Bochspwn Reloaded

Detecting Kernel Memory Disclosure with x86
Emulation and Taint Tracking

Mateusz Jurczyk
[email protected]
Google LLC
June 2018
Abstract
One of the responsibilities of modern operating systems is to enforce
privilege separation between user-mode applications and the kernel. This
includes ensuring that the influence of each program on the execution en-
vironment is limited by the defined security policy, but also that programs
may only access the information they are authorized to read. The latter
goal is especially difficult to achieve considering that the properties of C
– the main programming language used in kernel development – make
it highly challenging to securely pass data between different security do-
mains. There is a significant risk of disclosing sensitive leftover kernel
data hidden amidst the output of otherwise harmless system calls, unless
special care is taken to prevent the problem. Issues of this kind can help
bypass security mitigations such as KASLR and StackGuard, or retrieve
information processed by the kernel on behalf of the system or other users,
e.g. file contents, network traffic, cryptographic keys and so on.
In this paper, we introduce the concept of employing full system em-
ulation and taint tracking to detect the disclosure of uninitialized kernel
stack and heap/pool memory to user-space, and discuss how it was suc-
cessfully implemented in the Bochspwn Reloaded project based on the
open-source Bochs IA-32 emulator. To date, the tool has been used to
identify over 70 memory disclosure vulnerabilities in the Windows kernel,
and more than 10 lesser bugs in Linux. Further in the document, we
evaluate alternative ways of detecting such information leaks, and outline
data sinks other than user-space where uninitialized memory may also
leak from the kernel. Finally, we provide suggestions on related research
areas that haven’t been fully explored yet. Appendix A details several
further ideas for system-wide instrumentation (implemented using Bochs
or otherwise), which can be used to discover other programming errors in
OS kernels.
1
Contents
1 Introduction 4
2 Memory disclosure in operating systems 6

2.1 C language-specific properties . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Indeterminate state of uninitialized variables . . . . . . . 6
2.1.2 Structure alignment and padding bytes . . . . . . . . . . 8
2.1.3 Unions and diversely sized fields . . . . . . . . . . . . . . 10
2.1.4 sizeof considered harmful . . . . . . . . . . . . . . . . . 12
2.2 System-specific properties . . . . . . . . . . . . . . . . . . . . . . 13
2.2.1 Memory reuse in dynamic allocators . . . . . . . . . . . . 13
2.2.2 Fixed-sized arrays . . . . . . . . . . . . . . . . . . . . . . 14
2.2.3 Arbitrary syscall output buffer lengths . . . . . . . . . . . 15
2.3 Further contributing factors . . . . . . . . . . . . . . . . . . . . . 17
2.4 Severity and impact on system security . . . . . . . . . . . . . . 18
2.5 Prior research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.1 Microsoft Windows . . . . . . . . . . . . . . . . . . . . . . 19
2.5.2 Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Bochspwn Reloaded – detection with software x86 emulation 23

3.1 Core logic – kernel memory taint tracking . . . . . . . . . . . . . 24
3.1.1 Shadow memory representation . . . . . . . . . . . . . . . 25
3.1.2 Tainting stack frames . . . . . . . . . . . . . . . . . . . . 27
3.1.3 Tainting heap/pool allocations . . . . . . . . . . . . . . . 30
3.1.4 Clearing taint . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.5 Taint propagation . . . . . . . . . . . . . . . . . . . . . . 32
3.1.6 Bug detection . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Ancillary functionality . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Keeping track of kernel modules . . . . . . . . . . . . . . 41
3.2.2 Unwinding stack traces . . . . . . . . . . . . . . . . . . . 43
3.2.3 Address symbolization . . . . . . . . . . . . . . . . . . . . 44
3.2.4 Breaking into the kernel debugger . . . . . . . . . . . . . 45
3.2.5 Address space visualization . . . . . . . . . . . . . . . . . 46
3.3 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.1 CPU overhead . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.2 Memory overhead . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4.2 Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.5.2 Linux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4 Windows bug reproduction techniques 59
2
5 Alternative detection methods 62
5.1 Static analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 Manual code review . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3 Cross-version kernel binary diffing . . . . . . . . . . . . . . . . . 65
5.4 Differential syscall fuzzing . . . . . . . . . . . . . . . . . . . . . . 67
5.5 Taintless Bochspwn-style instrumentation . . . . . . . . . . . . . 71
6 Other data sinks 72

6.1 Mass-storage devices . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.1.1 Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.1.2 Testing and results . . . . . . . . . . . . . . . . . . . . . . 74
6.2 Outbound network traffic . . . . . . . . . . . . . . . . . . . . . . 75
7 Future work 76
8 Conclusion 78
9 Acknowledgments 78
A Other system instrumentation schemes 88

A.1 Enumeration of kernel attack surface . . . . . . . . . . . . . . . . 88
A.2 Deeply nested user-mode memory accesses . . . . . . . . . . . . . 88
A.3 Unprotected accesses to user-mode memory . . . . . . . . . . . . 90
A.4 Broad exception handlers . . . . . . . . . . . . . . . . . . . . . . 92
A.5 Dereferences of unsanitized user-mode pointers . . . . . . . . . . 93
A.6 Read-after-write conditions . . . . . . . . . . . . . . . . . . . . . 95
A.7 Double writes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
A.8 Ignored function return values . . . . . . . . . . . . . . . . . . . . 97
A.9 API misuse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3
1 Introduction
Modern operating systems running on x86/x86-64 CPUs are multi-threaded
and use a client—server model, where user-mode applications (clients) execute
independently of each other, and only call into the kernel (server ) once they
intend to operate on a resource managed by the system. The mechanism used
by code running in ring 3 to call into a predefined set of ring 0 functions is
known as system calls or syscalls in short. While CPU registers are used to
return the exit code and to pass the first few arguments on 64-bit platforms,
the primary I/O data exchange channel is user-mode memory. Consequently,
the kernel continuously operates on ring 3 memory during its activity. The life
of a typical system call is illustrated in Figure 1.
Shared Memory
User-mode Program System Kernel
(user-mode)
Write input data
Invoke system call
Read input data
Syscall logic
Write output data
Return to user space
Read output data
Figure 1: Life of a system call
The execution flow is simple in principle, but the fact that user-mode mem-
ory is a shared resource that can be read or written to at any point in time by
another thread makes the kernel’s interactions with the memory prone to race
conditions and other errors, if not implemented carefully. One way to achieve
a secure implementation is to have every system call handler adhere to the
following two rules (in that order):
1. Fetch data from each user-mode memory location at most once, with an
active exception handler in place, and save a local copy in the kernel space
for further processing.
2. Write to each user-mode memory location at most once, with an active
exception handler in place, only with data intended for the user-mode
client.
4
The usage of pointer annotations (such as user in Linux) and dedicated
functions for operating on client memory (e.g. copy from user, copy to user,
get user, put user in Linux) helps maintain a healthy user-mode interaction
hygiene, by forcing developers to make conscious decisions about when to per-
form these operations. On the other hand, the looser approach observed in the
Windows kernel (direct pointer manipulation) seems to provoke the presence of
more security issues, as demonstrated later in the paper.
Breaking each of the above requirements has its own unique consequences,
with varying degree of impact on system security:
• Double fetches – caused by reading a user-mode value more than once

within the same semantic context, and expecting it to remain consistent
between the reads. This introduces time of check to time of use race condi-
tions which may lead to buffer overflows, write-what-where conditions and
other memory safety violations, depending on the affected code. The work
discussed in this paper is largely based on the original Bochspwn project,
developed in 2013 to detect double fetch vulnerabilities [68, 69, 70].
• Read-after-write conditions – caused by writing to a user-mode mem-
ory address and subsequently reading from it with the assumption that
the written value has remained unchanged. These race conditions have
security impact similar to that of double fetches.
• Double writes – caused by writing sensitive information to user-mode
memory, and later overwriting it with legitimate data in the scope of
the same system call. They result in information disclosure of limited
duration, most commonly involving kernel-space addresses.
• Unprotected memory operations – caused by the lack of exception

handlers set up around user-mode memory accesses. They may be ex-
ploited to trigger unhandled kernel exceptions and consequently crash the
operating system.
The above bug classes are addressed in more detail in Appendix A. The
primary subject of this paper is the breaking of the following rule:
Write to each user-mode memory location at most once, [...], only

with data intended for the user-mode client.
Due to a combination of circumstances such as the nature of the C pro-

gramming language, current state of compilers, design of system allocators and
common code optimization techniques, it is difficult to avoid unintentionally
leaking kernel memory while interacting with user-mode programs. In the next
section, we discuss the origins of the problem, its impact on system security and
prior work done to address it.
5
2 Memory disclosure in operating systems
A disclosure of privileged system memory occurs when the kernel returns a larger
region of data than is necessary to store the relevant information contained
within. Frequently, the redundant bytes originate from a kernel memory region
which used to store data previously processed in another context, and are not
pre-initialized to ensure that the old values are not propagated to new data
structures.
In some cases, fault can be clearly attributed to insufficient initialization
of certain variables or structure fields in the code. At other times, information
leaks occur in the kernel even though the corresponding source code is seemingly
correct (sometimes only on specific CPU architectures). In either case, memory
disclosure between different privilege levels running C code is a problem hardly
visible to the naked eye.
2.1 C language-specific properties

In this section, we outline several aspects of the C language that are most
relevant to the problem in question.
2.1.1 Indeterminate state of uninitialized variables

Standalone variables of simple types such as char or int, as well as members of
larger data structures (arrays, structures and unions) remain in an indetermi-
nate state until first initialized, if they are stored on the stack (formally under
automatic storage duration) or heap (allocated storage duration). The relevant
citations from the C11 N1570 specification draft [14] are as follows (emphasis
added by author):
6.7.9 Initialization
...
10 If an object that has automatic storage duration is not initialized
explicitly, its value is indeterminate.
7.22.3.4 The malloc function

...
2 The malloc function allocates space for an object whose size is
specified by size and whose value is indeterminate.
7.22.3.5 The realloc function

...
2 The realloc function deallocates the old object pointed to by ptr
and returns a pointer to a new object that has the size specified by
size. The contents of the new object shall be the same as that of the
old object prior to deallocation, up to the lesser of the new and old
sizes. Any bytes in the new object beyond the size of the old object
have indeterminate values.
6
The part most applicable to system code is the one concerning stack-allocated
objects, as kernels typically have dynamic allocation interfaces with their own
semantics (not necessarily consistent with the C standard library, as discussed
in Section 2.2.1 “Memory reuse in dynamic allocators”).
To our best knowledge, none of the three most popular C compilers for
Windows and Linux (Microsoft C/C++ Compiler, gcc, LLVM) produce code
which pre-initializes otherwise uninitialized objects on the stack in release mode
(or equivalent). There are compiler switches enabling the poisoning of stack
frames with marker bytes (e.g. /RTCs in Microsoft Visual Studio), but they are
not used in production builds for performance reasons. As a result, uninitialized
objects on the stack inherit old values of their corresponding memory areas.
Let’s consider an example of a fictional Windows system call implementation,
which multiplies the input integer by two and returns the product (Listing 1).
It is evident that in the corner case of InputValue=0, the OutputValue variable
remains uninitialized and is copied in that form back to the client. Such a bug
would enable the disclosure of four bytes of kernel stack memory on each syscall
invocation.
1 NTSTATUS NTAPI NtMultiplyByTwo(DWORD InputValue, LPDWORD OutputPointer) {

2 DWORD OutputValue;
3
4 if (InputValue != 0) {
5 OutputValue = InputValue * 2;
6 }
7
8 *OutputPointer = OutputValue;
9 return STATUS_SUCCESS;
10 }
Listing 1: Memory disclosure via a local uninitialized variable
While possible, information leaks through standalone variables are not very
common in practice, as modern compilers will often detect and warn about
such problems, and being functional bugs, they may also be identified during
development or testing. However, a second example in Listing 2 shows that a
leak may also take place through a structure field.
In this case, the Reserved field is never explicitly used in the code, but is still
copied back to user-mode, and thus also discloses four bytes of kernel memory
to the caller. This example distinctly shows that having every field of every
structure returned to the client initialized on every code path is a difficult task,
and in many cases it is plainly counterintuitive to do so, if the field in question
does not play any practical role in the code.
Overall, the fact that uninitialized variables on the stack and in dynamic
allocations take the contents of data previously stored in their memory regions
lies at the core of the kernel memory disclosure problem.
7
1 typedef struct _SYSCALL_OUTPUT {
2 DWORD Sum;
3 DWORD Product;
4 DWORD Reserved;
5 } SYSCALL_OUTPUT, *PSYSCALL_OUTPUT;
6
7 NTSTATUS NTAPI NtArithOperations(
8 DWORD InputValue,
9 PSYSCALL_OUTPUT OutputPointer
10 ) {
11 SYSCALL_OUTPUT OutputStruct;
12
13 OutputStruct.Sum = InputValue + 2;
14 OutputStruct.Product = InputValue * 2;
15
16 RtlCopyMemory(OutputPointer, &OutputStruct, sizeof(SYSCALL_OUTPUT));
17
19 }
Listing 2: Memory disclosure via a reserved structure field
2.1.2 Structure alignment and padding bytes

The initialization of all members of an output structure is a good start to avoid
memory disclosure, but it is not always sufficient to guarantee that uninitialized
bytes will not appear in its low-level representation. Let’s consider the following
excerpts from the C11 specification draft:
6.5.3.4 The sizeof and Alignof operators

...
4 [...] When applied to an operand that has structure or union type,
the result is the total number of bytes in such an object, including
internal and trailing padding.
6.2.8 Alignment of objects

1 Complete object types have alignment requirements which place
restrictions on the addresses at which objects of that type may be
allocated. An alignment is an implementation-defined integer value
representing the number of bytes between successive addresses at
which a given object can be allocated. [...]
6.7.2.1 Structure and union specifiers

...
17 There may be unnamed padding at the end of a structure or
union.
8
Practically speaking, C compilers for the x86(-64) architectures apply natu-
ral alignment to structure fields of primitive types, which means that each such
field is N-byte aligned, where N is the given field’s width. Furthermore, entire
structures and unions are also aligned such that when they are declared in an
array, the alignment requirements of the nested fields are still met. In order to
accommodate the alignment, padding bytes are artificially inserted into struc-
tures where necessary1 . While not directly accessible in the source code, these
bytes also inherit old values from the underlying memory regions and may leak
information to user-mode, if they are not reset in time.
In the example shown in Listing 3, a SYSCALL OUTPUT structure is passed
back to the caller. It contains a 4-byte and an 8-byte field, separated by 4 bytes
of padding required to align LargeSum to an 8-byte boundary. Even though
both fields are properly initialized, the padding bytes are not explicitly set,
which again leads to a kernel stack memory disclosure. The specific layout of
the structure in memory is illustrated in Figure 2.
1 typedef struct _SYSCALL_OUTPUT {

2 DWORD Sum;
3 QWORD LargeSum;
5
6 NTSTATUS NTAPI NtSmallSum(
7 DWORD InputValue,
9 ) {
10 SYSCALL_OUTPUT OutputStruct;
11
12 OutputStruct.Sum = InputValue + 2;
13 OutputStruct.LargeSum = 0;
14
15 RtlCopyMemory(OutputPointer, &OutputStruct, sizeof(SYSCALL_OUTPUT));
16
18 }
Listing 3: Memory disclosure via structure padding
Leaks through alignment holes are relatively common, as a variety of complex

structures are used in user↔kernel communication in modern systems. This is
especially true for 64-bit platforms, where the width of pointers, size t and
similar types increases from 4 to 8 bytes, thus also increasing the number of
padding bytes necessary to align such structure fields.
As padding bytes cannot be directly addressed in the source code, it is
intuitive to use the memset or analogous function to reset the entire memory
1 Typically, a maximum of 7 padding bytes can be observed between two consecutive fields,
if they are of type char and long long (or equivalent), respectively. In some rare and extreme
cases, the padding can be as long as 15, 31 or even 63 bytes, for such esoteric types as 80-bit
long double, m256 or m512.
9
Sum LargeSum
3B 05 00 00 00 00 00 00 00 00 00 00
0x0 0x1 0x2 0x3 0x4 0x5 0x6 0x7 0x8 0x9 0xA 0xB 0xC 0xD 0xE 0xF
Figure 2: Layout of a structure with padding bytes
region prior to initializing any of its fields and copying it to user-mode, like so:
memset(&OutputStruct, 0, sizeof(OutputStruct));
However, Seacord R. C. argues [100] that it is a noncompliant solution,

because the padding bytes may still be clobbered after the memset call, e.g. as
a side effect of operating on the adjacent fields. The concern may be justified
by the following statement in the C specification draft:
6.2.6 Representations of types
6.2.6.1 General
...
6 When a value is stored in an object of structure or union type,
including in a member object, the bytes of the object representation
that correspond to any padding bytes take unspecified values. [...]
In practice, none of the C compilers we have tested read or write beyond the
memory regions of the explicitly declared fields. This sentiment also seems to
be shared by operating system vendors and kernel developers, who use memset
as their preferred fix of choice.
The aforementioned publication lists the following compliant ways of ad-
dressing the problem:
• Serializing the structure contents into a continuous, dense buffer of bytes,
and unserializing it back on the other side.
• Explicitly declaring the padding areas as dummy structure fields, and
manually zero-initializing them in the code.
• Enabling structure packing to completely disable memory alignment, thus
eliminating the problematic padding bytes.
Evaluating the pros and cons of each of the above solutions is left as an
exercise to the reader.
2.1.3 Unions and diversely sized fields

Unions are another intricate construct in the context of cross-privilege level
communication in the C language. The relevant parts of the C11 specification
draft explaining how unions are represented in memory are quoted on the next
page.
10
6.2.5 Types
...
20 Any number of derived types can be constructed from the ob-
ject and function types, as follows: [...] A union type describes an
overlapping nonempty set of member objects, each of which has an
optionally specified name and possibly distinct type.
6.7.2.1 Structure and union specifiers

...
6 As discussed in 6.2.5, a structure is a type consisting of a sequence
of members, whose storage is allocated in an ordered sequence, and
a union is a type consisting of a sequence of members whose storage
overlap.
...
16 The size of a union is sufficient to contain the largest of its mem-
bers. The value of at most one of the members can be stored in a
union object at any time.
The problem here is that if a union consists of several fields of various widths,
and only one of the smaller fields is set, then the remaining bytes allocated to
accommodate the larger ones are left uninitialized. Let’s examine an example
of an imaginary system call handler presented in Listing 4, together with the
memory layout of the SYSCALL OUTPUT union illustrated in Figure 3.
1 typedef union _SYSCALL_OUTPUT {

2 DWORD Sum;
3 QWORD LargeSum;
5
6 NTSTATUS NTAPI NtSmallSum(
7 DWORD InputValue,
9 ) {
10 SYSCALL_OUTPUT OutputUnion;
11
12 OutputUnion.Sum = InputValue + 2;
13
14 RtlCopyMemory(OutputPointer, &OutputUnion, sizeof(SYSCALL_OUTPUT));
15
17 }
Listing 4: Memory disclosure via a partially initialized union
As can be seen, the total size of the SYSCALL OUTPUT union is 8 bytes, due to
the width of the larger LargeSum field. However, the function only sets the value
of the smaller field, leaving the 4 trailing bytes uninitialized and subsequently
disclosed to the client application.
11
LargeSum
Sum
3B 05 00 00
0x0 0x1 0x2 0x3 0x4 0x5 0x6 0x7
Figure 3: Memory layout of a partially initialized union
A safe implementation should only set the Sum field in the user address space,
instead of copying the entire object with potentially unused regions of memory.
Another frequently seen fix is the usage of the memset function to reset the
kernel’s copy of the union prior to setting any of its fields and passing it back
to user-mode.
2.1.4 sizeof considered harmful

As shown in the two previous subsections, the usage of the sizeof operator
may (in)directly contribute to the disclosure of kernel memory, by provoking the
copying of more data than the kernel had previously initialized. This manifests a
broader problem of having to weight code cleanliness and brevity against system
security, often with the kernel developer only having a blurry understanding of
the outcome of each possible decision.
To put it in other words, the C language lacks the apparatus necessary to
securely move data from the kernel to userland – or more generally, between
any differently privileged security contexts. It provides no runtime metadata
that can help establish which bytes have been set in each data structure used
for user↔kernel communication. As a result, the burden is on the programmer
to determine which portions of each object contain relevant data and should be
passed to the caller. If done properly, this would require introducing a dedicated
safe-copy function for every output data structure used by the kernel, which
in turn would bloat the code size, likely degrade its readability, and overall be
a tedious and time-consuming task.
On the other hand, it is convenient to avoid the hassle and simply copy an
entire kernel memory region with a single memcpy call and a sizeof argument,
and let the client determine which parts of the output buffer it will actually
use. This seems to be the state of Windows and Linux today. When a partic-
ular instance of an information leak is reported, a patch with a memset call is
promptly submitted and distributed, but the deeper problem remains.
This is why we believe that even with tools for automated detection of kernel
memory disclosure (such as Bochspwn Reloaded), the bug class may persist for
many years, if systemic solutions are not implemented in the language specifica-
tion, compilers and operating systems to address the problem on a more general
level.
12
2.2 System-specific properties
There are certain kernel design decisions, programming practices and code pat-
terns which affect how prone operating systems are to memory disclosure vul-
nerabilities. They are briefly discussed in the following subsections.
2.2.1 Memory reuse in dynamic allocators

Current dynamic memory allocators (in both user and kernel-mode) are highly
optimized, as their performance has a considerable impact on the performance of
the entire operating system. One of the most important optimizations is memory
reuse – when a heap/pool allocation is freed, the corresponding memory is rarely
completely discarded, but is instead saved in a list of regions ready to be returned
when the next allocation request is received. To save CPU cycles, the memory
regions are not cleared by default between being used by two different callers.
As a result, it is not uncommon for two completely unrelated parts of the kernel
to subsequently operate on the same address range within a short time-frame.
It also means that leaks of dynamically allocated ring 0 memory may disclose
data processed by various components in the operating system. A more detailed
analysis of the nature of data found in different memory regions can be found
in Section 2.4 “Severity and impact on system security”.
In the paragraphs below, we provide a brief overview of the allocators used
in the Windows and Linux kernels and their most notable qualities.
Windows The core of the Windows kernel pool allocator is the ExAllocate-
PoolWithTag function [82], which may be used directly, or via several available
wrappers: ExAllocatePool{∅, Ex, WithQuotaTag, WithTagPriority}. None
of these routines reset the returned memory regions, neither by default, nor
through any input flags. On the contrary, all of them have the following warning
in their corresponding MSDN documentation entries:
Note Memory that function allocates is uninitialized. A kernel-

mode driver must first zero this memory if it is going to make it
visible to user-mode software (to avoid leaking potentially privileged
contents).
There are six primary pool types the callers can choose from: NonPagedPool,
NonPagedPoolNx, NonPagedPoolSession, NonPagedPoolSessionNx, PagedPool
and PagedPoolSession. Each of them has a designated region in the virtual
address space, and so memory allocations may only be reused within the scope
of their specific pool type. The reuse rate of memory chunks is very high, and
zeroed out areas are generally only returned when no suitable entry is found in
the lookaside lists, or the request is so large that fresh memory pages need to be
mapped to facilitate it. In other words, there are currently hardly any factors
preventing pool memory disclosure in Windows, and almost every such bug can
be exploited to leak sensitive data from various parts of the kernel.
13
Linux The Linux kernel offers three main interfaces for dynamically allocating
memory:
• kmalloc – a generic function used to allocate arbitrarily sized memory

chunks (continuous in both virtual and physical address space), backed by
the slab allocator.
• kmem cache create and kmem cache alloc – a specialized mechanism for
allocating objects of a fixed size (e.g. structures), also backed by the slab
allocator.
• vmalloc – a rarely used allocator returning regions that are not guaranteed
to be continuous in physical memory.
By themselves, these functions provide no assurance that the allocated re-

gions won’t contain old, potentially sensitive data, making kernel heap memory
disclosure feasible. However, there is a number of ways callers can request
zeroed-out memory:
• The kmalloc function has a kzalloc counterpart, which makes sure that
the returned memory is cleared.
• An optional GFP ZERO flag can be passed to kmalloc, kmem cache alloc
and several other functions to achieve the same result.
• The kmem cache create routine accepts a pointer to an optional construc-
tor function, called to pre-initialize each object before returning it to the
requestor. The constructor may be implemented as a wrapper around
memset to reset the given memory area.
We see the existence of these options as an advantage for kernel security,

as they encourage developers to make conscious decisions, and enable them to
express their intent of operating on a clean allocation at the time of requesting
memory, instead of having to add extra memset calls in every such instance.
2.2.2 Fixed-sized arrays

A number of resources in operating systems can be accessed through their tex-
tual names. The variety of named resources is notably vast in Windows, in-
cluding files and directories, registry keys and values, windows, fonts, atoms,
and so forth. For some of them, the name length is limited by design, and is
expressed by a constant such as MAX PATH (260) or LF FACESIZE (32). In such
cases, kernel developers often simplify the code by declaring buffers of the max-
imum allowed size and copying them around as a whole (e.g. using the sizeof
keyword) instead of operating only on the relevant part of the string. This is
especially convenient if the strings are members of larger structures – such ob-
jects can be freely moved around in memory without worrying about the pointer
management of any dependent allocations.
14
As can be expected, large buffers are rarely used to their full capacity, and
the remaining storage space is often not reset. This may lead to particularly
severe leaks of long, continuous areas of kernel memory. In the fictional example
shown in Listing 5, a system call uses a RtlGetSystemPath function to load the
system path into a local buffer, and if the call succeeds, all of the 260 bytes are
passed to the caller, regardless of the actual length of the string.
1 NTSTATUS NTAPI NtGetSystemPath(PCHAR OutputPath) {

2 CHAR SystemPath[MAX_PATH];
3 NTSTATUS Status;
4
5 Status = RtlGetSystemPath(SystemPath, sizeof(SystemPath));
6 if (NT_SUCCESS(Status)) {
7 RtlCopyMemory(OutputPath, SystemPath, sizeof(SystemPath));
8 }
9
10 return Status;
11 }
Listing 5: Memory disclosure via a partially initialized string buffer

The memory region copied back to user-mode in this example is illustrated
in Figure 4.
C:\\Windows\\System32\0
Figure 4: Memory layout of a partially initialized string buffer
A secure implementation should only return the requested path, with no

extra padding bytes. Once again, this example demonstrates how the evaluation
of the sizeof operator (used as a parameter to RtlCopyMemory) may be utterly
inconsistent in relation to the actual amount of data the kernel is supposed to
pass to userland.
2.2.3 Arbitrary syscall output buffer lengths

Most system calls accept pointers to user-mode output buffers together with
the size of the buffers. In most cases, information about the size should only be
used to determine if it is large enough to receive the syscall’s output data, but
should not otherwise influence how much memory is copied. However, we have
observed cases where the kernel would try to fill every byte of the user’s output
buffer, disregarding the amount of actual data there is to copy. An example of
such behavior is shown in Listing 6.
15
1 NTSTATUS NTAPI NtMagicValues(LPDWORD OutputPointer, DWORD OutputLength) {
2 if (OutputLength < 3 * sizeof(DWORD)) {
3 return STATUS_BUFFER_TOO_SMALL;
4 }
5
6 LPDWORD KernelBuffer = Allocate(OutputLength);
7
8 KernelBuffer[0] = 0xdeadbeef;
9 KernelBuffer[1] = 0xbadc0ffe;
10 KernelBuffer[2] = 0xcafed00d;
11
12 RtlCopyMemory(OutputPointer, KernelBuffer, OutputLength);
13 Free(KernelBuffer);
14
16 }
Listing 6: Memory disclosure via an arbitrary buffer output size

The purpose of the fictional system call is to provide the caller with three
special 32-bit values, occupying a total of 12 bytes. While the buffer size sanity
check in lines 2-4 is correct, the usage of the OutputLength argument should
end at that. Knowing that the output buffer is large enough to store the result,
the kernel could allocate a 12-byte buffer, fill it out accordingly and copy its
contents to the user-mode location. Instead, the syscall requests an allocation
with a user-controlled length and copies it back as a whole, even though there
is only a fixed number of bytes that the caller needs. The remaining bytes are
uninitialized and mistakenly disclosed to userland, as presented in Figure 5.
KernelBuffer[0] [1] [2] [3] [4] […]
EF BE AD DE FE 0F DC BA 0D D0 FE CA
Figure 5: Memory layout of a memory allocation with a user-controlled size
The scheme discussed in this section is particularly characteristic to Win-

dows. When implemented incorrectly, it may provide an attacker with an ex-
tremely useful memory disclosure primitive, for two reasons:
• An optimization frequently employed in Windows system calls is to use

stack-based buffers for small user-defined buffer sizes, and pool allocations
for larger ones. In combination with an infoleak bug, this could facilitate
the disclosure of both kernel stack and pool memory through a single
vulnerability.
• Being able to leak data from pool allocations of controlled lengths gives the
attacker significantly more options as for the types of sensitive information
they can acquire. As modern allocators tend to cache memory regions for
16
subsequent requests of the same length, it becomes possible to have the
leaked allocations overlap with ones that had contained a specific kind of
confidential data in the past.
As such, this is one of the most dangerous types of memory disclosure. It can
be addressed by keeping track of the number of bytes written to the temporary
kernel buffer, and passing only this amount of data back to the client.
2.3 Further contributing factors

In the previous section, we described some of the reasons why introducing infor-
mation disclosure bugs to kernels is trivial, and how it can happen involuntarily.
Here, we show the challenges behind identifying these bugs later in the devel-
opment, which contributed to the lack of recognition of the problem for many
years, with dozens of kernel infoleaks piling up in the Windows kernel.
Lack of visible consequences As a general rule, information disclosure flaws

are more difficult to detect than memory corruption ones. The latter usually
visibly manifest themselves in the form of crashes or other software misbehavior,
especially when coupled with mechanisms such as AddressSanitizer, PageHeap
or Special Pool. On the contrary, information disclosure doesn’t cause any ob-
servable problems, and cannot be easily recognized by programmatic solutions.
The confidentiality of any piece of information is highly subjective, and it takes
a human to determine if the presence of some information in a specific security
context is legitimate or not. On the other hand, with gigabytes of data being
transferred from ring 0 to ring 3, from program memory to files on disk, from
one machine to another through the network, and between a number of other
security contexts, it is hardly possible to manually review all this traffic to find
every instance of memory disclosure taking place. As a result, many bugs may
actively leak data for years before they are noticed, if they are noticed at all.
The fact that infoleaks produce no tangible feedback to kernel developers
also means that they cannot learn from their own mistakes, and may repeat the
same insecure code patterns in a number of places if they are not aware of the
bug class and are not actively trying to prevent it.
Leaks hidden behind system API Typical client applications implement

their functionality using high-level system API, especially in Windows. The
API is often responsible for converting input parameters into internal structures
accepted by the system calls, and similarly converting the syscalls’ output to
a format understood by the caller. In the process, memory disclosed by the
kernel may be discarded by the user-mode system library, and thus may never
be passed back to the program. This further reduces the visibility of infoleaks,
and decreases their chance of being noticed during regular software development.
The situation is illustrated in Figure 6.
17
User-mode Program User-Mode System API System Kernel
Call API function
Convert arguments and invoke syscall
Syscall logic
Write output with leaks and return
Extract
meaningful data
Disclosed memory
Return the specific requested values lost here
Figure 6: Leaked kernel memory discarded by user-mode API
2.4 Severity and impact on system security

Due to their very nature, kernel→user memory leaks are strictly local infor-
mation disclosure bugs. There is no memory corruption or remote exploitation
involved, and an attacker needs to be able to execute arbitrary code on the af-
fected machine as a prerequisite. On the other hand, most disclosures are silent
and don’t leave any footprints in the system, so it is possible to trigger the bugs
indefinitely, until the objective of the exploit is accomplished. The severity of
each issue must be evaluated on a case-by-case basis, as it depends on the extent
of the leak and the type of data that can be exposed.
In general, bugs of this class seem to be mostly useful as single links in
longer local elevation of privileges exploit chains. The biggest secret-based mit-
igation currently in use is KASLR (Kernel Address Space Layout Randomiza-
tion), and the locations of various objects in kernel address space are among
the most frequently leaked types of data. A real-life example is a Windows
kernel exploit found in the Hacking Team dump in July 2015 [40], which took
advantage of a pool memory disclosure to derandomize the base address of
the win32k.sys driver, subsequently used for the exploitation of another vul-
nerability. Coincidentally, the flaw was discovered around the same time by
Matt Tait of Google Project Zero, and was later fixed in the MS15-080 bulletin
as CVE-2015-2433 [75].
Stacks The distinction between stacks and heaps/pools is that stacks gener-
ally store information directly related to the control flow, such as addresses of
kernel modules, dynamic allocations, stacks, and the secret values of stack cook-
18
ies installed by exploit mitigations such as StackGuard on Linux [24] and /GS
on Windows [83]. These are consistent pieces of information immediately useful
for potential attackers who intend to combine them with memory corruption
exploits. However, the variety of data they offer is limited2 , which leads us to
believe that they don’t present much value as standalone vulnerabilities.
Pools/Heaps Kernel pools and heaps contain addresses of executable images

and dynamic allocations too, but they also include a range of data processed by
various components in the system, such as disk drivers, file system drivers, net-
work drivers, video drivers and so forth. This may allow attackers to effectively
sniff the activity of privileged services and other users in the system, potentially
revealing sensitive data which has value on its own, beyond only facilitating
the exploitation of another bug. The problem of reliably leaking specific types
of data from the kernel (e.g. file contents, network traffic or passwords) is still
open and largely unexplored.
2.5 Prior research

Various degrees of work have been put to address memory disclosure in different
operating systems. In the subsections below we outline the documented efforts
undertaken for Windows and Linux in the areas of detection and mitigation.
2.5.1 Microsoft Windows

Detection There are very few publicly available sources discussing the prob-
lem in Windows before 2015. In July 2015, Matt Tait reported a disclosure of
uninitialized pool memory through the win32k!NtGdiGetTextMetrics system
call [75], which was also revealed a few weeks later to be known by the Hack-
ing Team security firm [40]. To our best knowledge, this was the first piece of
evidence that vulnerabilities of this type had been successfully used as part of
0-day local privilege escalation exploit chains for Windows.
Also in 2015, WanderingGlitch of HP’s Zero Day Initiative was credited by
Microsoft for the discovery of eight kernel memory disclosure vulnerabilities [78].
Some of them were later discussed by the researcher at the Ruxcon 2016 security
conference, during a talk titled “Leaking Windows Kernel Pointers” [109].
Lastly, in June 2017 fanxiaocao and pjf of IceSword Lab (Qihoo 360) pub-
lished a blog post titled “Automatically Discovering Windows Kernel Informa-
tion Leak Vulnerabilities” [27], where they described the design of their tool,
which they used to identify a total of 14 infoleaks in 2017 (8 of which collided
with our findings). In many ways, the project was similar to Bochspwn Reloaded
in how it poisoned kernel stack and pool allocations, and automatically detected
leaks by examining all kernel→userland memory writes. The key differences
were that a virtual machine managed by VMware was used instead of a full
2 As long as old heap pages are not reused for stacks without clearing, which was still the
case in macOS in October 2017 [34], and in Linux until April 2018 [41].
19
CPU emulator, and memory taint tracking was not implemented. More infor-
mation on this approach can be found in Section 5.5 “Taintless Bochspwn-style
instrumentation”. To our best knowledge, aside from Bochspwn Reloaded, this
was the first attempt to automatically identify kernel memory disclosure in a
closed-source operating system.
While not directly related to memory disclosure, it is also worth noting that
in 2010/2011, several types of ring 0 addresses were found to be leaked through
uninitialized return values of several win32k system calls [104, 51]. The problem
was supposed to be mitigated in Windows 8 [42], but in 2015, Matt Tait spotted
that the fix had been incomplete and discovered three further issues [74].
Mitigation One of the few generic mitigations we are aware of is that since
June 2017, the Windows I/O manager resets the output memory regions for all
buffered I/O operations handled by kernel drivers [39]. This change killed an
entire class of infoleaks where the IOCTL handlers left uninitialized chunks of
bytes in the buffer, or didn’t initialize the buffer at all.
Another minor security improvement is the fact that in Visual Studio 15.5
and later, POD (plain old data) structures that are initialized at declaration
using a “= {0}” directive are filled with zeros as a whole. This is different from
the previous behavior, where the first member was set-by-value to 0 and the
rest of the structure was reset starting from the second field, thus potentially
leaking the contents of the padding bytes between the first and second member.
2.5.2 Linux
In contrast to Windows, the Linux community has been vocal about memory
disclosure for many years, with the biggest spike of interest starting around 2010.
Since then, a number of research projects have been devised, focusing on au-
tomatic detection of existing kernel infoleaks and on reducing or completely
nullifying the impact of presumed, yet unknown issues. We believe that the gap
in the current state of the art between Windows and Linux is primarily caused
by the open-source nature of the latter, which enables easy experimentation
with a variety of approaches – static, dynamic, and a combination of both.
Detection Throughout the last decade, there have been dozens of kernel in-
foleak fixes committed to the Linux kernel every year. According to Chen
et al. [22], there were 28 disclosures of uninitialized kernel memory fixed in the
period from January 2010 to March 2011. An updated study by Lu K. [44] shows
that further 59 such vulnerabilities were reported between January 2013 and
May 2016. A large portion of the findings can be attributed to a small group of
researchers. For example, Rosenberg and Oberheide collectively discovered more
than 25 memory disclosure vulnerabilities in Linux in 2009-2010 [25, 26, 35],
mostly from the kernel stack. They subsequently demonstrated the usefulness
of such disclosures in the exploitation of grsecurity/PaX-hardened Linux ker-
nels in 2011 [37, 36]. Kulikov found over 25 infoleaks in 2010-2011 with manual
20
analysis and Coccinelle [107]. Similarly, Krause identified and patched 21 kernel
memory disclosure issues in 2013 [73], and more than 50 such bugs overall.
There are several tools readily available to detect infoleaks and other uses
of uninitialized memory in Linux, mostly designed with kernel developers in
mind. The most basic one is the -Wuninitialized compiler flag supported
by both gcc and LLVM, capable of detecting simple instances of uninitialized
reads within the scope of single functions. A more advanced option is the
kmemcheck debugging feature [108], which can be considered a kernel-mode
equivalent of Valgrind’s memcheck. At the cost of a significant CPU and mem-
ory overhead, the dynamic checker detects all uses of uninitialized memory
occurring in the code. The feature was recently removed from the mainline
kernel [17, 76], as the project is now considered inferior to the new and more
powerful KernelAddressSanitizer [12] and KernelMemorySanitizer checkers [19].
In the last few months, KMSAN coupled with the coverage-based syzkaller sys-
tem call fuzzer [11] has identified 19 reads of uninitialized memory [7], including
three leaks of kernel memory to userspace.
There have also been some notable efforts to use static analysis to detect
Linux kernel infoleaks. In 2014-2016, Peiró et al. demonstrated successful use
of model checking with the Coccinelle engine [3] to identify stack-based mem-
ory disclosure in Linux kernel v3.12 [97, 98]. The model checking was based on
taint tracking objects in memory from stack allocation to copy to user calls,
and yielded eight previously unknown vulnerabilities. In 2016, Lu et al. imple-
mented a project called UniSan [45, 44] – advanced, byte-granular taint tracking
performed at compile time to determine which stack and dynamic allocations
could leak uninitialized memory to one of the external data sinks (user-mode
memory, files and sockets). While the tool was primarily meant to mitigate
infoleaks by clearing all potentially unsafe allocations, the authors randomly
chose and analyzed a sample of about 20% of them (350 of about 1800), and
reported 19 new vulnerabilities in the Linux and Android kernels as a result.
Finally, several authors have proposed a technique of multi-variant program
execution to identify use of uninitialized memory. The basic premise of the ap-
proach is to concurrently run several replicas of the same software, capture their
output and compare it. If all legitimate sources of entropy are virtualized to re-
turn stable data across all replicas, then any differences in the output may only
be caused by leaking or using uninitialized memory. The non-determinism may
originate from entropy introduced by ASLR, or from different marker bytes used
to initialize new stack/heap allocations. The method was implemented for user-
mode programs in the DieHard [21] and BUDDY [44] projects in 2006 and 2017,
respectively. A similar approach was discussed by North in 2015 [96]. Lastly,
authors of SafeInit [94] also stated that their tool was meant as a software hard-
ening mechanism, but could be combined with a multi-variant execution system
to achieve bug detection capability. The technique was extensively evaluated
for client applications, but to our knowledge it hasn’t been successfully imple-
mented for the Linux kernel. In Section 5.4 “Differential syscall fuzzing”, we
present how a similar concept was shown to be effective in identifying infoleaks
in a subset of Windows system calls.
21
Mitigation Generic mitigations against kernel memory disclosure generally
revolve around zeroing old memory regions to prevent leftover data from being
inherited by new, unrelated objects. The ultimate advantage of this method is
that it addresses the problem on a fundamental level, completely eliminating the
threat of uninitialized memory, and killing existing and future kernel infoleaks
at once. Only a small fraction of memory allocated by the kernel is ever copied
to user-mode; on the other hand, resetting all memory areas prior or after usage
incurs a significant overhead. Finding the optimal balance between system
performance and the degree of protection against memory disclosure is currently
the main point of discussion.
In mainline kernel, the CONFIG PAGE POISONING and CONFIG DEBUG SLAB op-
tions can be set to enable the poisoning of all freed dynamic allocations with
marker bytes. The result is that each piece of data is overwritten at the mo-
ment it is discarded by the caller, which prevents it from being leaked back
to user-mode later on. Since all allocations are subject to poisoning, the op-
tions come with a considerable performance hit, and they don’t protect stack
allocations, which seem to constitute a majority of Linux kernel infoleaks.
The grsecurity/PaX project [4, 9] provides further hardening features. Set-
ting the PAX MEMORY SANITIZE flag causes the kernel to erase memory pages and
slab objects when they are freed, except for low-risk slabs whitelisted for perfor-
mance reasons. Furthermore, the PAX MEMORY STRUCTLEAK option is designed to
zero-initialize stack-based objects such as structures, if they are detected to be
copied to userland. It may prevent leaks through uninitialized fields and padding
bytes, but it is a relatively lightweight feature that may be subject to false nega-
tives. A more exhaustive, but also more costly option is PAX MEMORY STACKLEAK,
which erases the used portion of the kernel stack before returning from each sys-
tem call. This eliminates any disclosure of stack memory shared between two
subsequent syscalls, but doesn’t affect leaks of data put on the stack by the vul-
nerable syscall itself. Currently, the Kernel Self Protection Project is making
efforts to port the STACKLEAK feature to the mainline kernel [18, 38].
Other researchers have also proposed various variants of zeroing objects at
allocation and deallocation times in the Linux kernel. Secure deallocation [23]
(Chow et al., 2005) reduces the lifetime of data in memory by zeroing all regions
at deallocation or within a short, predictable period of time. A prototype of the
concept was implemented for Linux user-mode applications and the kernel page
allocator. Split Kernel [43] (Kurmus and Zippel, 2014) protects the system from
exploitation by untrusted processes by clearing entire kernel stack frames upon
entering each hardened function. SafeInit [94] (Milburn et al., 2017) clears
all stack and dynamic allocations before they are used in the code, to ulti-
mately eliminate information leaks and use of uninitialized memory. UniSan [45]
(Lu et al., 2016) reduces the overhead of SafeInit by performing advanced mem-
ory taint tracking at compile time, to conservatively determine which allocations
are safe to be left without zero-initialization, while still clearing the remaining
stack and heap-based objects.
As shown above, Linux has been subject to extensive experimentation in the
area of data lifetime and kernel memory disclosure.
22
3 Bochspwn Reloaded – detection with software
x86 emulation
Bochs [1] is an open-source IA-32 (x86) PC emulator written in C++. It in-
cludes emulation of the Intel x86 CPU, common I/O devices and a custom
BIOS, making up a complete, functional virtual machine fully emulated in soft-
ware. It can be compiled to run in a variety of configurations, and similarly
it can correctly host most common operating systems, including Windows and
GNU/Linux. In our research, we ran Bochs on Windows 10 64-bit as the host
system.
Among many of the project’s qualities, Bochs provides an extensive instru-
mentation API, which makes it a prime choice for performing DBI (Dynamic
Binary Instrumentation) against OS kernels. At the time of this writing, there
are 31 supported instrumentation callbacks invoked by the emulator on many
occassions during the emulated computer’s run time, such as:
• before starting the machine and after shutting it down,

• before and after the execution of each CPU instruction,
• on virtual and physical memory access,

• on exceptions and interrupts,
• on reading from and writing to I/O ports, and many more.
From a technical standpoint, the instrumentation is implemented in the form

of C++ callback procedures built into the emulator executable. These callbacks
may read and modify the CPU context and other parts of the execution environ-
ment (e.g. virtual memory), making it possible to inspect and alter the system
execution flow in any desired way, as well as modify the processor’s behavior
and extend it with new functionality.
Basic documentation of the interface and the meaning of each particular call-
back can be found in the instrument/instrumentation.txt file in the Bochs
source tree, and code examples are located in instrument/example{0,1,2}.
The source code of the original Bochspwn tool from 2013 designed to iden-
tify double fetch issues in system kernels is also available for reference on
GitHub [71].
In the following subsections, we describe how the core logic of memory dis-
closure detection and other ancillary features were implemented in the Bochs
instrumentation.
23
3.1 Core logic – kernel memory taint tracking
The fundamental idea behind Bochspwn Reloaded is memory taint tracking
performed over the entire kernel address space of the guest system. In this case,
taint is associated with every kernel virtual address, and represents information
about the initialization state of each byte residing in ring 0.
The high-level logic used to detect disclosure of uninitialized (tainted) kernel
memory is as follows:
• Set taint on all new stack and heap/pool-based allocations.

• Remove taint from memory overwritten with constant values and registers.
• Remove taint from freed allocations (optionally).
• Propagate taint across memory during copy operations (memcpy calls).

• Identify infoleaks by verifying the taint of data copied from kernel-mode
to user-mode addresses.
By implementing the above functionality, the instrumentation should be

able to detect and signal most instances of uninitialized kernel memory being
disclosed to userland. One of the obvious limitations of the approach is the
fact that tainting is only applied to virtual memory and not CPU registers,
which are effectively always considered untainted. This means that every time
data in memory is copied indirectly through a register (e.g. with a sequence of
mov reg, [mem]; mov [mem], reg instructions), the taint information is dis-
carded in the process, potentially yielding undesired false-negatives.
The design decision to only taint memory and not the CPU context was
primarily motivated by performance reasons. Instrumenting just the memcpy
calls has low overhead; for example on x86 platforms, it only requires the
special handling of the rep movs{b,w,d,q} instructions (as discussed in Sec-
tion 3.1.5, “Taint propagation”). On the contrary, in order to propagate taint
across registers, the instrumentation would have to cover all or most instructions
operating on them, which would significantly slow down the already slow emula-
tion of the guest system. Furthermore, Bochs doesn’t provide any documented
API to intercept register access, so more complex changes to the source code of
the emulator would be necessary. Lastly, we believe that a majority of kernel
infoleaks occur through large data structures and not primitive variable types,
which suggests that the extent of false-negatives caused by the limited scope
of the tainting should be negligible. Overall, our evaluation of the cost/benefit
ratio of both methods seems to favor the more simplistic scheme.
Below, we discuss the implementation details of each part of the tainting
mechanism used in Bochspwn Reloaded.
24
3.1.1 Shadow memory representation
The taint information of the guest kernel address space is stored in so-called
shadow memory, a memory region in the Bochs process that maps each byte
(or chunk of bytes) to the corresponding metadata. As the scope of the metadata
and its low-level representation is different for 32-bit and 64-bit target systems,
we address them separately in the following paragraphs.
Shadow memory in x86 In the default configuration of 32-bit versions of

Windows, the kernel occupies addresses between 0x80000000 and 0xffffffff,
which makes a total of 2 GB. On Linux, in case of the 3G/1G memory split, the
kernel is assigned 1 GB of the address space. Coupled with the fact that the
Bochs emulator may be compiled as a 64-bit process, this means that as long as
the amount of metadata is kept at a reasonable size, it is possible to statically
allocate the shadow memory for the entire kernel address range.
In order to maximize the verbosity of Bochspwn Reloaded reports, we intro-
duced a number of extra metadata information classes in addition to the taint
bit, related to the memory allocation that each address belongs to:
• size – size of the memory allocation,

• base address – address of the start of the allocation,
• tag/flags – on Windows, the tag of the pool allocation; on Linux, the
flags of the heap allocation,
• origin – address of the instruction that requested the allocation.
Each of the values is 32 bits wide. In total, the above fields together with the
taint boolean consume 17 bytes. If a separate descriptor entry was allocated for
each byte of the kernel address space, the overhead would be 34 GB for Windows
guests and 17 GB for Linux. To optimize the memory usage, we increased the
granularity of the above allocation-related metadata from 1 to 8 bytes, which
reduced the effective overhead from a factor of 17 to 3. This was sufficient to run
the instrumentation on a typical modern workstation equipped with a standard
amount of RAM. A summary of the information classes is shown in Table 1.
Memory usage
Information class Type Granularity
Windows Linux
Taint uint8 1 byte 2 GB 1 GB
Allocation size uint32 8 bytes 1 GB 512 MB
Allocation base uint32 8 bytes 1 GB 512 MB
Allocation tag/flags uint32 8 bytes 1 GB 512 MB
Allocation origin uint32 8 bytes 1 GB 512 MB
Total 6 GB 3 GB
Table 1: Summary of metadata information classes stored for x86 guest systems
25
Shadow memory in x86-64 Representing taint information for 64-bit sys-
tems is more difficult, primarily because the address range subject to shad-
owing is as large as the user-mode address space of the Bochs emulator it-
self. Valid kernel memory addresses fall between 0xffff800000000000 and
0xffffffffffffffff, which is a 128 terabyte area that cannot be mapped to a
statically allocated region, both due to physical memory and virtual addressing
limitations. As a result, this technical challenge called for a substantial rework
of the metadata storage.
To start things off, we removed some of the less critical information related
to memory allocations, specifically the size, base address and tag/flags. While
certainly useful to have, we could still triage all potential reports without these
information classes, and maintaining them for x86-64 targets would be a sig-
nificant burden. However, we deemed the allocation origin important enough
to stay, as it was often the only way to track the control flow back to the vul-
nerable code area, e.g. when universal interfaces (such as ALPC in Windows)
were used to send data to user space. As allocating the data structure stat-
ically was no longer an option, we converted the 0x10000000-item long array
to a std::unordered map<uint64, uint64> hash map container, semantically
nearly identical to its 32-bit counterpart.
The last information class that needed to be considered was the taint itself.
It was possible to also use a hash map in this case, but it was not optimal and
would have significantly slowed the instrumentation down. Instead, we made
use of the fact that the taint state for each kernel byte could be represented
as a single bit. As a result, the taint information of the kernel address space
was packed into bitmasks, thus mapping the 0x800000000000-byte (128 TB)
region into a 0x100000000000-byte (16 TB) shadow memory. While an area of
this size still cannot be allocated all at once, it can be reserved in the virtual
address space. During the run time of the emulator, the specific pages accessed
by the instrumentation are mapped on demand, resembling a mechanism known
as memory overcommittment.
Memory overcommittment is not supported on Windows – our host system
of choice – but it is possible to implement it on one’s own using exception
handling:
1. Reserve the overall shadow memory area with a VirtualAlloc API call
and a MEM RESERVE flag.
2. Set up an exception handler using the AddVectoredExceptionHandler
function.
3. In the exception handler, check if the accessed address falls within the
shadow memory, and exception code equals EXCEPTION ACCESS VIOLATION.
If this is the case, commit the memory page in question and return with
EXCEPTION CONTINUE EXECUTION.
With the modifications explained above, the shadow memory was success-
fully ported to work with 64-bit guest systems.
26
Double-tainting In addition to setting taint on allocations in the shadow
memory, our instrumentation also fills the body of new memory regions with
fixed marker bytes3 – 0xaa for heap/pools and 0xbb for stack objects. While
not essential for the correct functioning of the infoleak detection, it is a useful
debugging feature.
First of all, the mechanism enables the instrumentation to verify the cor-
rectness of its own taint tracking by cross checking the information from two
different sources. If a specific region is uninitialized according to the shadow
memory, but in fact it no longer contains the marker bytes, this indicates that
the memory was overwritten in a way our tool was not aware of (e.g. by a disk
controller or another external device). In such cases, the instrumentation can
correct the taint early on, instead of propagating the wrong information further
in memory. Similarly, the behavior guarantees a nearly 100% true-positive ratio
of the reported bugs, as they are verified both against the taint information and
the guest virtual memory.
Furthermore, the markers are easily recognizable under kernel debuggers
attached to the guest systems, which often aids in understanding the current
system state and thus makes it easier to establish the root cause of the dis-
covered bugs. Lastly, the mechanism may potentially expose other types of
vulnerabilities in the process, such as use of uninitialized memory.
3.1.2 Tainting stack frames

Allocating memory from the stack is typically a platform-agnostic operation
facilitated by the CPU architecture. On x86, it is achieved with a sub esp
instruction, where the second operand is either an immediate value or another
register. The equivalent sub rsp instruction is used on x86-64. This is the
case in both Windows and Linux. In order to account for any unexpected code
constructs that could also decrease ESP to allocate memory, we additionally
instrumented the add esp and and esp instructions.
The overall logic of the stack tainting implemented in Bochspwn Reloaded
is shown in pseudo-code in Listing 7. The before execution callback is used to
detect instances of kernel instructions modifying the stack pointer. Later, after
the instruction executes, a subsequent callback checks if ESP was decreased and
if so, the entire memory region in the range of ESPnew – ESPold is tainted. The
current instruction pointer is also saved, so that the memory can be traced back
to the place where it was allocated in case it is detected to be disclosed to user-
mode. This universal approach is effective for both Linux and a majority of
the Windows kernel. The few remaining corner cases in Windows that require
special handling are discussed in the next paragraph.
3 Unless the caller explicitly requests a zeroed-out area, e.g. by passing the GFP ZERO flag
to kmalloc in Linux.
27
1 void bx_instr_before_execution(CPU cpu, instruction i) {
2 if (!cpu.protected_mode ||
3 !os::is_kernel_address(cpu.eip) ||
4 !os::is_kernel_address(cpu.esp)) {
5 return;
6 }
7
8 if (i.opcode == SUB || i.opcode == ADD || i.opcode == AND) {
9 if (i.op[0] == ESP) {
10 globals::esp_changed = true;
11 globals::esp_value = cpu.esp;
12 }
13 }
14 }
15
16 void bx_instr_after_execution(CPU cpu, instruction i) {
17 if (globals::esp_changed && cpu.esp < globals::esp_value) {
18 set_taint(/*from= */cpu.esp,
19 /*to= */globals::esp_value - 1,
20 /*origin=*/cpu.eip);
21 }
22
23 globals::esp_changed = false;
24 }
Listing 7: Pseudo-code of the stack tainting logic
Microsoft Windows One system-specific assembly code construct not cov-

ered by the previously described logic is the implementation of the chkstk
built-in function in 32-bit builds of Windows. The routine is equivalent to
alloca in that it is used for requesting dynamic allocations from the stack. The
relevant part of the code is shown in Listing 8; the highlighted xchg eax, esp
instruction is used to save the updated stack pointer in ESP. In order to account
for this non-standard way of modifying ESP, we included BX IA XCHG ERXEAX in
the list of opcodes possibly manifesting a stack allocation.
As chkstk is a generic function, saving its address doesn’t reveal the actual
requestor of the memory region. In order to work around this, our instrumen-
tation reads the return address from [EAX] after the xchg instruction executes,
and uses it to identify the creator of the allocation.
There is also another circumstance when the instruction directly modify-
ing ESP cannot be used as the allocation’s origin. The SEH prolog4 and
SEH prolog4 GS procedures are used in 32-bit Windows to create stack frames
for functions with exception handlers, including most top-level syscall entry
points. A shortened version of the SEH prolog4 function assembly is shown
in Listing 9. To handle this case, Bochspwn Reloaded obtains the address of
the procedure caller from [EBP-8], and uses it as the origin of the stack frame.
28
.text:00548D94 public __chkstk
.text:00548D94 __chkstk proc near
.text:00548D94
[...]
.text:00548DA8
.text:00548DA8 cs10:
.text:00548DA8 cmp ecx, eax
.text:00548DAA jb short cs20
.text:00548DAC mov eax, ecx
.text:00548DAE pop ecx
.text:00548DAF xchg eax, esp
.text:00548DB0 mov eax, [eax]
.text:00548DB2 mov [esp+0], eax
.text:00548DB5 retn
.text:00548DB6
[...]
.text:00548DBD __chkstk endp
Listing 8: Built-in chkstk function on 32-bit versions of Windows
.text:00557798 __SEH_prolog4 proc near

.text:00557798
.text:00557798 arg_4 = dword ptr 8
.text:00557798
.text:00557798 push offset __except_handler4
[...]
.text:005577AC lea ebp, [esp+8+arg_4]
.text:005577B0 sub esp, eax
.text:005577B2 push ebx
[...]
.text:005577D3 lea eax, [ebp-10h]
.text:005577D6 mov large fs:0, eax
.text:005577DC retn
.text:005577DC __SEH_prolog4 endp
Listing 9: Built-in SEH prolog4 function on 32-bit versions of Windows
29
3.1.3 Tainting heap/pool allocations
Contrary to stack frames and automatic objects, detecting and tainting dynamic
allocations is a highly system-specific task, which must be implemented for each
tested OS dedicatedly. As a general rule, the instrumentation should intercept
the addresses of all newly allocated regions and their lengths, optionally together
with further information such as the origin, tag/flags etc. The specifics of the
Windows and Linux kernel allocators are outlined in the paragraphs below.
Microsoft Windows As noted in Section 2.2.1 “Memory reuse in dynamic al-

locators”, the core of the Windows kernel pool allocator is the ExAllocatePool-
WithTag function [82]. While there are many available wrappers which offer ex-
tended functionality, all of them eventually call into ExAllocatePoolWithTag to
request the dynamic memory. Thanks to this design, it is sufficient to hook into
this single function to obtain information about nearly all dynamic allocations
taking place in the kernel.
In 32-bit versions of Windows, the function follows the stdcall calling con-
vention, which means that the arguments are stored on the stack, and the return
value is passed back via the EAX register. This control flow is convenient for our
instrumentation, as it allows us to set a single “breakpoint” on the RET instruc-
tion(s) at the end of the allocator, and every time it is hit, we can read both
the details of the request (allocation origin, size and tag) and the address of the
allocated memory. At that point, the instrumentation marks the entire region
as tainted in the shadow memory.
The corresponding implementation for 64-bit builds of Windows is more
complex, due to the fact that parameters to ExAllocatePoolWithTag are passed
through the RCX, RDX and R8 registers. These registers are considered volatile,
which means that by the time the allocator returns, they may be modified and
no longer contain the input values. In other words, there isn’t a single point in
the control flow where both the request information and the allocation address
are guaranteed to be known at the same time. To address the problem, we hook
into the allocator twice – in the prologue and epilogue of the function. In the
first hook, we read the allocation origin and length, and save them in a hash
map keyed by RSP. In the second one, we read the allocation address from RAX,
load the request information from the hash map (the stack pointer is the same
when entering and leaving the function), and taint the memory accordingly.
One corner case that requires special attention are alternate kernel allocators
used by specific subsystems in the kernel. Even though they are typically based
on ExAllocatePoolWithTag, they may implement additional optimizations such
as caching, which bypass our taint tracking algorithm. One example of such an
allocator is the win32k!AllocFreeTmpBuffer function, whose pseudo-code is
illustrated in Listing 10. The function maintains a global pointer to a long-
lasting 0x1000-byte allocation, which is used to satisfy requests whenever it is
not currently in use. As the memory region is allocated only once and never
freed, but it is reused by multiple parts of the graphic subsystem, it is not subject
to proper memory disclosure detection. The solution to this problem is to modify
30
1 PVOID AllocFreeTmpBuffer(SIZE_T Size) {
2 PVOID Result;
3
4 if (Size > 0x1000 ||
5 (Result = InterlockedExchange(gpTmpGlobalFree, NULL)) == NULL) {
6 Result = AllocThreadBufferWithTag(Size, ’pmTG’);
7 }
8
9 return Result;
10 }
Listing 10: Pseudo-code of the win32k!AllocFreeTmpBuffer allocator
AllocFreeTmpBuffer to always call into AllocThreadBufferWithTag, or to

change the value of the pointer under gpTmpGlobalFree to NULL. Both actions
have the same outcome and are easily achieved with the WinDbg debugger
attached to the guest system; the latter has the advantage that it works with
64-bit builds of Windows with PatchGuard enabled.
Linux Compared to Windows, intercepting all kernel dynamic allocations in

Linux from the level of DBI is relatively complex for several reasons. First of
all, the x86 kernel is compiled with the regparm=3 option, which causes the first
three arguments of each function to be passed through volatile registers instead
of the stack. As a result, the values of these parameters are no longer available
when the function returns – if our instrumentation requires access to both the
arguments and return value, it needs to hook into the function in question twice,
in the prologue and epilogue(s). Furthermore, there are three distinct allocators
(kmalloc, vmalloc, kmem cache alloc), and some of them have many entry-
points, e.g. kmalloc, kmalloc order and kmalloc track caller. Lastly, in
case of kmem cache, the allocation size must be saved during the creation of the
cache, as it is not explicitly provided later while requesting memory. If the cache
has a constructor specified, then the taint must be applied when the constructor
is called, and not when the memory is returned to the caller.
In response to the issues raised above, the Linux-specific part of our instru-
mentation installed a number of hooks in the emulated kernel:
• kmalloc and vmalloc
– Prologue: save the size and flags arguments.

– Epilogue: set taint on the allocated memory.
• kfree, vfree and kmem cache free – untaint freed allocations.
31
• kmem cache create
– Prologue: save the cache size and the constructor function pointer.
– Epilogue: if the function succeeded, save the address of the newly
created cache and set a breakpoint on the constructor function, if
present.
• kmem cache destroy – remove the cache from the internal structures and
clear the breakpoint on the cache’s constructor, if present.
• kmem cache alloc
– Prologue: save the cache and flags arguments.

– Epilogue: set taint on the allocated memory.
• Dynamically detected cache constructors – set taint on the memory region

received in the argument.
By intercepting the kernel execution in those locations, it was possible to

apply taint tracking to dynamic allocations being requested during the run time
of the Linux kernel.
3.1.4 Clearing taint

By principle, Bochspwn Reloaded untaints memory every time it is overwritten
by any instruction which is not part of a memcpy operation. This includes writes
with constant values and registers, which by themselves are not subject to taint
tracking. Therefore, virtually any write to memory marks it as initialized as
long as it is not receiving existing data from another source.
Optionally, memory can also be untainted upon being freed, for example
on a add esp instruction increasing the address of the stack pointer, or when
a ExFreePoolWithTag / kfree function is called to destroy a dynamic object.
However, doing so is not required, and Solar Designer suggested [102] that it
could be beneficial to re-taint freed allocations in order to potentially detect
use-after-free and similar vulnerabilities. In our testing, we untainted freed
allocations in 32-bit Windows and Linux guests, but didn’t handle any free-
like functions in 64-bit Windows.
3.1.5 Taint propagation

In Bochspwn Reloaded, taint is propagated when the instrumentation recognizes
that data is copied between two kernel-mode locations in the guest memory,
i.e. a memcpy operation is in progress. While primitive variables are out of
scope due to the fact that CPU registers are not subject to taint tracking, it is
paramount for the tool to effectively detect the copying of larger memory blobs.
32
On the x86 and x86-64 platforms, the instruction dedicated to copying con-
tinuous memory blocks is rep movsb4 , which moves ECX bytes at address DS:ESI
to address ES:EDI (equivalent 64-bit registers are used on x86-64). It is very
frequently used in kernels, both as part of the standard library memcpy im-
plementation and as an inlined form of the function. From the perspective of
instrumentation, the instruction is very convenient, as it specifies the source,
desination and length of the copy all at the same time. This makes it trivial
to propagate taint in the shadow memory, and is the main reason why taint
propagation in our tool was built around the special handling of rep movs.
In the paragraphs below, we explain how practical the core idea was with
regards to the actual binary code of each of the tested systems, and what further
steps we took to maximize the scope of the detected memory copying activity.
Microsoft Windows In Windows, there are generally four functions which

can be used to copy data in memory: memcpy, memmove, RtlCopyMemory and
RtlMoveMemory.
On 32-bit builds of the system, RtlCopyMemory is not an actual function, but
merely a macro around memcpy. The RtlMoveMemory routine is based exclusively
on rep movsb and rep movsd, which makes it fully compatible with our taint
tracking approach. Finally, memcpy and memmove are separate functions but
have exactly the same code, so they can be considered as one. They are also
mostly based on the desired rep movs instruction, with the exception of several
conditions:
1. the source and destination buffers overlap,
2. the destination pointer is not 4-byte aligned,
3. the copy length is below 32 bytes,
4. the copy length is not 4-byte aligned.
Practically speaking, we believe that conditions (1) and (2) can be safely
disregarded, as they occur very rarely during normal system execution and are
unlikely to affect our infoleak detection performance. Condition (4) is also
negligible, as it is similarly rare and only affects the copying of the trailing 1-3
bytes. Condition (3) is the most undesired one, as it may lead to the loss of
taint corresponding to numerous small objects in the kernel address space.
In order to address the problem, it is possible to patch the memcpy and
memmove functions in the WinDbg debugger attached to the guest system, such
that they use rep movs regardless of the length of the operation. On binary
level, it is a matter of modifying a single byte in the functions’ body, changing
the operand of a cmp instruction from 0x08 to 0x00, thus nullifying the con-
dition check. The relevant assembly code from the nt!memcpy function found
in Windows 7 is shown in Listing 11. Thanks to the patch, the rep movsd
instruction at address 0x439293 is always reached.
4 The corresponding rep movsw, rep movsd and rep movsq variants operate on units of
2, 4 and 8 bytes respectively.
33
.text:00439260 _memcpy proc near
.text:00439260
.text:00439260 arg_0 = dword ptr 8
.text:00439260 arg_4 = dword ptr 0Ch
.text:00439260 arg_8 = dword ptr 10h
.text:00439260
.text:00439260 55 push ebp
.text:00439261 8B EC mov ebp, esp
[...]
.text:00439278 3B F8 cmp edi, eax
.text:0043927A 0F 82 7C+ jb CopyDown
.text:00439280
.text:00439280 CopyUp:
.text:00439280 F7 C7 03+ test edi, 3
.text:00439286 75 14 jnz short CopyLeadUp
.text:00439288 C1 E9 02 shr ecx, 2
.text:0043928B 83 E2 03 and edx, 3
.text:0043928E 83 F9 08 cmp ecx, 8
.text:00439291 72 29 jb short CopyUnwindUp
.text:00439293 F3 A5 rep movsd
.text:00439295 FF 24 95+ jmp ds:off_4393AC[edx*4]
Listing 11: Patched memcpy function prologue
It is important to note that both memcpy and memmove functions should

be modified to achieve the best result. Moreover, depending on the specific
Windows version, each driver may include its own copies of these functions.
In our testing, we patched the code in the two most important modules –
ntoskrnl.exe and win32k.sys – where applicable. Alternatively, one could
also entirely overwrite the memory copy routines to exclusively use the rep movsb
instruction, or implement the patching logic in the Bochs instrumentation in-
stead of performing it manually in a kernel debugger.
On 64-bit Windows builds, it is more difficult to apply the previously dis-
cussed logic. Contrary to their 32-bit counterparts, the memory copy functions
no longer use rep movs in any form, but instead they are optimized with the
usage of mov and equivalent SSE instructions (movdqu, movntdq etc.). These
instructions move memory indirectly through registers, and are therefore not
compliant with our taint tracking logic. Example excerpts from the nt!memcpy
64-bit assembly code on Windows 7 and 10 are shown in Listings 12 and 13.
As we had already decided against taint tracking of CPU registers, which was
the only generic solution to the problem, we adopted a more implementation-
specific concept to dynamically detect all instances of the memcpy functions at
run time, and propagate taint based on their arguments. In the context of this
idea, it is convenient that in Windows x64, the four aforementioned memory
copy functions are all handled by a single implementation, as illustrated in
Listing 14. Furthermore, the binary representation of the functions’ code in all
34
.text:0000000140095720 mcpy90:
.text:0000000140095720 mov r9, [rdx+rcx]
.text:0000000140095724 mov r10, [rdx+rcx+8]
.text:0000000140095729 movnti qword ptr [rcx], r9
.text:000000014009572D movnti qword ptr [rcx+8], r10
.text:0000000140095732 mov r9, [rdx+rcx+10h]
.text:000000014009573C movnti qword ptr [rcx+10h], r9
.text:0000000140095741 movnti qword ptr [rcx+18h], r10
.text:000000014009574B mov r10, [rdx+rcx+28h]
.text:0000000140095750 add rcx, 40h
.text:0000000140095754 movnti qword ptr [rcx-20h], r9
.text:0000000140095759 movnti qword ptr [rcx-18h], r10
.text:000000014009575E mov r9, [rdx+rcx-10h]
.text:0000000140095763 mov r10, [rdx+rcx-8]
.text:0000000140095768 dec eax
.text:000000014009576A movnti qword ptr [rcx-10h], r9
.text:000000014009576F movnti qword ptr [rcx-8], r10
.text:0000000140095774 jnz short mcpy90
Listing 12: Example block of the nt!memcpy function on Windows 7 64-bit
.text:0000000140189700 lcpy40:
.text:0000000140189700 movdqu xmm0, xmmword ptr [rdx+rcx]
.text:0000000140189705 movdqu xmm1, xmmword ptr [rdx+rcx+10h]
.text:000000014018970B movntdq xmmword ptr [rcx], xmm0
.text:000000014018970F movntdq xmmword ptr [rcx+10h], xmm1
.text:0000000140189714 add rcx, 40h
.text:0000000140189718 movdqu xmm0, xmmword ptr [rdx+rcx-20h]
.text:000000014018971E movdqu xmm1, xmmword ptr [rdx+rcx-10h]
.text:0000000140189724 movntdq xmmword ptr [rcx-20h], xmm0
.text:0000000140189729 movntdq xmmword ptr [rcx-10h], xmm1
.text:000000014018972E dec eax
.text:0000000140189730 jnz short lcpy40
Listing 13: Example block of the nt!memcpy function on Windows 10 64-bit
35
.text:0000000140095600 ; Exported entry 1365. RtlCopyMemory
.text:0000000140095600 ; Exported entry 1541. RtlMoveMemory
.text:0000000140095600 ; Exported entry 2063. memcpy
.text:0000000140095600 ; Exported entry 2065. memmove
.text:0000000140095600
.text:0000000140095600 ; Attributes: library function
.text:0000000140095600
.text:0000000140095600 public memmove
.text:0000000140095600 memmove proc near
.text:0000000140095600
.text:0000000140095600 mov r11, rcx
.text:0000000140095603 sub rdx, rcx
.text:0000000140095606 jb mmov10
.text:000000014009560C cmp r8, 8
Listing 14: Shared implementation of memory copy functions in 64-bit

