Apple XNU Kernel - Memory Corruption due to Integer Overflow in __offsetof Usage in posix_spawn on 32-bit Platforms







posix_spawn is a complex syscall which takes a lot of arguments from userspace. The third argument
is a pointer to a further arguments descriptor in userspace with the following structure (on 32-bit):

  struct user32__posix_spawn_args_desc {
    uint32_t  attr_size;  /* size of attributes block */
    uint32_t  attrp;    /* pointer to block */
    uint32_t  file_actions_size;  /* size of file actions block */
    uint32_t  file_actions;  /* pointer to block */
    uint32_t  port_actions_size;  /* size of port actions block */
    uint32_t  port_actions;  /* pointer to block */
    uint32_t  mac_extensions_size;
    uint32_t  mac_extensions;
    uint32_t  coal_info_size;
    uint32_t  coal_info;
    uint32_t  persona_info_size;
    uint32_t  persona_info;

port_actions then points to another structure in userspace of this type:

  struct _posix_spawn_port_actions {
    int      pspa_alloc;
    int      pspa_count;
    _ps_port_action_t   pspa_actions[];

and finally _ps_port_action_t looks like this:

  struct _ps_port_action {
    pspa_t      port_type;
    exception_mask_t  mask;
    mach_port_name_t  new_port;
    exception_behavior_t  behavior;
    thread_state_flavor_t  flavor;
    int      which;

Note that pspa_actions is a zero-sized array. pspa_count is supposed to be the number of entries
in this array.

The following constraints are checked in posix_spawn in kern_exec.c:

  if (px_args.port_actions_size != 0) {
    /* Limit port_actions to one page of data */
    if (px_args.port_actions_size < PS_PORT_ACTIONS_SIZE(1) ||
        px_args.port_actions_size > PAGE_SIZE) {
      error = EINVAL;
      goto bad;

PS_PORT_ACTIONS_SIZE is defined like this:

  #define  PS_PORT_ACTIONS_SIZE(x)  \
    __offsetof(struct _posix_spawn_port_actions, pspa_actions[(x)])

if port_actions_size passes this then we reach the following code:

  MALLOC(px_spap, _posix_spawn_port_actions_t,
        px_args.port_actions_size, M_TEMP, M_WAITOK);
  if (px_spap == NULL) {
    error = ENOMEM;
    goto bad;

  imgp->ip_px_spa = px_spap;

  if ((error = copyin(px_args.port_actions, px_spap,
                      px_args.port_actions_size)) != 0)
    goto bad;

This allocates a kernel heap buffer to hold the port_actions buffer and copies from userspace into it.

The code then attempts to check whether the pspa_count valid is correct:

  /* Verify that the action count matches the struct size */
  if (PS_PORT_ACTIONS_SIZE(px_spap->pspa_count) != px_args.port_actions_size) {
    error = EINVAL;
    goto bad;

There is an integer overflow here because offsetof is just simple arithmetic. With a carefully chosen
value for pspa_count we can make it very large but when it's passed to the PS_PORT_ACTIONS_SIZE macro
the result is equal to port_actions_size. Nothing bad has happened yet but we can now get pspa_count
to be much larger than it should be.

Later on we reach the following code:

  if (px_spap->pspa_count != 0 && is_adaptive) {
    portwatch_count = px_spap->pspa_count;
    MALLOC(portwatch_ports, ipc_port_t *, (sizeof(ipc_port_t) * portwatch_count), M_TEMP, M_WAITOK | M_ZERO);
  } else {
    portwatch_ports = NULL;

  if ((error = exec_handle_port_actions(imgp, &portwatch_present, portwatch_ports)) != 0)

We can cause another integer overflow here, sizeof(ipc_port_t) is 4 (on 32-bit) so with a carefully chosen value of pspa_count
we can cause the integer overflow here and earlier too whilst still passing the checks.

exec_handle_port_actions then uses portwatch ports like this:

  for (i = 0; i < pacts->pspa_count; i++) {
    act = &pacts->pspa_actions[i];

    if (MACH_PORT_VALID(act->new_port)) {
      kr = ipc_object_copyin(get_task_ipcspace(current_task()),
                             act->new_port, MACH_MSG_TYPE_COPY_SEND,
                             (ipc_object_t *) &port);
      switch (act->port_type) {
        if (portwatch_ports != NULL && IPC_PORT_VALID(port)) {
         *portwatch_present = TRUE;
        /* hold on to this till end of spawn */
        portwatch_ports[i] = port;

note that pspa_actions was allocated earlier also based on the result of an integer overflow.
This means we can cause an OOB write to portwatch_ports only if we can successfully read suitable valid
values OOB of pspa_actions. That's why this PoC first fills a kalloc.1024 buffer with suitable values before
freeing it and then hoping that it will get reallocated as pspa_actions (but less thatn 1024 bytes will be written)
such that we control what's read OOB and the ipc_object_copyin will succeed.

This seems to be pretty reliable. You can use this to build a nice primitive of a heap overflow with pointers
to ipc_port structures.

I don't believe there are any iOS 11 32-bit iPod/iPhone/iPad/AppleTV devices but the new Apple Watch Series 3
is running essentially the same kernel but has a 32-bit CPU. This PoC is provided as an Apple watch app
and has been tested on Apple Watch Series 3 (Watch3,2) running WatchOS 4.0.1. I also tested on an older 32-bit iOS 9 device.

Apple Watch Series 3 now has its own LTE modem and can be used without an iPhone making it a suitably interesting target for exploitation
by itself.

Note that all the uses of offsetof in those posix_spawn macros are quite wrong, I think you might be able to get
a kernel memory disclosure with one of them also on 64-bit platforms. The fix is to add correct bounds checking.

Please also note that this really shouldn't be attack surface reachable from an app sandbox. The MAC hook in posix_spawn
is very late and there's a *lot* of code which you can hit before it.

Proof of Concept: