DirectX/HyperV; An Offensive View

A Black Hat talk follow up

Blog DirectX/HyperV; An Offensive View

| 8 min read

Contact us

This year I attended Black Hat USA. The available talks were diverse, all of them inviting and some of them particularly attractive for my current field of work, which is currently mainly focused on advanced topics on Red Teaming and Exploit Development.

One of the talks I found most interesting was DirectX: The new Hyper-V Attack surface, presented by Zhenhao Hong (@rthhh17). In that talk, four vulnerabilities were presented (CVE-2021-43219, CVE-2022-21898, CVE-2022-21912 and CVE-2022-21918) regarding bugs like Null Pointer Dereference, Arbitrary Address Read and Arbitrary Address Write, which included a few lines of the PoC (Proof-Of-Concept) code to trigger each vulnerability.

Also, it was presented an overview of the architecture of Hyper-V DirectX components and a proposed fuzzing methodology to find new vulnerabilities.

In this post(s) I will try to follow up with that research and overcome expected shortcomings of the talk due to time restrictions:

  • There is no public access to the PoC codes.
  • There is no public access to the fuzzing artifacts.
  • The infrastructure to perform research on that specific environment was also not covered.
  • Hyper-V → DirectX integration is a work-in-progress for Microsoft, so many of the things mentioned in that talk are no longer working in the current version of Windows 11.

Setting up environment

We have already covered a post to set up a basic environment to perform remote kernel debugging. It involved creating a virtual machine, enabling debug mode using a network connection and plugging in the debugger. That could be done using a single computer.

This case is different. We need to debug a DirectX GPU adapter on a Windows machine acting as hypervisor using Hyper-V with a VM running Linux. Enabling virtualized extensions (VT-x) in a Windows VM can enable a nested Hyper-V, but the DirectX adapter will not be visible. After testing different scenarios, I ended needing to use two laptops.

The first laptop will be the debuggee (host1). In that laptop, the latest version of Windows 11 was installed:

dxgkrnl

In that machine, WSL was installed along with Kali as guest VM:

dxgkrnl

The latest stable kernel used on WSL for the guest machines is 5.10.102.1-microsoft-standard-WSL2. However, I wanted to use the latest version available of WSL, so I built it to use it. To the date of the exercise, the latest version was 5.15.57.1.

dxgkrnl

Make sure that you have partitionable GPUs on the host using Get-VMHostPartitionableGpu:

dxgkrnl

In a past article, we could be able to perform remote debugging using a network connection. I tried to do that, but failed because the physical network adapter didn't support debugging:

dxgkrnl

I had to use another approach. Luckily, Windows has several ways to be debugged. In this case, I chose to use USB3 debugging. To do that, I had to:

  • Find a USB3 port on my debuggee laptop with debugging support. That could be done using USBView from Windows SDK:

dxgkrnl

  • Enable debug options.

dxgkrnl

  • Plug the debugger and the debuggee using a quality USB3 cable.

dxgkrnl

In the end, the lab environment looked like this:

dxgkrnl

We are ready now!

An updated Hyper-V DirectX data flow

The following graph was presented by Zhenhao Hong (@rthhh17) which nicely describes the DirectX components and how are they accessed by a VM on Hyper-V:

dxgkrnl

The following is a detailed and updated flow of these interactions:

  1. There is a Linux driver called dxgkrnl.ko which exposes a set of IOCTL commands to interact with the host's DirectX adapters.
  2. When a IOCTL is called, there is another driver called hv_vmbus.ko which uses the VMBUS to create a packet and a bus channel between the VM, the hypervisor and the kernel of the host machine.
  3. The IOCTL payload is contained in a structure called DXGADAPTER_VMBUS_PACKET which contains the command (DXGK_VMBCOMMAND) and the command options to be sent.
  4. The host machine implements the receiving and processing counterpart in the dxgkrnl.sys driver.
  5. The procedure dxgkrnl!VmBusProcessPacket is the VMBUS receiving method that handles the DXGADAPTER_VMBUS_PACKET payload.
  6. If the DXGK_VMBCOMMAND is a global command (listed on enum dxgkvmb_commandtype_global), a function pointer (indirect call) is set to a method with the form dxgkrnl!DXG_HOST_GLOBAL_VMBUS::<command>, for example dxgkrnl!DXG_HOST_GLOBAL_VMBUS::VmBusDestroyProcess. Otherwise, the flow skips to point 7.
  7. If the DXGK_VMBCOMMAND is not a global command packet, it is processed by dxgkrnl!VmBusExecuteCommandInProcessContext which also uses indirect calls (function pointers) to compute the target handling method of that specific IOCTL request command. In this case, the handler has the form dxgkrnl!DXG_HOST_VIRTUALGPU_VMBUS::<command>, for example dxgkrnl!DXG_HOST_VIRTUALGPU_VMBUS::VmBusCreateDevice.
  8. The handling method casts the DXGADAPTER_VMBUS_PACKET packet using dxgkrnl!CastToVmBusCommand<DXGKVMB_COMMAND_<command>> (for example, dxgkrnl!CastToVmBusCommand<DXGKVMB_COMMAND_DESTROYPROCESS>) to filter the data as needed to this specific command handler.
  9. The handler performs boilerplate checks and perform the desired action. In some cases, it delivers the packet to a function with the pattern dxgkrnl!*Internal (for example, dxgkrnl!SignalSynchronizationObjectInternal) or dxgkrnl!Dxgk<command>Impl (for example, dxgkrnl!DxgkCreateDeviceImpl) which has the required interfaces to deliver the packet to the MMS (Microsoft Media System) components of DirectX that resides on the dxgmms1.sys and dxgmms2.sys drivers.
  10. The MMS system is finally in charge to talk with the corresponding GPU driver, which exposes the adapter that can either be virtual or physical.
  11. In the end, the response is sent back to the VM via dxgkrnl!VmBusCompletePacket.

It's a complex process if you read it, but let's look at it in action. Let's see an example performing only one command: Create Device. Here is the sample code.

Get started with Fluid Attacks' Red Teaming solution right now

/* Hyper-V -> DirectX Interaction Sample Code Compile as: cc -ggdb -Og -o sample1 sample1.c Author: Andres Roldan LinkedIn: https://www.linkedin.com/in/andres-roldan/ Twitter: @andresroldan */ #define _GNU_SOURCE 1 #include <stdio.h> #include <stdint.h> #include <stdbool.h> #include <stdlib.h> #include <string.h> #include <errno.h> #include <sys/stat.h> #include <fcntl.h> #include <unistd.h> #include <sys/ioctl.h>
#include "/home/aroldan/WSL2-Linux-Kernel-linux-msft-wsl-5.15.y/include/uapi/misc/d3dkmthk.h"

int open_device() {
    int fd;
    fd = open("/dev/dxg", O_RDWR);
    if (fd < 0) {
        printf("Cannot open device file...\n");
        exit(1);
    }
    printf("Opened /dev/dxg: 0x%x\n", fd);
    return fd;
}

void create_device(int fd) {
    int ret;
    struct d3dkmt_createdevice ddd = { 0 };
    struct d3dkmt_adapterinfo adapterinfo = { 0 };
    struct d3dkmt_enumadapters3 enumada = { 0 };

    enumada.adapter_count = 0xff;
    enumada.adapters = &adapterinfo;
    ret = ioctl(fd, LX_DXENUMADAPTERS3, &enumada);
    if (ret) {
        printf("Error calling LX_DXENUMADAPTERS3: %d: %s\n", ret, strerror(errno));
        exit(1);
    }
    printf("Adapters found: %d\n", enumada.adapter_count);

    ddd.adapter = adapterinfo.adapter_handle;
    printf("Adapter handle: 0x%x\n", ddd.adapter.v);
    printf("Creating device\n");
    ret = ioctl(fd, LX_DXCREATEDEVICE, &ddd);
    if (ret) {
        printf("Error calling LX_DXCREATEDEVICE: %d: %s\n", ret, strerror(errno));
        exit(1);
    }
    printf("Device created: 0x%x\n", ddd.device);
}

int main() {
    int fd;
    struct d3dkmthandle device;

    fd = open_device();
    create_device(fd);
    close(fd);
}

It's a straightforward code:

  1. Opens a handle to /dev/dxg.
  2. Uses that handle to enumerate the adapters available.
  3. Creates the device handle.

The output on the console should be something like:

aroldan@host1:~$ cc -ggdb -Og -o sample1 sample1.c
aroldan@host1:~$ ./sample1
Opened /dev/dxg: 0x3
Adapters found: 2
Adapter handle: 0x40000000
Creating device
Device created: 0x40000000

Let's first look at the Linux side of things. First, I'm going to pause the execution on gdb at *create_device+176 (sample1.c:43) which is when the IOCTL calling the command LX_DXCREATEDEVICE is performed:

dxgkrnl

If we see the value of the variable ddd.device before the call, you should see something like this:

dxgkrnl

After the IOCTL, we can see that the device handle is now populated:

dxgkrnl

dxgkrnl

Now, let's check at the host running Windows. We should be able to witness the creation of the device handler (0x40000000) on a dxgkrnl!VmBusCompletePacket response. We're going to need to set a few breakpoints to check the flow. First, let's put a breakpoint at dxgkrnl!VmBusProcessPacket

dxgkrnl

Inspecting dxgkrnl!VmBusProcessPacket we can see at dxgkrnl!VmBusProcessPacket+0x568 an indirect call being performed. This is where dxgkrnl!VmBusProcessPacket handles DXGK_VMBCOMMAND global commands. You can find indirect calls (function pointers) in kernel space because they are wrapped by calls to _guard_dispatch_icall_fptr, which is added when the kernel is compiled with CFG.