Windows
kernel images across the system is the same, meaning that we could successfully
identify all copies of memcpy in memory by recognizing a single unique signature
of the routine’s prologue. In our testing, we used the first 16 bytes of the
function code, which corresponded to the four initial assembly instructions.
Such a signature proved to uniquely identify the procedure in question, enabling
our tool to track the memory taint on 64-bit Windows platforms.
One currently unresolved problem is the fact that with each newer version
of Windows, an increasing number of memcpy instances with constant length
are compiled as inlined sequences of mov instructions, instead of rep movs or
direct calls into the library function. This is clearly visible in the numbers of
references to the function – on Windows 7 64-bit (January 2018 patch), the
win32k.sys module calls into memcpy at 1133 unique locations in the code.
However, the combined drivers on Windows 10 Fall Creators Update 64-bit
(win32k.sys, win32kbase.sys and win32kfull.sys) only invoke the function
696 times. The remaining instances were replaced with mov instructions located
directly in the client functions, causing our tool to lose track of a large part of
the kernel memory taint. We hope that the issue was partially mitigated by the
fact that we performed the testing against Windows 7 and 10, so bugs dating
back to Windows 7 should have been successfully detected and fixed in both
versions, even if the reduced effectiveness of the taint tracking would prevent
their discovery on Windows 10. Nonetheless, this circumstance consistutes a
significant problem in the current scheme of the Bochspwn Reloaded project.
36
175 /*
176 * No 3D Now!
177 */
178
179 #ifndef CONFIG KMEMCHECK 0
180
181 #if (__GNUC__ >= 4)
182 #define memcpy(t, f, n) __builtin_memcpy(t, f, n)
183 #else
184 #define memcpy(t, f, n) \
185 (__builtin_constant_p((n)) \
186 ? __constant_memcpy((t), (f), (n)) \
187 : __memcpy((t), (f), (n)))
188 #endif
189 #else
190 /*
191 * kmemcheck becomes very happy if we use the REP instructions
unconditionally,
192 * because it means that we know both memory operands in advance.
193 */
194 #define memcpy(t, f, n) memcpy((t), (f), (n))
195 #endif
Listing 15: Patch applied to arch/x86/include/asm/string 32.h to redirect

memcpy to the Bochspwn-compliant memcpy in line 194
Linux Linux being an open-source kernel, it gives us full control over how
memcpy is compiled. Following brief experimentation, we determined that only
minor code modifications were necessary to make it compliant with our tool,
as the function’s assembly code was largely influenced by the kernel configura-
tion flags. More specifically, we set the CONFIG X86 GENERIC option to y and
CONFIG X86 USE 3DNOW to n, and applied the patch shown in Listing 15 to un-
conditionally redirect memcpy invocations to the memcpy function comprising
of the rep movsd and rep movsb instructions. These actions were sufficient
to ensure that the resulting memory-copying assembly worked with the taint
propagation mechanism used in our instrumentation.
3.1.6 Bug detection

Bug detection in Bochspwn Reloaded may be considered a part of the taint prop-
agation functionality. Whenever the instrumentation detects a kernel→kernel
data copy, it propagates the corresponding taint in the shadow memory. When
a kernel→userland copy takes place, the tool checks the source memory region
for uninitialized bytes and if any are found, a vulnerability is reported. As a re-
sult, detection of memory disclosure is an extension of taint propagation, and
works correctly as long as memcpy-like operations are properly instrumented.
37
There are no system-specific considerations related to bug detection in Win-
dows. In Linux, we additionally implemented support for identifying informa-
tion leaks through simple variable types, and more general detection of use of
uninitialized memory. Both efforts are documented in the paragraphs below.
Leaks through primitive types in Linux In Linux, there are two interfaces
facilitating the writing of data from the kernel into user space – copy to user
and put user. The copy to user function is an equivalent of memcpy, and is
subject to the regular bug detection logic if the kernel is compiled with the
CONFIG X86 INTEL USERCOPY option set to n. On the other hand, put user is
designed to allow the kernel to write values of simple types into ring 3, such as
characters, integers or pointers. It uses direct pointer manipulation, and at the
binary level, the data in question is copied through registers, so it is not subject
to kernel infoleak detection based on the rep movs instruction. Considering that
put user is used extensively in the Linux kernel, and that we had the power to
recompile the software to adjust it to our needs, we implemented an additional
bug detection mechanism dedicated to put user, which required changes in the
source code of both the Linux kernel and the Bochs instrumentation.
The main problem related to the sanitization of data written with put user
is the fact that it is passed through value and not reference (pointer). Con-
sequently, the first argument of the macro may be a constant, variable, struc-
ture/union field, array item, function return value, or an expression involving
components of any of the above types. Therefore, it is not clear which particular
memory region should be sanitized in each specific case, and it is difficult to iso-
late only the simple cases on the level of either kernel code or instrumentation.
While most CPU architectures supported by Linux have their own imple-
mentation of the put user macro, there is also a generic version declared in
include/asm-generic/uaccess.h. Listing 16 shows the body of put user,
an internal macro which sits at the core of put user. As we can see in line 147,
the expression passed by the caller to be written to user space is evaluated and
stored in a local, helper variable called x. While analyzing the code, we de-
cided to sanitize all memory reads executed as part of the expression evaluation,
as long as they occured in the context of the current function (i.e. not in nested
function calls). To that end, we modified the macro to wrap the initialization of
the x variable with two assembly instructions – prefetcht1 and prefetcht2.
The diff of the change is presented in Listing 17. As a result, all accesses to mem-
ory that was passed to user-mode through put user were placed between the
two artificially inserted instructions. Examples of disassembled code snippets
from a Linux kernel compiled in this manner are shown in Listing 18.
In Bochs, the prefetcht1 and prefetcht2 instructions aren’t emulated with
any dedicated logic, but are instead handled by the BX CPU C::NOP method.
This makes them prime candidates to be used as hypercalls – special opcodes
that don’t have any effect on the execution of the guest system, but are used
to communicate with the emulator. In this case, our instrumentation detects
the execution of prefetcht1 and treats it as a signal to start sanitizing all
38
145 #define __put_user(x, ptr) \
146 ({ \
147 __typeof__(*(ptr)) __x = (x); \
148 int __pu_err = -EFAULT; \
149 __chk_user_ptr(ptr); \
150 switch (sizeof (*(ptr))) { \
151 case 1: \
152 case 2: \
153 case 4: \
154 case 8: \
155 __pu_err = __put_user_fn(sizeof (*(ptr)), \
156 ptr, &__x); \
157 break; \
158 default: \
159 __put_user_bad(); \
160 break; \
161 } \
162 __pu_err; \
163 })
Listing 16: Declaration of the generic put user macro

(include/asm-generic/uaccess.h)
145 #define __put_user(x, ptr) \

146 ({ \
147 __typeof__(*(ptr)) __x = (x); \
148 int __pu_err = -EFAULT; \
149 __chk_user_ptr(ptr); \
150 asm("prefetcht1 (%eax)"); \
151 x = (x); \
152 asm("prefetcht2 (%eax)"); \
153 switch (sizeof (*(ptr))) { \
Listing 17: Changes applied to the put user macro
39
.text:C1027F72 prefetcht1 byte ptr [eax]
.text:C1027F75 mov eax, [ebp+var_B4]
.text:C1027F7B mov [ebp+var_AC], eax
.text:C1027F81 prefetcht2 byte ptr [eax]
[...]
.text:C1035910 prefetcht1 byte ptr [eax]
.text:C1035913 mov eax, [ebp+var_14]
.text:C1035916 mov edx, edi
.text:C1035918 call getreg
.text:C103591D mov [ebp+var_10], eax
.text:C1035920 prefetcht2 byte ptr [eax]
[...]
.text:C1071AD7 prefetcht1 byte ptr [eax]
.text:C1071ADA mov edx, [ebp+var_1C]
.text:C1071ADD shl edx, 8
.text:C1071AE0 or edx, 7Fh
.text:C1071AE3 mov [ebp+var_10], edx
.text:C1071AE6 prefetcht2 byte ptr [eax]
Listing 18: Instances of the compiled put user macro with added marker
instructions
kernel-mode memory references at the current value of ESP. If any uninitialized

memory is read while operating in this mode, the potential bug is reported
in the same way as any typical kernel infoleak. Accordingly, the prefetcht2
instruction disables the strict sanitization mechanism.
Thanks to this design, Bochspwn Reloaded was capable of detecting refer-
ences to uninitialized variables passed – directly or indirectly, as part of larger
expressions – to the put user macro. The approach is not free from occasional
false positives, as the uninitialized memory may not always be propagated to
user space (e.g. if it is an unused function argument). However, bug candidates
flagged by the tool can be manually investigated to determine the root cause
and severity of the findings, which is facilitated by the open-source nature of the
Linux kernel. Overall, we consider it an effective way of instrumenting put user
calls, which seems an otherwise difficult task for dynamic binary instrumenta-
tion.
Use of uninitialized memory in Linux As discussed in Section 3.5.2, we

initially didn’t have much success with the kernel infoleak detection used against
Linux. In order to test the correctness of the taint tracking and to generate more
output to analyze, we extended the scope of detection to all reads of uninitial-
ized memory. On a technical level, this was achieved by enabling the strict
sanitization described in the previous section for the entirety of the kernel ex-
ecution, and not just for blocks of code between the prefetcht instructions.
In this mode, every unique read of uninitialized memory was flagged as a po-
tential bug.
40
Considering that our instrumentation didn’t recognize if the uninitialized
data had any influence on the kernel control flow, the output logs included
a number of false-positives where leftover memory was read (e.g. while being
copied), but never actually used in a meaningful way. On the upside, the vol-
ume of the reports turned out to be manageable, and the false-positives were
relatively easy to filter out upon brief analysis of the kernel code.
It is also important to note that while uses of uninitialized memory are typ-
ically real bugs, they are often functional and not security problems. For exam-
ple, a majority of the issues identified by Bochspwn Reloaded in Linux had very
little to no security impact. Nonetheless, proposed patches for all discovered
bugs were submitted and subsequently accepted by the kernel developers.
3.2 Ancillary functionality

The core capabilities of the project discussed above are designed to enable ef-
fective detection of kernel infoleaks, but provide little extra information about
the context of the bugs or the overall system state. In order to produce verbose
reports that could be used to deduplicate vulnerabilities and quickly understand
their root cause and impact, we implemented a series of additional features in
the instrumentation. While they are not critical to the correct functioning of
the tool, they have proved highly useful while analyzing and reproducing the
identified flaws. The technical details behind these ancillary mechanisms are
explained in the sections below.
3.2.1 Keeping track of kernel modules

Keeping track of kernel modules is the first step to symbolizing raw addresses
from the guest system. The most important modules to track are ntoskrnl.exe
in Windows and vmlinux in Linux. Based on their location in memory, the in-
strumentation can obtain the addresses of all other drivers loaded in the system.
One reliable way to acquire the base addresses of these core images is to
intercept all writes to Model-Specific Registers (MSR) in search of addresses
residing at known offsets inside of the desired kernel executables. An example of
such MSRs are registers storing the addresses of system call entrypoints, used to
transfer control flow upon the execution of SYSENTER and SYSCALL instructions.
In protected mode, the register in question is 0x176 (SYSENTER EIP MSR), while
in long mode, it is 0xC0000082 (IA32 LSTAR).
There are several advantages of using MSRs to determine the kernel base:
• The system call interface is initialized early during system boot.
• The syscall entrypoints are located at fixed offsets in the kernel images.
• Bochs provides a BX INSTR WRMSR instrumentation callback which receives
notifications about all MSR writes taking place in the emulated system.
41
Once the base address of the primary image is established, listing other
loaded modules is a matter of traversing through simple system-specific linked
lists of driver descriptors in the guest memory. In Bochspwn Reloaded, the
traversing is performed every time the instrumentation encounters an address
that cannot be associated with any of the currently known modules.
In Windows, a static nt!PsLoadedModuleList variable points to the head
of a doubly-linked list consisting of LDR MODULE structures. In Linux, the cor-
responding head pointer is named modules, and the list is made of module
structures. In both cases, each such structure describes a single kernel driver,
including its name, base address and size in memory. The layouts of these lists
are illustrated in Figures 7 and 8.
ntoskrnl.exe
LDR_MODULE LDR_MODULE LDR_MODULE
BaseAddress BaseAddress BaseAddress

SizeOfImage SizeOfImage SizeOfImage
BaseDllName BaseDllName BaseDllName
Flink Flink Flink

Blink Blink Blink
PsLoadedModuleList
Figure 7: Windows kernel module linked list layout
vmlinux
struct module struct module struct module
next next next

prev prev prev
name name name
base base base

modules size size size
Figure 8: Linux kernel module linked list layout
42
3.2.2 Unwinding stack traces
Collecting full stack traces upon detecting kernel infoleaks helps both dedupli-
cate the bugs (to avoid flooding logs with multiple instances of the same issue),
and understand how the execution flow reached the affected code. Depending
on the bitness of the guest system, the goal can be achieved in different ways.
On 32-bit builds of Windows and Linux, the stack trace has a simple and
consistent form – consecutive stack frames of nested functions are chained to-
gether through stack frame pointers (saved values of the EBP register), with the
current stack frame being pointed to by EBP itself. At any point of execution,
it is possible to iterate through this chain to unwind the call stack and save the
observed return addresses, as shown in Figure 9. This logic was implemented
in our tool for 32-bit guest systems.
ESP
locals
foo()
EBP
saved ebp
return address
locals
bar()
saved ebp
return address
locals
syscall()
saved ebp
return address
Trap frame
Figure 9: Example of a typical stack layout in x86 kernel code
In 64-bit Windows, the RBP register is not saved as part of the stack frame
creation anymore, and thus the callstack cannot be traversed by the instrumen-
tation without access to additional debug information. The necessary informa-
tion is provided by Microsoft’s debug symbols (.pdb files) corresponding to the
kernel modules. Regardless of stack trace unwinding, our tool loads symbol files
for every new detected driver for the purpose of address symbolization (see Sec-
tion 3.2.3). In the presence of these debug symbols, obtaining the full callstack
can be achieved using the StackWalk64 API function [90] from the Debug Help
Library (DbgHelp). As part of the input, the function expects to receive the
full CPU context, which is available through the internal BX CPU C object. An-
other required primitive is a pointer to a custom ReadMemoryRoutine function,
invoked by DbgHelp to read the virtual memory of the target process/kernel.
In our case, it is a simple wrapper around read lin mem [72], a helper function
43
for reading guest system memory. After putting these pieces together, retriev-
ing the full call stack can be implemented as a straightforward loop over the
StackWalk64 call.
Support for 64-bit builds of Linux was not implemented, so we didn’t study
the problem in that configuration.
3.2.3 Address symbolization

The translation of raw addresses from the guest address space to meaningful
symbols is essential for producing useful output reports. In this section, we de-
scribe how address symbolization was implemented on the Windows and Linux
platforms.
Microsoft Windows The symbols for nearly all Windows system files are
available on the Microsoft symbol server [85] and can be downloaded using the
SymChk tool (symchk.exe), shipped with Debugging Tools for Windows. The
command line syntax for downloading the symbols for a specific file to a chosen
directory is as follows:
C:\> symchk.exe <driver> /s srv*<directory>*http://msdl.microsoft.com/

download/symbols
Prior to starting each Bochspwn Reloaded session, we extracted all ker-

nel modules from the guest filesystem, and downloaded their corresponding
.pdb files to a single directory on the host machine. Later during Bochs run
time, when our instrumentation discovered any new driver loaded in the em-
ulated system, it would instantly look up symbols for that driver using the
SymLoadModule64 function [92]. Thanks to this, the Debug Help Library was
always up to date with regards to the locations of kernel images in memory
and their associated debug information. When the bug reports were gener-
ated, translating addresses into function names and offsets was achieved with a
SymFromAddr [91] call, which performed the overall symbolization process and
returned with the desired information.
Linux In the Linux-specific part of Bochspwn Reloaded, we didn’t perform

live address symbolization while the reports were produced by the tool. Instead,
kernel addresses were printed in verbatim, and the log file was later subject to
post-processing by a custom Python script. The script matched all kernel-mode
addresses, symbolized them using the standard addr2line utility, and substi-
tuted them accordingly in the output file. This was possible by compiling the
Linux kernel with ASLR disabled (CONFIG RANDOMIZE BASE=n), which guaran-
teed that the addresses in the instrumented system memory were compatible
with the static vmlinux file on disk.
44
3.2.4 Breaking into the kernel debugger
The textual reports generated by our tool are verbose, but not always sufficient
to understand the underlying vulnerabilities. One example of such scenario
is when the leaked uninitialized data travels a long way (i.e. is copied across
multiple locations in memory) before arriving at the final memcpy to user space.
In cases like this, it is useful to attach a kernel debugger to the emulated system
and learn about the memory layout and contents, kernel objects involved in the
disclosure, the user-mode caller that invoked the affected system call, and any
other important details about the state of the execution environment.
Both Windows and Linux support remote kernel debugging through COM
ports. In the Bochs emulator on a Windows host, guest COM ports can be redi-
rected to named Windows pipes by including the following line in the bochsrc
configuration file:
com1: enabled=1, mode=pipe-server, dev=\\.\pipe\name
With the above configuration set up, it is possible to attach the WinDbg or
gdb debuggers to the tested systems. However, this by itself is not enough to
put the debuggers to good use, as long as we can’t break the execution of the
OS precisely at the moment of each new detected disclosure. To achieve this,
we need assistance from the Bochs instrumentation.
In the x86(-64) architectures, breakpoints are installed by placing an int3
instruction (opcode 0xcc) at the desired location in code. This is no different
in the emulated environment. To stop the kernel execution and pass control
to the debugger, the instrumentation only needs to write the 0xcc byte to EIP
or RIP, depending on the system bitness. This suffices to break into the kernel
debugger, but since the overwritten first byte of the next instruction is never
restored in the above logic, the user has to do it themselves by looking up the
value of the original byte in the executable image of the driver in question.
To address this inconvenience, our instrumentation declares a callback for
the bx instr interrupt event invoked by Bochs every time an interrupt is
generated, including the #BP trap triggered by the injected breakpoint. In
that handler, we restore the previously saved value of the overwritten byte, thus
bringing the original instruction back to its original form even before control
is transferred to the kernel debugger. From the perspective of the emulated
system, the #BP exception is generated for no apparent reason, as the int3
instruction only lasts in memory for as long as it is needed to be fetched and
executed by Bochs, and disappears shortly after.
Listing 19 presents an example WinDbg log from a brief investigation of a
bug identified by Bochspwn Reloaded. The debugger informs us that a break-
point exception was triggered, but the disassembly of memory under RIP shows
the original sub rdx, rcx instruction, as further confirmed with the u com-
mand. To find out more about the circumstances of the leak, we check the
value of the RCX register (the destination argument of memmove), to make sure
that it points into user-mode memory. To follow up, we dump the memory
45
area between RDX and RDX+R8-1, where RDX is the source address and R8 is the
number of bytes to copy. At offsets 0x4 through 0x7, we can observe the 0xaa
values, which is the filler byte for pool allocations. We can therefore assume
that these bytes are the subject of the disclosure, and as we look closely at the
contents of the buffer, we can also deduce that it is a UNICODE STRING structure
followed by the corresponding textual data. Lastly, we display the stack trace
to establish how the control flow reached the current point. After our analysis
is completed, we can continue the system execution with the g command.
3.2.5 Address space visualization

The taint information stored by Bochspwn Reloaded at any given time of the
guest system run time may be used to display the layout of the kernel address
space as an image. Graphical representation provides valuable insight into the
inner workings of the dynamic allocator and memory manager. It also makes
it possible to observe memory consumption under various conditions, as well as
compare the behavior of different versions of an operating system. Finally, it
also proves useful for security research – for example, memory visualization may
help in fine-tuning pool spraying techniques, or calibrating a pool-massaging
algorithm designed to facilitate the exploitation of use-after-free vulnerabilities.
Several projects for visualizing process and kernel address spaces have been
devised so far. MemSpyy [99] and MemoryDisplay [106] display maps of the
virtual memory of Windows userland applications. KernelMAP [67] displays the
locations of drivers, objects, locks and CPU structures in the Windows kernel
memory. An improved version of the tool called MemMAP [15] also shows
kernel stacks and GDI objects, as well as memory maps of ring 3 programs.
Radocea and Wicherski further demonstrated successful visualization of page
tables on Android, OS X and iOS [20]. The main advantage of performing the
visualization in an x86 emulator is that the address space can be observed at
very early boot stages, which would be otherwise impossible to achieve from
within the guest system itself.
As an exercise, we implemented the feature in Bochspwn Reloaded for 32-bit
guests. Every N seconds, a separate thread in the emulator would make a snap-
shot of the current state of the shadow memory, marking each of the existing
memory pages5 as free, stack or heap. The snapshots were then converted to
bitmaps of a 1024x{256,512} resolution, where each pixel represented a single
page. Depending on their type, the pixels were colored black (unused), green
(stack) or red (heap/pool). Figure 10 shows the Windows 7 kernel memory
layout after 40 minutes of booting up and running several initial ReactOS tests.
Figure 11 shows the layout on Windows 10 after 120 minutes of run time. Fi-
nally, Figure 12 shows the memory state of Ubuntu 16.04 after 60 minutes of run
time, which included booting up, running the Trinity fuzzer [13] and starting a
few initial tests from the Linux Test Project [8].
5 There are 0x80000 (524,288) kernel pages on Windows, and 0x40000 (262,144) kernel
pages on Linux in the 3G/1G memory split configuration.
46
kd> g
Break instruction exception - code 80000003 (first chance)
nt!memmove+0x3:
fffff800‘026fc603 482bd1 sub rdx,rcx
kd> u
nt!memmove+0x3:
fffff800‘026fc603 482bd1 sub rdx,rcx
fffff800‘026fc606 0f829e010000 jb nt!memmove+0x1aa
fffff800‘026fc60c 4983f808 cmp r8,8
fffff800‘026fc610 7262 jb nt!memmove+0x74
fffff800‘026fc612 f6c107 test cl,7
fffff800‘026fc615 7437 je nt!memmove+0x4e
fffff800‘026fc617 f6c101 test cl,1
fffff800‘026fc61a 740c je nt!memmove+0x28
kd> ? rcx
Evaluate expression: 2554392 = 00000000‘0026fa18
kd> db /c 8 rdx rdx+r8-1

fffff8a0‘00ff1d20 2e 00 30 00 aa aa aa aa ..0.....
fffff8a0‘00ff1d28 30 1d ff 00 a0 f8 ff ff 0.......
fffff8a0‘00ff1d30 5c 00 44 00 65 00 76 00 \.D.e.v.
fffff8a0‘00ff1d38 69 00 63 00 65 00 5c 00 i.c.e.\.
fffff8a0‘00ff1d40 48 00 61 00 72 00 64 00 H.a.r.d.
fffff8a0‘00ff1d48 64 00 69 00 73 00 6b 00 d.i.s.k.
fffff8a0‘00ff1d50 56 00 6f 00 6c 00 75 00 V.o.l.u.
fffff8a0‘00ff1d58 6d 00 65 00 32 00 00 00 m.e.2...
kd> k
# Child-SP RetAddr Call Site
00 fffff880‘03d0b8c8 fffff800‘02a75319 nt!memmove+0x3
01 fffff880‘03d0b8d0 fffff800‘02938426 nt!IopQueryNameInternal+0x289
02 fffff880‘03d0b970 fffff800‘0294cfa8 nt!IopQueryName+0x26
03 fffff880‘03d0b9c0 fffff800‘0297713b nt!ObpQueryNameString+0xb0
04 fffff880‘03d0bac0 fffff800‘0271d283 nt!NtQueryVirtualMemory+0x5fb
05 fffff880‘03d0bbb0 00000000‘77589ada nt!KiSystemServiceCopyEnd+0x13
[...]
kd> g
Listing 19: WinDbg log after an infoleak in nt!IopQueryNameInternal is

flagged by Bochspwn Reloaded (CVE-2018-0894)
47
Figure 10: Windows 7 kernel address space layout; green: stack pages, red:
pool pages
Figure 11: Windows 10 kernel address space layout; green: stack pages, red:
pool pages
Figure 12: Ubuntu 16.04 kernel address space layout; green: stack pages, red:
heap pages
48
3.3 Performance
In this section, we examine the general performance of the instrumentation by
comparing its CPU and memory usage to the same guest operating systems
run in a regular virtual machine (VirtualBox) and a non-instrumented Bochs
emulator. The aim of this section is to provide a general overview of the over-
head associated with the proposed approach; the numbers presented here are
approximate and should not be considered as accurate benchmarks.
The testing was performed on a workstation equipped with an Intel Xeon
E5-1650 v4 @ 3.60 GHz CPU, 64 GB of DDR4-2400 RAM, and a Samsung 850
PRO 512 GB SSD (SATA 3); the guest systems were assigned 1 processor core
and 2 GB of physical memory.
3.3.1 CPU overhead

The boot times of the systems tested as part of this research in three different
setups are shown in Table 2. Boot time is defined as the time period from
cold boot to a fully loaded log-on screen. The data was obtained by testing
each configuration in several (typically three) attempts and accepting the lowest
value, in hope to minimize the influence of the host machine’s background load
on the results.
Unsurprisingly, it is visible that operating systems run significantly slower
in a software emulator compared to a virtual machine; in our case, the slow-
down factor was observed to be around 13-18X. Under such non-instrumented
emulators, the guests usually ran at a rate of 80-150 million IPS (Instructions
Per Second ) with reasonably responsive graphical interfaces. As the boot times
all fall within a 10-minute mark, we consider the execution speed provided by
Bochs to be sufficient for research purposes.
In order to keep the inevitable overhead modest, it is essential for any ad-
ditional instrumentation to be implemented as efficiently as possible, especially
when using frequently invoked callbacks such as bx instr before execution
or bx instr lin access. In Bochspwn Reloaded, the average emulation speed
drops to around 30-50 MIPS, which seems to correctly correlate with the in-
crease in measured boot times of around 2.2 to 2.65X compared to unmodified
Bochs. Even with the aggregate slowdown multiplier between 32-50X in rela-
tion to a regular VM, all tested environments remained responsive, and the test
cases and programs designed to extend the kernel code coverage successfully
completed within 24 hours of booting the emulated systems up.
Windows 7 Windows 10 Ubuntu 16.10

x86 x64 x86 x64 x86
VirtualBox 00:07 00:11 00:24 00:22 00:25
Bochs 01:33 02:20 05:47 06:52 06:38
Bochspwn Reloaded 03:45 05:55 12:43 18:04 17:35
Table 2: Time from cold boot to log-on screen in tested configurations (mm:ss)
49
Windows 7 Windows 10 Ubuntu 16.10
x86 x64 x86 x64 x86
Bochs 2126
Bochspwn Reloaded 8300 4220 8290 6628 5205
Table 3: Peak memory consumption of the bochs.exe process between cold

boot and log-on screen of the guest systems (in MB)
3.3.2 Memory overhead

The peak memory usage of the bochs.exe emulator process for both non-
instrumented and instrumented Bochs is shown in Table 3, measured as the
“Peak Private Bytes” value shown by Process Explorer [10] when the log-on
screen was loaded by the guest OS. In the non-instrumented case, memory con-
sumption was fixed at around 2126 MB for all tested systems, the bulk of which
corresponds to the 2 GB of emulated physical memory assigned to the guests.
Any memory allocated beyond that point can be accounted as an overhead of
the Bochspwn Reloaded project.
In the 32-bit versions of our instrumentation, the shadow memory and other
auxilliary metadata are allocated statically and induce a constant overhead of
6 GB for Windows and 3 GB for Linux, as explained in Section 3.1.1. This
is correctly reflected in the observed data. On the contrary, the memory con-
sumption is variable in case of emulated x64 platforms, where both the taint
and allocation origin descriptors are allocated on demand. For Windows 7 and
Windows 10 64-bit, the memory overhead at the time of the log-on screen was
2094 MB and 4502 MB, respectively. Somewhat unintuitively, the presented
virtual memory usage is higher for the x86 architecture than x64, but it should
be noted that (a) the overhead on x64 could grow beyond the 6 GB bound-
ary during system run time, while it remains constant on x86, and (b) accesses
to metadata on x64 come with extra computational costs related to the custom
memory overcommittment mechanism and the overhead of hash map operations.
In summary, the memory requirements of the project are significant, but can
be met by any modern workstation with at least 16 GB of RAM installed.
3.4 Testing
The effectiveness of any instrumentation-based vulnerability detection is as good
as the code coverage achieved against the tested software. With this in mind,
we attempted to maximize the coverage of the analyzed kernels using publicly
available tools and methodology, while running them inside our tool. In the
subsections below, we explain how the testing was carried out on the Windows
and Linux platforms.
50
As part of the research project, we ran Bochspwn Reloaded against Windows 7
and Windows 10, both 32 and 64-bit builds. The middle version of the operating
system – Windows 8.1 – was excluded from the analysis. The assumption behind
this approach was that the testing of Windows 7 would reveal bugs that had
been internally fixed by Microsoft in newer systems but not backported to the
older ones, while instrumentation of Windows 10 would uncover bugs in the
most recently introduced kernel code. In this context, there was very little
attack surface in Windows 8.1 that wouldn’t be already covered by the testing
of the two other versions, and as such, we considered it redundant to analyze
all three major releases of the OS.
In terms of bitness, in our experience and by intuition, a majority of kernel
infoleaks are cross-platform and affect both x86 and x64 builds of the code.
This is related to the fact that most root causes of the bugs, such as explicitly
uninitialized variables, structure fields, arrays etc. are bitness-agnostic and re-
produce in both execution modes. The subset of issues limited to x86 is very
narrow, as there are no fundamental reasons for the presence of such 32-bit only
disclosures other than low level platform-specific code (e.g. exception handling).
On the other hand, x64-only bugs may exist due to the fact that the width of
certain data types (size t, pointers etc.) and their corresponding alignment
requirements increase from 4 to 8 bytes, which in turn:
• creates new alignment holes in structures; one example being the standard
UNICODE STRING structure,
• extends the size of unions, which may misalign their fields and introduce
new uninitialized bytes, as is the case in IO STATUS BLOCK.
Throughout most of 2017, we developed instrumentation for 32-bit guest

systems, which yielded a total of 48 CVEs assigned by Microsoft. At the end of
that year, we implemented support for 64-bit platforms, which uncovered further
17 x64-specific issues. The results of the project are detailed in Section 3.5.
The testing process of Windows in the emulated environment involved the
following steps aimed to maximize the kernel code coverage:
• Booting up the system.

• Starting and navigating through a number of default programs, acces-
sories and administrative tools, including File Explorer, Internet Explorer,
Calculator, Notepad, WordPad, Command Line, Registry Editor, Paint,
Windows Media Player, XPS Viewer, Control Panel and so forth.
• Running more than 800 winetests and rostests [2] that are part of the
ReactOS project. They are API tests designed to check that the ReactOS
implementations behave in the same way that Microsoft APIs do. Due
to the fact that they invoke a variety of system interfaces and test them
extensively, they are the perfect means to extend the kernel code coverage.
51
• Running and navigating through the code samples [32] for the “Windows
Graphics Programming: Win32 GDI and DirectDraw” book [31].
• Running around 30 NtQuery test suites developed specifically to uncover
memory disclosure in the specific subset of the system calls for querying
information about objects in the system. These test programs are further
discussed in Section 5.4 “Differential syscall fuzzing”.
• Shutting down the system.
As shown, Windows was tested on a best effort basis, and there is much room
for improvement in terms of coupling existing dynamic binary instrumentation
schemes with new ways of exploring a more substantial portion of the kernel
code. However, we believe that the steps we took allowed us to identify most of
the easily discoverable bugs that other parties could likely run across.
3.4.2 Linux
The Linux platform subject to examination was Ubuntu Server 16.10 32-bit
with a custom-compiled kernel v4.8. Support for x64 builds of the kernel was
never implemented. As part of the testing, we executed the following actions in
the system:
• Booting up the system.

• Logging in via SSH.
• Running several standard command-line utilities operating on processes,

file system and the network.
• Running the full set of the Linux Test Project (LTP) [8] unit tests.
• Running the iknowthis [105] syscall fuzzer for a day.
• Running the Trinity [13] syscall fuzzer for a day.

• Shutting down the system.
Currently, the state-of-the-art Linux syscall fuzzer is syzkaller [11]. It is

compatible with the various Sanitizers (most importantly KASAN [12]), uses
code coverage information to guide the fuzzing, and has identified dozens of bugs
in the Linux kernel to date [5]. Unfortunately, syzkaller couldn’t be run inside
Bochspwn Reloaded, as it only supported the x64 and arm64 architectures while
our tool was limited to x86 builds of the system.
3.5 Results
In this section, we present a summary of previously unknown vulnerabilities
discovered by running Bochspwn Reloaded against Windows and Linux.
52
Throughout the development of the project in 2017 and early 2018, we ran mul-
tiple iterations of the instrumentation on the then-latest builds of Windows 7
and 10. All issues found in each session were promptly reported to Microsoft in
accordance with the Google Project Zero disclosure policy. The first identified
vulnerability was reported on March 1, 2017 [59], and the last one was sent
to the vendor on January 22, 2018 [53]. In that time period, we progressively
improved the tool and introduced new features, which enabled us to regularly
uncover new layers of infoleaks. This is reflected in the history of the bug re-
ports – for example, the first 12 reported issues were pool-based disclosures,
because the handling of stack allocations was added a few weeks later. Simi-
larly, we initially developed instrumentation for 32-bit guest systems, and only
implemented support for 64-bit platforms at the end of 2017. This explains why
all of the x64-specific leaks we discovered were patched between February and
April 2018.
In total, we filed 73 issues describing Windows kernel memory disclosure to
user-mode in the Project Zero bug tracker. Out of those, 69 were closed as
“Fixed”, 2 as “Duplicate” and 2 as “WontFix”. The duplicate reports were
caused by the fact that some leaks reported as separate issues were determined
by the vendor to be caused by a single vulnerability in the code. The WontFix
cases were valid bugs, but turned out to be only reachable from a privileged
dwm.exe process and hence didn’t meet the bar to be serviced in a security
bulletin.
Microsoft classified the problems as 65 unique security flaws. The discrep-
ancy between the number of CVEs and “Fixed” bug tracker entries stems from
the fact that the vendor combined several groups of issues into one, e.g. if they
considered the bugs to have a common logic root cause. In cases where this
was inconsistent with our assessment and we viewed them as separate bugs, we
marked the corresponding tracker issues as “Fixed” instead of “Duplicate”.
Two example reports generated for the CVE-2017-8473 and CVE-2018-0894
vulnerabilities are shown in Listings 20 and 21. A complete summary of mem-
ory disclosure vulnerabilities found by the tool is presented in Tables 4 and 5,
while Figure 13 illustrates the distribution of disclosed memory types. We ex-
pect that stack leaks are more prevalent than pool leaks due to the fact that
a majority of temporary objects constructed by the kernel in system call han-
dlers are allocated locally. Furthermore, Figure 14 shows a classification of the
bugs based on the kernel modules they were found in. The core ntoskrnl.exe
executable image was the most suspectible to infoleaks, likely due to its large
attack surface and our intensified testing of the NtQuery syscall family. The
graphical win32k.sys driver was also affected by a significant number of prob-
lems, both in regular syscall handlers and a mechanism known as user-mode
callbacks. Lastly, several individual issues were found in other drivers such as
partmgr.sys, mountmgr.sys or nsiproxy.sys, mostly in the handling of user-
accessible IOCTLs with complex output structures.
53
------------------------------ found uninit-access of address 94447d04
[pid/tid: 000006f0/00000740] { explorer.exe}
READ of 94447d04 (4 bytes, kernel--->user), pc = 902df30f
[ rep movsd dword ptr es:[edi], dword ptr ds:[esi] ]
[Pool allocation not recognized]

Allocation origin: 0x90334988 (win32k.sys!__SEH_prolog4+00000018)
Destination address: 1b9d380

Shadow bytes: 00 ff ff ff Guest bytes: 00 bb bb bb
Stack trace:
#0 0x902df30f (win32k.sys!NtGdiGetRealizationInfo+0000005e)
#1 0x8288cdb6 (ntoskrnl.exe!KiSystemServicePostCall+00000000)
Listing 20: Report of the CVE-2017-8473 bug detected on Windows 7 32-bit
------------------------------ found uninit-copy of address fffff8a000a63010
[pid/tid: 000001a0/000001a4] { wininit.exe}

COPY of fffff8a000a63010 ---> 1afab8 (64 bytes), pc = fffff80002698600
[ mov r11, rcx ]
Allocation origin: 0xfffff80002a11101
(ntoskrnl.exe!IopQueryNameInternal+00000071)
--- Shadow memory:
00000000: 00 00 00 00 ff ff ff ff 00 00 00 00 00 00 00 00 ................
00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
--- Actual memory:
00000000: 2e 00 30 00 aa aa aa aa 20 30 a6 00 a0 f8 ff ff ..0..... 0......
00000010: 5c 00 44 00 65 00 76 00 69 00 63 00 65 00 5c 00 \.D.e.v.i.c.e.\.
00000020: 48 00 61 00 72 00 64 00 64 00 69 00 73 00 6b 00 H.a.r.d.d.i.s.k.
00000030: 56 00 6f 00 6c 00 75 00 6d 00 65 00 32 00 00 00 V.o.l.u.m.e.2...
--- Stack trace:
#0 0xfffff80002698600 (ntoskrnl.exe!memmove+00000000)
#1 0xfffff80002a11319 (ntoskrnl.exe!IopQueryNameInternal+00000289)
#2 0xfffff800028d4426 (ntoskrnl.exe!IopQueryName+00000026)
#3 0xfffff800028e8fa8 (ntoskrnl.exe!ObpQueryNameString+000000b0)
#4 0xfffff8000291313b (ntoskrnl.exe!NtQueryVirtualMemory+000005fb)
#5 0xfffff800026b9283 (ntoskrnl.exe!KiSystemServiceCopyEnd+00000013)
Listing 21: Report of the CVE-2018-0894 bug detected on Windows 7 64-bit
54
Both
Kernel pools
24
39
Kernel stack
Figure 13: Distribution of disclosed memory types between flaws discovered

in Windows
Other drivers
35 23
ntoskrnl.exe
win32k.sys
Figure 14: Distribution of affected kernel modules between flaws discovered

in Windows
55
CVE ID Component Fix Date Leaked bytes x64 only
CVE-2017-0258 ntoskrnl.exe May 2017 8
CVE-2017-0259 ntoskrnl.exe May 2017 60
CVE-2017-8462 ntoskrnl.exe June 2017 1
CVE-2017-8469 partmgr.sys June 2017 484
CVE-2017-8484 win32k.sys June 2017 5
CVE-2017-8488 mountmgr.sys June 2017 14
CVE-2017-84896 ntoskrnl.exe June 2017 6 or 72
CVE-2017-8490 win32k.sys June 2017 6672
CVE-2017-8491 volmgr.sys June 2017 8
CVE-2017-8492 partmgr.sys June 2017 4
CVE-2017-8564 nsiproxy.sys July 2017 13
CVE-2017-02997 ntoskrnl.exe August 2017 2
CVE-2017-8680 win32k.sys September 2017 Arbitrary
CVE-2017-11784 ntoskrnl.exe October 2017 192
CVE-2017-11785 ntoskrnl.exe October 2017 56
CVE-2017-11831 ntoskrnl.exe November 2017 25
CVE-2018-0746 ntoskrnl.exe January 2018 12
CVE-2018-08108 win32k.sys February 2018 4 X
CVE-2018-0813 win32k.sys March 2018 4 X
CVE-2018-0894 ntoskrnl.exe March 2018 4 X
CVE-2018-0899 videoprt.sys March 2018 20 X
CVE-2018-0972 ntoskrnl.exe April 2018 8
Table 4: A summary of discovered Windows pool memory disclosure bugs
6 The CVE was assigned to a generic mitigation of zeroing the Buffered I/O out-
put buffer [39]. It fixed two bugs in IOCTLs handled by the \Device\KsecDD and
\\.\WMIDataDevice devices, filed in the Google Project Zero bug tracker as issues 1147
and 1152.
7 A patch for the vulnerability first shipped in June 2017. After it was proven ineffective,
Microsoft released a revised version of the fix in August of the same year.
8 The CVE collectively covers four different memory disclosure bugs found in win32k.sys
user-mode callbacks – one pool-based and three stack-based leaks. They were filed in the
Google Project Zero bug tracker as issues 1467, 1468, 1485 and 1487.
9 The vulnerability disclosed uninitialized pool memory on Windows 7, and stack memory
on Windows 10.
56
CVE ID Component Fix Date Leaked bytes x64 only
CVE-2017-0167 win32k.sys April 2017 20
CVE-2017-0245 win32k.sys May 2017 4
CVE-2017-8470 win32k.sys June 2017 50
CVE-2017-8471 win32k.sys June 2017 4
CVE-2017-8472 win32k.sys June 2017 7
CVE-2017-8473 win32k.sys June 2017 8
CVE-2017-8475 win32k.sys June 2017 20
CVE-2017-8477 win32k.sys June 2017 104
CVE-2017-8677 win32k.sys September 2017 8
CVE-2017-11853 win32k.sys November 2017 12
CVE-2018-0810 win32k.sys February 2018 4 or 8
CVE-2018-0832 ntoskrnl.exe February 2018 4
CVE-2018-0896 msrpc.sys March 2018 8 X
CVE-2018-0968 ntoskrnl.exe April 2018 4 X
CVE-2018-0970 ntoskrnl.exe April 2018 4 or 16
CVE-2018-0975 ntoskrnl.exe April 2018 4 or 56
Table 5: A summary of discovered Windows stack memory disclosure bugs
57
------------------------------ found uninit-access of address f5733f38
========== READ of f5733f38 (4 bytes, kernel--->kernel), pc = f8aaf5c5
[ mov edi, dword ptr ds:[ebx+84] ]
[Heap allocation not recognized]
Allocation origin: 0xc16b40bc: SYSC_connect at net/socket.c:1524
Shadow bytes: ff ff ff ff Guest bytes: bb bb bb bb
Stack trace:
#0 0xf8aaf5c5: llcp_sock_connect at net/nfc/llcp_sock.c:668
#1 0xc16b4141: SYSC_connect at net/socket.c:1536
#2 0xc16b4b26: SyS_connect at net/socket.c:1517
#3 0xc100375d: do_syscall_32_irqs_on at arch/x86/entry/common.c:330
(inlined by) do_fast_syscall_32 at arch/x86/entry/common.c:392
Listing 22: Report of a bug in llcp sock connect on Ubuntu 16.10 32-bit
3.5.2 Linux
We implemented support for 32-bit Linux kernels in April 2017, including the
same infoleak detection logic that had been successfully used for Windows. As
a result of instrumenting Ubuntu 16.10 for several days, we detected a single
minor bug – disclosure of 7 uninitialized kernel stack bytes in the processing
of specific IOCTLs in the ctl ioctl function (drivers/md/dm-ioctl.c). The
routine handles requests sent to the /dev/control/mapper device, which is only
accessible by the root user, significantly reducing the severity of the issue. We
identified the problem on April 20, but before we were able to submit a patch,
we learned that it had been independently fixed by Adrian Salido in commit
4617f564c0 [16] on April 27.
The lack of success in identifying new infoleaks in Linux can be explained by
the vast extent of work done to secure the kernel in the past. Accordingly, we
decided to extend the detection logic to include all references to uninitialized
memory. By making this change, we intended to uncover other, possibly less
dangerous bugs, where uninitialized memory was used in a meaningful way, but
not directly copied to the user space. Thanks to this approach, we discovered
further 15 bugs – one disclosure of uninitialized stack memory through AF NFC
sockets (see the corresponding report presented in Listing 22) and 14 lesser,
functional issues in various subsystems of the kernel, with limited security im-
pact. The combined results of this effort are enumerated in Table 6.
During and after the triage of the output reports, we noticed that some
of our findings collided with the work of independent researchers and devel-
opers; i.e. several bugs had been fixed days or weeks prior to our discovery,
or reported by other parties shortly after we submitted patches. Most colli-
sions occurred with the KernelMemorySanitizer project [6], which was actively
developed and used to test Linux in the same period. Consequently, we only
submitted 11 patches for the 16 uncovered bugs, as the rest had already been
addressed. This example is very illustrative of the velocity of improvements
applied to Linux, and the scope of work done to eliminate entire vulnerability
classes.
58
File Fix commit Collision Mem. Type
net/nfc/llcp sock.c 608c4adfca X Stack L
drivers/md/dm-ioctl.c 4617f564c0 X Stack L
net/bluetooth/l2cap sock.c
net/bluetooth/rfcomm/sock.c d2ecfa765d Stack U
net/bluetooth/sco.c
net/caif/caif socket.c 20a3d5bf5e Stack U
net/iucv/af iucv.c e3c42b61ff Stack U
net/nfc/llcp sock.c f6a5885fc4 Stack U
net/unix/af unix.c defbcf2dec Stack U
kernel/sysctl binary.c 9380fa60b1 X Stack U
fs/eventpoll.c c857ab640c X Stack U
kernel/printk/printk.c 5aa068ea40 X Heap U
net/decnet/netfilter/dn rtmsg.c dd0da17b20 Heap U
net/netfilter/nfnetlink.c f55ce7b024 Heap U
fs/ext4/inode.c 2a527d6858 X Stack U
net/ipv4/fib frontend.c c64c0b3cac X Heap U
fs/fuse/file.c 68227c03cb Heap U
arch/x86/kernel/alternative.c fc152d22d6 Stack U
Table 6: A summary of discovered uninitialized memory bugs in Linux.

L: kernel memory disclosure; U: use of uninitialized memory.
4 Windows bug reproduction techniques

Bug reproduction is an integral part of hunting for memory disclosure vulnerabil-
ities, as it (a) helps make sure that a problem identified by Bochspwn Reloaded
is an actual bug in the kernel and not in the instrumentation, and (b) helps
vendors clearly observe the problem in the same way we see it in our test envi-
ronment. For these reasons, we have developed a unified set of system configura-
tion steps and methods of writing our proof-of-concept programs to achieve the
most deterministic behavior possible. In the paragraphs below, we discuss our
approaches to reproducing Windows pool-based and stack-based disclosures. In
both cases, we used a VirtualBox VM with the desired guest operating system,
typically Windows 7 or 10, 32-bit or 64-bit.
In a number of cases, attempting reproduction was preceded by long hours
spent on analyzing Bochspwn Reloaded reports and the internal state of the
system in the kernel debugger, in order to fully understand the root cause of
the bugs first.
Kernel pools Reproducing pool-based leaks in Windows is a relatively easy

task. The system supports a special option in the kernel called Special Pool [79],
which helps detect pool buffer overflows and underflows by allocating each mem-
ory chunk at the beginning or end of a separate page. It can be controlled with
a default program called Driver Verifier Manager (verifier.exe).
59
One useful property of the mechanism is that it fills the body of every new
allocation with a unique, repeated marker byte. If leaked to user-mode, these
variable marker bytes can be seen at common offsets of the PoC program output.
The only requirement is to determine which kernel module requests the affected
allocation (which is usually answered by Bochspwn Reloaded) and enable special
pool for that module. Then, we should see output similar to the following after
starting the proof-of-concept twice (on the example of CVE-2017-8491):
D:\>VolumeDiskExtents.exe
00000000: 01 00 00 00 39 39 39 39 ....9999
00000008: 00 00 00 00 39 39 39 39 ....9999
00000010: 00 00 50 06 00 00 00 00 ..P.....
00000018: 00 00 a0 f9 09 00 00 00 ........
D:\>VolumeDiskExtents.exe
00000000: 01 00 00 00 2f 2f 2f 2f ....////
00000008: 00 00 00 00 2f 2f 2f 2f ....////
00000010: 00 00 50 06 00 00 00 00 ..P.....
00000018: 00 00 a0 f9 09 00 00 00 ........
It is worth noting that special pool is only applied to allocations of up to

around 4080 bytes, which should however not be a significant restriction for
practical purposes. As the size of the special pool increases with the amount of
RAM, we recommend assigning the test VM with enough physical memory to
ensure that all special pool requests can be fulfilled.
Kernel stack Reliable reproduction of kernel stack disclosure is more trou-

blesome than leaks from the pools, as there is no official or documented way to
fill kernel stack frames with marker bytes at the time of allocation. These mark-
ers are essential, because otherwise the leaked bytes contain non-deterministic
leftover memory which is often either uninteresting or plainly filled with zeros.
This makes it difficult to distinguish legitimate syscall output from uninitialized
data, let alone hope that the proof-of-concept output observed by the vendor
will be similar to ours.
In order to address this problem, we decided to use a kernel stack spraying
technique [50, 46]. As mentioned in Section 2.2.3, many system call handlers
in Windows load small chunks of input data into stack-based buffers, and only
request pool allocations when larger memory blocks are required. This opti-
mization reduces the number of allocator invocations by using the capacity of
the stack. As it turns out, the meaning of small and large is flexible in the
Windows kernel, with some syscalls using local helper buffers of up to several
kilobytes in size before turning to pool allocations. This allows user-mode pro-
grams to fill a vast portion of the kernel stack with fully controlled data, which
proves very useful for demonstrating memory disclosure vulnerabilities.
60
Candidates of syscalls capable of spraying the kernel stack can be identified
by looking for functions with large stack frames whose names start with “Nt”.
In our reproducers, we used the following services:
• nt!NtMapUserPhysicalPages – Sprays up to 4096 bytes on 32-bit systems

and 8192 bytes on 64-bit systems.
• win32k!NtGdiEngCreatePalette – Sprays up to 1024 bytes of the ker-
nel stack, and may return a new GDI palette object which needs to be
destroyed later on.
Both system calls can be invoked through their corresponding wrappers in

the ntdll.dll and gdi32.dll libraries, which is useful for cross-version system
compatibility. The overall process of using kernel stack spraying to achieve
reliable reproduction is illustrated in Figure 15. Upon implementing the idea,
it should be simple to spot the markers in the syscall output where kernel data
is leaking, as shown on the example of CVE-2017-8473:
D:\>NtGdiGetRealizationInfo.exe
00000000: 10 00 00 00 03 01 00 00 ........
00000008: 2e 00 00 00 69 00 00 46 ....i..F
00000010: 41 41 41 41 41 41 41 41 AAAAAAAA
1. Spray the kernel stack with 2. Trigger the bug, and observe marker bytes at uninitialized
a recognizable pattern offsets
Kernel stack Kernel stack User-mode memory
41 41 41 41 41 41 41 41 41 41 41 41
41 41 41 41 41 41 41
00 41
50 41
A8 41
00 41 41 00
41 50
41 A8
41 00
41 41 41
41 41 41 41 41 41 41 41 41
9B 41
01 41
00 41
00 41 41 41
9B 41
01 41
00 41
00
41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41
00 00 19 00 48 45 00 00 19 00 48 45
41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41
00 00 98 44 00 00 00 00 98 44 00 00
41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41
30 0A 00 00 00 05 30 0A 00 00 00 05
41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41
41 41 41 41 41 41 41 41 41 41 41
00 41
00 41 41 41 41 00
41 00
41
41 41 41 41 41 41 41 41 41 41 41
00 41
00 41 41 41 41 41
00 41
00
41 41 41 41 41 41 41 41 41 41 41 41
41 41 41 41 41 41 41 41 41 41 41 41
41 41 41 41 41 41 41 41 41 41 41 41
Figure 15: Kernel stack spraying used for memory disclosure reproduction
61
5 Alternative detection methods
Kernel memory disclosure can be detected with a variety of techniques, which
complement each other in how they approach the problem. Full system emula-
tion is one of them, and while effective, it is limited by the reachable kernel code
coverage. In the following sections, we discuss alternative methods of identifying
infoleaks, outlining their feasibility, pros and cons, and the results of evaluating
some of them against the Windows kernel.
5.1 Static analysis

Static analysis was shown to be successful in finding Linux kernel bugs in the
past, both by analyzing the source code for simple, known-vulnerable pat-
terns [97, 98], and by running full-fledged memory tainting during compilation
to identify execution paths which may lead to memory disclosure [45]. The
prime advantage of the method over dynamic techniques is the fact that it is
capable of processing entire code bases without the need to actually execute
them. On the other hand, depending on the design decisions and implementa-
tion, they may yield multiple false positives and/or false negatives. For exam-
ple, the UniSan authors admitted [45] that due to limited time resources and a
conservative approach taken by their project, they were only able to manually
analyze about 20% of all allocations marked as potentially leaking uninitialized
memory. Overall, we consider static analysis to be a powerful technique that
complements other dynamic methods well.
Given the closed-source nature of Windows, existing tools devised for Linux
cannot be tested against Microsoft’s operating system by anyone other than the
vendor itself. It is difficult to estimate how much work is done internally in this
area by Microsoft due to limited public documentation of such efforts. However,
we are hoping that Microsoft will continue to invest in static analysis and will
adopt the current state-of-the-art techniques, as they are in the best position
to design and run advanced source-code analysis on the Windows kernel.
Binary-level static analysis is another viable option, but much more difficult
to implement effectively, as a significant portion of useful information – such as
structure definitions and variable types – is stripped away during compilation.
For example, we expect that Peiró’s model checking [97] could be ported to
Windows with relative ease, to search for stack-based objects which are only
partially initialized in their parent function and subsequently copied to user-
mode. However, we haven’t tested this method in practice.
5.2 Manual code review

Earlier in the paper, we stated that memory disclosure in software written in C
may be hardly visible to the naked eye. This is especially true for uninitial-
ized structure fields, padding bytes and unions with diversely sized members.
Nonetheless, there are also certain vulnerable code constructs specific to Win-
dows – discussed in Sections 2.2.2 and 2.2.3 – which are easier to spot during
62
code review. In May 2017, we decided to test this assumption by performing
a cursory manual analysis of the Windows kernel in search of low hanging fruit
issues that could have been missed by Bochspwn Reloaded due to incomplete
code coverage.
On Linux, a good starting point is to locate references to the copy to user
function and begin the audit from there, progressively going back through the
code to track the initialization of each part of the copied objects. On Windows,
there is no such dedicated function, and all user↔kernel memory operations are
achieved with direct pointer manipulation or regular memcpy calls. However, us-
age of the ProbeForWrite function [89] may be a valuable hint, because pointer
sanitization often directly preceeds writing data back to userland. To reduce
the scope of the review, we decided to focus on the top-level syscall handlers,
since they typically have a concise execution flow and easy to understand life-
time of objects in memory. Table 7 shows the combined numbers of direct and
inlined memcpy calls (as detected by IDA Pro) in Nt functions on 32-bit builds
of Windows 7 and 10, with the January 2018 patch installed. Assuming that
approximately half of the calls were used to copy output data, we assessed the
extent of required manual analysis to be manageable within several days.
Windows 7 Windows 10
Core (ntoskrnl.exe etc.) 139 165
Graphics (win32k.sys etc.) 288 405
Table 7: Numbers of memcpy calls in Windows syscall handlers
While examining the system call implementations, we discovered two new

vulnerabilities, now assigned CVE-2017-8680 and CVE-2017-8681. They were
canonical examples of the “arbitrary output length” and “fixed-sized array” vari-
ants discussed earlier in this paper. The pseudo code of the affected win32k.sys
routines is shown in Listings 23 and 24.
In the case of the first bug, the cbBuffer argument was used to request a
pool allocation of that size in line 6, and the entire memory region was sub-
sequently copied back to userland in line 12. The problem was that the dy-
namic memory was not zero-initialized upon allocation, and more importantly,
that the syscall disregarded the actual number of bytes written to the buffer
by GreGetGlyphOutlineInternal, instead only operating on the user-provided
size. The leak had relatively high severity, as it allowed the attacker to disclose
arbitrarily many bytes from allocations of controlled lengths. It also demon-
strated that spotting a system call parameter passed directly to the 3rd argu-
ment of memcpy is frequently a red flag, as it indicates that a controlled number
of bytes is returned to the caller instead of only the relevant output.
In the second bug, a temporary local array of 256 bytes (128 wide characters)
was used to store a textual monitor description, later copied to the caller. The
buffer was not pre-initialized, and was copied to user-mode as a whole regardless
of the actual length of the string. In our test environment, this resulted in a
disclosure of exactly 128 bytes from the kernel stack, but the number could be
63
1 DWORD NTAPI NtGdiGetGlyphOutline(
2 ...,
3 DWORD cbBuffer,
4 LPVOID lpvBuffer
5 ) {
6 LPVOID KernelBuffer = Allocate(cbBuffer);
7
8 DWORD Status = GreGetGlyphOutlineInternal(..., KernelBuffer, cbBuffer);
9
10 if (Status != GDI_ERROR) {
11 ProbeForWrite(lpvBuffer, cbBuffer, 1);
12 memcpy(lpvBuffer, KernelBuffer, cbBuffer);
13 }
14
15 Free(KernelBuffer);
16
17 return Status;
18 }
Listing 23: Vulnerable win32k!NtGdiGetGlyphOutline function pseudo code
1 NTSTATUS NTAPI NtGdiGetPhysicalMonitorDescription(

2 HANDLE hMonitor,
3 DWORD dwSizeInChars,
4 LPWSTR szDescription
5 ) {
6 WCHAR KernelBuffer[PHYSICAL_MONITOR_DESCRIPTION_SIZE];
7
8 if (dwSizeInChars != PHYSICAL_MONITOR_DESCRIPTION_SIZE) {
9 return STATUS_INVALID_PARAMETER;
10 }
11
12 NTSTATUS Status = CMonitorAPI::GetMonitorDescription(
13 hMonitor, PHYSICAL_MONITOR_DESCRIPTION_SIZE, KernelBuffer);
14
15 if (NT_SUCCESS(Status)) {
16 ProbeForWrite(szDescription, sizeof(KernelBuffer), 1);
17 memcpy(szDescription, KernelBuffer, sizeof(KernelBuffer));
18 }
19
20 return Status;
21 }
Listing 24: Vulnerable win32k!NtGdiGetPhysicalMonitorDescription

function pseudo code
64
30 ms_exc.registration.TryLevel = -2;
31 if ( v16 )
32 memset(v16, 0, a5);
33 v10 = v16;
34 v15 = GreGetGlyphOutlineInternal(a1, a2, a3, &v13, a5, v16, &v14, a8);
Listing 25: Diff of win32k!NtGdiGetGlyphOutline in Windows 7 and 10
different depending on the model of the physical monitor.

Both vulnerabilities were fixed by Microsoft in the September 2017 bulletin
by adding adequate memset calls to erase the affected memory regions prior to
operating on them. In conclusion, our experiment showed that manual code
review may be an effective approach to finding kernel infoleaks which could be
otherwise missed by other techniques.
5.3 Cross-version kernel binary diffing

At the time of this writing, there are three main versions of Windows under
active support – Windows 7, 8.1 and 10 [80]. While Windows 7 and 8.1 combined
still have just over 50% of the worldwide Windows market share [103], Microsoft
is known for introducing a number of structural security improvements and
sometimes even ordinary bugfixes only to the most recent Windows platform.
This makes it potentially possible to detect 0-day vulnerabilities in the older
systems merely by spotting subtle changes in the corresponding code in different
versions of Windows.
When filing an issue for CVE-2017-8680 in the Project Zero bug tracker [60],
we realized that the bug only affected Windows 7 and 8.1, while it had been
internally fixed by Microsoft in Windows 10. Listing 25 shows the difference
between the vulnerable and patched versions of the code, as decompiled by
Hex-Rays and compared by Diaphora. The patch is evident, as a new memset
call added in a top-level syscall handler is very likely a fix for a memory disclosure
issue.
Based on this discovery, we suspected that there could be more such extra
memset references in newer systems, revealing unpatched bugs in older ones.
To verify this, we compared [52] decompiled listings of the ntoskrnl.exe and
win32k.sys modules10 between Windows 7, 8.1 and 10, looking for discrepancies
in the usage of memset. The numbers of added invocations of the function in
respective kernels are summarized in Table 8.
10 Including tm.sys, win32kbase.sys and win32kfull.sys on Windows 10.
65
ntoskrnl.exe win32k.sys
functions syscalls functions syscalls
Windows 7 vs. 10 153 8 89 16
Windows 8.1 vs. 10 127 5 67 11
Table 8: New memset calls in Windows 10 in relation to older systems
We then manually reviewed all instances of memset calls added to system

call handlers. As a result, we discovered two new vulnerabilities:
• CVE-2017-8684 – A stack-based disclosure of about 88 bytes in win32k!

NtGdiGetFontResourceInfoInternalW, affecting Windows 7 and 8.1. It
was a variant of the “arbitrary output length” pattern discussed in Sec-
tion 2.2.3.
• CVE-2017-8685 – A stack-based disclosure of 1024 bytes in win32k!Nt-
GdiEngCreatePalette, affecting Windows 7.
The diffs that enabled us to recognize the vulnerabilities are shown in List-
ings 26 and 27.
14 v12 = 0;
15 v13 = 0;
16 memset(&v14, 0, 0x5Cu);
17 v11 = 0;
18 ms_exc.registration.TryLevel = 0;
Listing 26: Diff of win32k!NtGdiGetFontResourceInfoInternalW between

Windows 7 and Windows 10
15 v16[0] = 0;
16 memset(&v16[1], 0, 0x3FCu);
17 v14 = 0;
18 if ( a2 > 0x10000 )
19 return 0;
Listing 27: Diff of win32k!NtGdiEngCreatePalette between Windows 7 and

Windows 8.1
In July 2017, we learned that another flaw could have been also discovered
the same way – CVE-2017-11817, a leak of over 7 kB of uninitialized kernel
pool memory to NTFS metadata, discussed in Section 6.1.2. The vulnera-
bility was present in Windows 7, but a memset function had been added to
Ntfs!LfsRestartLogFile in Windows 8.1 and later, potentially exposing the
bug to researchers proficient in binary diffing. When reported to Microsoft, it
was patched in the October 2017 Patch Tuesday.
Interestingly, we have also observed the opposite situation, where a memset
call was removed from the kernel in newer systems, thus introducing a bug.
66
This was the case for CVE-2018-0832, a stack-based disclosure of four bytes
in nt!RtlpCopyLegacyContextX86 [64]. The buffer created with alloca was
correctly zero-initialized in Windows 7, but not in Windows 8.1 or 10. The root
cause behind adding the regression is unclear.
As demonstrated above, most fixes for kernel infoleaks are obvious both
when seen in the source code and in assembly. The binary diffing required to
identify inconsistent usage of memset doesn’t require much low-level expertise
or knowledge of operating system internals. Therefore, we hope that these bugs
were some of the very few instances of such easily discoverable issues, and we
encourage software vendors to make sure of it by applying security improvements
consistently across all supported versions of their software.
5.4 Differential syscall fuzzing

In Section 2.5, we mentioned that several authors have proposed multi-variant
program execution to identify use of uninitialized memory. The concept was
implemented in various forms in the DieHard [21], BUDDY [44] and SafeInit [94]
projects. We have found that a similar technique can be used to effectively
identify Windows kernel infoleaks. If each newly allocated stack and pool-based
region is filled with a single-byte pattern that changes over time, then a user-
mode program may analyze the output of two subsequent syscall invocations
looking for unequal, but repeated bytes at common offsets. Every occurrence of
such a pattern most likely manifests an information disclosure vulnerability.
Filling stacks and pools with marker bytes may be accomplished in a few
different ways. For pools:
1. The analyzed system may be run under the Bochs emulator, with an
instrumentation that sets the bytes of all new allocations once they are
requested.
2. As explained in Section 4 “Windows bug reproduction techniques”, the
desired effect is a part of the default behavior of the Special Pool option in
Driver Verifier [79]. When the feature is enabled for any specific module,
all pool regions returned to that module are filled with a marker byte
(which changes after each new allocation).
3. Low-level hooks may be installed on kernel functions such as ExAllocate-
PoolWithTag to briefly redirect code execution and set the allocation’s
bytes. This option was used by fanxiaocao and pjf [27]. It may be prob-
lematic on 64-bit Windows platforms due to the mandatory Patch Guard
mechanism.
In turn, the following methods may be applied to the stack:
1. Again, the analyzed system may be run under Bochs, with an instrumenta-
tion responsible for writing markers to allocated stack frames (as described
in Section 3.1.2 “Tainting stack frames”).
67
2. Stack-spraying primitives (as detailed in Section 4 “Windows bug repro-
duction techniques”) may be used prior to invoking the tested system calls,
each time with a different value used for the spraying.
3. Low-level hooks may be installed on system call entry points such as
nt!KiFastCallEntry, to poison the kernel stack before passing execution
to the syscall handlers. This option was used by fanxiaocao and pjf [27],
and similarly to their pool hooks, it may not work well with Patch Guard
on 64-bit system builds.
Options (1) have the best allocation coverage, especially in terms of stack
poisoning, but come at the cost of a significant slowdown. Options (2) have
marginally worse coverage, but can be used on bare metal or in virtual ma-
chines, without digging in low-level system internals. Options (3) can be useful
when a custom memory-poisoning mechanism is needed, and system perfor-
mance plays a substantial role. We have successfully tested methods (1) and (2)
and confirmed their effectiveness in discovering Windows security flaws.
Besides poisoning all newly requested allocations, it is necessary to develop
a user-mode harness to invoke the tested syscalls and analyze their output.
This is strongly related to the more general field of effective Windows system
call fuzzing, which is still an open problem. Windows 10 RS3 32-bit supports
as many as 460 native syscalls [66] and 1174 graphical ones [65]. Ideally, the
harness should be aware of the prototypes of all system calls in order to invoke
them correctly, reach their core functionality and interpret the output. To our
best knowledge, there isn’t currently any existing framework that would enable
us to run this kind of analysis. As a result, we decided to take a more basic
approach and focus on a specific subset of the system calls.
As we were looking for disclosure of uninitialized memory, we were primarily
interested in services designed to query information and return it back to the
caller. One such family of syscalls consists of kernel functions whose names
start with the NtQuery prefix. Their purpose is to obtain various types of
information regarding different objects and resources present in the system.
There is currently a total of 60 such syscalls, with each type of kernel object
having a corresponding service, e.g. NtQueryInformationProcess for processes,
NtQueryInformationToken for security tokens, NtQuerySection for sections
and so forth. Conveniently, a majority of these system calls share a common
definition. An example prototype of the NtQueryInformationProcess handler
is shown in Listing 28.
NTSTATUS WINAPI NtQueryInformationProcess(

_In_ HANDLE ProcessHandle,
_In_ PROCESSINFOCLASS ProcessInformationClass,
_Out_ PVOID ProcessInformation,
_In_ ULONG ProcessInformationLength,
_Out_opt_ PULONG ReturnLength
);
Listing 28: Prototype of the nt!NtQueryInformationProcess system call
68
The parameters, in the original order, are:
• ProcessHandle – a handle to the queried object of a specific type, in this

case a process.
• ProcessInformationClass – a numeric identifier of the requested infor-
mation in the range of [0..N ], where N is the maximum supported informa-
tion class on the given operating system. The names of some information
classes can be found in Windows SDK header files, but few of them are
officially documented. The classes are also subject to change between dif-
ferent versions of the operating system. In our experiment, we regarded
the argument as a “black box” and brute-forced all possible values in the
conservative range of [0..255]. Non-supported classes were filtered out
based on the STATUS INVALID INFO CLASS return code.
• ProcessInformation – a pointer to a user-mode buffer receiving the out-

put data.
• ProcessInformationLength – length of the supplied output buffer. Every
information class expects the length to be equal to the size of the output
object(s), or to fall within a specific range. As the extent of correct values
is limited, the argument can also be successfully brute-forced, or alterna-
tively set according to the value returned through ReturnLength.
• ReturnLength – a pointer to a variable which receives the size of the
requested information.
While the basic premise is shared across all NtQuery services, some of them
may diverge slightly from the prototype shown above. For example, they may
accept textual paths instead of object handles, use the IO STATUS BLOCK struc-
ture instead of a simple ReturnLength variable, or not take the information
class as an argument, because only one type of data is returned.
We briefly reverse-engineered all 60 system calls in question, and determined
that 31 of them were either simple enough that they didn’t require dynamic test-
ing, or could only be accessed by users with administrative rights. Accordingly,
we developed test cases (in the form of standalone test programs) for the remain-
ing 29 non-trivial syscalls. By running them on Windows 7 and 10 (32/64-bit)
in both regular VMs and in Bochspwn Reloaded, we discovered infoleaks in a
total of 14 services across 23 different information classes. The results of the
experiment are summarized in Table 9. A majority of the bugs were stack-based
leaks caused by uninitialized fields and padding bytes in the output structures.
We believe that the experiment demonstrates that differential syscall fuzzing
may be effective in identifying kernel memory disclosure, and that certain groups
of system calls are more suspectible to the problem than others due to their
design and purpose.
69
System call (NtQuery...) Information class
AttributesFile —
FileBothDirectoryInformation (3)
DirectoryFile
FileIdBothDirectoryInformation (37)
FullAttributesFile —
JobObjectBasicLimitInformation (2)
JobObjectExtendedLimitInformation (9)
InformationJobObject
JobObjectNotificationLimitInformation (12)
JobObjectMemoryUsageInformation (28)
ProcessVmCounters (3)
InformationProcess ProcessImageFileName (27)
ProcessEnergyValues (76)
InformationResourceManager ResourceManagerBasicInformation (0)
InformationThread ThreadBasicInformation (0)
InformationTransaction TransactionPropertiesInformation (1)
InformationTransactionManager TransactionManagerRecoveryInformation (4)
InformationWorkerFactory WorkerFactoryBasicInformation (7)
Object ObjectNameInformation (1)
SystemPageFileInformation (18)
SystemInformation MemoryTopologyInformation (138)
SystemPageFileInformationEx (144)
MemoryBasicInformation (0)
MemoryMappedFilenameInformation (2)
VirtualMemory
MemoryImageInformation (6)
MemoryPrivilegedBasicInformation (8)
VolumeInformationFile FileFsVolumeInformation (1)
Table 9: Vulnerable combinations of NtQuery syscalls and information classes
70
5.5 Taintless Bochspwn-style instrumentation
A full system instrumentation similar to Bochspwn Reloaded is capable of de-
tecting some instances of disclosure of uninitialized memory even without the
notion of shadow memory and taint tracking, as long as it is able to examine
all user-mode memory writes originating from the kernel. One way to achieve
this is to analyze all syscall output data in search of information that shouldn’t
normally be found there and thus manifests infoleaks – for example, valid ring 0
addresses. As mentioned earlier in the paper, kernel addresses are among the
most common types of data found in uninitialized memory, and since they are
also relatively easy to recognize (especially on 64-bit platforms), they may be
used as highly reliable bug indicators. On the downside, the approach can-
not detect leaks that don’t contain any addresses, or leaks whose individual
continuous chunks are smaller than the width of a pointer.
To improve the above design, it is possible to poison the stack and heap/pool
allocations with a specific marker byte, and search for sequences of that byte
in data written by the kernel to ring 3. This is trivial to achieve from the level
of Bochs instrumentation, but might be more difficult when running the tested
system in a regular virtual machine. The various available avenues of poisoning
newly created kernel memory areas are detailed in Section 5.4 “Differential
syscall fuzzing”. An important quality of this technique is that it overcomes
any potential limitations of taint tracking and propagation, as it is based solely
on the analysis of actual memory contents in the guest system. On the other
hand, it may be prone to false-positives, in cases where legitimate syscall output
data happens to contain a sequence of the marker bytes. However, that risk can
be significantly reduced by running several instances of the instrumentation with
different marker bytes, and cross-checking the results to only analyze reports
which reproduce across all of the sessions.
The taintless approach was successfully employed by fanxiaocao and pjf of
IceSword Lab to discover 14 Windows kernel infoleaks in 2017 [27], and by the
grsecurity team to identify an unspecified number of bugs in the Linux kernel
in 2013 [33]. We also used a variant of this method to hunt for disclosure of
uninitialized memory to mass-storage devices, as discussed in Section 6.1.
71
6 Other data sinks
In addition to the user-mode address space, uninitialized kernel memory may
become available to unauthorized code through other data sinks, such as mass-
storage devices (internal and external hard drives, USB flash drives, DVDs etc.)
and the network. The UniSan tool [45] accounted for those possibilities in Linux
by including the sock sendmsg and vfs write functions in the list of sinks
together with copy to user, hence universally intercepting most exit points
where data escapes the kernel. The idea proved effective, as more than 50%
of new kernel vulnerabilities discovered by the project (10 out of 19) leaked
memory through the socket sink.
In our experimentation, we focused on recognizing disclosure of Windows
kernel memory through common file systems, using an enhanced variant of the
taintless technique discussed in Section 5.5. In the following subsections, we
outline the inner workings of this side project and present the results, includ-
ing one particularly interesting vulnerability found in the ntfs.sys file system
driver (CVE-2017-11817).
6.1 Mass-storage devices

Storage devices are a peculiar type of data sinks for kernel infoleaks, as any
realistic attack which involves them requires physical access to the target ma-
chine and/or user interaction. Imagine, for example, that person A shares some
files with person B on a flash drive. In the presence of kernel memory leaks to
that drive, person A may believe that they are only sharing the contents of the
few explicitly copied files, while in fact they may also be disclosing fragments
of leftover data from their operating system. Furthermore, such unknowingly
leaked memory saved to disks of various forms could be used as a source of
forensic evidence, potentially revealing some of the actions performed by the
user while having the storage devices connected to their computer. The severity
and practical usefulness of such bugs is limited in comparison to leaks over the
network or the kernel/user memory boundary, but we believe they still require
proper attention from operating system vendors.
Some operating systems, such as Windows, automatically mount the parti-
tions found on storage devices attached to the machine, even when the system is
locked. Consequently, any memory that is leaked immediately upon mounting
the volume opens up the possibility to automatically exfiltrate data through
USB or other physical ports, bypassing the requirement of user interaction.
6.1.1 Detection
A canonical approach to sanitizing memory saved to disk in the Windows kernel
would be to develop a file system filter driver, and have it intercept all write
operations performed on the mounted volumes. Then, the driver would invoke
a specially designed hypercall, which would send a signal to the instrumentation
indicating that the taint of a specific memory area needs to be checked. However,
72
instead of going that route, we decided to test a more experimental method of
taintless detection, based on poisoning all new allocations in the system with
a known marker byte, and scanning the guest’s disk image in search of those
markers. The technique was potentially more effective as it was not subject
to the innacuracy of taint propagation, but it came at the cost of losing some
valuable context information – we no longer learned about the call stack and
system state at the exact time of the leak taking place. The only information
available to us was the location of the disclosed bytes on disk, and the type of
the leaked memory, if we used different markers for the stack and the pools.
This put us in an inconvenient spot where we could detect leaks, but it was
difficult to analyze them and determine their root cause. In order to solve this
problem, we tried to use the marker bytes to encode more information than
just the fact that they originated from a kernel allocation – specifically, the
addresses of the code that made the allocations. As the testing was performed
on x86 versions of Windows, the pointers were 4-byte wide. Thus, a 16-byte
pool buffer allocated at address 0x8b7ad4ab would no longer be padded with
the 0xaa byte:
aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa aa
but instead would be filled with a repeated address of its origin in little-endian
format:
ab d4 7a 8b ab d4 7a 8b ab d4 7a 8b ab d4 7a 8b
Addresses of the allocation origins encountered by the instrumentation were

saved to an output file, and later used by a separate script to periodically scan
the disk image in search of kernel infoleaks. One evident shortcoming was that
only leaks of a minimum of 4 bytes could be detected this way, but we considered
this an acceptable trade-off for the additional information gained for the larger
leaks.
The above method is probabilistic and heavily relies on the disk not con-
taining the “magic” sequences of bytes in unrelated contexts. While the design
is susceptible to false positives, we took steps to reduce their extent. Firstly, we
changed the modes of the virtual disks subject to testing from flat to volatile
in the bochsrc configuration file. This caused Bochs to save all disk modifica-
tions to a separate changelog file, thus enabling us to look for the markers in
the disk delta instead of the entire multi-gigabyte image, which vastly decreased
the chance of finding unrelated sequences of bytes colliding with our markers.
Secondly, we observed that values resembling valid kernel-mode addresses
still occurred often in the changelog file, introducing noise in the output of our
tool. Through trial and error, we determined that collisions happened much
less frequently when the origins were first combined (xored) with another fixed
32-bit value, in our case 0x3bdd78ec for pool markers and 0x7345c6f1 for stack
markers. By applying this logic to the previous example, the allocation would
be pre-initialized with bytes shown on the next page.
73
47 ac a7 b0 47 ac a7 b0 47 ac a7 b0 47 ac a7 b0
With the above scheme implemented, we were able to successfully identify

uninitialized kernel memory written to persistent storage. Thanks to the en-
coded origin addresses, we could also investigate the related code and usually
understand the root cause of the issues. While the available information was
still limited, it allowed Microsoft to fix all of the reported bugs. Any remaining
false positives were filtered out either by ignoring addresses of kernel functions
completely unrelated to file system handling, or manually inspecting bytes sur-
rounding the marker sequences in the changelog file to determine if they were
legitimate leaks or accidental collisions.
6.1.2 Testing and results

In this experiment, we tested 32-bit builds of Windows 7 and 10, combined
with file systems such as FAT, FAT32, exFAT and NTFS. We scanned both
the system volume and an additional, separate partition created and attached
to the guest system to determine which leaks only occur on the system drive,
and which ones can be triggered by connecting an external device. Once the
emulated OS was booted, we performed a series of operations on the file systems,
including:
• Creating and deleting files of various sizes,

• Creating and deleting deeply nested directory structures,
• Renaming files and directories,
• Traversing the existing directory structure,
• Reading from and writing to existing files,
• Enabling and disabling compression and encryption (where applicable),
• Modifying the hidden, read-only, and security attributes of files and direc-
tories.
As a result of the testing, we didn’t identify any issues in the handling of file
systems from the FAT family. On the other hand, we discovered a multitude of
both stack-based and pool-based leaks to NTFS volumes, originating from 10
different locations in the ntfs.sys driver (see Table 10). The size of the leaks
was generally modest and ranged between 4-12 bytes in continuous blocks of
uninitialized memory. Most of them affected all currently supported versions
of Windows. They were initially submitted to Microsoft in a single report on
July 7, 2017, followed by an update with more details on the bugs on August 25.
A summary of our communication with MSRC can be found in the correspond-
ing bug #1325 in the Project Zero bug tracker [58]. The issues were collectively
fixed in a single bulletin as CVE-2017-11880 in November 2017.
74
Leaked allocation origin Memory Windows
ntfs!NtfsInitializeReservedBuffer+0x20 Pool 7
ntfs!NtfsAddAttributeAllocation+0xb16 Pool 7-10
ntfs!NtfsCheckpointVolume+0xdcd Pool 7-10
ntfs!NtfsDeleteAttributeAllocation+0x12d Pool 7-10
ntfs!CreateAttributeList+0x1c Pool 7-10
ntfs!NtfsCreateMdlAndBuffer+0x95 Pool 7-10
ntfs!NtfsDeleteAttributeAllocation+0xf Stack 7-10
ntfs!NtfsWriteLog+0xf Stack 7-10
ntfs!NtfsAddAttributeAllocation+0xf Stack 7-10
ntfs!NtfsCreateAttributeWithAllocation+0xf Stack 7-10
Table 10: A summary of minor kernel infoleaks found in the ntfs.sys driver
One other, particularly dangerous vulnerability found by the project was a

pool-based disclosure of about 7500 bytes in the ntfs!LfsRestartLogFile func-
tion [61], addressed by Microsoft separately in October 2017 as CVE-2017-11817.
The severity of the flaw stemmed from the fact that it resulted in saving a very
substantial amount of old kernel memory to the $LogFile pseudo-file, and it
didn’t require user interaction – the bug was triggered automatically every time
a new NTFS volume was mounted in the system. Coupled with the fact that
Windows mounts all connected storage devices even when the system is locked,
this enabled attackers with physical access to the victim’s powered-on machine
to exfiltrate uninitialized kernel memory through the USB port. Remarkably,
the issue only affected Windows 7, as it had been internally fixed in Windows 8
and later by adding a memset call to reset the affected memory area. The prob-
lems associated with inconsistent patching across supported system versions are
explained in Section 5.3 “Cross-version kernel binary diffing”.
6.2 Outbound network traffic

Network traffic is another example of a potential data sink for uninitialized ker-
nel memory. It is arguably more important than mass-storage devices, as any
such network-related bug could allow fully remote information disclosure, not
otherwise possible through the other discussed channels. On Windows, the de-
tection of such leaks could be implemented by developing a network filter driver
to mark all packets passing through for sanitization by the instrumentation. An-
other option is a taintless approach similar to the one discussed in the previous
section; poisoning all kernel allocations with a recognizable byte pattern, and
searching for those patterns in the network traffic captured during the testing
of the target system. We didn’t perform any network-related experimentation
during this project, and as such it remains an open research area.
75
7 Future work
Disclosure of uninitialized memory is far from a solved problem, and there
are multiple avenues for improvement on all levels of work – both specific to
Bochspwn Reloaded and the wider topic of mitigating this class of vulnerabili-
ties in software. They are discussed in the paragraphs below.
Accurate taint tracking. In our implementation, taint was propagated only

for memory copied with the memcpy function or its inlined equivalent (the
rep movs instruction). This limitation may generate false negatives in cases
where data is copied indirectly through CPU registers, which is becoming a
common optimization scheme in modern compilers. This is especially relevant
to Windows 10, where a considerable fraction of memcpy calls with constant
sizes are compiled as a series of mov instructions between memory and registers.
By extending the taint tracking logic to also cover registers without compro-
mising on performance, the instrumentation could track memory and detect
information leaks more effectively.
Extended code coverage. In dynamic binary instrumentation, the extent

of code executed in the tested software marks the upper bound for the volume
of issues that can be identified with this method. In the recent years, coverage-
guided fuzzers have been trending in the security community, with the two
canonical examples being american fuzzy lop [77] for user-mode applications
and syzkaller [11] for the Linux kernel. Such fuzzers are able to progressively
build a growing set of test cases reaching new code paths in the program, and
thus greatly supplement any accompanying instrumentation. Likely due to the
closed-source nature of Windows and related difficulty of implementation, there
currently isn’t a corresponding way of achieving vast kernel code coverage in
that system. Any such project would instantly increase the bug detection rate
of Bochspwn Reloaded and similar tools by a large margin.
Other data sinks. As discussed in Section 6, kernel memory may be dis-

closed not only through virtual memory, but also mass-storage devices and the
network. While we briefly experimented with detecting infoleaks to file systems
on Windows, the work is not complete, and there are still a number of untested
system/data sink combinations and detection techniques left to be evaluated.
Other operating systems. During our experimentation, we tested 32-bit

and 64-bit builds of Windows and a 32-bit build of Linux. The binary-level
approach is most suitable for systems without publicly available source code,
but as the Linux results show, it can also prove useful for platforms which in
theory have more powerful debugging tools at their disposal. Therefore, we
believe that Bochspwn Reloaded could be also succesfully used against other
environments such as Linux x64, macOS and BSD-family systems.
76
Other security domains. We discussed how the characteristics of the C pro-
gramming language contribute to the difficulty of securely passing data between
different security domains in Section 2.1. In principle, the outlined problems are
not specific to interactions between user space and the kernel, and may appear
in any software where low-level objects are passed through shared memory be-
tween components with different privilege levels. In particular, we expect that
both inter-process communication channels (used in sandboxing) and virtual-
ization software may also be suspectible to similar infoleaks, and the concepts
described in this paper should be applicable to these areas.
Other instrumentation types. To date, the two of our most productive

full system instrumentation projects were designed to identify double fetch and
memory disclosure bugs in OS kernels. However, we believe that the tech-
nique has much broader potential, and a number of other execution patterns
can be formulated to detect security bugs or pinpoint sensitive areas of code.
The ideas for various instrumentation types we recognized while working on
Bochspwn Reloaded are documented in Appendix A. For some of them, such
as “unprotected accesses to user-mode memory” or “double writes”, we devel-
oped functional prototypes and used them to uncover several initial bugs, which
shows that they are practical and may help expose further issues in the future.
Exploitation. Despite the amount of attention given to memory disclosure

flaws, especially in the Linux community, little work has been done to investigate
the exploitability of such bugs to achieve the disclosure of specific types of
information, e.g. cryptographic keys, file contents, network traffic and so on.
More research on the subject of manipulating the heap and pools to align the
disclosed allocations with desired freed objects would help better understand the
true severity of kernel infoleaks, beyond their utility in bypassing the KASLR
and StackGuard mitigations.
Adoption of exploit mitigations. Extensive research has been performed

on the subjects of detecting and mitigating infoleaks in the Linux kernel, as
outlined in Section 2.5 “Prior research”. We consider the most outstanding
contributions to be UniSan [45], KMSAN [6] and grsecurity/PaX [4, 9]. Histori-
cally, most such projects were developed as independent tools and patches, and
they were not included in the mainline kernel. We believe that Linux security
would greatly benefit from integrating some of these features into the official
code; more importantly, we also hope that the bug class will receive more recog-
nition from Microsoft in the future. As the creator of Windows, the vendor is
in a prime position to implement, evaluate and possibly deploy the techniques
known in other operating systems. Due to the closed-source nature of Windows,
many of these techniques are not otherwise available to external parties.
77
8 Conclusion
Information disclosure in general, and kernel memory disclosure in particular
are a very specific class of software vulnerabilities, different from the more con-
ventional types such as buffer overflows or use-after-free conditions. They don’t
reveal themselves through crashes or hangs, even though they may be triggered
thousands of times every minute during normal system run time. As they are
usually a by-product of functional user↔kernel communication and very difficult
to spot in the code, they may remain unnoticed for many years.
In this paper, we elaborated on the different root causes and factors con-
tributing to memory disclosure in modern software, and showed that the Win-
dows operating system was affected by kernel infoleaks present in a range of
drivers and system calls. To address the problem, we designed and implemented
a Bochs-based instrumentation to automatically identify such issues during sys-
tem run time. We then proceeded to use it to discover and report over 70 unique
vulnerabilities in the Windows kernel and over 10 lesser bugs in Linux through-
out 2017 and the beginning of 2018. To expand more on the general subject, we
also evaluated alternative techniques for exposing similar leaks, and tackled the
problem of recognizing uninitialized memory escaping the kernel to mass-storage
devices. Finally, we reviewed other applications of system instrumentation to
uncover kernel bugs and sensitive parts of the code.
We are optimistic that with continued work in the areas of detection and
mitigation by OS vendors and compiler developers, the entire class of kernel
infoleaks may be completely eliminated in the foreseeable future.
9 Acknowledgments
We would like to thank Gynvael Coldwind, Jann Horn, Joe Bialek, Matt Miller,
Mathias Krause, Brad Spengler and Solar Designer for reviewing this paper or
parts of it and providing valuable feedback.
78
References
[1] bochs: The Open-Source IA-32 Emulation Project. http://
bochs.sourceforge.net/.
[2] Building Modules. https://www.reactos.org/wiki/Building Modules.
[3] Coccinelle. http://coccinelle.lip6.fr/.
[4] grsecurity. https://grsecurity.net/.

[5] KernelAddressSanitizer: Found Bugs. https://github.com/google/
kasan/wiki/Found-Bugs.
[6] KernelMemorySanitizer. https://github.com/google/kmsan.
[7] KMSAN (KernelMemorySanitier) – Trophies. https://github.com/

google/kmsan#trophies.
[8] Linux Test Project. https://linux-test-project.github.io/.
[9] PaX. https://pax.grsecurity.net/.
[10] Process Explorer. https://docs.microsoft.com/en-us/sysinternals/

downloads/process-explorer.
[11] syzkaller - kernel fuzzer. https://github.com/google/syzkaller.
[12] The Kernel Address Sanitizer (KASAN). https://www.kernel.org/doc/
html/latest/dev-tools/kasan.html.
[13] Trinity. http://codemonkey.org.uk/projects/trinity/.

[14] ISO/IEC 9899:201x Committee Draft N1570, April 2011.
[15] a d 13. MemMAP v0.1.2. http://www.woodmann.com/collaborative/
tools/index.php/MemMAP.
[16] Adrian Salido. dm ioctl: prevent stack leak in dm ioctl call. https:
//git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/
commit/?id=4617f564c06117c7d1b611be49521a4430042287.
[17] Alexander Levin. kmemcheck: remove annotations. https:
commit/?id=4950276672fce5c241857540f8561c440663673d.
[18] Alexander Popov. How STACKLEAK improves Linux kernel security.
https://linuxpiter.com/en/materials/2344. Linux Piter 2017.
[19] Alexander Potapenko. KernelMemorySanitizer against uninitialized mem-
ory. http://www.linuxplumbersconf.org/2017/ocw/proposals/4825.
79
[20] Alexandru Radocea, Georg Wicherski. Visualizing Page Ta-
bles for Local Exploitation: Hacking Like in the Movies.
https://media.blackhat.com/us-13/US-13-Wicherski-Hacking-
like-in-the-Movies-Visualizing-Page-Tables-Slides.pdf,
https://media.blackhat.com/us-13/US-13-Wicherski-Hacking-
like-in-the-Movies-Visualizing-Page-Tables-WP.pdf. Black Hat
USA 2013.
[21] Berger, Emery D and Zorn, Benjamin G. DieHard: probabilistic memory
safety for unsafe languages. In Acm sigplan notices, volume 41, pages
158–168. ACM, 2006.
[22] Chen, Haogang and Mao, Yandong and Wang, Xi and Zhou, Dong and
Zeldovich, Nickolai and Kaashoek, M Frans. Linux kernel vulnerabilities:
State-of-the-art defenses and open problems. In Proceedings of the Second
Asia-Pacific Workshop on Systems, page 5. ACM, 2011.
[23] Chow, Jim and Pfaff, Ben and Garfinkel, Tal and Rosenblum, Mendel.
Shredding Your Garbage: Reducing Data Lifetime Through Secure Deal-
location. In USENIX Security Symposium, pages 22–22, 2005.
[24] Cowan, Crispan and Pu, Calton and Maier, Dave and Walpole, Jonathan
and Bakke, Peat and Beattie, Steve and Grier, Aaron and Wagle, Perry
and Zhang, Qian and Hinton, Heather. Stackguard: Automatic adaptive
detection and prevention of buffer-overflow attacks. In USENIX Security
Symposium, volume 98, pages 63–78. San Antonio, TX, 1998.
[25] Dan Rosenberg. Subject: CVE request: multiple kernel stack memory
disclosures. http://www.openwall.com/lists/oss-security/2010/09/
25/2.
[26] Dan Rosenberg. Vulnerabilities. http://www.vulnfactory.org/vulns/.
Listed under “memory disclosure”.
[27] fanxiaocao and pjf of IceSword Lab (Qihoo 360). Automatically Dis-
covering Windows Kernel Information Leak Vulnerabilities. http:
//www.iceswordlab.com/2017/06/14/Automatically-Discovering-
Windows-Kernel-Information-Leak-Vulnerabilities en/.
[28] fanxiaocao of IceSword Lab (Qihoo 360). great! I am also got multi
case of “double-write”. yet I report about 20 kernel pool address leak to
MS. but they change the bar. https://twitter.com/TinySecEx/status/
943410888119218176.
[29] fanxiaocao of IceSword Lab (Qihoo 360). new type of info-leak. https:
//twitter.com/TinySecEx/status/943417169953505282.
[30] fanxiaocao of IceSword Lab (Qihoo 360). you see , i am also
found this case . haha! https://twitter.com/TinySecEx/status/
943411731845410816.
80
[31] Feng Yan. Windows Graphics Programming: Win32 GDI and DirectDraw.
Prentice Hall Professional, 2001.
[32] Feng Yuan. Source code for Windows Graphics Programming: Win32
GDI and DirectDraw. https://blogs.msdn.microsoft.com/fyuan/
2007/03/21/source-code-for-windows-graphics-programming-
win32-gdi-and-directdraw/.
[33] grsecurity. Probably didn’t find anything in the typical Linux userland
interface because 2013 we did some similar instrumentation to make some
leaks fall out - modified the magic for STACKLEAK/SANITIZE to a value
we told a fuzzer to never provide to the kernel,inspected copy to user for
it. https://twitter.com/grsecurity/status/991450745642905602.
[34] Jann Horn. MacOS getrusage stack leak through struct padding. https:
//bugs.chromium.org/p/project-zero/issues/detail?id=1405.
[35] Jon Oberheide. Advisories. https://jon.oberheide.org/advisories/.
Listed under “Stack Disclosure”.
[36] Jon Oberheide, Dan Rosenberg. Stackjackin’ 2: Electric Booga-

loo. https://jon.oberheide.org/blog/2011/07/06/stackjackin-2-
electric-boogaloo/.
[37] Jon Oberheide, Dan Rosenberg. Stackjacking Your Way to grsec/PaX By-
pass. https://jon.oberheide.org/blog/2011/04/20/stackjacking-
your-way-to-grsec-pax-bypass/.
[38] Jonathan Corbet. Preventing kernel-stack leaks. https://lwn.net/
SubscriberLink/748642/4cdb84b99ce171e6/.
[39] Joseph Bialek. Anyone notice my change to the Windows IO Man-
ager to generically kill a class of info disclosure? BufferedIO output
buffer is always zero’d. https://twitter.com/JosephBialek/status/
875427627242209280.
[40] Juan Vazquez. Revisiting an Info Leak. https://blog.rapid7.com/2015/
08/14/revisiting-an-info-leak/.
[41] Kees Cook. fork: unconditionally clear stack on fork. https:

commit/?id=e01e80634ecdde1dd113ac43b3adad21b47f3957.
[42] Ken Johnson, Matt Miller. Exploit Mitigation Improvements in Win-
dows 8. https://media.blackhat.com/bh-us-12/Briefings/M Miller/
BH US 12 Miller Exploit Mitigation Slides.pdf. Black Hat USA
2012, Slide 35.
81
[43] Kurmus, Anil and Zippel, Robby. A tale of two kernels: Towards ending
kernel hardening wars with split kernel. In Proceedings of the 2014 ACM
SIGSAC Conference on Computer and Communications Security, pages
1366–1377. ACM, 2014.
[44] Lu, Kangjie. Securing software systems by preventing information leaks.
PhD thesis, Georgia Institute of Technology, 2017.
[45] Lu, Kangjie and Song, Chengyu and Kim, Taesoo and Lee, Wenke.
UniSan: Proactive kernel memory initialization to eliminate data leak-
ages. In Proceedings of the 2016 ACM SIGSAC Conference on Computer
and Communications Security, pages 920–932. ACM, 2016.
[46] Lu, Kangjie and Walter, Marie-Therese and Pfaff, David and Nürnberger,
Stefan and Lee, Wenke and Backes, Michael. Unleashing use-before-
initialization vulnerabilities in the Linux kernel using targeted stack spray-
ing. In NDSS’17, Network and Distributed System Security Symposium,
2017.
[47] Mateusz Jurczyk. A story of win32k!cCapString, or unicode strings gone
bad. http://j00ru.vexillium.org/?p=1609.
[48] Mateusz Jurczyk. FreeType 2.5.3 CFF CharString parsing heap-based
buffer overflow in ”cff builder add point”. https://bugs.chromium.org/
p/project-zero/issues/detail?id=185.
[49] Mateusz Jurczyk. FreeType 2.5.3 multiple unchecked function calls return-
ing FT Error. https://bugs.chromium.org/p/project-zero/issues/
detail?id=197.
[50] Mateusz Jurczyk. nt!NtMapUserPhysicalPages and Kernel Stack-
Spraying Techniques. http://j00ru.vexillium.org/?p=769.
[51] Mateusz Jurczyk. Subtle information disclosure in WIN32K.SYS syscall
return values. http://j00ru.vexillium.org/?p=762.
[52] Mateusz Jurczyk. Using Binary Diffing to Discover Windows Kernel Mem-
ory Disclosure Bugs. https://googleprojectzero.blogspot.com/2017/
10/using-binary-diffing-to-discover.html.
[53] Mateusz Jurczyk. Windows Kernel 64-bit stack memory disclosure
in nt!NtQueryVirtualMemory (MemoryImageInformation). https://
bugs.chromium.org/p/project-zero/issues/detail?id=1519.
[54] Mateusz Jurczyk. Windows Kernel Local Denial-of-Service
#1: win32k!NtUserThunkedMenuItemInfo (Windows 7-10).
http://j00ru.vexillium.org/?p=3101.
[55] Mateusz Jurczyk. Windows Kernel Local Denial-of-Service #2:
win32k!NtDCompositionBeginFrame (Windows 8-10). http:
//j00ru.vexillium.org/?p=3151.
82
[56] Mateusz Jurczyk. Windows Kernel Local Denial-of-Service #3:
nt!NtDuplicateToken (Windows 7-8). http://j00ru.vexillium.org/?p=
3187.
[57] Mateusz Jurczyk. Windows Kernel Local Denial-of-Service
#4: nt!NtAccessCheck and family (Windows 8-10). http:
//j00ru.vexillium.org/?p=3225.
[58] Mateusz Jurczyk. Windows Kernel multiple stack and pool memory dis-
closures into NTFS file system metadata. https://bugs.chromium.org/
p/project-zero/issues/detail?id=1325.
[59] Mateusz Jurczyk. Windows Kernel pool memory disclosure due to output
structure alignment in win32k!NtGdiGetOutlineTextMetricsInternalW.
https://bugs.chromium.org/p/project-zero/issues/detail?id=
1144.
[60] Mateusz Jurczyk. Windows Kernel pool memory disclosure in
win32k!NtGdiGetGlyphOutline. https://bugs.chromium.org/p/
project-zero/issues/detail?id=1267.
[61] Mateusz Jurczyk. Windows Kernel pool memory disclosure into
NTFS metadata ($LogFile) in Ntfs!LfsRestartLogFile. https://
bugs.chromium.org/p/project-zero/issues/detail?id=1352.
[62] Mateusz Jurczyk. Windows Kernel Reference Count Vulnerabilities - Case

