Apple macOS 10.13.2 - Double mach_port_deallocate in kextd due to Failure to Comply with MIG Ownership Rules







Here's a kextd method exposed via MIG (

  kern_return_t _kextmanager_unlock_kextload(
      mach_port_t server,
      mach_port_t client)
      kern_return_t mig_result = KERN_FAILURE;
      if (gClientUID != 0) {
          OSKextLog(/* kext */ NULL,
              kOSKextLogErrorLevel | kOSKextLogIPCFlag,
              "Non-root kextutil doesn't need to lock/unlock.");
          mig_result = KERN_SUCCESS;
          goto finish;
      if (client != (mach_port_t)dispatch_source_get_handle(_gKextutilLock)) {
          OSKextLog(/* kext */ NULL,
              kOSKextLogErrorLevel | kOSKextLogIPCFlag,
              "%d not used to lock for kextutil.", client);
          goto finish;

      mig_result = KERN_SUCCESS;
      // we don't need the extra send right added by MiG
      mach_port_deallocate(mach_task_self(), client);
      return mig_result;

If the client has UID 0 but passes an invalid client port this code will
drop a UREF on client port then return KERN_FAILURE.

Returning KERN_FAILURE in MIG means all resources will be released which will
cause client to be passed to mach_port_deallocate again, even though only 
one UREF was taken.

You'll have to use a debugger attached to kextd to see this behaviour.

This class of bug is exploitable; please see the writeup for mach_portal from 2016
where I exploited a similar issue []
The TL;DR is that an attacker can drop an extra UREF on any send rights in kextd for which the
attacker also has a send right; you could use this to cause a name for a privileged service
to be deallocated then cause the name to be reused to name a port you control.

Exploitation of this would be a privesc from unentitled root to root with and entitlements,
which at least last time I looked was equal to kernel code execution.

tested on MacOS 10.13.2

Proof of Concept: