hvpp

hvpp is a lightweight Intel x64/VT-x hypervisor written in C++ focused primarily on virtualization of already running operating system.

Motivation

Although several open-source research hypervisors aimed at simplicity already exist, in my opinion this field is still somewhat unexplored and needs more open-source projects. This can especially help those who have just started exploring virtualization technologies and are looking for small/reference projects. If you're one of them, my bets are that you're really disappointed right now, because all you've got are barely dozen of (great!) projects and huge pile of Intel Manual pages.

C++ has been chosen as a language for this project because of two reasons:

• The virtualization architecture can be better expressed in OOP concepts (with such objects as VCPU, EPT).
• I didn't find other projects which would use modern C++17 features, except for bareflank. Although bareflank is compilable under Windows, Linux and UEFI, on Windows, it uses cygwin to cross-compile the hypervisor. Since the hypervisor is a self contained ELF binary, the Windows kernel is missing the debug symbols for the hypervisor, which prevents easy debugging with WinDbg.

Even though this project is primarily developed for Windows, I've decided to not use traditional Windows Driver naming convention (aka DrvCamelCase). Instead, traditional C++ snake_case is used. The reason is that hypervisor is very "stand-alone" and doesn't depend on many OS functions. Therefore I decided to treat it as a regular C++ project.

If you want to challenge yourself in low-level programming, my advice would be to try and write a simple hypervisor. During the process you'll get invaluable knowledge and experience and you'll probably discover many new things. For instance here's a selection of some things I've learned thanks to writing this project:

• what are MTRR good for
• NT APIs like
  • MmGetPhysicalMemoryRanges() (and \REGISTRY\MACHINE\HARDWARE\RESOURCEMAP)
  • KeIpiGenericCall - which can be used instead of undocumented KeGenericCallDpc (but at the cost of higher IRQL)
• that TraceLogging API (built upon ETW) exist
• how Meltdown vulnerability has been mitigated in Windows
• that VMWare uses not only I/O port 0x5658 for backdoor, but also port 0x5659 and it uses almost all GP registers for both input and output parameters
• that bugs in other hypervisors exist (and found a way how to fix them)
• that VirtualBox uses ring 1 for software-based virtualization (i.e. when VT-x/AMD-V isn't available)
• how guest memory has been virtualized before EPT (by so called Shadow-paging, trapping page-faults and mov cr3, <source> instructions in the guest)

Also - as obvious as it might sound - I'd like to point out that if you decide to write your own VT-x hypervisor, you'll NEED Intel® 64 and IA-32 architectures software developer’s manual combined volumes: 1, 2A, 2B, 2C, 2D, 3A, 3B, 3C, 3D, and 4. So download the PDF - together with Adobe Acrobat Reader - because trust me, you don't want to read and navigate through 5000 pages with browser's built-in PDF reader.

Features

• EPT with identity mapping with usage of 2MB pages for the first 512GB of physical memory (see ept.cpp). This results in faster translations of the memory. It also means splitting particular 2MB pages into 4kb pages might be desired if EPT hooking is required. This process is actually not complicated at all and this repository includes example on how to achieve that.
• Simple pass-through VM-exit handler, which can handle:
  • exceptions or NMIs
  • CPUID, (WB)INVD, INVLPG, RDTSC(P), MOV CR, MOV DR, IN/OUT, RDMSR, WRMSR, SGDT, SIDT, LGDT, LIDT, SLDT, STR, LLDT, LTR, XSETBV and INVPCID instructions
  • VMCALL instruction (used for termination of hvpp)
  • VMCLEAR, VMLAUNCH, VMRESUME, VMPTRLD, VMPTRST, VMREAD, VMWRITE, VMFUNC, VMXOFF, VMXON, INVEPT and INVVPID instructions raise #UD (invalid opcode exception)
• Ability to run in VMWare (tested even with KVM - I advise to turn off Hyper-V Enlightenments, as it can cause conflicts). VMWare's nested virtualization makes development and debugging of hypervisors much easier.
• Simple custom memory manager (see mm.cpp). The need for custom memory manager emerges from the fact that you should think twice before calling any kernel function from VM-exit handler, because many of them can be called at IRQL <= DISPATCH_LEVEL (such as ExAllocatePoolWithTag). But in VM-exit handler, interrupts are disabled and your IRQL is effectively HIGH_LEVEL.
• Detailed code comments, which should explain what the code does and why - sometimes with direct references to Intel Manual for further reading.
• TraceLogging API (which builds on ETW) - the main benefit is it can be used for really high frequency logging (10'000+ per second) and it can be used from any IRQL - which makes it a perfect candidate even for logging in VM-exit handlers.
  • hvpp already includes tracing VM-exit handler (see vmexit_stats.cpp).
• Various reimplemented classes and functions - such as bitmaps and spinlocks - to avoid calling kernel functions.
• Included simple application (hvppctrl) which should demonstrate CPUID instruction interception, hiding hooks in user-mode applications via EPT and communication with hvpp via VMCALL

Code workflow

Note: hvpp is compiled as a static library, which is linked with the hvppdrv project.

• Bootstrap of the driver (hvpp, driver.cpp)
  • preallocate enough memory and initialize the hvpp memory manager
  • initialize the logger
• Bootstrap of the hypervisor (hvppdrv, main.cpp)
  • create vmexit_handler instance
• Start the hypervisor with provided VM-exit handler: hypervisor::start(vmexit_handler& handler)
  • initialize virtual cpu (VCPU) for each logical processor
  • assign provided vmexit_handler instance to each VCPU
  • launch all VCPUs via IPI (inter-processor interrupt): vcpu_t::start()
  • setup VMXON region and VMCS: vcpu_t::vmx_enter()
  • vmexit_handler::setup() is called, which allows anyone to initialize the VM-exit handler and/or modify the VMCS before the launch (see vmexit_custom_handler::setup() in hvppdrv, vmexit_custom.cpp)
• Hypervisor is now running and handling VM-exits via provided VM-exit handler
• Stop the hypervisor: hypervisor::stop()
  • destroy each VCPU via IPI: vcpu_t::stop()
  • vmexit_handler::teardown() is called and switches into VMX mode (vmexit_passthrough_handler::teardown() does it by VMCALL instruction)
  • in VMX mode, vcpu_t::vmx_leave() is called - it leaves VMX mode with VMXOFF instruction

Compilation

Compile hvpp using Visual Studio 2017. Solution file is included. The only required dependency is WDK.

Usage

You can run hvpp on Windows 7 or higher. Windows 10 is recommended though, because it supports TraceLogging.

Enable Test-Signing boot configuration option (note that you'll need administrative privileges to use bcdedit and sc commands):

bcdedit /set testsigning on

Register driver with Service Control Manager (yes, it's important to leave these spaces):

sc create hvpp type= kernel binPath= "C:\full\path\to\hvppdrv.sys"

Now you should restart your computer for testsigning to take effect, otherwise you'll be unable to start the driver. But before you do, you might want to prepare DebugView from SysInternals and traceview.exe tool from the WDK (note that traceview will work properly only on Windows 10).

After restart, launch DebugView and TraceView. In TraceView:

• go to File -> Create New Log Session, click on Add Provider
  • pick Manually Entered Control GUID or Hashed Name
  • paste 916fcd3e-673b-4610-aaba-0b71e28acd40 (arbitrarily chosen, see lib/win32/tracelog.cpp)
  • click OK
• in the next dialog, leave the Source Of WPP Format Information set to Auto
  • click OK
• after that, click Next, which will bring you to Log Session Options dialog
  • in Log Session Name editbox you can give this logging session any name you like, e.g. HvppSession or you can leave it as it is
  • if you desire to analyze this logging session when it's stopped (e.g. with Windows Performance Analyzer) you can optionally enable Log Trace Event Data To File, which saves whole logging session into an .ETL file
  • click Finish

TraceView is now set-up and ready to show tracelogs from hvpp. You can launch hvpp now:

sc start hvpp

hvpp now performs various checks and enters VMX mode if they pass. In case of success you should see message hvpp started in the DebugView.

Run hvppctrl:

hvppctrl.exe

• hvppctrl performs CPUID instruction with EAX = 0x70707668 ('hvpp') which hvpp should intercept and return string hello from hvpp in EAX, EBX, ECX and EDX registers (see vmexit_custom.cpp). hvppctrl should print this string.

• hvppctrl tries to "stealthily" hook ntdll!ZwClose function using EPT. The exact process is described further below.

• hvppctrl performs IOCTL, which should instruct hvpp to set one-time breakpoint when IN/OUT instruction manipulating with port 0x64 (keyboard) is executed.

