Abusing Exceptions for Code Execution, Part 2
In this article, we'll explore how the concepts behind Exception Oriented Programming can be abused when exploiting stack overflow vulnerabilities on Windows.
Full disclosure- Microsoft hired me following part 1 of this series. This research was conducted independently, and a vast majority of it was completed before I joined. Obviously, no internal information was used, and everything was built on public resources.
In Abusing Exceptions for Code Execution, Part 1, I introduced the concept of Exception Oriented Programming (EOP), which was a method of executing arbitrary operations by chaining together code from legitimate modules. The primary benefit of this approach was that the attacker would never need their shellcode to be in an executable region of memory, as the technique relied on finding the instructions of their shellcode in existing code.
The last article primarily focused on abusing this technique when you already have some form of code execution. Although powerful for obfuscation and evasion, the use cases provided would only be relevant when an attacker had already compromised an environment. For example, how does EOP compare to existing exploitation techniques such as Return Oriented Programming (ROP)? In this article, we'll explore how the concepts behind Exception Oriented Programming can be abused when exploiting stack overflow vulnerabilities on Windows.
Background
Before we can get into how EOP can help exploit stack-based attacks, it's important to know the history of the mitigations we are up against. I assume you already have familiarity with the OS-agnostic basics, such as ASLR and DEP.
Security Cookies
Security cookies (aka "stack canaries") are a compiler mitigation introduced around two decades ago. Here is a helpful summary from Microsoft's documentation:
On functions that the compiler recognizes as subject to buffer overrun problems, the compiler allocates space on the stack before the return address. On function entry, the allocated space is loaded with a security cookie that is computed once at module load. On function exit, and during frame unwinding on 64-bit operating systems, a helper function is called to make sure that the value of the cookie is still the same. A different value indicates that an overwrite of the stack may have occurred. If a different value is detected, the process is terminated.
Security cookies are relatively straightforward. By placing a "random" cookie next to the return address on the stack, attackers exploiting stack overflow vulnerabilities face a significant problem- how do you modify the return address without failing the cookie check?
Over the years, there has been lots of work put into bypassing these security cookies. I found this helpful overview from the Corelan team written in 2009. Let's review some of the techniques they discuss that are still relevant to this day:
- This mitigation is irrelevant if you have an overflow vulnerability in a function that does not have a security cookie check (i.e. because there are no string buffers).
- If you have an information disclosure primitive, you could attempt to leak the security cookie for the current function from the stack or the security cookie in the
.data
section.- For example, if you had a string buffer and a method of getting the application to "print" that string, you could overflow the buffer up to the security cookie such that there is no NULL terminator. When the string is "printed", all the bytes of the cookie until a NULL terminator would be returned as a part of the string.
- If you already have an arbitrary "write-what-where" primitive and know the location of the security cookie, you can overwrite it with your own, allowing you to predict the "correct" value to place on the stack.
- You can still overwrite local variables on the stack to hijack control flow.
- For example, if a pointer was stored on the stack (before the overflow'd variable) used in a desirable operation like
memcpy
after the overflow occurs, you could overwrite this pointer without corrupting the security cookie. - Another example would be objects with "virtual tables" on the stack that we can overwrite. If an object's virtual table is used after the overflow occurs, an attacker could influence the target of those virtual calls. Of course, this would likely be subject to control-flow integrity mitigations like Control Flow Guard (or xFG) on Windows.
- For example, if a pointer was stored on the stack (before the overflow'd variable) used in a desirable operation like
Outside of these approaches, there has been extensive research into abusing exception handling. Before mitigations such as SafeSEH and SEHOP, which we will discuss soon, attackers in the context of 32-bit applications could modify "exception registration records" on the stack. The Corelan team covered this path of exploitation in a separate blog. More recently, however, @_ForrestOrr wrote in detail about SEH hijacking in his article about memory corruption bugs on Windows.
SEH Hijacking and the Mitigations Against It
In 32-bit applications, exception registration records contain a pointer to the "next" SEH record on the stack and a pointer to the exception handler itself. Back in the day, attackers could hijack control flow even with security cookies by:
- Replacing the exception handler on the stack with their own.
- Triggering an exception before the security cookie check.
This would allow the attacker to call an arbitrary handler with partial control over the passed arguments.
SafeSEH
To protect against this technique, Microsoft introduced a mitigation called SafeSEH. At a high level, "legitimate" exception handlers are built into the binary at compile-time. Although an attacker can still replace the exception handler on the stack, if it is not in the module's list of exception handlers, a STATUS_INVALID_EXCEPTION_HANDLER
exception is raised.
SEHOP
SEH Overwrite Protection (SEHOP) is another mitigation that would protect 32-bit applications from having their exception handlers overwritten- without requiring them to be recompiled. This approach works by adding an exception registration record at the bottom of the chain and making sure it is "reachable" when an exception occurs. Remember that besides the exception handler, the registration record contains a pointer to the "next" SEH record. If an attacker corrupts this "next" pointer, the chain is broken, and this final item is not reachable, preventing the attack. Of course, if an attacker can predict the "next" pointer successfully, this mitigation can be evaded.
64-bit Applications
64-bit applications are already protected against this attack by default, which we briefly mentioned in the last article of this series:
Nowadays SEH exception handling information is compiled into the binary, specifically the exception directory, detailing what regions of code are protected by an exception handler. When an exception occurs, this table is enumerated during an "unwinding process", which checks if the code that caused the exception or any of the callers on the stack have an SEH exception handler.
Since the exception handlers are built into the binary itself, there is no exception registration record on the stack that an attacker can corrupt. This prevents the existing approaches to SEH hijacking entirely.
The Exception Directory
Let's talk more about how the exception directory works in 64-bit applications.
The location of the exception directory can be retrieved by parsing the optional header of the binary, specifically the IMAGE_DIRECTORY_ENTRY_EXCEPTION
data directory. This directory is an array of IMAGE_RUNTIME_FUNCTION_ENTRY
structures. You can calculate the number of entries by dividing the directory size by the size of the IMAGE_RUNTIME_FUNCTION_ENTRY
structure.
Each entry contains a begin address, end address, and the offset of an UNWIND_INFO
structure. The begin/end addresses specify the region of code that the given entry provides information for. The UNWIND_INFO
structure is represented by the following:
typedef struct _UNWIND_INFO {
unsigned char Version : 3;
unsigned char Flags : 5;
unsigned char SizeOfProlog;
unsigned char CountOfCodes;
unsigned char FrameRegister : 4;
unsigned char FrameOffset : 4;
UNWIND_CODE UnwindCode[1];
/* UNWIND_CODE MoreUnwindCode[((CountOfCodes+1)&~1)-1];
* union {
* OPTIONAL unsigned long ExceptionHandler;
* OPTIONAL unsigned long FunctionEntry;
* };
* OPTIONAL unsigned long ExceptionData[];
*/
} UNWIND_INFO, * PUNWIND_INFO;
The commented region is still present in the structure, but its location depends on the size of the dynamic UnwindCode
array. Note that most functions in an application will have a dedicated entry. This is because even if the function does not need to handle exceptions, each entry contains essential information for how the function should be unwound. For example, if function A contains an exception handler, calls function B, which does not, and an exception occurs, we still need to be able to unwind the stack to get to function A's handler.
Of note, the Flags
field of the structure can contain the following:
UNW_FLAG_NHANDLER
- The function has no handler.UNW_FLAG_EHANDLER
- The function has an exception handler that should be called.UNW_FLAG_UHANDLER
- The function has a termination handler that should be called when unwinding an exception.UNW_FLAG_CHAININFO
- The FunctionEntry member is the contents of a previous function table entry.
We can tell if a given function contains an exception handler by checking if the Flags
field specifies UNW_FLAG_EHANDLER
.
The structure's dynamic UNWIND_CODE
array represents the operations needed to "unwind"/undo the changes a given function's prolog has made to the stack. We will talk about these operations in a later section when they become relevant.
typedef struct _UNWIND_INFO {
...
/* UNWIND_CODE MoreUnwindCode[((CountOfCodes+1)&~1)-1];
* union {
* OPTIONAL unsigned long ExceptionHandler;
* OPTIONAL unsigned long FunctionEntry;
* };
* OPTIONAL unsigned long ExceptionData[];
*/
} UNWIND_INFO, * PUNWIND_INFO;
Going back to the definition of the UNWIND_INFO
structure, note that there is a single field dedicated to the offset of the "exception handler". When you create some code with an SEH try/except block, the address of your exception handler is not what goes into this field. Instead, every language (including C/C++) is responsible for defining a "language-specific" handler. In our case, ExceptionHandler
points to the __C_specific_handler
. These handlers have the following type definition:
typedef EXCEPTION_DISPOSITION (*PEXCEPTION_ROUTINE) (
IN PEXCEPTION_RECORD ExceptionRecord,
IN ULONG64 EstablisherFrame,
IN OUT PCONTEXT ContextRecord,
IN OUT PDISPATCHER_CONTEXT DispatcherContext
);
typedef struct _SCOPE_TABLE_AMD64 {
DWORD Count;
struct {
DWORD BeginAddress;
DWORD EndAddress;
DWORD HandlerAddress;
DWORD JumpTarget;
} ScopeRecord[1];
} SCOPE_TABLE_AMD64, *PSCOPE_TABLE_AMD64;
The __C_specific_handler
handler for C/C++ leverages the ExceptionData
field to store a SCOPE_TABLE
structure. Like the IMAGE_RUNTIME_FUNCTION_ENTRY
parent structure, each ScopeRecord
has a begin/end address, but this time we have a handler and jump target as well. The begin/end address specifies the scope or "region of code" to "protect". The last two fields store offsets to your SEH exception filter and exception handler.
int main() {
__try {
*(int*)0 = 0xDEADBEEF;
} __except(MyExceptionFilter()) {
printf("My exception handler!\n");
}
return 0;
}
Exception filters go inside the parenthesis for your __except
block. In this example, MyExceptionFilter
is responsible for determining whether the __except
handler block should be called for a given exception. Exception filters often perform conditional checks, such as whether the exception code matches something specific. Filters can return the following results:
EXCEPTION_CONTINUE_EXECUTION
- Indicates that execution should continue where the exception occurred.EXCEPTION_CONTINUE_SEARCH
- Continue the search for an exception handler.EXCEPTION_EXECUTE_HANDLER
- Execute the handler block. In the example code,My exception handler!
would only be printed ifMyExceptionFilter
returned this value.
Exception filters and handlers are defined in the ScopeRecord
structure as the HandlerAddress
and JumpTarget
offsets. If the HandlerAddress
is 1, then this means execute the exception handler (JumpTarget
) for all exceptions.
You can find the source code for __C_specific_handler
included with the MSVCRT since it needs to support static compilation into binaries. On my installation of Visual Studio, the relevant source file is located at C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\crt\src\amd64\chandler.c
.
The Exception Dispatching Process
Before we continue, I want to clarify a fundamental concept we need to understand about the UNW_FLAG_EHANDLER
vs UNW_FLAG_UHANDLER
flags in the UNWIND_INFO
structure.
When an exception occurs, RtlDispatchException
is the first function to be called. RtlDispatchException
will create a temporary copy of the CONTEXT
record (containing the state of registers, etc.) and "virtually unwind" the stack searching for exception handlers to call. Unwinding means undoing the modifications done to the stack (and registers) by the prolog/epilog of functions in the call stack. If a function has a corresponding UNWIND_INFO
structure with the UNW_FLAG_EHANDLER
flag, its exception handler is called.
If the handler returns EXCEPTION_CONTINUE_EXECUTION
, execution continues right where the exception occurred (which means the exception was "handled"). Note that the changes made to the temporary CONTEXT
copy will not be reflected if execution were to continue (i.e. if virtual unwind modified Rcx
register, that doesn't change Rcx
when execution continues).
If the handler returns EXCEPTION_CONTINUE_SEARCH
, the virtual unwinding process continues, looking for the next function with the UNW_FLAG_EHANDLER
flag.
In the context of the C-specific language handler, the exception filter specified by the "handler address" can return the two results above and EXCEPTION_EXECUTE_HANDLER
. In this case, even though we are in the context of RtlDispatchException
, __C_specific_handler
will call RtlUnwind
to unwind execution to the handler specified by the "jump target".
RtlUnwind
is incredibly similar to RtlDispatchException
, but it has a few notable differences. First, the context record modified by the unwinding process in this function will be reflected when execution continues. This is because RtlUnwind
is intended to get both the stack and registers into the state corresponding to the target exception handler's function. So, for example, if you had an exception handler in a parent function of where the exception occurred, RtlUnwind
is responsible for making sure that Rsp
is corrected such that you can access any local variables from the context of your parent function as well as the values in its nonvolatile registers.
The second significant difference is that only "termination" or unwind handlers are called, aka functions with an UNWIND_INFO
structure specifying the UNW_FLAG_UHANDLER
flag. A good example of an unwind handler would be a __try
/__finally
block intended to free resources rather than catch an exception.
RtlUnwind
will unwind the stack, calling relevant unwind handlers until the "target frame" is reached. The target frame is generally the stack frame for the exception handler it is trying to unwind to. When reached, RtlUnwind
passes the modified context record to RtlRestoreContext
, which is responsible for continuing execution at the target handler.
We're going to cover RtlDispatchException
and RtlUnwind
further in later sections, but if you'd like to learn more outside of this article, check out the publicly leaked Windows Research Kernel (WRK), which contains the source code for RtlDispatchException and RtlUnwind.
Although we've gone over how Structured Exception Handling works "under the hood" to an extent, much was simplified for the purposes of this article. If you're interested in learning more, I would encourage you to check out this article by Ry Auscitte, who goes into even more detail.
We've explored the existing mitigations against stack-based attacks and how Structured Exception Handling works. In the following sections, we'll discuss the practical approaches I propose to simplify the exploitation of stack overflow vulnerabilities.
Bypassing Security Cookies
Example
void GetString() {
char tempBuffer[16];
scanf("%s", tempBuffer);
printf(tempBuffer);
}
int main() {
__try {
GetString();
}
__except (EXCEPTION_EXECUTE_HANDLER) {
printf("Something bad happened!\n");
}
return 0;
}
Let's start with an example of a simple vulnerable application I've compiled using Clang/LLVM in Visual Studio. The buffer overflow is simple, scanf
reads a string into the tempBuffer
stack variable without any bound checks.
If we take a look at the function in IDA Pro, there is a significant challenge preventing us from exploiting this overflow primitive- the security cookie check at the epilogue of the function. Since the return address is right "after" the security cookie on the stack, we couldn't modify it without also corrupting the cookie. This would prevent us from gaining code execution since the program would crash before the ret
instruction.
What can we do? Existing approaches to these scenarios include all of the methods we discussed to bypass security cookies, such as trying to leak it. For example, if the attacker had access to the stdout
of this program, they could use printf
to leak the security cookie on the stack. However, the issue they'd run into is that the program would exit soon after. Even if they could trigger another execution, a new random cookie would be generated.
This is where our new methodology can start to shine. Our exploit fails if we corrupt the return address and hit the __security_cookie_check
. What if we... corrupted the stack and triggered an exception (i.e with a bad format string)?
int main() {
__try {
GetString();
}
__except (EXCEPTION_EXECUTE_HANDLER) {
printf("Something bad happened!\n");
}
return 0;
}
Since main
has a "catch-all" exception handler, the program would print Something bad happened!
and return. The security cookie check would never be reached because the exception redirected execution to the handler! Also, because main
does not have any stack variables itself, it has no security cookie check. If we used our overflow to corrupt the return address of main
rather than GetString
, once main
returns after the exception is handled, we'd gain control over what code is executed!
This example relies on an overflow vulnerability, a method of triggering an exception, and a parent function having a "catch-all" exception handler. These first two requirements are relatively straightforward as 1) security cookies are only relevant for overflow scenarios, and 2) causing an exception is relatively easy if you can corrupt local variables.
What about the last requirement that the context in which you have an overflow vulnerability also contains a parent function with a catch-all handler? That is a much higher bar. Fortunately, this is where we can abuse how unwinding works.
How did the unwinding process determine that it should call main
's exception handler? When main
called GetString
, the Rip
register was pushed on the stack as the return address GetString
should return to. The unwinding process uses this return address on the stack while searching for a parent function that contains an exception handler.
If we have a leak of the location for any module in the process that contains a C-specific exception handler where 1) the exception filter returns EXCEPTION_EXECUTE_HANDLER
and 2) the handler will ret
without a security cookie check, then we don't need a desirable parent in our stack at all!
By replacing the parent return address on the stack with an address protected by a "desirable" exception handler which meets the previous requirements, the unwinding process will pass the exception to the fake parent's handler, which will ret
into an address we control on the stack.
Finding these "desirable" handlers can be easier than it may seem. For example, if we statically compile the previous barebone example application, we already have several candidates to choose from.
[RUNTIME_FUNCTION]
...
[UNWIND_INFO]
...
Flags: UNW_FLAG_EHANDLER
Unwind codes: .ALLOCSTACK 0x18
[SCOPE_TABLE]
Scope 0
BeginAddress: 0x16cb
EndAddress: 0x1755
HandlerAddress: 0x10314
push rbp
mov rbp, rdx
mov rax, qword ptr [rcx]
xor ecx, ecx
cmp dword ptr [rax], 0xc0000005
sete cl
mov eax, ecx
pop rbp
ret
JumpTarget: 0x1755
xor al, al
add rsp, 0x18
ret
Above is an excerpt from a tool I wrote to dump C-specific exception handlers. These structures correspond to the __scrt_is_nonwritable_in_current_image
function in the MSVCRT. Look at the exception filter's disassembly. If we can generate an access violation exception (i.e. by reading/writing an invalid pointer), the exception handler (jump target) would be executed, returning to any address of our choice.
Theory
As we covered earlier, using exceptions to escape security cookies is not new. In the past, however, the methods have involved x86-specific weaknesses, such as the exception handler pointer being stored on the stack. This new approach works with 64-bit applications by leveraging existing legitimate exception handlers.
Taking a step back and looking at this attack as a generic methodology. If you...
- Have a stack overflow primitive.
- Can trigger an exception before a security cookie check.
- Know the location of any module in the process that contains a region of code protected by an exception handler...
- Whose C-specific exception filter (handler address) returns
EXCEPTION_EXECUTE_HANDLER
for the exception you can generate. - Whose C-specific exception handler (jump target)
ret
's without a security cookie check.
- Whose C-specific exception filter (handler address) returns
- Meet specific compiler requirements (discussed later).
You can spoof your call stack to include a region of code protected by the desirable exception handler, trigger an exception, and bypass the security cookie check entirely.
Are All Compilers Impacted?
SEH Security Cookie Check
When I was parsing the exception handlers registered for modules such as ntdll.dll
, I was confused to see that only 203 out of 713 exception handlers were set to the expected __C_specific_handler
function. Here is a breakdown of the handlers for my version of ntdll.dll
:
- 713 total runtime function entries with a registered exception or termination handler
- 203 entries matched
__C_specific_handler
- 454 entries matched
__GSHandlerCheck
? - 48 entries matched
__GSHandlerCheck_SEH
? - 5 entries matched
LdrpICallHandler
- 1 entry matched
KiUserApcHandler
- 1 entry matched
RtlpExceptionHandler
- 1 entry matched
RtlpUnwindHandler
The two __GSHandlerCheck
functions caught my eye. What were these exception handlers being used for?
__GSHandlerCheck
- This handler takes an undocumented structure from theUNWIND_INFO
'sExceptionData
field and passes it to__GSHandlerCheckCommon
. If this call succeeds,__GSHandlerCheck
returnsEXCEPTION_EXECUTE_HANDLER
.__GSHandlerCheckCommon
parses this undocumented structure to find the location of the security cookie for the function the exception was occurring in. Then, it emulates the cookie check usually found in the epilog of a function by XOR'ing the cookie from the stack and jumping to__security_check_cookie
.
__GSHandlerCheck_SEH
- This function does nearly the same thing as__GSHandlerCheck
, except after checking the security cookie, it calls__C_specific_handler
.
Taking a look at the functions that __GSHandlerCheck
and __GSHandlerCheck_SEH
were assigned to revealed that all of them had security cookie checks built into them. The __GSHandlerCheck_SEH
variant appeared to be used in functions that also had an exception handler, whereas __GSHandlerCheck
was used in functions with only a security cookie check.
MSVC Mitigation
void GetString() {
char tempBuffer[16];
scanf("%s", tempBuffer);
printf(tempBuffer);
}
This was a smart mitigation by Microsoft. The purpose behind these exception handlers is to prevent attackers from being able to escape a security cookie check by causing an exception. For example, take a look at what happens when I compile the previous GetString
function using the MSVC++ compiler:
Although GetString
does not use an exception handler, it is built with one anyway. The disassembly above shows that the unwind handler is defined as __GSHandlerCheck
. Even if an attacker could cause an exception in GetString
(i.e. with a bad format string), before unwinding the stack, __GSHandlerCheck
would be called, and a security cookie check would occur- preventing the bypass. Additionally, there are __GSHandlerCheck
variants for several other common "language-specific" handlers such as __GSHandlerCheck_EH
for C++'s __CxxFrameHandler3
.
Outside of our example, this is an effective mechanism that makes abuse of exceptions in applications that use the MSVC compiler significantly more difficult. With this mitigation, an attacker would need to predict the security cookie of the function they can cause an overflow in. Note that this doesn't mean an attacker knows the security cookie for all functions.
If an attacker could get around the initial security cookie, they could likely leverage ROP. There are some advanced attacks with exceptions we'll discuss that can provide more powerful primitives than ROP, however. Additionally, as we'll review in a later section, exceptions can be a solid alternative to ROP in environments that use the MSVC compiler and hardware mitigations like Intel's Control Flow Enforcement Technology (CET).
What About Other Compilers?
A noteworthy caveat in my last paragraph is that Microsoft's mitigation makes our lives harder only in applications that use MSVC. What about applications created with other compilers for Windows?
Clang/LLVM
For the GetString
original example, I used Clang/LLVM in Visual Studio, which does not use the __GSHandlerCheck
mitigation for functions with security cookie checks. This means that any application compiled with Clang/LLVM at the time of writing is vulnerable.
GCC
Although GCC does not support SEH-style __try
/__except
blocks, it still uses SEH for C++ exceptions. We can replace our main function with a C++ try
/catch
block to compile the application.
As you can see, there is no unwind block for GetString
, demonstrating that applications compiled for GCC are also vulnerable to this attack.
Honorable Mentions
A side note- several compiled languages outside of C/C++ for Windows, like Rust and GoLang, do not have an equivalent to the __GSHandlerCheck
mitigation. But, of course, these languages are designed to be inherently safe against these vulnerabilities in the first place, assuming developers don't use the unsafe
functionality.
Most Applications Use MSVC, Though… Right?
It may seem as if the threat of this attack is reduced on Windows because of the MSVC mitigation. However, although many applications are compiled with MSVC, Clang/GCC is still frequently used, especially for cross-platform applications.
I'll give you a great example. What if I told you that the top three browsers on Windows are vulnerable to this attack?
Google Chrome, Firefox, and (ironically) Microsoft Edge use Clang/LLVM, which does not have a __GSHandlerCheck
mitigation equivalent (yet). This means that if there was a stack overflow vulnerability in the browser you might be reading this article on, an attacker could potentially abuse exceptions to escape security cookie checks!
The wrong takeaway would be that developers should use MSVC over its alternatives or that this is somehow the fault of Clang/GCC's developers. Yes, applications compiled with Clang/GCC are not protected against this attack at the time of writing, but that can change. Microsoft should proactively work with compiler developers to share the mitigations developed for MSVC. This would only help make the ecosystem more secure as a whole.
… Microsoft Has Known About This for How Long?
One question that caught my curiosity was, "How long has Microsoft known about the attack of abusing exceptions to escape security cookies?". Several old versions of 64-bit binaries I looked at seemed to contain the __GSHandlerCheck
function.
For a more conclusive answer around a date, I sought out several old versions of Windows and checked if their ntdll
binaries contained __GSHandlerCheck
. I was shocked to find an ntdll
binary signed in 2008 for Windows Vista with this mitigation in place. This suggests that Microsoft has known about this attack for at least 15 years!
Now that we've introduced the trivial implementation of Exception Oriented Programming for stack overflow vulnerabilities, let's revisit the SEH unwinding process and explore advanced attacks.
An Alternative to ROP
Background
Unwind Operations
typedef struct _UNWIND_INFO {
unsigned char Version : 3;
unsigned char Flags : 5;
unsigned char SizeOfProlog;
unsigned char CountOfCodes;
unsigned char FrameRegister : 4;
unsigned char FrameOffset : 4;
UNWIND_CODE UnwindCode[1];
/* UNWIND_CODE MoreUnwindCode[((CountOfCodes+1)&~1)-1];
* union {
* OPTIONAL unsigned long ExceptionHandler;
* OPTIONAL unsigned long FunctionEntry;
* };
* OPTIONAL unsigned long ExceptionData[];
*/
} UNWIND_INFO, * PUNWIND_INFO;
In an earlier section about the exception directory and unwind info structure, we skipped over the UNWIND_CODE
structure as it was irrelevant at the time.
typedef union _UNWIND_CODE {
struct {
unsigned char CodeOffset;
unsigned char UnwindOp : 4;
unsigned char OpInfo : 4;
};
unsigned short FrameOffset;
} UNWIND_CODE, *PUNWIND_CODE;
UNWIND_CODE
entries specify the operations required to "unwind" (or undo) the changes a given function's prolog has made to registers or the stack. Here are the various documented operations an UNWIND_CODE
structure can contain:
UWOP_PUSH_NONVOL
- This operation specifies that a nonvolatile integer register was pushed to the stack. TheOpInfo
field specifies what register was pushed. For example, if the prolog of a function containedpush rbp
, you would see a correspondingUWOP_PUSH_NONVOL
operation.UWOP_ALLOC_LARGE
/UWOP_ALLOC_SMALL
- These operations specify that a specific size was allocated on the stack. You would expect to see these operations for instructions likesub rsp, 0xABC
.UWOP_SET_FPREG
- Specifies the frame pointer register and some offset ofrsp
. This operation is only used in functions that need a frame pointer in the first place, such as those that need dynamic stack allocations. An example instruction for this operation would includelea rbp, [rsp+offset]
.UWOP_SAVE_NONVOL
/UWOP_SAVE_NONVOL_FAR
- These operations specify that a nonvolatile integer register was saved on the stack using amov
instruction rather than apush
. Similar toUWOP_PUSH_NONVOL
, the specific register is contained in theOpInfo
field.UWOP_SAVE_XMM128
/UWOP_SAVE_XMM128_FAR
- These operations are used to save XMM register values.UWOP_PUSH_MACHFRAME
- This is a special type of operation that indicates the function is a hardware interrupt or exception handler that receives a "machine frame" from the stack. This frame contains information about the state of various registers at the time the interrupt/exception occurred. An example of a function with this operation in user-mode includesntdll!KiUserExceptionDispatcher
.
With a basic understanding of how the dispatcher can unwind the effects of various functions, let's go through an example.
Dumping the Exception Directory of NTDLL
As a small demo of everything we've learned, we can use the Python pefile package to enumerate the exception directory of any PE module. Here is a small script that will print the runtime function entries of a binary specified by the first argument.
import sys
import pefile
pe = pefile.PE(sys.argv[1])
for runtime_function in pe.DIRECTORY_ENTRY_EXCEPTION:
print("\n".join(runtime_function.struct.dump()))
if hasattr(runtime_function, "unwindinfo") and \
runtime_function.unwindinfo is not None:
print("\n\t".join(runtime_function.unwindinfo.dump()))
The ntdll.dll
on my machine produced 4884 unique runtime function entries. This doesn't mean that there are 4884 functions in ntdll.dll
with an exception handler- entries often only contain the operations needed to unwind a given function.
[RUNTIME_FUNCTION]
0x168A40 0x0 BeginAddress: 0x4C270
0x168A44 0x4 EndAddress: 0x4C45F
0x168A48 0x8 UnwindData: 0x146D50
[UNWIND_INFO]
0x143F50 0x0 Version: 0x1
0x143F50 0x0 Flags: 0x3
0x143F51 0x1 SizeOfProlog: 0x1F
0x143F52 0x2 CountOfCodes: 0x8
0x143F53 0x3 FrameRegister: 0x0
0x143F53 0x3 FrameOffset: 0x0
0x143F64 0x14 ExceptionHandler: 0x9CC44
Flags: UNW_FLAG_EHANDLER, UNW_FLAG_UHANDLER
Unwind codes: .ALLOCSTACK 0x70; .PUSHREG R15; .PUSHREG R14; .PUSHREG R13; .PUSHREG R12; .PUSHREG RDI; .PUSHREG RSI; .PUSHREG RBX
I've noted a couple of times that the unwind operations are there to "undo" the prolog of the function. I'd like to show a practical example of this. Above, we have the runtime function entry for ntdll!RtlQueryAtomInAtomTable
. Look at the instructions in the epilog of the function (intended to "restore" the changes of the prolog) and see if you notice a pattern with the unwind operations:
RtlQueryAtomInAtomTable proc near
; __unwind { // __GSHandlerCheck_SEH
; PROLOG
; ...
; FUNCTION CONTENT
; ...
mov rcx, [rsp+0A8h+var_48]
xor rcx, rsp ; StackCookie
call __security_check_cookie
; EPILOG
add rsp, 70h ; .ALLOCSTACK 0x70
pop r15 ; .PUSHREG R15
pop r14 ; .PUSHREG R14
pop r13 ; .PUSHREG R13
pop r12 ; .PUSHREG R12
pop rdi ; .PUSHREG RDI
pop rsi ; .PUSHREG RSI
pop rbx ; .PUSHREG RBX
retn
; }
RtlQueryAtomInAtomTable endp
The instructions in the epilog match the unwind operations and occur in the same order too! This is why runtime function entries are critical to the unwinding process. They effectively tell you how to restore the state of the stack and registers at any point in time, even if you're in the middle of executing a function. Without this context, writing a reliable unwinding mechanism to support arbitrary continuation at an exception handler would be much more challenging.
What About CET / Shadow Stacks?
An interesting mitigation we have not yet covered is Hardware-enforced Stack Protection, otherwise known as Control-flow Enforcement Technology (CET) for Intel CPUs and shadow stacks for AMD CPUs.
At a high level, when a function is called and a return address is pushed on the regular stack, a copy of that return address is also pushed on a "shadow stack" region. When the function returns, the address on the normal stack, which could have been corrupted by an attacker, is compared with the value on the shadow stack. If these values don't match, the program is terminated.
This is an opt-in mitigation, meaning you won't find it turned on by default. In 2021, Google Security wrote a blog about the work that went into enabling shadow stacks for Chrome. Although shadow stacks certainly aren't commonplace for most applications, it's a mitigation we may see increased adoption of in the future as it becomes more standardized.
This article is not intended to be a comprehensive look into how shadow stacks work in practice. If you're curious and want to learn more about specific implementation details, check out "RIP ROP: CET Internals in Windows 20H1" by Yarden Shafir and Alex Ionescu.
Going back to our trivial implementation of Exception Oriented Programming, let's say we are in the context of a process with Hardware-enforced Stack Protection. Even if we did escape a security cookie check by throwing an exception into a handler without one, when the handler returns, our corrupted return address on the stack would not match what's on the shadow stack and thus prevent our attack.
Core Concepts
In a classical ROP attack, the epilogues of functions are chained together to perform various operations, such as modifying registers and the stack, before returning to a function, simulating a call. Security cookies initially posed a significant challenge to ROP, as you typically couldn't modify the return address without also corrupting the cookie. With the next generation of system mitigations like shadow stacks, ROP is only becoming more challenging of an attack to leverage in real-world scenarios.
The trivial approach of escaping a security cookie check by throwing an exception only scratches the surface of what is possible through the unwinding process.
Here is a quick reminder about how exception dispatching works from a high level:
RtlDispatchException
is called, which "virtually unwinds" the stack and calls the exception handlers for functions with theUNW_FLAG_EHANDLER
flag in theirUNWIND_INFO
structure.- If a handler returns
EXCEPTION_CONTINUE_EXECUTION
, the virtual unwinding process is halted, and execution continues where the exception occurred, with the original state of the registers. - If a handler returns
EXCEPTION_CONTINUE_SEARCH
, the virtual unwinding process continues for other exception handlers. - If in the context of the C-specific language handler and the exception filter (handler address) returns
EXCEPTION_EXECUTE_HANDLER
,RtlUnwind
is called.
- If a handler returns
- If
RtlUnwind
is called (i.e. by__C_specific_handler
), the state of the stack/registers is unwound to continue execution at the C-specific exception handler (jump target). Unlike "virtual" unwinding, changes made to theCONTEXT
structure by unwind operations will be reflected when execution continues at the handler.
There is an enormous amount of attack surface here. Sure, in the context of 64-bit applications, we will generally be limited to legitimate exception handlers and UNWIND_INFO
structures. This is similar to how we are stuck with executing the epilog's of legitimate functions with ROP as "gadgets". As an attacker with a stack overflow vulnerability, however, since we can overwrite the call stack with whatever we want, we have complete control over what legitimate functions are used in the unwinding process and the order in which they are used.
How? In our trivial example, we modified the caller of our function with a cookie check to be a legitimate function that has an exception handler without a cookie check. Why stop at adding only one function to the call stack? Why not leverage unwind operations to modify registers to use untrusted values from the stack we control? This is where things start to get interesting. Let's break these attacks down.
What Can We Do in RtlDispatchException
with Control over the Stack?
The first half of exception dispatching is to "virtually unwind" the stack in RtlDispatchException
, calling exception handlers as we encounter them. As an attacker with influence over the stack, we can dictate the legitimate functions that RtlDispatchException
will use when "virtually unwinding". So what does that let us do?
In the context of Windows binaries, we often deal with the C-specific language handler, where functions can specify what exceptions they want to catch with an exception filter. Remember that the C-specific exception handlers (jump targets) are triggered via RtlUnwind
, which is outside the scope of RtlDispatchException
.
Unfortunately, combined with the fact that the CONTEXT
record we can influence when "virtual unwinding" occurs isn't used anywhere outside of this function, there is not much we can do in RtlDispatchException
alone other than trigger a "desirable" C-specific exception handler (jump target) for unwinding.
To trigger a legitimate function's C-specific exception handler (jump target), we face two main requirements:
- We need to know the address of the legitimate function (i.e. by knowing where the module is located).
- There is a slight exception we'll discuss later: we can leverage a partial return address overwrite to access the functions in the same module as what's already on our call stack.
- The function's exception filter (handler address) needs to be 1 or return
EXCEPTION_EXECUTE_HANDLER
for our exception code.
Once we meet these requirements, the next step is to figure out where to write our "fake return address" on the stack. Our goal is to trick the unwinding process into thinking the function with our desired exception handler "called" the function where we generated an exception; hence it should be the one to handle our exception.
For our first parent caller, knowing the location to write is easy- it's just where the actual return address for the previous function is. Once we overwrite the first return address, however, we need to do some math.
[RUNTIME_FUNCTION]
...
[UNWIND_INFO]
...
Unwind codes: .ALLOCSTACK 0x70; .PUSHREG R15; .PUSHREG R14; .PUSHREG R13; .PUSHREG R12; .PUSHREG RDI; .PUSHREG RSI; .PUSHREG RBX
Let's go through a quick example with ntdll!RtlQueryAtomInAtomTable
. Imagine we replaced the first caller on the stack with an address to RtlQueryAtomInAtomTable
. Where would we write the second caller's return address?
Each function's prolog is going to impact the stack differently. We have to account for each unwind operation that modifies Rsp
. In this case, we have a stack allocation of 0x70 bytes (sub rsp, 0x70
) and 7 registers being pushed (push r??
). Assume our offset is relative to the location of the previous return address + 0x8
. To calculate the total stack allocation for RtlQueryAtomInAtomTable
, we do 0x70 + 0x8*7 = 0xA8
, where 0x70 is our stack allocation, and 0x8*7 accounts for each register that was pushed. This means our next return address would be written to Rsp + 0xA8
following the previous one.
As I mentioned earlier, you can look at the source code of RtlDispatchException in the leaked Windows Research Kernel yourself. Additionally, here is the specific code that will tell you exactly how each unwind operation modifies Rsp
.
Now that we know how to do this math, we can chain together an unlimited amount of functions on the call stack- while accounting for how they impact Rsp
. However, before we get into RtlUnwind
, there is one more small step to understand.
typedef struct _SCOPE_TABLE_AMD64 {
DWORD Count;
struct {
DWORD BeginAddress;
DWORD EndAddress;
DWORD HandlerAddress;
DWORD JumpTarget;
} ScopeRecord[1];
} SCOPE_TABLE_AMD64, *PSCOPE_TABLE_AMD64;
Our first requirement was to know the location of the legitimate function protected by our desirable exception handler. Writing this function's address to the call stack alone is not sufficient. Remember that the scope record structure specifies what part of the function is protected by a given filter/handler with the BeginAddress
/EndAddress
fields. To ensure we trigger our desired JumpTarget
handler, we need to add at least the BeginAddress
of the relevant scope to our image base rather than only the function offset.
To recap, RtlDispatchException
will virtually unwind through each function in our fake call stack. Once the entry with our desirable exception handler is reached, __C_specific_handler
is called. Then, since our return address is within the scope bounds, the exception filter is called, which returns EXCEPTION_EXECUTE_HANDLER
(or is 1, which means the same thing). This will trigger a final call to RtlUnwind
, responsible for unwinding the stack and resuming execution at our exception handler!
The following section is where we'll get into some fun bits and explore why we'd want to include other functions (including those without exception handlers) before our final target.
What Can We Do in RtlUnwind
with Control over the Stack?
Now that we understand how to create a fake call stack with any function we want and how to trigger RtlUnwind
from the context of an exception, let's get into some primitives.
A quick reminder about the differences between RtlDispatchException
and RtlUnwind
. In RtlDispatchException
, the CONTEXT
structure being modified does not impact anything other than the virtual unwind process itself, whereas in RtlUnwind
, changes will be reflected when execution continues. Also, in RtlDispatchException
, only functions with UNW_FLAG_EHANDLER
in their unwind info structure will have their ExceptionHandler
called, whereas in RtlUnwind
, the same is true with the UNW_FLAG_UHANDLER
flag.
The four primary "primitives" we can leverage in RtlUnwind
are:
- We can resume execution at any C-specific exception handler (jump target) whose location we know and whose exception filter we've passed.
- We can read untrusted values into registers from the stack and these values will be reflected once execution is resumed at our C-specific exception handler.
- We can influence the offset on the stack we resume execution at.
- We can execute any termination/unwind handler we want on our way to the unwind destination.
- For example, before reaching our desired exception handler (jump target), we could trigger the
__finally
blocks for any function we know the address of. This can lead to scenarios like a use-after-free or double-free.
- For example, before reaching our desired exception handler (jump target), we could trigger the
We've covered how to do #1 already in the previous section. What about the other primitives?
Modify Any Register We Want
Earlier I said there were reasons we would want to include functions on our call stack prior to the function with our desirable exception handler. We can leverage the unwind operations of any function we know the location of to modify registers before we resume execution at a desirable exception handler.
How? Many functions do not have an exception handler defined, but they still have a runtime function / unwind info entry to allow them to be unwound. The reason unwind operations can provide us with very powerful primitives is because they are designed to restore untrusted data from the stack into registers to which we otherwise wouldn't have access.
For example, take the UWOP_PUSH_NONVOL
operation that represents PUSH REGISTER
instructions. Let's say our fake call stack had two functions, one that has no handler with a UWOP_PUSH_NONVOL
unwind operation and one that has our desired exception handler. When the first function is unwound, RtlVirtualUnwind
in RtlUnwind
will replace the value of REGISTER
with an untrusted value from the stack. By including this function in our call stack, we have direct control over the value of REGISTER
when execution resumes at our exception handler. Even more powerful- we can chain together multiple functions with desirable unwind operations on the call stack to ultimately influence the value of almost every register!
Although you can see what each unwind operation does in the leaked Windows Research Kernel, I've created a small summary of the primitives they give us below:
UWOP_PUSH_NONVOL
- Pulls an untrusted value from the stack and places it into a register specified by theOpInfo
field.UWOP_ALLOC_LARGE
/UWOP_ALLOC_SMALL
- IncrementsRsp
by a constant value.UWOP_SET_FPREG
- SetRsp
to the register specified by theFrameRegister
field. Subtract 16 * theFrameOffset
field. Both fields are from the unwind info structure.UWOP_SAVE_NONVOL
/UWOP_SAVE_NONVOL_FAR
- Read a value from a constant offset on the stack into the register specified by theOpInfo
field.UWOP_PUSH_MACHFRAME
- Pulls an untrusted value from the stack and places it intoRip
.Rsp
is then replaced with an untrusted value from offset 0x18 of the current stack.- Note that
Rip
is replaced at the end ofRtlUnwind
(unless your exception code isSTATUS_UNWIND_CONSOLIDATE
), so no, you can't just restore execution at some arbitrary address from the stack.
- Note that
As you can see, these operations provide powerful primitives to modify the state of registers before restoring at a legitimate exception handler. Of course, you would need to find these operations already present in an existing function's unwind info structure, but given that every function has an unwind info structure, you have a lot to choose from.
Another side effect to worry about is "what if I only want to replace a few registers, but my unwind info structure contains operations I don't want?". Fortunately, there is a trick we can use to get around this.
typedef union _UNWIND_CODE {
struct {
unsigned char CodeOffset;
unsigned char UnwindOp : 4;
unsigned char OpInfo : 4;
};
unsigned short FrameOffset;
} UNWIND_CODE, *PUNWIND_CODE;
When unwind operations are enumerated, it's not as simple as "just enumerate every operation if the exception address is in this function". For example, what happens if an exception occurs in the prolog?
To account for this exists the CodeOffset
field of each UNWIND_CODE
(unwind operation). An unwind operation is only executed if the address inside the function minus the function address itself is greater than or equal to the CodeOffset
. This way, if an exception occurred in the prolog, only unwind operations corresponding to instructions that have already been executed would be processed.
This functionality is helpful because we can specify which unwind operation we want to start with!
[UNWIND_INFO]
Unwind codes:
0x10: .ALLOCSTACK 0x70
0xc: .PUSHREG R15
0xa: .PUSHREG R14
0x8: .PUSHREG R13
0x6: .PUSHREG R12
0x4: .PUSHREG RDI
0x3: .PUSHREG RSI
0x2: .PUSHREG RBX
For example, above, the unwind operations for ntdll!RtlQueryAtomInAtomTable
are prepended with their CodeOffset
fields. If we wanted to only replace the value of RBX
, we place the address of RtlQueryAtomInAtomTable
+ 0x2 on the stack. This works because RtlpUnwindPrologue will assume an exception occurred at the PUSH RBX
instruction, thus only process that unwind operation.
Influence the Stack Pointer
Another useful primitive is the ability to modify the stack pointer. Here are the most prominent reasons this would be helpful:
- If we want to return to a legitimate function on the call stack that we haven't modified after faking the functions "below it", we need to align
Rsp
with that legitimate function's return address on the stack.- If we are executing a desirable exception handler (jump target) who will
ret
into a legitimate caller, we'd need to account for the changes the handler will make to the stack.
- If we are executing a desirable exception handler (jump target) who will
- If we are using a partial return address overwrite (i.e. because we don't have a leak), then controlling
Rsp
would let us choose which return address on the legitimate call stack we want to perform an overwrite on.- Besides having a more comprehensive selection of modules to choose from, maybe we control the local variables for some caller and can find an exception handler that reads from these local variables.
The first method of influencing Rsp
does require leaks, but it's relatively straightforward. As previously discussed, each function's unwind operations will impact Rsp
differently. By placing legitimate functions on the call stack, we can use these unwind operations to increment Rsp
by any amount we want. We can control this even further by using the previous trick of placing our return address in the middle of a prolog (to exclude certain unwind operations).
The second method of influencing Rsp
without leaks is slightly more nuanced. When I was reading part 3 of Ken Johnson's series explaining how 64-bit exception handling worked on Windows, this paragraph caught my eye:
If RtlUnwindEx encounters a "leaf function" during the unwind process (a leaf function is a function that does not use the stack and calls no subfunctions), then it is possible that there will be no matching RUNTIME_FUNCTION entry for the current call frame returned by RtlLookupFunctionEntry. In this case, RtlUnwindEx assumes that the return address of the current call frame is at the current value of Rsp (and that the current call frame has no unwind or exception handlers). Because the x64 calling convention enforces hard rules as to what functions without RUNTIME_FUNCTION registrations can do with the stack, this is a valid assumption for RtlUnwindEx to make (and a necessary assumption, as there is no way to call RtlVirtualUnwind on a function with no matching RUNTIME_FUNCTION entry). The current call frame's value of Rsp (in the context record describing the current call frame, not the register value of rsp itself within RtlUnwindEx) is dereferenced to locate the call frame's return address (Rip value), and the saved Rsp value is then adjusted accordingly (increased by 8 bytes).
Ken described the logic that occurs during unwinding when a return address is retrieved from the stack that does not have a corresponding function entry. This is helpful for our purposes because what it means is that if we put an invalid address on the stack, the unwinding process will consider it a "leaf function" (since it can't find a function entry) and skip over it! This allows us to increment Rsp
by 8 by including invalid addresses in our call stack.
I ran into an issue when I was testing this out myself. I had placed one constant address on the stack repeatedly, hoping that Rsp
would keep incrementing by 8. What happened instead was after the first instance of the invalid constant, unwinding failed. I found the answer while reading through the WRK leak of RtlDispatchException:
//
// If the old control PC is the same as the return address,
// then no progress is being made and the function tables are
// most likely malformed.
//
if (ControlPc == *(PULONG64)(ContextRecord1.Rsp)) {
break;
}
What was happening was that the unwinding process included a check to make sure the previous control point did not match the next control point. So, to fix the behavior of incrementing Rsp
as many times as I wanted, all I had to do was swap between two invalid constants. For example, if I wanted to increment Rsp
by 0x20, my call stack would look like the following:
0x1111111111111111
0x2222222222222222
0x1111111111111111
0x2222222222222222
This is helpful when we don't have a leak because we can create a "sled" to get to any other legitimate return address on the stack.
Summary
To summarize the advanced variants of Exception Oriented Programming here is what we can do as an attacker with a stack overflow vulnerability:
- We can restore execution at any C-specific exception handler (jump target) we know the location of.
- We can directly control the state of registers when an exception is handled.
- This could be leveraged to corrupt the caller's state when our exception handler returns to a given function.
- We can call the termination handler (
__finally
block) in any function we know the location of.- This could be used to trigger a use-after-free or double-free scenario.
Without any leaks, we would be limited to partial return address overwrites. Fortunately, we can influence Rsp
such that we can overwrite any address on the stack rather than only our direct parent. However, the primary method of controlling what values are used in unwind operations would be to overwrite the return address of a function we control specific local variables in. This is because we cannot overwrite the local variables without knowing the complete return address.
Does This Work with CET / Shadow Stacks?
One of the considerable benefits of these attacks is that we only hit a return instruction relevant to shadow stacks when our desired exception handler returns. Otherwise, the unwinding process blindly trusts addresses from the stack we control.
Still, the return instruction in the exception handler could pose a problem; what can we do to ensure we don't crash at that point?
First, whenever a new function on your call stack is enumerated by the unwinding process, a return address is "popped" from the shadow stack. This is important because when you unwind to a function N
return addresses away, you need to get rid of those N
return addresses on the shadow stack to ensure that the next ret
instruction will match what's on the shadow stack.
This is a useful "feature" that we can abuse because we can cause corruption in the state of an application by returning to any function in the call stack we want to. By chaining together fake exception handlers, we can "pop" the shadow stack until we reach a desired parent function and use a jump target's ret
instruction to return to it early.
CET does not verify that your return address is at the same stack offset where it was initially placed. Therefore, as long as the value matches what's on the shadow stack, it does not matter where on the stack the return address is retrieved from.
This design weakness has been known for at least two years, however. Another security researcher, Saar Amar, highlighted how an attacker could cause a "type confusion" condition even with CET by unwinding to a desired function already on the call stack.
For example, imagine a parent function responsible for initializing a structure. By returning to a function above it mid-way during the initialization process, that parent may end up using an incomplete structure.
If you have a leak of the module location, you can predict the legitimate return address on the shadow stack (after the popping occurs) and put it on the normal stack where the following return address will be retrieved from. If you don't know the location of the legitimate function and can avoid overflowing it on the stack, you can create a fake call stack that will increment Rsp
right up to that legitimate address.
Blast from the Past
Background
Part 5 of Ken Johnson's series on 64-bit exception handling discussed how certain edge cases known as "collided unwinds" were addressed. To give a high-level overview of what collided unwinds are, here is a good quote from Ken's blog:
A collided unwind occurs when an unwind handler initiates a secondary unwind operation in the context of an unwind notification callback. In other words, a collided unwind is what occurs when, in the process of a stack unwind, one of the call frames changes the target of an unwind. This has several implications and requirements in order to operate as one might expect:
1. Some unwind handlers that were on the original unwind path might no longer be called, depending on the new unwind target.
2. The current unwind call stack leading into RtlUnwindEx will need to be interrupted.
3. The new unwind operation should pick up where the old unwind operation left off. That is, the new unwind operation shouldn't start unwinding the exception handler stack; instead, it must unwind the original stack, starting from the call frame after the unwind handler which initiated the new unwind operation.
An even more straightforward way of thinking about a collided unwind is that it occurs when an unwind handler calls RtlUnwind
itself. To solve this problem, Ken describes Microsoft's "elegant" solution: any call to an unwind handler is executed through a helper function, RtlpExecuteHandlerForUnwind
.
typedef struct _DISPATCHER_CONTEXT {
ULONG64 ControlPc;
ULONG64 ImageBase;
PRUNTIME_FUNCTION FunctionEntry;
ULONG64 EstablisherFrame;
ULONG64 TargetIp;
PCONTEXT ContextRecord;
PEXCEPTION_ROUTINE LanguageHandler;
PVOID HandlerData;
struct _UNWIND_HISTORY_TABLE *HistoryTable;
ULONG ScopeIndex;
ULONG Fill0;
} DISPATCHER_CONTEXT, *PDISPATCHER_CONTEXT;
When RtlpExecuteHandlerForUnwind
is called by RtlUnwindEx
, it saves a pointer to the current DISPATCHER_CONTEXT
structure on the stack. As we can see above, this structure contains the entire internal state of RtlUnwindEx
.
In an earlier section, we dumped the exception handlers for functions in ntdll, where the __GSHandlerCheck*
variants were discovered. One of the other exception handlers I skipped over was RtlpUnwindHandler
. This exception handler is actually used to protect RtlpExecuteHandlerForUnwind
.
This is where the "elegant" solution kicks in. When an unwind handler calls RtlUnwindEx
and the unwinding process occurs again, RtlUnwindEx
calls the unwind handler of RtlpExecuteHandlerForUnwind
, which is RtlpUnwindHandler
. What does this handler do? It overwrites the current DISPATCHER_CONTEXT
structure with the saved DISPATCHER_CONTEXT
structure and returns ExceptionCollidedUnwind
. When an unwind handler returns this result, RtlUnwindEx
will use the overwritten values to replace its internal unwinding state. This allows the unwinding process to resume from where it was left off without any fuss.
An Overpowered Primitive
... there is not much we can do in RtlDispatchException
alone other than trigger a "desirable" C-specific exception handler (jump target) for unwinding ...
When I was covering what we could do with control over the stack in RtlDispatchException
, remember how I said, "not much"?
I lied.
While reading the leaked source for RtlDispatchException
, I noticed that it too contained code to handle the ExceptionCollidedUnwind
result returned by exception handlers. This may have been added to cover the unlikely edge case where an exception occurs, RtlUnwindEx
is called, an unwind handler is called, and another exception occurs.
This is where I got a wild idea- as an attacker with a stack overflow vulnerability, we can modify the call stack to whatever we want, right? If we knew the location of ntdll, couldn't we make it seem like the function we are causing an exception in was called by RtlpExecuteHandlerForUnwind
?
If this was true, knowing that RtlpUnwindHandler
grabs the DISPATCHER_CONTEXT
structure pointer from the stack, if we had any memory location in the process that was attacker-controlled, couldn't we overwrite the entire internal state of RtlDispatchException
with our own values?
//
// The dispostion is collided unwind.
//
// A collided unwind occurs when an exception dispatch
// encounters a previous call to an unwind handler. In
// this case the previous unwound frames must be skipped.
//
case ExceptionCollidedUnwind:
ControlPc = DispatcherContext.ControlPc;
ImageBase = DispatcherContext.ImageBase;
FunctionEntry = DispatcherContext.FunctionEntry;
EstablisherFrame = DispatcherContext.EstablisherFrame;
RtlpCopyContext(&ContextRecord1,
DispatcherContext.ContextRecord);
RtlVirtualUnwind(UNW_FLAG_EHANDLER,
ImageBase,
ControlPc,
FunctionEntry,
ContextRecord1,
&HandlerData,
&EstablisherFrame,
NULL);
ContextRecord1.Rip = ControlPc;
ExceptionRoutine = DispatcherContext.LanguageHandler;
HandlerData = DispatcherContext.HandlerData;
HistoryTable = DispatcherContext.HistoryTable;
ScopeIndex = DispatcherContext.ScopeIndex;
Repeat = TRUE;
break;
Knowing the location of ntdll and some memory we control certainly isn't trivial, but the impact is astronomical. For example, in the (slightly corrected) WRK excerpt above, when the ExceptionCollidedUnwind
result is returned, the overwritten DispatcherContext
variable is used to update the current Rip
, image base, runtime function entry, CONTEXT
record, the UNWIND_HISTORY_TABLE
, and even a pointer that specifies an ExceptionRoutine to immediately call. All of this is controlled by an attacker with a stack overflow vulnerability.
See the Repeat
variable getting set to TRUE
? What happens is right after the break
, the while loop for calling an exception handler repeats without calling RtlLookupFunctionEntry
, and the attacker-controlled ExceptionRoutine
is passed to RtlpExecuteHandlerForException
, whose the second argument (RDX
/EstablisherFrame
) is entirely controlled by the attacker as well.
To add insult to injury, the call inside of RtlpExecuteHandlerForException
to the attacker-controlled ExceptionRoutine
is done without a Control Flow Guard (or xFG) check, meaning we can call into the middle of any function or any unaligned address.
I can't help but draw parallels to the x86 SEH hijacking attacks from ~15+ years ago, where an attacker could overflow the stack and replace the exception handler that would be called. With this primitive, we can achieve the same result with even more control.
To give you an idea of what's possible with the variables we can modify:
- We can call any function anywhere we want (via
LanguageHandler
) with the second argument (RDX
/EstablisherFrame
) completely controlled. - When
RtlVirtualUnwind
is called following theExceptionCollidedUnwind
result, we have full control over theControlPc
,ImageBase
,RUNTIME_FUNCTION
structure, andUNWIND_INFO
structure it uses.- This means we can execute any unwind operations we want. We'll talk more about how this can be abused in practice later.
- If we meet certain conditions (discussed later) and specify an exception handler that returns
EXCEPTION_CONTINUE_SEARCH
, we can continuously hijack the dispatching/unwinding process since we control theUNWIND_HISTORY_TABLE
structure.- The history table is used as a cache to store previously retrieved
RUNTIME_FUNCTION
entries. By creating a malicious history table, we can specify our ownRUNTIME_FUNCTION
structure for any function we know the location of.
- The history table is used as a cache to store previously retrieved
- If we set
LanguageHandler
to the C-specific exception handler in ntdll, we can define a customSCOPE_TABLE
structure to call any ntdll functions we want consecutively (as long as the functions return 0).
Enough about what we can do; let's go through a practical example!
Planning Our Attack
To leverage this primitive in a stack overflow attack, we need to meet two major requirements:
- We need to know the location of ntdll.
- We need to know the location of attacker-controlled memory. This memory can be anywhere in the process.
Although we have significant control over the unwinding process, we still have many quirks and challenges to overcome. To best describe these limitations, I'll explain what occurs in RtlDispatchException
when the ExceptionCollidedUnwind
result is returned.
- The
ControlPc
,ImageBase
,FunctionEntry
, andContextRecord
variables inRtlDispatchException
are updated to attacker-controlled values.- The only variable we haven't explicitly covered is
ControlPc
, which represents the currentRip
value. During each step in the unwinding process, this is updated to the return address from the call stack.
- The only variable we haven't explicitly covered is
RtlVirtualUnwind
is called. As an attacker, we have complete control over the runtime function and unwind info structures used for the virtual unwind.- This means we can specify any unwind operations we want. Note that the
Rsp
register in our context record must be a valid pointer no matter what- although it doesn't need to point at the stack (i.e. it could point at our controlled memory). - These unwind operations are incredibly powerful. For example, we can easily chain arbitrary reads by setting
Rsp
to some offset inside of ntdll and leveragingUWOP_PUSH_NONVOL
to read an address fromRsp
,UWOP_ALLOC_*
to incrementRsp
by any offset, etc. - At the end of this virtual unwind, our
Rsp
andRip
registers in the context record are updated. By default,Rip
is read by dereferencingRsp
, andRsp
is incremented depending on the unwind operations (not true in cases likeUWOP_PUSH_MACHFRAME
). - If
Rip
matches the currentControlPc
, the unwinding process is halted because it is assumed the stack is corrupted.
- This means we can specify any unwind operations we want. Note that the
- Once the virtual unwind is complete, the
EstablisherFrame
,HandlerData
,ExceptionRoutine
, andHistoryTable
are all updated to attacker-controlled values. - The exception handler loop continues and the function we specified in
LanguageHandler
is called. The first argument (RCX
) is a pointer to the exception record, the second argument (RDX
) is the establisher frame we control entirely, the third argument (R8
) is a pointer to the context record, and the fourth argument (R9
) is a pointer to the dispatcher context structure. - The return value of this function is checked again.
- If the result is
ExceptionContinueExecution
, execution continues where the exception occurred. - If the result is
ExceptionContinueSearch
, the search for an exception handler continues. - If the result is
ExceptionNestedException
, then the exception flags are updated and the search continues. - If the result is
ExceptionCollidedUnwind
, we start over from step 1. - Any other result leads to a non-continuable exception. This effectively means that any exception routine you specify must return a value between 1-2 to ensure the search is continued.
- If the result is
- If the search is continued, our first major challenge occurs. The
Rsp
record is validated to be inside the low/high limit of the stack.- I'll talk about a complex way we can pass this check in a later section.
ControlPc
is updated toRip
, and the unwind loop continues assumingRsp
is in the stack's bounds.- At the beginning of the loop,
ControlPc
and our controlledHistoryTable
are passed toRtlLookupFunctionEntry
to find a corresponding runtime function entry.- Since we control the history table, if we can predict the value of
ControlPc
, we can define our ownImageBase
andRUNTIME_FUNCTION
pointers.
- Since we control the history table, if we can predict the value of
RtlVirtualUnwind
is called again and practically the same logic we mentioned in step 2 occurs. The significant differences worth mentioning are:- The
EstablisherFrame
is updated to the value ofRsp
. If theFrameRegister
field in the unwind info structure is populated, it is instead set to the value of that register (minus0x10
* theFrameOffset
field). - If the
UNW_FLAG_EHANDLER
flag is set, theExceptionRoutine
is calculated by adding theExceptionHandler
field from the unwind info structure to theImageBase
.
- The
- After this virtual unwind is where our second major challenge occurs.
EstablisherFrame
is validated to be within the bounds of the stack.- During the first loop,
Rsp
must have already been a value that points at the stack; henceEstablisherFrame
would likely be updated to that valid stack pointer. - The problem is that we no longer have arbitrary control over the second argument to the exception handler. Any
RDX
value we specify must be within the stack's low/high limits (also stored on the stack).
- During the first loop,
- Assuming an exception handler was defined, the logic from step 3 starts over again.
As a reminder, you can read the source code for all of this logic in the leaked WRK.
The most significant barrier to simply looping over and over again, calling a different exception handler of our choice each time, is that after the initial ExceptionCollidedUnwind
result is handled, we quickly lose control over the second argument and we need to somehow guarantee that Rsp
is within the bounds of the stack.
Of course, this is easy if the attacker-controlled memory you know the location of is the stack. But requiring the location of attacker-controlled memory and requiring that it is on the stack is lame. This is quite a powerful primitive; we should be able to perform this attack regardless of whether our controlled memory is in the stack, the heap, or anywhere else.
Initially, I spent ~2 weeks developing a complex exploit chain to execute arbitrary shellcode without needing to step outside the bounds of the unwinding process. It worked with CET and CFG in strict mode but had some stability issues. What I realized was during this process, I had discovered several quite powerful primitives such that if our goal is only to execute code remotely, there were much simpler (and more stable) methods of gaining arbitrary code execution. Let's discuss using some of these primitives to create a stable proof-of-concept to execute arbitrary code.
Allowing Functions to Return Zero
A frustrating challenge I faced while writing an exploit was that any exception handler I called had to return 1 (ExceptionContinueSearch
) or 2 (ExceptionNestedException
). Triggering 3 (ExceptionCollidedUnwind
) would enter an infinite loop, as it would keep calling the function repeatedly (since the dispatcher context would remain the same).
} else {
ExceptionFilter =
(PEXCEPTION_FILTER)(ScopeTable->ScopeRecord[Index].HandlerAddress + ImageBase);
Value = (ExceptionFilter)(&ExceptionPointers, EstablisherFrame);
}
//
// If the return value is less than zero, then dismiss the
// exception. Otherwise, if the value is greater than zero,
// then unwind to the target exception handler. Otherwise,
// continue the search for an exception filter.
//
if (Value < 0) {
return ExceptionContinueExecution;
} else if (Value > 0) {
// RtlUnwind is called.
...
}
// Loop continues.
...
// Eventually, ExceptionContinueSearch is returned.
return ExceptionContinueSearch;
While reading the code for the __C_specific_handler
, which is included with the MSVCRT (C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\crt\src\amd64\chandler.c
), I discovered that if the exception filter it called returned zero, it will continue enumerating the scope table.
This meant we could call arbitrary functions by setting our exception handler to __C_specific_handler
and crafting a malicious scope table. As long as the return value of our fake exception filter was zero, our search would continue without issue. Given that many functions in the ntdll module return an NTSTATUS
value and STATUS_SUCCESS
is zero, this significantly increased the number of functions we could call.
Execute Multiple Exception Handlers Consecutively
The following primitive was found quickly after the last- the ability to execute as many functions as I wanted consecutively inside a module I knew the location of. If you think about the RtlDispatchException
unwinding process, it may seem as if to call multiple functions, we'd need to perform an entire unwind loop to specify a new exception routine pointer.
typedef struct _SCOPE_TABLE_AMD64 {
DWORD Count;
struct {
DWORD BeginAddress;
DWORD EndAddress;
DWORD HandlerAddress;
DWORD JumpTarget;
} ScopeRecord[1];
} SCOPE_TABLE_AMD64, *PSCOPE_TABLE_AMD64;
With a malicious scope table structure, we can define multiple scope records that overlap. Nothing stops the BeginAddress
/EndAddress
fields from specifying the same scope. As long as our exception filters (handler address) return zero, the table is entirely enumerated, and we can call functions consecutively. One relevant limitation is that we are stuck with the same second argument (RDX
) value across all consecutive function calls.
What Functions Do We Call?
The collided unwind primitive is powerful sure, but what functions can we call to get remote code execution only controlling the second argument?
Even though we don't control the other arguments entirely, they're worth calling out. For example, the fact that the first argument (RCX
) will be a consistent pointer to some writable memory region can still be helpful.
I discovered most of the valuable functions we can abuse for this attack by spending hours reviewing published headers for ntdll. For example, two great resources I used were the Process Hacker Native API header collection and the source code of ReactOS. What I did with these headers was use the return type and SAL annotations, which specify whether arguments are written to or are used as input, to find potentially "desirable" functions.
For example, if a function's return type was NTSTATUS
, a zero return value was guaranteed as long as the function succeeded. SAL annotations let me search for functions that met specific criteria like "find functions where the second argument I control is written to".
It's not worth going through every function I found potentially valuable, as there were quite a lot. So instead, I'll focus on those we leverage in our minimal PoC.
RtlInitUnicodeStringEx
NTSTATUS
NTAPI
RtlInitUnicodeStringEx(
_Out_ PUNICODE_STRING DestinationString,
_In_opt_z_ PCWSTR SourceString
);
The first function I want to call out is RtlInitUnicodeStringEx
, which takes our completely controlled second argument and initializes a UNICODE_STRING
structure in the buffer specified by the first argument.
Remember how I said the fact that RCX
was a writable location is still helpful? In the context of the __C_specific_handler
, the first argument is an EXCEPTION_POINTERS
structure which contains our exception and context record pointers. Fortunately, it doesn't matter what is stored initially, as RtlInitUnicodeStringEx
doesn't care.
I couldn't use RtlInitUnicodeString
because it leverages the RAX
register (return value) as a "counter" representing the number of characters written, and we needed a return value of zero. RtlInitUnicodeStringEx
, on the other hand, wraps this call and returns zero as long as our source string is not larger than SHRT_MAX
.
This function lets us initialize RCX
to point to a UNICODE_STRING
containing whatever value we want. Many functions inside ntdll accept a UNICODE_STRING
pointer; hence this was useful for expanding the accessible attack surface.
LdrLoadDll
What do we do with a UNICODE_STRING
? One interesting attack idea I had was, "if all we want to do is execute arbitrary code remotely, does it really need to be shellcode?". For example, what's stopping us from loading a malicious DLL from our remote server?
NTSTATUS
NTAPI
LdrpLoadDll(
PUNICODE_STRING DllName,
PVOID DllPathState,
ULONG Flags,
PLDR_MODULE* ModuleOut
);
LdrLoadDll
wouldn't work directly as the DllPath
needed to be a PWSTR
, but all LdrLoadDll
did was wrap a call to LdrpLoadDll
, which did accept a UNICODE_STRING
as its first argument.
Although this initially seemed like a good candidate, I ran into several issues during testing. So, to make my life easier, I created a test program that would emulate the conditions inside __C_specific_handler
. For example, I called RtlInitUnicodeStringEx
on a heap buffer containing random data, and I passed two arbitrary heap pointers in the Flags
and ModuleOut
arguments.
Flags
being a pointer complicated things as different heap addresses would trigger different logic inside LdrpLoadDll
. I had similar issues with the DllPathState
containing a wide string.
Looking for alternatives, I used IDA Pro to check for cross-references to LdrpLoadDll
. To my surprise, I found a function named LdrpLoadWow64
which took a single UNICODE_STRING
argument. At a high level, the function:
- Copies the
UNICODE_STRING
argument into a stack buffer. - Appends
wow64.dll
to this stack buffer. - Uses
LdrpInitializeDllPath
on the stack buffer, which "normalizes" a given path forLdrpLoadDll
. - Calls
LdrpLoadDll
to load the DLL from the finalized path.
There are a few more operations this function does after the DLL is loaded, such as trying to parse certain exports, but that doesn't matter if we can get our arbitrary DLL loaded.
This appeared to be a great candidate because it took care of all the strange arguments LdrpLoadDll
took, which we had little control over. If we could initialize RCX
with a path to a malicious directory, LdrpLoadWow64
would append wow64.dll
to it and load a DLL from that path!
Preparing the Demo
For the demo, I developed a sample "vulnerable application" compiled using Visual Studio's Clang/LLVM build tools. This example program has a small network protocol designed to fulfill our requirements for the attack:
- The application allows a remote caller to request a leak of the
ntdll
base address and a pointer to a heap region allocated with a caller-specified size. - The application allows a remote caller to provide an arbitrary buffer to copy into the previously allocated heap buffer.
- The application allows a remote caller to provide an arbitrary buffer to unsafely copy into a stack variable and then trigger an access violation exception.
Obviously, the requirements for this attack will be more complex in the real world. The method by which you meet the requirements will change depending on the context of the application you are exploiting. Therefore, in our sample PoC, these primitives are accessible through a simple interface, allowing us to focus on the unwinding process.
To perform the attack, I created a small Python script requiring a target IP and several offsets of functions inside your target's ntdll. Note that this exploit code was written to work against any target application, not just the one we developed for this demo. Of course, you would need to implement code to fulfill the requirements for the attack, but the nice part of the LdrpLoadWow64
methodology is that it is stable across different versions of Windows.
To up the stakes, I configured my target's "Exploit Protection" settings to require strict control flow guard and strict hardware-enforced stack protection. Given that our exploit never needs to return to a value that isn't on the shadow stack, this shouldn't cause issues, but it's always better to confirm these assumptions.
Demo
We start the demo by running the vulnerable application on our victim machine. To show you the attack step-by-step, I'm using IDA Pro remote debugging with WinDbg. After setting relevant breakpoints, we can start the exploit script on our remote attacker machine.
The PoC requests a leak for ntdll
and some heap memory, copies the malicious heap buffer to the target, and triggers the overflow.
On our victim side, we can see an exception occurring inside the ProcessPacket
function after the overflow buffer is overwritten as intended.
At this point, we can verify that hardware-enforced stack protection is enabled by running the WinDbg command rM 8002
. This is a trick I learned from the Google Security blog discussing CET to see the shadow stack pointer (ssp
) and whether CET is enabled (cetumsr
). Since cetumsr
is 1
, we know that our previous exploit protection option took effect.
Once we pass this exception to the target, our corrupted call stack leads RtlDispatchException
to call RtlpUnwindHandler
, which is the exception handler for RtlpExecuteHandlerForUnwind
. This is where the dispatcher context structure in RtlDispatchException
is overwritten with our malicious dispatcher context pointer stored on the stack.
Next, __C_specific_handler
calls RtlInitUnicodeStringEx
, which initializes the ExceptionPointers
variable to a UNICODE_STRING
pointing at the wide string we specified in our EstablisherFrame
.
The last step of the attack is a call to LdrpLoadWow64
. By dumping the unicode string in RCX
right before the call to LdrpLoadDll
, we can see that it has been initialized to our initial directory path + wow64.dll
.
Once we continue, we see the wow64.dll
file being requested from our attacker server and a message box generated by our malicious DLL, completing the attack!
What Else Can We Do With This Attack?
Although we've gone through a minimal proof-of-concept, much more is possible with the CollidedUnwind
primitive. Here are some highlights:
- Using
__C_specific_handler
, we can trigger a call toRtlUnwind
while controlling the function's internal state.RtlUnwind
is more potent thanRtlDispatchException
as the changes you make to the context record will be reflected once execution continues at the unwind target. - In
RtlDispatchException
, a significant barrier to continuing the virtual unwind loop is that yourRsp
must be within the stack's bounds. There are quite a few ways to get around this issue.- One method was to use unwind operations to read the
TlsExpansionBitmap
in ntdll, which contained a pointer to the PEB. From here, we can go up to the TEB of our process and copy theStackBase
orStackLimit
fields intoRsp
, allowing us to pass the bounds check.
- One method was to use unwind operations to read the
- Another challenge I faced with complex variants was the unwind info structure in
RUNTIME_FUNCTION
is located by adding a 32-bit address to the image base we controlled. For us to call arbitrary exception handlers, the image base would need to be the location of ntdll. For us to entirely control the unwind info structure, the image base would need to be the base of our heap.- An interesting middle ground was that I could use any unwind info structure in ntdll to call any exception handler. This worked by subtracting the exception handler offset I didn't control from the target exception handler I wanted to call. This new value would be my image base, and I would specify an unwind info offset in my runtime function to account for this difference (i.e. real unwind info location minus my fake image base).
What About Linux?
We discussed how exception handling works on Windows and how we can weaponize structured exception handling to gain code execution. However, we didn't touch on how other operating systems like Linux are impacted by the same theory.
A few weeks ago, academic researchers from Vrije Universiteit Amsterdam and the University of California, Santa Barbara, released a paper titled, "Let Me Unwind That For You: Exceptions to Backward-Edge Protection". Their paper complements this article exceptionally well as they investigated how an attacker can abuse the unwinding process, but with a strong Linux and C++ focus.
If you enjoyed this article, I strongly encourage you to check out their work.
Tools
To access the tools and proof-of-concept mentioned in this article, please visit the Exception Oriented Programming repository.
Conclusion
In this article, we explored how an attacker can abuse fundamental weaknesses in the design of structured exception handling on Windows in the context of stack-based attacks. We started with the trivial approach to bypassing security cookies by triggering an exception into a handler that would return to our target. We investigated how although the MSVC compiler has mitigated this attack for fifteen years, much of the Windows ecosystem is still unprotected.
Next, we dove deep into the internals of the unwinding process on Windows, discovering how the unwind operations of legitimate functions can be weaponized by attackers to corrupt the state of an application. Finally, we discussed how Exception Oriented Programming is a compelling alternative to ROP when it comes to newer system mitigations such as hardware-enforced stack protection.
In our last section, we abused edge cases to achieve the modern-day equivalent of an SEH hijacking attack. We weaponized collided unwinds to gain code execution with strict hardware-enforced stack protection and control flow guard enabled.
I look forward to seeing if we can implement mitigations against these attacks in other compilers (and operating systems). There is quite a lot of opportunity to limit the attack surface currently exposed by the unwinding process, and I look forward to working with the broader compiler development community to see what can be done.
I hope you enjoyed reading this article as much as I enjoyed writing it. We covered a significant amount, and I appreciate those that stuck around.