Microsoft Windows 8.1 Kernel Patch Protection Analysis & Attack Vectors

[Nytro]

Microsoft Windows 8.1 Kernel Patch Protection Analysis & Attack Vectors

Kernel Patch Protection (also known as "patchguard") is a Windows mechanism designed to

control the integrity of vital code and data structures used by the operating system. It was

introduced in Windows 2003 x64 and has been constantly improved in further Windows

versions. In this article we present a descriptive analysis of the patchguard for the latest

Windows 8.1 x64 OS, and primarily focus on patchguard initialization and attack vectors related

to it.

It is natural that kernel patch protection is being developed incrementally, so the initialization

process is common for all versions of Windows that have patchguard. There are a lot of papers

published about kernel patch protection on Windows, which describe the process of its

initialization, so you may use references at the end of this article to obtain details.

Initialization sources

As widely known, the main component of patchguard is initialized in a misleadingly named

function "KiFilterFiberContext". It will be the starting point of our investigation. Looking for

cross-references doesn't help us much for pointing out its call site, but several articles help us

by stating that patchguard initialization is called indirectly in a function

"KeInitAmd64SpecificState". By indirectly we mean here not just an indirect call, but the usage

of exception handlers. It is a very common trick often found in patchguard-related functions, as

we'll see further. So, we have an initialization function call stack:

[FONT=Courier New] (call) (call) (exception)[/FONT]

[FONT=Courier New]... -->  Phase1InitializationDiscard --> ; KeInitAmd64SpecificState ->  KiFilterFiberContext[/FONT]

This type of initialization is described in more detail in [1]. By the way, this one is always called

on the last CPU core, if it matters.

However, it is not the only way that kernel uses to initialize patchguard. With a 4% probability

patchguard context can also be initialized from a function also misleadingly called

"ExpLicenseWatchInitWorker":

[FONT=Courier New]...  -->  Phase1InitializationDiscard -->  sub_14071815C (obviously  with a stripped symbol because this one processes Windows license type  for a current PC) -->  ExpLicenseWatchInitWorker[/FONT]

The pseudocode of this function looks like this:

VOID ExpLicenseWatchInitWorker()

{

PVOID KiFilterParam;

NTSTATUS (*KiFilterFiberContext)(PVOID pFilterparam);

BOOLEAN ForgetAboutPG;

// KiServiceTablesLocked == KiFilterParam

KiFilterParam = KiInitialPcr.Prcb.HalReserved[1];

KiInitialPcr.Prcb.HalReserved[1] = NULL;

KiFilterFiberContext = KiInitialPcr.Prcb.HalReserved[0];

KiInitialPcr.Prcb.HalReserved[0] = NULL;

ForgetAboutPG = (InitSafeBootMode != 0) | (KUSER_SHARED_DATA.KdDebuggerEnabled

-> -> 1);

// 96% of cases will fail

if ( __rdtsc() % 100 -> 3 )

ForgetAboutPG |= 1;

if ( !ForgetAboutPG && KiFilterFiberContext(KiFilterParam) != 1 )

KeBugCheckEx(SYSTEM_LICENSE_VIOLATION, 0x42424242, 0xC000026A, 0, 0);

}

As you may notice, there is a small "present" in the “HalReserved” processor control block field

left for this initialization case. Tracing down the guy who left it leads us to the very beginning of

system startup:

[FONT=Courier New]... -->  KiSystemStartup -->  KiInitializeKernel -->  KeCompactServiceTable -->  KiLockServiceTable -v ??????[/FONT]

We have to pause here, because there is no code that puts data into HalReserved fields

directly. As instead, it is done using the exception handler. And it is done in a different way

from "KeInitAmd64SpecificState", because it doesn't trigger any exceptions. What it does

instead is – it directly looks up the current instruction pointer, finds the corresponding function

and it's exception handler manually, and then calls it. The exception handler of

"KiLockServiceTable" function is an unnamed stub to the "KiFatalExceptionFilter".

[FONT=Courier New]?????? --->  KiFatalExceptionFilter[/FONT]

“KiFatalExceptionFilter” in turn looks up an exception handler for "KiServiceTablesLocked"

function. And surprisingly it is the "KiFilterFiberContext"! Also, a parameter that is passed to

"KiFilterFiberContext" is located right after the "KiServiceTablesLocked" function. It is a small

structure:

typedef struct _KI_FILTER_FIBER_PARAM

{

NTSTATUS (*PsCreateSystemThread)(); // a pointer to PsCreateSystemThread function

KSTART_ROUTINE sub_140235C44; // unnamed checker subroutine

KDPC KiBalanceSetManagerPeriodicDpc; // global DPC struct

} KI_FILTER_FIBER_PARAM, *PKI_FILTER_FIBER_PARAM;

"KiFatalExceptionFilter" stores these pointers to “HalReserved” fields.

Creating patchguard context

Let's get back to the "KiFilterFiberContext" function. It's pseudocode is given below:

BOOLEAN KiFilterFiberContext(PVOID pKiFilterParam)

{

BOOLEAN Result = TRUE;

DWORD64 dwDpcIdx1 = __rdtsc() % 13;

DWORD64 dwRand2 = __rdtsc() % 10;

DWORD64 dwMethod1 = __rdtsc() % 6;

AntiDebug();

// Let's call sub_1406D6F78 KiInitializePatchGuardContext since it does

initialize patchguard context

Result = KiInitializePatchGuardContext(dwDpcIdx, dwMethod1, (dwRand2 < 6) + 1, pKiFilterParam, TRUE);

// A 50% chance to create two patchguard contexts

if (dwRand2 < 6)

{

DWORD64 dwDpcIdx2 = __rdtsc() % 13;

DWORD64 dwMethod2 = __rdtsc() % 6;

do

{

dwMethod2 = __rdtsc() % 6;

}

while ((dwMethod1 != 0) && (dwMethod1 == dwMethod2));

Result = KiInitializePatchGuardContext(dwDpcIdx2, dwMethod2, 2,

pKiFilterParam, FALSE);

}

AntiDebug();

return Result;

}

It is rather clear, and with provided code we can assume that up to 4 patchguard contexts can

be active on a running system simultaneously. Remember this one because wherever it is

called, we can be 100% sure that a new patchguard context is being initialized.

The function that creates and initializes patchguard context is so-called

"KiInitializePatchGuardContext". It is a huge obfuscated function. I guess it is suitable to

reference Alex's Ionescu tweet about it:

"I love the new #Windows 8 Patch Guard. Fixes so many of the obvious holes in downlevel, and the new hyper-inlined obfuscation makes me cry."

You bet it! IDA Pro's decompiler works on it ~20 min on 3770 Core i7 CPU and spews out 26K

lines of code. It is not worth dealing with it as a single unit. Luckily, you can bite out small

pieces of information that give you a clue about methods that the new patchguard uses. That's

why we did not reverse engineer it entirely, as instead we took and analyzed several parts in it.

Feel free to explore this function yourself, and you may discover new wonderful things!

It takes 5 parameters on Windows 8.1:

1. Index of DPC routine to be called from a created patchguard DPC for checking the

patchguard context. It may be one of these:

// These ones don't use exception handlers to fire checks

KiTimerDispatch (copied to random pool allocation)

KiDpcDispatch (copied into patchguard context)

// These use exception handlers to fire patchguard checks

ExpTimerDpcRoutine

IopTimerDispatch

IopIrpStackProfilerTimer

PopThermalZoneDpc

CmpEnableLazyFlushDpcRoutine

CmpLazyFlushDpcRoutine

KiBalanceSetManagerDeferredRoutine

ExpTimeRefreshDpcRoutine

ExpTimeZoneDpcRoutine

ExpCenturyDpcRoutine

Also those 10 DPCs are regular system DPCs with useful payload, but when they encounter a

DeferredContext which has non-canonical address, they fire a corresponding

KiCustomAccessRoutine function.

These functions are only called when an appropriate scheduling method is used (0, 1, 2, 5)

2. Scheduling method:

These are the methods that are used to fire a patchguard DPC object that is created inside

"KiInitializePatchGuardContext" function.

• KeSetCoalescableTimer (0). A timer object is created with a random fire period between 2 minutes and 2 minutes and 10 seconds.
• Prcb.AcpiReserved (1). In this case a patchguard DPC is fired when a certain ACPI event occurs, f.e. transitioning to idle state. In this case "HalpTimerDPCRoutine" checks if 2 minutes have passed since last queued by itself DPC, and queues another one, taken from Prcb.AcpiReserved field.
• Prcb.HalReserved (2). Here a patchguard DPC is queued when HAL timer clock interrupt occurs, in the "HalpMcaQueueDpc". It is also done with 2 minutes period at least. Queued patchguard DPC is taken from Prcb.HalReserved field.
• PsCreateSystemThread (3). In this case, patchguard DPC routine is not used, as instead a system thread is created. The thread procedure is taken from KI_FILTER_FIBER_PARAM structure. Patchguard DPC in turn is used just as a container of the address of a newly created patchguard context.
• KeInsertQueueApc (4). This time a regular kernel APC is queued to the one of the system threads with "KiDispatchCallout" APC procedure. No patchguard DPC is fired also. System thread is chosen based on its start address, i.e. it must be equal to either PopIrpWorkerControl or CcQueueLazyWriteScanThread.
• KiBalanceSetManagerPeriodicDpc (5). Patchguard DPC is stored in a global variable named "KiBalanceSetManagerPeriodicDpc". It is queued in "KiUpdateTimeAssist" function and "KeClockInterruptNotify" function within every "KiBalanceSetManagerPeriod" ticks.

3. This parameter can be either 1 or 2. We are not sure about how it affects "KiInitializePatchGuardContext" function, but it is somehow connected to the quantity of checks

being done during patchguard context verification routine execution.

4. A pointer to KI_FILTER_FIBER_PARAM structure. It is noticeable that a method chosen inside

"KiInitializePatchGuardContext" is selected based on the presence of this parameter. If it is

present, a method bit mask is tested with 0x29 (101001b) which allows methods 0, 3 and 5.

Otherwise methods 0, 1, 2 and 4 are available. That makes sense, because methods 3 and 5

require a valid KI_FILTER_FIBER_PARAM structure.

5. Boolean parameter which tells if NT kernel functions checksums have to be recalculated.

As you might guess, the only scheduling method that can be initialized twice is 0, so

"KiFilterFiberContext" takes this fact into account when chooses a method for a second call of

"KiInitializePatchGuardContext".

Firing a patchguard check

Methods that fire patchguard DPC

The main principle of patchguard check routine is to launch a patchguard context verification

routine on a DPC level, and then queue a work item that will check vital system structures on a

passive level with a proceeding context recreation and rescheduling. The verification work item

uses a copy of "FsRtlUninitializeSmallMcb" function. You can check this one out, if you want to

figure out how the check works.

For the methods which use DPC activation there is a common code inside 10 listed DPC

routines, which checks "DeferredContext" for being a non-canonical address. If it is OK, DPC

just executes its payload. Otherwise one of 10 "KiCustomAccessRoutineX" functions is called.

When "KiCustomAccessRoutineX" is called, (last 2 bits + 1) of "DeferredContext" are taken and

used to roll along "KiCustomRecurseRoutineX". These recursive routines are cycled

incrementing X value. When the roll is over, "KiCustomRecurseRoutineX" tries to dereference a

DeferredContext value as a pointer, which inevitably generates #GP exception since this

address is non-canonical.

// Inside DPC routine

if ( (DeferredContext >> 47) < 0xFFFFFFFFFFFFFFFFui64 && DeferredContext >> 47 != 0 )

// Is DeferredContext a canonical address

{

...

KiCustomAccessRoutineX(DeferredContext);

...

}

void KiCustomAccessRoutine9(DWORD64 DeferredContext)

{

return KiCustomRecurseRoutine9((DeferredContext & 3) + 1, DeferredContext);

}

void KiCustomRecurseRoutine9(DWORD dwRoll, DWORD64 DeferredContext)

{

DWORD dwNextRoll;

DWORD64 go_go_GP;

dwNextRoll = dwRoll - 1;

if ( dwNextRoll )

KiCustomRecurseRoutine0(dwNextRoll, DeferredContext);

Microsoft Windows 8.1 Kernel Patch Protection Analysis Page 11 / 18

go_go_GP = *DeferredContext; // #GP

}

// DPC routine call sequence

ExpTimerDpcRoutine -> KiCustomAccessRoutine0 -> KiCustomRecurseRoutine0 ...

KiCustomRecurseRoutineN

IopTimerDispatch -> KiCustomAccessRoutine1 -> KiCustomRecurseRoutine1 ...

KiCustomRecurseRoutineN

IopIrpStackProfilerTimer -> ; KiCustomAccessRoutine2 -> KiCustomRecurseRoutine2 ...

KiCustomRecurseRoutineN

PopThermalZoneDpc -> KiCustomAccessRoutine3 -> KiCustomRecurseRoutine3 ...

KiCustomRecurseRoutineN

CmpEnableLazyFlushDpcRoutine -> KiCustomAccessRoutine4 -> KiCustomRecurseRoutine4 ...

KiCustomRecurseRoutineN

CmpLazyFlushDpcRoutine -> KiCustomAccessRoutine5 -> KiCustomRecurseRoutine5 ...

KiCustomRecurseRoutineN

KiBalanceSetManagerDeferredRoutine -> KiCustomAccessRoutine6 ->

KiCustomRecurseRoutine6 ... KiCustomRecurseRoutineN

ExpTimeRefreshDpcRoutine -> KiCustomAccessRoutine7 -> KiCustomRecurseRoutine7 ...

KiCustomRecurseRoutineN

ExpTimeZoneDpcRoutine -> KiCustomAccessRoutine8 -> KiCustomRecurseRoutine8 ...

KiCustomRecurseRoutineN

ExpCenturyDpcRoutine -> KiCustomAccessRoutine9 -> KiCustomRecurseRoutine9 ...

KiCustomRecurseRoutineN

Here comes vectored exception handling again. If you look up all the exception handlers for

these DPC routines, you'll discover that there are several nested __try\__except and

__try\__finally blocks. For example, "ExpTimerDpcRoutine" looks something like this:

...

__try

{

__try

{

__try

{

__try

{

KiCustomAccessRoutine0(DeferredContext);

}

__finally

{

FinalSub1();

}

}

__except (FilterSub1()) // patchguard context decryption occurs here

{

// Nothing

}

}

__finally

{

FinalSub2();

}

}

__except (FilterSub2())

{

// Nothing

}

...

ExpCenturyDpcRoutine, ExpTimeZoneDpcRoutine, ExpTimeRefreshDpcRoutine,

KiBalanceSetManagerDeferredRoutine, CmpLazyFlushDpcRoutine, CmpEnableLazyFlushDpcRoutine,

PopThermalZoneDpc, ExpTimerDpcRoutine … -> _C_specific_handler

IopIrpStackProfilerTimer , IopTimerDispatch … -> _GSHandlerCheck_SEH (GS check + _C_specific_handler)

Depending on the DPC routine, decryption routine (based on KiWaitAlways and KiWaitNever

variables) may reside in one of the exception filters, exception handlers or termination handlers.

Further patchguard context verification occurs also inside decryption routine, right after the

decryption.

As for "KiTimerDispatch" and "KiDpcDispatch" DPC routines - they call patchguard context

verification directly. Also, depending on the DPC routine a different type of patchguard context

encryption is used (or not used at all).

Other methods

Method 3 creates a system thread. System thread procedure sleeps between 2 minutes and 2

minutes and 10 seconds using "KeDelayExecutionThread" or "KeWaitForSingleObject" on a

kernel object, which is always not signaled. After the wait is timed out it decrypts patchguard

context and executes verification routine.