Let's put another breakpoint there:

dxgkrnl

Now, let's put a breakpoint at dxgkrnl!VmBusExecuteCommandInProcessContext:

dxgkrnl

In that function at dxgkrnl!VmBusExecuteCommandInProcessContext+0x1f0 we can also find an indirect call being performed. We can set a new breakpoint in that place:

dxgkrnl

Finally, a breakpoint at dxgkrnl!VmBusCompletePacket will be set:

dxgkrnl

We should now have five breakpoints as follows:

dxgkrnl

I'm going to reference the steps described above in the following execution flow.

When we run the sample code again, it hits our first breakpoint (step 5):

dxgkrnl

When we resume the execution, the next breakpoint is hit at the dxgkrnl!_guard_dispatch_icall_fptr call, which is an indirect call to the first command's handler. In this case, the handling function was resolved as dxgkrnl!DXG_HOST_GLOBAL_VMBUS::VmBusCreateProcess (step 6):

dxgkrnl

If we resume the execution, a call to dxgkrnl!VmBusCompletePacket is performed to send to the caller the result of the dxgkrnl!DXG_HOST_GLOBAL_VMBUS::VmBusCreateProcess command:

dxgkrnl

When we resume the execution twice, first the breakpoint at dxgkrnl!VmBusProcessPacket is hit (step 5) as expected, but the next breakpoint hit is at dxgkrnl!VmBusExecuteCommandInProcessContext, which means that the incoming command is not a global command (step 7):

dxgkrnl

Now, when we resume the execution, the next breakpoint is hit at dxgkrnl!VmBusExecuteCommandInProcessContext+0x1f0 which contains the indirect call resolved to a non-global command. In this case, we see the command we sent (LX_DXCREATEDEVICE) for creating a device:

dxgkrnl

In that method, at dxgkrnl!DXG_HOST_VIRTUALGPU_VMBUS::VmBusCreateDevice+0x8d we can see a call to dxgkrnl!CastToVmBusCommand<DXGKVMB_COMMAND_CREATEDEVICE> which will extract the needed parts of the DXGADAPTER_VMBUS_PACKET (step 8):

dxgkrnl

Here's the decompiled code:

dxgkrnl

Later on that function, at dxgkrnl!DXG_HOST_VIRTUALGPU_VMBUS::VmBusCreateDevice+0x3b0, we can see a call to dxgkrnl!DxgkCreateDeviceImpl which do the dirty job (step 9):

dxgkrnl

And finally, when we continue the execution, the breakpoint at dxgkrnl!VmBusCompletePacket is hit. According to this article, the second parameter of the function dxgkrnl!VmBusCompletePacket is the data to be sent back to the caller (step 11). It means that if we check the double word data pointed by the rdx register, we should see the device handler (0x40000000) returned as we saw before in the Linux VM output:

dxgkrnl

Great!

You can download the sample1.c file here.

Conclusion

The Hyper-V DirectX interaction is not officially documented. You can understand most of the internals by reading the WSL code, performing reverse engineering of the Windows drivers and doing kernel debugging. In the next article, we will see that most of the dxgkrnl commands are not stateless and some of them depends on creating certain kernel objects first. We will also see how to leverage this architecture using an offensive approach.

If you want to follow this series and receive our next blog posts, don't hesitate to subscribe to our weekly newsletter.

Subscribe to our blog

Sign up for Fluid Attacks' weekly newsletter.

Recommended blog posts

You might be interested in the following related posts.

Photo by James Lee on Unsplash

A lesson of this global IT crash is to shift left

Photo by CardMapr on Unsplash

Users put their trust in you; they must be protected

Photo by Towfiqu barbhuiya on Unsplash

Ensuring compliance and security in the banking sector

Photo by Andre Taissin on Unsplash

With great convenience comes increased risk

Photo by FlyD on Unsplash

Software supply chain management in financial services

Photo by Robs on Unsplash

Consequential data breaches in the financial sector

Photo by Towfiqu barbhuiya on Unsplash

Data protection in the financial sector, tips and more

Start your 21-day free trial

Discover the benefits of our Continuous Hacking solution, which hundreds of organizations are already enjoying.

Start your 21-day free trial
Fluid Logo Footer

Hacking software for over 20 years

Fluid Attacks tests applications and other systems, covering all software development stages. Our team assists clients in quickly identifying and managing vulnerabilities to reduce the risk of incidents and deploy secure technology.

Copyright © 0 Fluid Attacks. We hack your software. All rights reserved.