Description of "stealth hooking" process

• locates ZwClose function in ntdll.dll
  • disassembles first 16 bytes of this function and prints them
  • printed instructions should indicate that this function is NOT hooked yet
  • calls this function (with NULL parameter, this function call will most likely fail with some NTSTATUS error code, which it ignores)
  • prints value of HookCallCount and it's expected value (explained below)
• hooks ntdll!ZwClose fuction using Detours
  • disassembles first 16 bytes of this function and prints them
  • printed instructions should now indicate that the function IS hooked (by jmp being first instruction)
  • calls this function (with NULL parameter)
  • instead of original function, the hook function will be called - on each call, it increments variable HookCallCount
  • prints value of HookCallCount and it's expected value - it should be 1 now, as the hooked function has been called for the first time now
• calls hvpp by VMCALL instruction and RCX = 0xc1 (arbitrarily chosen), RDX = AddressOfReadPage and R8 = AddressOfExecutePage - this instructs hvpp to hide the hook
  • disassembles first 16 bytes of this function and prints them
  • printed instructions should now indicate that the function hook is HIDDEN (by showing original content of memory - no jmp)
  • calls this function (with NULL parameter)
  • despite what the memory returned when we read it, the hook function will be called again and the HookCallCount will be incremented again
  • prints value of HookCallCount and it's expected value - it should be 2
• calls hvpp by VMCALL instruction and RCX = 0xc2 (arbitrarily chosen) - this instructs hvpp to unhide the hook
  • disassembles first 16 bytes of this function and prints them
  • printed instructions should now indicate that the function hook is NOT HIDDEN and it should show jmp as a first instruction again
  • calls this function (with NULL parameter)
  • because the function is still hooked, the hook function will be called and HookCallCount will be incremented again
  • prints value of HookCallCount and it's expected value - it should be 3
• unhooks ntdll!ZwClose function
  • disassembles first 16 bytes of this function and prints them
  • printed instructions should now indicate that the function is NOT hooked - it should show the same content as when this function wasn't hooked
  • calls this function (with NULL parameter)
  • original function will be called, therefore the HookCallCount should not be incremented now
  • prints value of HookCallCount and it's expected value - it should be still 3

At the same time you should see tracelog messages in the TraceView - they are generated on each VMCALL and on each EPT Violation.

When you decide you want to turn off the hvpp, just execute:

sc stop hvpp

Remarks

• hvpp is designed to virtualize already running OS - i.e. it's not capable of running multiple guests like VMWare or VirtualBox. It also lacks support for any nested VMX operations.
• hvpp is designed to run only on 64bit Intel processors, which support VT-x and EPT. This makes the code more simple and less bloated.
• hvpp is designed to run only on Windows - future work might focus on Linux.
• hvpp currently doesn't exit VMX mode on sleep or hibernate (S3 and S4 power states) - Intel Manual says we should do so - this is known limitation.

License

This software is open-source under the MIT license. See the LICENSE.txt file in this repository.

Detours is licensed under MIT license (a copy of the license is included here).

udis86 is licensed under the terms of the 2-clause "Simplified BSD License" (a copy of the license is included here).

Similar work

SimpleVisor: https://github.com/ionescu007/SimpleVisor

HyperPlatform: https://github.com/tandasat/HyperPlatform

HyperBone: https://github.com/DarthTon/HyperBone

Bareflank: https://github.com/Bareflank/hypervisor

ksm: https://github.com/asamy/ksm

MoRE: https://github.com/ainfosec/MoRE

hyperdbg: https://github.com/rmusser01/hyperdbg

virtdbg: https://github.com/upring/virtdbg

BluePill: https://invisiblethingslab.com/resources/bh07/nbp-0.32-public.zip

Phrack #69: http://www.phrack.org/issues/69/15.html

NOVA Microhypervisor: https://github.com/udosteinberg/NOVA

Finally, I'd especially like to suggest reading interesting writings from Satoshi Tanda (github, twitter):

• http://tandasat.github.io/HyperPlatform/
• http://tandasat.github.io/HyperPlatform/userdocument/index.html
• http://standa-note.blogspot.com/2015/08/writing-hypervisor-for-kernel-mode-code.html

And notes from LordNoteworthy (github, twitter):

• https://github.com/LordNoteworthy/cpu-internals
• https://github.com/LordNoteworthy/windows-internals

If you find this project interesting, you can buy me a coffee

  BTC 3GwZMNGvLCZMi7mjL8K6iyj6qGbhkVMNMF
  LTC MQn5YC7bZd4KSsaj8snSg4TetmdKDkeCYk