(DWORD)proc - ((DWORD)this+sizeof(MemFunToStdCallThunk)

should be sufficient here.

A very ugly (non-portable etc. etc.), but small and self-contained example follows:

#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <unistd.h>
#include <sys/mman.h>

struct thunk {
    uint32_t mov;
    uint32_t this_ptr;
    uint8_t jmp;
    uint32_t rel;
} __attribute__((packed));

class Test {
public:
    virtual int foo(void) {
        printf("foo! %p\n", (void*)this);
        return 42;
    }
};

int main()
{
    Test test;
    printf("%d\n", test.foo());

    thunk t;
    t.mov = 0x042444C7;
    t.this_ptr = (uint32_t)&test;
    t.jmp = 0xe9;
    t.rel = ((uint32_t)(void*)&Test::foo) - ((uint32_t)&t + sizeof(thunk));

    uint32_t addr = (uint32_t)&t;
    long page_size = sysconf(_SC_PAGE_SIZE);
    if (mprotect((void*)(addr & -page_size), 2*page_size, PROT_READ|PROT_WRITE|PROT_EXEC)) {
        perror("mprotect");
        return 1;
    }

    union {
        void* p;
        int (*foo)(int);
    } u;
    u.p = &t;
    printf("%d\n", u.foo(0));
    return 0;
}

Answer 2

A reasonable approach is something like this:

struct Foo {
   void doit();
};

extern "C" {
   void callback(void *handle) {
      reinterpret_cast<Foo*>(handle)->doit();
   }
}

The assembly of callback looks like this here (x64):

callback:
    jmpq    _ZN3Foo4doitEv

Answer 3

You can't pass pointer-to-member pointers to C callbacks directly, but there are portable tricks (i.e. not restricted to one target OS) that work very well.

The easiest way to do that is just to use a wrapper non-member function whose only purpose is to call your member function.

void wrapper()
{
  object->callWhatever();
}

You can pass wrapper() as a function pointer.

See also for example Cast member function for create_pthread() call for how to handle cases where you get a void* parameter with the callback and want to use that to store (directly or not) a reference/pointer to the object you want to operate on.

Answer 4

I want to use class member functions as callbacks, I don't use libsigc, because it's slow. In ATL, we can use member function for C-style callback(http://www.codeproject.com/KB/cpp/SoloGenericCallBack.aspx), so can we implement c++ thunk in linux?

You probably can. However, there is no need to.

Most asynchronous APIs allow to pass a void* argument when registering for an asynchronous event. When the event gets reported this void* is reported as well and can be used to call a member function of an object. (Vague language because APIs like epoll_wait() don't actually call you back, where as pthread_create() does.).