Microsoft Windows NT #GP Trap Handler Allows Users to Switch Kernel Stack ------------------------------------------------------------------------- In order to support BIOS service routines in legacy 16bit applications, the Windows NT Kernel supports the concept of BIOS calls in the Virtual-8086 mode monitor code. These are implemented in two stages, the kernel transitions to the second stage when the #GP trap handler (nt!KiTrap0D) detects that the faulting cs:eip matches specific magic values. Transitioning to the second stage involves restoring execution context and call stack (which had been previously saved) from the faulting trap frame once authenticity has been verified. This verification relies on the following incorrect assumptions: - Setting up a VDM context requires SeTcbPrivilege. - ring3 code cannot install arbitrary code segment selectors. - ring3 code cannot forge a trap frame. This is believed to affect every release of the Windows NT kernel, from Windows NT 3.1 (1993) up to and including Windows 7 (2009). Working out the details of the attack is left as an exercise for the reader. Just kidding, that was an homage to Derek Soeder :-) - Assumption 0: Setting up a VDM context requires SeTcbPrivilege. Creating a VDM context requires EPROCESS->Flags.VdmAllowed to be set in order to access the authenticated system service, NtVdmControl(). VdmAllowed can only be set using NtSetInformationProcess(), which verifies the caller has SeTcbPrivilege. If this is true, the caller is very privileged and can certainly be trusted. This restriction can be subverted by requesting the NTVDM subsystem, and then using CreateRemoteThread() to execute in the context of the subsystem process, which will already have this flag set. - Assumption 1: ring3 code cannot install arbitrary code segment selectors. Cpl is usually equal to the two least significant bits of cs and ss, and is a simple way to calculate the privilege of a task. However, there is an exception, Virtual-8086 mode. Real mode uses a segmented addressing scheme in order to allow 16-bit addresses to access the 20-bit address space. This is achieved by forming physical addresses from a calculation like (cs << 4) + (eip & 0xffff). The same calculation is used to map the segmented real address space onto the protected linear address space in Virtual-8086 mode. Therefore, I must be permitted to set cs to any value, and checks for disallowed or privileged selectors can be bypassed (PsSetLdtEnties will reject any selector where any of the three lower bits are unset, as is the case with the required cs pair). - Assumption 2: ring3 code cannot forge a trap frame. Returning to usermode with iret is a complicated operation, the pseudocode for the iret instruction alone spans several pages of Intel's Software Developers Manual. The operation occurs in two stages, a pre-commit stage and a post-commit stage. Using the VdmContext installed using NtVdmControl(), an invalid context can be created that causes iret to fail pre-commit, thus forging a trap frame. The final requirement involves predicting the address of the second-stage BIOS call handler. The address is static in Windows 2003, XP and earlier operating systems, however, Microsoft introduced kernel base randomisation in Windows Vista. Unfortunately, this potentially useful exploit mitigation is trivial to defeat locally as unprivileged users can simply query the loaded module list via NtQuerySystemInformation(). -------------------- Affected Software ------------------------ All 32bit x86 versions of Windows NT released since 27-Jul-1993 are believed to be affected, including but not limited to the following actively supported versions: - Windows 2000 - Windows XP - Windows Server 2003 - Windows Vista - Windows Server 2008 - Windows 7 -------------------- Consequences ----------------------- Upon successful exploitation, the kernel stack is switched to an attacker specified address. An attacker would trigger the vulnerability by setting up a specially formed VDM_TIB in their TEB, using a code sequence like this: /* ... */ // Magic CS required for exploitation Tib.VdmContext.SegCs = 0x0B; // Pointer to fake kernel stack Tib.VdmContext.Esi = &KernelStack; // Magic IP required for exploitation Tib.VdmContext.Eip = Ki386BiosCallReturnAddress; NtCurrentTeb()->Reserved4[0] = &Tib; /* ... */ Followed by /* ... */ NtVdmControl(VdmStartExecution, NULL); /* ... */ Which will reach the following code sequence via the #GP trap handler, nt!KiTrap0D. Please note how the stack pointer is restored from the saved (untrusted) trap frame at 43C3E6, undoubtedly resulting in the condition described above. /* ... */ .text:0043C3CE