Method 4 inserts an APC with "KiDispatchCallout" function as a kernel routine and

"EmpCheckErrataList" as a normal routine. Patchguard context decryption and validation occurs

upon APC delivery to the target waiting thread, which happens almost immediately. A 2 minutes

wait is located inside the verifier work item routine in this method.

One more piece of a puzzle

That would be it about patchguard initialization, but looking for the cross-references to

KUSER_SHARED_DATA.KdDebuggerEnabled lead me to a suspicious function named

"CcInitializeBcbProfiler". It is full of bit rotations and magic numbers, which forced me to check

whether it is related to patchguard mechanism.

[FONT=Courier New]... ->  Phase1InitializationDiscard -->  CcInitializeCacheManager -->  CcInitializeBcbProfiler[/FONT]

It seems to have the same roots!

With 50% chance it queues DPC with "CcBcbProfiler" routine or a work item with an unnamed

work item routine (which is almost identical to the "CcBcbProfiler" routine). This mechanism

picks one random function from NT kernel module and checks its consistency every 2 minutes.

It is interesting that all of the patchguard-related functions are located nearby, one after

another starting from "FsRtlMdlReadCompleteDevEx". It tells us that they are likely to be

located in a single compilation unit. This fact gives us a hope that all of the patchguard

initialization paths have been covered in this article.

Attacks

Now, as we covered patchguard initialization, we know what wires of a patchguard bomb can

be cut to defuse it! However, there are several ways depending on a patchguard DPC

scheduling method. Since we cover a specific version of patchguard, i.e. Windows 8.1, we are

going to use precomputed offsets for accessing the private kernel structures' fields.

The common defusing principle is firstly to check if verification routine is in progress, and wait a

bit if it is true. Then do the following:

• KeSetCoalescableTimer (0). Scan through the Prcb timer table and disable the one with suitable DPC object.
  
• AcpiReserved field (1). Zero this field out, so the DPC won't be fired again.
  
• HalReserved field (2). Same here.
  
• PspCreateSystemThread (3). Enumerate all threads in a system and unwind their stacks. Then check if a start routine from “KiServiceTablesLocked” structure is present in a call stack. If it is there, it's a patchguard thread. Disable it while it is in a wait state setting the wait time to infinite.
  
• APC (4). Take the current Prcb NUMA Node and its worker thread pool. Scan through its sleeping worker threads unwinding the stacks until "ExpWorkerThread" function. If there are functions that are not to be found in NT image runtime function data, try to unwind them sequentially with runtime data for "FsRtlMdlReadCompleteDevEx" and "FsRtlUninitializeSmallMcb". If succeeded, than it is a patchguard worker. Disable it setting the wait time to infinity.
  
• KiBalanceSetManagerPeriodicDpc (5). Zero this struct out.
  

By disabling a timer we mean setting its due time to infinity, so it never fires. And by suitable

DPC object we mean a DPC object with a deferred context set to a non-canonical address.

Furthermore, you can additionally check this pointer to be valid after XORing its value with a

quad-word following right after KDPC struct and ANDing it with 0xFFFF800000000000.

As for the "CcBcbProfiler" piece, we consider it not to be relevant since there is a small chance

that it will check exactly the needed function.

Summary

A quality of Windows 8.1 kernel patch protection mechanism is extremely high. There are a lot

of interesting anti-debugging tricks used again dynamic analysis, f.e. resetting IDT before

accessing debug registers (which leads you to hanging if you set break on debug registers

access), overall obfuscation like using macroses for generating pseudo-random values, loop

unrolling etc. It is also extremely difficult to do a static analysis since a lot of indirect function

calls are used including the usage of exception handlers.

It is a really nice tool to keep the system safe. Therefore we hope that as a developer you

won't face situations when you need to disable this cool mechanism!

Authors: Mark Ermolov, Artem Shishkin, Positive Research

Sursa: Positive Research Center: Microsoft Windows 8.1 Kernel Patch Protection Analysis & Attack Vectors

Edited September 17, 2014 by Nytro

[S1mon]

interesant