Ki386BiosCallReturnAddress proc near .text:0043C3CE mov eax, large fs:KPCR.SelfPcr .text:0043C3D4 mov edi, [ebp+KTRAP_FRAME.Esi] .text:0043C3D7 mov edi, [edi] .text:0043C3D9 mov esi, [eax+KPCR.NtTib.StackBase] .text:0043C3DC mov ecx, 84h .text:0043C3E1 mov [eax+KPCR.NtTib.StackBase], edi .text:0043C3E4 rep movsd .text:0043C3E6 mov esp, [ebp+KTRAP_FRAME.Esi] .text:0043C3E9 add esp, 4 .text:0043C3EC mov ecx, [eax+KPCR.PrcbData.CurrentThread] .text:0043C3F2 mov [ecx+KTHREAD.InitialStack], edi .text:0043C3F5 mov eax, [eax+KPCR.TSS] .text:0043C3F8 sub edi, 220h .text:0043C3FE mov [eax+KTSS.Esp0], edi .text:0043C401 pop edx .text:0043C402 mov [ecx+KTHREAD.Teb], edx .text:0043C405 pop edx .text:0043C406 mov large fs:KPCR.NtTib.Self, edx .text:0043C40D mov ebx, large fs:KPCR.GDT .text:0043C414 mov [ebx+3Ah], dx .text:0043C418 shr edx, 10h .text:0043C41B mov byte ptr [ebx+3Ch], dl .text:0043C41E mov [ebx+3Fh], dh .text:0043C421 sti .text:0043C422 pop edi .text:0043C423 pop esi .text:0043C424 pop ebx .text:0043C425 pop ebp .text:0043C426 retn 4 /* ... */ Possibly naive example code for triggering this condition is availble from the link below. http://lock.cmpxchg8b.com/c0af0967d904cef2ad4db766a00bc6af/KiTrap0D.zip The code has been tested on Windows XP, Windows Server 2003/2008, Windows Vista and Windows 7. Support for other affected operating systems is left as an exercise for the interested reader. ------------------- Mitigation ----------------------- If you believe you may be affected, you should consider applying the workaround described below. Temporarily disabling the MSDOS and WOWEXEC subsystems will prevent the attack from functioning, as without a process with VdmAllowed, it is not possible to access NtVdmControl() (without SeTcbPrivilege, of course). The policy template "Windows Components\Application Compatibility\Prevent access to 16-bit applications" may be used within the group policy editor to prevent unprivileged users from executing 16-bit applications. I'm informed this is an officially supported machine configuration. Administrators unfamiliar with group policy may find the videos below instructive. To watch a demonstration of this policy being applied to a Windows Server 2003 domain controller, see the link below. http://www.youtube.com/watch?v=XRVI4iQ2Nug To watch a demonstration of this policy being applied to a Windows Server 2008 domain controller, see the link below. http://www.youtube.com/watch?v=u8pfXW7crEQ To watch a demonstration of this policy being applied to an unjoined Windows XP Professional machine, see the link below. http://www.youtube.com/watch?v=u7Y6d-BVwxk On Windows NT4, the following knowledgebase article explains how to disable the NTVDM and WOWEXEC subsystems. http://support.microsoft.com/kb/220159 Applying these configuration changes will temporarily prevent users from accessing legacy 16-bit MS-DOS and Windows 3.1 applications, however, few users require this functionality. If you do not require this feature and depend on NT security, consider permanently disabling it in order to reduce kernel attack surface. ------------------- Solution ----------------------- Microsoft was informed about this vulnerability on 12-Jun-2009, and they confirmed receipt of my report on 22-Jun-2009. Regrettably, no official patch is currently available. As an effective and easy to deploy workaround is available, I have concluded that it is in the best interest of users to go ahead with the publication of this document without an official patch. It should be noted that very few users rely on NT security, the primary audience of this advisory is expected to be domain administrators and security professionals. Please note, Microsoft typically do not credit researchers who do not wait for permission to publish their research, and therefore I am unlikely to be referenced in their forthcoming update. ------------------- Credit ----------------------- This bug was discovered by Tavis Ormandy. ------------------- Greetz ----------------------- Greetz to Julien, Neel, Redpig, Lcamtuf, Spoonm, Skylined, asiraP, LiquidK, ScaryBeasts, spender and all my other elite colleagues. Check out some photography while at ring0 @ http://flickr.com/meder. ------------------- References ----------------------- Derek Soeder has previously reported some legendary NT bugs, including multiple vdm bugs that, while unrelated to this issue, make fascinating reading. - http://seclists.org/fulldisclosure/2004/Oct/404, Windows VDM #UD LocalPrivilege Escalation - http://seclists.org/fulldisclosure/2004/Apr/477, Windows VDM TIB Local Privilege Escalation - http://seclists.org/fulldisclosure/2007/Apr/357, Zero Page Race Condition Privilege Escalation