Study. http://j00ru.vexillium.org/slides/2012/zeronights.pdf.
ZeroNights 2012.
[63] Mateusz Jurczyk. Windows Kernel ring-0 address leak via a double-
write in NtQueryVirtualMemory(MemoryMappedFilenameInformation).
https://bugs.chromium.org/p/project-zero/issues/detail?id=
1456.
[64] Mateusz Jurczyk. Windows Kernel stack memory disclosure
in nt!RtlpCopyLegacyContextX86. https://bugs.chromium.org/p/
[65] Mateusz Jurczyk. Windows WIN32K.SYS System Call

Table (NT/2000/XP/2003/Vista/2008/7/8/10). http://
j00ru.vexillium.org/syscalls/win32k/32/.
[66] Mateusz Jurczyk. Windows X86 System Call Ta-
ble (NT/2000/XP/2003/Vista/2008/7/8/10). http://
j00ru.vexillium.org/syscalls/nt/32/.
[67] Mateusz Jurczyk. x86 Kernel Memory Space Visualization (KernelMAP
v0.0.1). http://j00ru.vexillium.org/?p=269.
83
[68] Mateusz Jurczyk, Gynvael Coldwind. Bochspwn: Exploiting Ker-
nel Race Conditions Found via Memory Access Patterns. http://
j00ru.vexillium.org/slides/2013/syscan.pdf. SyScan 2013.
[69] Mateusz Jurczyk, Gynvael Coldwind. Bochspwn: Identifying 0-
days via system-wide memory access pattern analysis. http://
j00ru.vexillium.org/slides/2013/bhusa.pdf. Black Hat USA 2013.
[70] Mateusz Jurczyk, Gynvael Coldwind. Identifying and Exploiting Win-
dows Kernel Race Conditions via Memory Access Patterns. http://
vexillium.org/dl.php?bochspwn.pdf.
[71] Mateusz Jurczyk, Gynvael Coldwind. kfetch-toolkit. https://
github.com/j00ru/kfetch-toolkit.
[72] Mateusz Jurczyk, Gynvael Coldwind. read lin mem() func-
tion. https://github.com/j00ru/kfetch-toolkit/blob/master/
instrumentation/mem interface.cc.
[73] Mathias Krause. CVE Requests (maybe): Linux kernel: various info
leaks, some NULL ptr derefs. http://www.openwall.com/lists/oss-
security/2013/03/05/13.
[74] Matt Tait. Google Project Zero Bug Tracker. https:
//bugs.chromium.org/p/project-zero/issues/list?can=1&q=
id%3A390%2C435%2C453.
[75] Matt Tait. Kernel-mode ASLR leak via uninitialized memory returned
to usermode by NtGdiGetTextMetrics. https://bugs.chromium.org/p/
[76] Michal Hocko. kmemcheck: rip it out for real. https://git.kernel.org/
pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=
f335195adf043168ee69d78ea72ac3e30f0c57ce.
[77] Michal Zalewski. American Fuzzy Lop. http://lcamtuf.coredump.cx/
afl/.
[78] Microsoft. Acknowledgments 2015. https://docs.microsoft.com/en-
us/security-updates/Acknowledgments/2015/acknowledgments2015.
[79] Microsoft. Special Pool (MSDN). https://docs.microsoft.com/en-us/
windows-hardware/drivers/devtest/special-pool.
[80] Microsoft. Windows lifecycle fact sheet. https://
support.microsoft.com/en-us/help/13853/windows-lifecycle-
fact-sheet.
[81] Microsoft (MSDN). ExAllocatePoolWithQuotaTag function.
https://docs.microsoft.com/en-us/windows-hardware/drivers/
ddi/content/wdm/nf-wdm-exallocatepoolwithquotatag.
84
[82] Microsoft (MSDN). ExAllocatePoolWithTag function. https:
//docs.microsoft.com/en-us/windows-hardware/drivers/ddi/
content/wdm/nf-wdm-exallocatepoolwithtag.
[83] Microsoft (MSDN). /GS (Buffer Security Check). https://
msdn.microsoft.com/en-us/library/8dbf701c.aspx.
[84] Microsoft (MSDN). Handling Exceptions. https://docs.microsoft.com/

en-us/windows-hardware/drivers/kernel/handling-exceptions.
[85] Microsoft (MSDN). Microsoft public symbol server. https:
//docs.microsoft.com/en-us/windows-hardware/drivers/debugger/
microsoft-public-symbols.
[86] Microsoft (MSDN). MmSecureVirtualMemory function. https:
//docs.microsoft.com/en-us/windows-hardware/drivers/ddi/
content/ntddk/nf-ntddk-mmsecurevirtualmemory.
[87] Microsoft (MSDN). ObReferenceObjectByHandle function.
ddi/content/wdm/nf-wdm-obreferenceobjectbyhandle.
[88] Microsoft (MSDN). ProbeForRead function. https://
docs.microsoft.com/en-us/windows-hardware/drivers/ddi/
content/wdm/nf-wdm-probeforread.
[89] Microsoft (MSDN). ProbeForWrite function. https://

docs.microsoft.com/en-us/windows-hardware/drivers/ddi/
content/wdm/nf-wdm-probeforwrite.
[90] Microsoft (MSDN). StackWalk64 function. https://
msdn.microsoft.com/en-us/library/windows/desktop/ms680650(v=
vs.85).aspx.
[91] Microsoft (MSDN). SymFromAddr function. https://
vs.85).aspx.
[92] Microsoft (MSDN). SymLoadModule64 function. https://

vs.85).aspx.
[93] Microsoft (MSDN). Windows Kernel-Mode Object Manager.
kernel/windows-kernel-mode-object-manager.
[94] Milburn, Alyssa and Bos, Herbert and Giuffrida, Cristiano. SafeInit:
Comprehensive and Practical Mitigation of Uninitialized Read Vulnerabil-
ities. In Proceedings of the 2017 Annual Network and Distributed System
Security Symposium (NDSS)(San Diego, CA), 2017.
85
[95] MITRE. CWE-252: Unchecked Return Value. https://cwe.mitre.org/
data/definitions/252.html.
[96] North, John. Identifying Memory Address Disclosures. 2015.
[97] Peiró, Salva and Muñoz, M and Masmano, Miguel and Crespo, Alfons.
Detecting stack based kernel information leaks. In International Joint
Conference SOCO14-CISIS14-ICEUTE14, pages 321–331. Springer, 2014.
[98] Peiró, Salva and Munoz, M and Crespo, Alfons. An analysis on the im-
pact and detection of kernel stack infoleaks. Logic Journal of the IGPL,
24(6):899–915, 2016.
[99] pj4533. MemSpyy. https://www.codeproject.com/Articles/21090/
MemSpyy.
[100] Robert C. Seacord. The CERT C Coding Standard, Second Edition: 98
Rules for Developing Safe, Reliable, and Secure Systems. Addison-Wesley
Professional, 2014. DCL39-C. Avoid information leakage when passing a
structure across a trust boundary.
[101] Sebastian Apelt. Pwn2Own 2014 - Escaping the sandbox through
AFD.sys. http://siberas.blogspot.de/2014/07/pwn2own-2014-
escaping-sandbox-through.html.
[102] Solar Designer. Finally went through the Bochspwn Reloaded slides.
Kudos! Feedback: rather than “Remove taint on free”, you could re-
taint & detect UAF+leak. https://twitter.com/solardiz/status/
879288169174315008.
[103] StatCounter GlobalStats. Desktop Windows Version Market Share
Worldwide. http://gs.statcounter.com/os-version-market-share/
windows/desktop/worldwide.
[104] Tavis Ormandy. Fun Rev. Challenge: On 32bit Windows7, explain where
the upper 16bits of eax come from after a call to NtUserRegisterClassEx-
WOW(). https://twitter.com/taviso/status/16853682570.
[105] Tavis Ormandy. iknowthis Linux System Call Fuzzer. https://
github.com/rgbkrk/iknowthis.
[106] Tess Ferrandez. Show me the memory: Tool for visualizing virtual mem-
ory usage and GC heap usage. https://blogs.msdn.microsoft.com/
tess/2009/04/23/show-me-the-memory-tool-for-visualizing-
virtual-memory-usage-and-gc-heap-usage/.
[107] Vasiliy Kulikov. Re: Linux kernel proactive security hardening. http:
//seclists.org/oss-sec/2010/q4/129.
[108] Vegard Nossum. Getting started with kmemcheck. https://
www.kernel.org/doc/html/v4.12/dev-tools/kmemcheck.html.
86
[109] Wandering Glitch. Leaking Windows Kernel Pointers. https:
//ruxcon.org.au/assets/2016/slides/RuxCon%20-%20Leaking%
20Windows%20Kernel%20Pointers.pdf.
[110] Weimin Wu. An Analysis of A Windows Kernel-Mode Vulnerabil-
ity (CVE-2014-4113). https://blog.trendmicro.com/trendlabs-
security-intelligence/an-analysis-of-a-windows-kernel-mode-
vulnerability-cve-2014-4113/.
87
A Other system instrumentation schemes
During the development of Bochspwn and Bochspwn Reloaded, we have consid-
ered a number of alternative ways in which full system instrumentation could
help identify security flaws, or at least signal sensitive areas of code that should
receive more attention. For some of the ideas, we implemented functional proto-
types which did uncover new bugs in the Windows kernel. While none of these
experimental tools had as much success as the two main projects, we decided
to discuss them in this appendix for completeness. We hope that the concepts
outlined in the following subsections may serve as a source of inspiration for
researchers aiming to take up the subject of kernel instrumentation through
software emulation. They are mostly discussed in the context of Windows, but
many of them also apply to Linux and other operating systems.
A.1 Enumeration of kernel attack surface

All interactions between user-mode and the kernel may be considered attack
surface in itself – due to the numerous pitfalls described earlier in this paper –
or more generally, reliable indicators of attack surface related to the processing of
input data provided by the caller. Thanks to pointer annotations and dedicated
copy functions, all such interactions in the Linux kernel may be enumerated
using static code analysis.
The same objective cannot be achieved as easily on Windows, due to the
lack of annotations and specialized functions, and the unavailability of the ker-
nel source code. Without performing a thorough analysis of the assembly sur-
rounding a specific memory access and all code paths leading up to it, it is often
impossible to determine if an instruction operates on a user-mode pointer, a
kernel-mode pointer, or both. There are reliable symptoms revealing the usage
of ring 3 memory, such as references to the ProbeForRead and ProbeForWrite
functions or the MmUserProbeAddress constant, but they are often inaccurate,
as pointer sanitization may take place in a completely different location than
the actual reference to that pointer.
To effectively enumerate the Windows kernel attack surface, a simple dy-
namic instrumentation may be used to detect and log all accesses to the user
address space originating from the kernel. As with every instrumentation, its
effectiveness is limited by the kernel code coverage reached during testing, but
even just booting up the system is sufficient to generate enough logs for weeks
of Windows kernel auditing.
A.2 Deeply nested user-mode memory accesses

In a majority of cases, reads and writes from/to user-mode memory should occur
in top-level syscall handlers – Nt functions in Windows and sys functions in
Linux. As the first point of contact between ring 3↔0, the routines should create
local copies of input data, allocate temporary kernel buffers for output data, and
take full responsibility for the interactions with userland, instead of burdening
88
internal kernel functions with operating on user-mode pointers. With the lack
of explicit pointer annotations in Windows, no kernel routine may definitely
know if a pointer it receives as an argument is a user-mode one or not.
Passing along a user-controlled pointer into nested functions may blur the
understanding of which part of the code is accountable for sanitizating the
address. This may have dire consequences for system security, as illustrated in
Listing 29.
1 NTSTATUS Bar(LPDWORD Output) {

2 Output[0] = 0xdeadbeef;
3 Output[1] = 0xbadc0ffe;
4 Output[2] = 0xcafed00d;
6 }
7
8 NTSTATUS Foo(DWORD Type, LPDWORD Output) {
9 if (Type == 2) {
10 return Bar(Output);
11 }
12
13 [...]
14 }
15
16 NTSTATUS NTAPI NtMagicValues(DWORD Family, DWORD Type, LPDWORD Output) {
17 if (Family == 1) {
18 return Foo(Type, Output);
19 }
20
21 [...]
22 }
Listing 29: Example of passing an input pointer to nested kernel functions
In this example, the Output parameter is never validated before being writ-
ten to. The top-level NtMagicValues handler doesn’t sanitize the pointer be-
cause it doesn’t directly operate on it. The Foo() function doesn’t do it, because
it assumes that any argument it receives will already have been checked by the
caller. Finally, the Bar() function doesn’t do it, because it is a simple internal
function that has no notion of different types of pointers. As a whole, this re-
sults in an arbitrary kernel memory overwrite – an easily exploitable security
flaw – all because of the ambiguity caused by passing user-mode pointers to
internal kernel functions which do not expect it.
Potential issues of this kind may be flagged by logging all user-mode memory
references taking place within relatively deep callstacks. While not all instances
of such behavior manifest actual bugs, heavily nested accesses to the ring 3
address space may suggest that the relevant code is not aware of the nature of
the referenced pointer. This in turn increases the likelihood of the presence of
a vulnerability, and warrants manual follow-up analysis.
89
A.3 Unprotected accesses to user-mode memory
Kernels should typically never assume the validity of user-mode pointers, unless
the address range in question is explicitly locked in memory. In Windows, this
imposes the need to wrap each user-mode memory reference with an adequate
exception handler. The absence of such handler at any place where a controlled
pointer is accessed may be exploited to trigger an unhandled exception and crash
the operating system. The need to safely handle all exceptions arising from the
usage of ring 3 memory is reflected in the documentation of the ProbeForRead
function [88]:
Drivers must call ProbeForRead inside a try/except block. If the

routine raises an exception, the driver should complete the IRP with
the appropriate error. Note that subsequent accesses by the driver to
the user-mode buffer must also be encapsulated within a try/except
block: a malicious application could have another thread deleting,
substituting, or changing the protection of user address ranges at
any time (even after or during a call to ProbeForRead or ProbeFor-
Write).
Moreover, the “Handling Exceptions” MSDN article [84] includes an ade-

quate code example shown in Listing 30.
try {
...
ProbeForWrite(Buffer, BufferSize, BufferAlignment);
/* Note that any access (not just the probe, which must come first,
* by the way) to Buffer must also be within a try-except.
*/
...
} except (EXCEPTION_EXECUTE_HANDLER) {
/* Error handling code */
...
}
Listing 30: Example of a secure implementation accessing user-mode memory

in the Windows kernel
From a technical standpoint, it is relatively easy to detect user-mode accesses

being performed when no exception handlers are set up. Since the exception
handling mechanism is largely different in 32-bit and 64-bit versions of Windows,
we will focus on the subjectively easier goal of instrumenting the x86 platform.
However, please note that each unsafe memory access identified on a 32-bit
version of the system should reproduce on a 64-bit build, and vice-versa.
In x86 Windows builds, the handler records are chained together in a SEH
(Structured Exception Handling) chain starting at the fs:[0] address, where
each handler is described by the EH3 EXCEPTION REGISTRATION structure shown
in Listing 31.
90
struct _EH3_EXCEPTION_REGISTRATION {
struct _EH3_EXCEPTION_REGISTRATION *Next;
PVOID ExceptionHandler;
PSCOPETABLE_ENTRY ScopeTable;
DWORD TryLevel;
};
Listing 31: Definition of the EH3 EXCEPTION REGISTRATION structure
The structures reside in the stack frames of their corresponding functions and
are initialized by calling the SEH prolog4( GS) procedures. During execution,
entering the try{} blocks is denoted by writing their zero-based indexes to the
TryLevel fields in the aforementioned structures, and later overwriting them
with −2 (0xfffffffe) when execution leaves the blocks and exception handling
is disabled. Below is an example of a try/except block encapsulating the
writing of a single DWORD value to user space:
PAGE:00671CF3 mov [ebp+ms_exc.registration.TryLevel], 1

PAGE:00671CFA mov eax, [ebp+var_2C]
PAGE:00671CFD mov ecx, [ebp+arg_14]
PAGE:00671D00 mov [ecx], eax
PAGE:00671D02 mov [ebp+ms_exc.registration.TryLevel], 0FFFFFFFEh
It is possible for the Bochs instrumentation to iterate through the SEH

chain by following the Next pointers, and determine which handlers are en-
abled and which functions they correspond to. If there are no exception records
present, or all of them have their TryLevel fields set to −2, then an excep-
tion occurring right at that moment could potentially crash the operating sys-
tem. It should be noted, however, that not all non-guarded accesses to user-
mode memory are dangerous by definition – regions previously secured with
MmSecureVirtualMemory [86] and special areas such as TEB or PEB are not
affected.
In February 2017, we implemented the detection scheme and ran it against
Windows 7 and 10. After analyzing the output, we identified several bugs
allowing local attackers to trigger unhandled exceptions in the kernel and crash
the operating system with a Blue Screen of Death. As denial-of-service issues
don’t meet the bar to be fixed by Microsoft in a security bulletin, we posted the
details and proof-of-concept code of our findings on our blog [54, 55, 56, 57].
Affected syscall Windows 7 Windows 8.1 Windows 10

win32k!NtUserThunkedMenuItemInfo X X X
win32k!NtDCompositionBeginFrame X X
nt!NtDuplicateToken X X
nt!NtAccessCheck etc. X X
Table 11: Summary of DoS bugs caused by unprotected access to user space
91
A.4 Broad exception handlers
In Windows, for every try{} block of code modifying any global data structures
in the kernel, there should be a corresponding except{} block which reverts all
persistent changes made prior to the exception. This is relatively easy to achieve
with a flat structure of the code, when all relevant operations are explicitly vis-
ible inside the try{}. On the other hand, the rule is more difficult to enforce
when nested calls are used, thus obfuscating the code flow and potentially fa-
cilitating the interruption of functions which don’t anticipate being preempted
in the middle of execution. In certain situations, this can lead to leaving global
objects in the system in an inconsistent state, which may open up security
vulnerabilities. One example of such bug is CVE-2014-1767 [101], a dangling
pointer flaw in the afd.sys network driver, which was used during the pwn2own
competition to elevate privileges in the system as part of a longer exploit chain.
1 typedef struct _DATA_ITEM {

2 LIST_ENTRY ListEntry;
3 DWORD Value;
4 } DATA_ITEM, *PDATA_ITEM;
5
6 PDATA_ITEM DataListHead;
7
8 VOID KeAddValueInternal(PDWORD ValuePtr) {
9 PDATA_ITEM NewDataItem = ExAllocatePool(PagedPool, sizeof(DATA_ITEM));
10 PDATA_ITEM OldListHead = DataListHead;
11
12 DataListHead = NewDataItem;
13
14 NewDataItem->Value = *ValuePtr;
15 NewDataItem->ListEntry.Flink = OldListHead;
16 NewDataItem->ListEntry.Blink = OldListHead->ListEntry.Blink;
17
18 OldListHead->ListEntry.Blink = NewDataItem;
19 }
20
21 NTSTATUS NTAPI NtAddValue(PDWORD UserValuePtr) {
22 NTSTATUS Status = STATUS_SUCCESS;
23
24 try {
25 ProbeForRead(UserValuePtr, sizeof(DWORD), sizeof(DWORD));
26 KeAddValueInternal(UserValuePtr);
27 } except (EXCEPTION_EXECUTE_HANDLER) {
28 Status = GetExceptionCode();
29 }
30
31 return Status;
32 }
Listing 32: Example of a vulnerability caused by a broad exception handler
92
Let’s consider an example shown in Listing 32. The NtAddValue top-level
syscall handler validates the UserValuePtr pointer but doesn’t read its value,
instead passing it to the internal KeAddValueInternal function. The latter
routine is not aware of the type of the pointer it receives, so it simply imple-
ments its self-contained logic – allocates a new object in memory, inserts it into
a doubly-linked list and initializes it with the input data. A problem arises
when accessing ValuePtr in line 14 fails and generates an ACCESS VIOLATION
exception, thus effectively aborting the execution of KeAddValueInternal and
jumping straight into the handler in line 28. Due to the fact that the nested
function was interrupted, it has already saved the newly allocated object as the
head of the list, but hasn’t yet initialized the object’s LIST ENTRY structure.
On the other hand, NtAddValue doesn’t know the internal logic of the functions
it invokes, so it doesn’t revert any changes made by them. Consequently, the
linked list becomes corrupted with uninitialized pointers, leaving the system
unstable and prone to privilege escalation attacks.
Candidates for such issues can be automatically identified by system instru-
mentation by examining all kernel accesses to user-mode memory11 and checking
the location of the first enabled exception handler in the stack trace for each
of them. If the closest handler is not in the current function (or even worse,
several functions below in the call stack), it is an indicator of a broad exception
handler that could be unprepared to correctly restore the kernel to the state
before the exception. That said, not all broad handlers are bugs, so each case
needs to be examined on its own to determine if any changes made up to the
current point of execution are not accounted for in the corresponding exception
handler.
A.5 Dereferences of unsanitized user-mode pointers

To maintain system security, every user-mode pointer passed to the Windows
kernel by a client application needs to be sanitized with a ProbeForRead [88]
or ProbeForWrite [89] function (or equivalent) before being operated on. Miss-
ing pointer checks may create memory disclosure/corruption primitives such as
arbitrary read/write, making them one of the most severe types of kernel vul-
nerabilities. If system instrumentation could recognize pointer sanitization at
the OS runtime, it could then audit each access to user space to determine if
the referenced address has been already sanitized in the context of the current
system call, and report the bug if not.
One way to catch instances of pointer sanitization is to detect the execution
of the ProbeFor functions and record the address ranges found in their argu-
ments. The problem with this approach is that there are very few references to
these functions in the core kernel images, especially to ProbeForRead. Instead
of issuing direct calls to the function, its usages are typically inlined to the
assembly equivalent of the pseudo-code shown on the next page.
11 Or any other events that are known to generate exceptions, such as calls to
ExAllocatePoolWithQuotaTag [81].
93
if ((Address & (~Alignment)) != 0) {
ExRaiseDatatypeMisalignment();
}
if ((Address + Length < Address) ||

(Address + Length >= MmUserProbeAddress)) {
// Trigger an ACCESS_VIOLATION exception.
}
In some cases, the comparison with MmUserProbeAddress is performed against

the value in memory, and in others it is achieved through a register. Small,
constant values of Length are often ignored. Furthermore, the win32k.sys
driver on Windows 7 and 8.1 doesn’t use the MmUserProbeAddress constant
exported from ntoskrnl.exe, but instead has its own copy stored in a static
W32UserProbeAddress variable initialized in DriverEntry. To make things
even more complicated, the constant is no longer used in the core kernel image
of Windows 10 64-bit, as it was replaced with an immediate 0x7fffffff0000
operand in all of the relevant cmp instructions.
To address the above problems and reliably detect all instances of user-mode
pointer sanitization, we propose the following logic:
• Intercept all cmp instructions with a register or memory as the first operand.
• Resolve the values of both operands of the instruction.
• If the value of the second operand is equal to MmUserProbeAddress (for
example 0x7fff0000 in x86 builds), mark the address in the first operand
as sanitized in the scope of the current syscall.
According to our experimentation, following the above steps in the Bochs

instrumentation is sufficient to recognize pointer checking. Unfortunately, this
still only provides us with very limited information in the form of just one end
of the probed user-mode memory region (beginning or end, depending on the
code construct). One solution to the problem of incomplete data is to extend
the area marked as sanitized to the entire memory page the address belongs
to, or even several pages around it. While this approach may significantly
reduce the volume of false-positive results, it shifts the balance towards more
false-negatives. For example, in this scheme, the sanitization of a single stack-
based pointer would automatically cause all other pointers residing within the
same page to also be considered as verified. This limitation can be partially
mitigated by enabling the PageHeap mechanism for all user-mode programs
run in the guest system, which should at least prevent heap-based pointers
from overlapping with each other on the memory page level.
For testing purposes, we developed a basic prototype of the above scheme
and tested it against Windows 7 and 10, but only identified references to unsan-
itized addresses in code accessible by system administrators or very early dur-
ing system boot (e.g. in the win32k!WmsgpConnect function reachable through
94
win32k!RegisterLogonProcess). However, there is much room for improve-
ment in this field, and we believe that the instrumentation could successfully
detect security issues provided an effective way to determine both ends of each
probed ring 3 address range.
A.6 Read-after-write conditions

When the kernel writes a value to user-mode memory, there is no guarantee that
it won’t be modified the very next moment by a concurrent user thread. Thus,
reading from a ring 3 memory region that was previously written to in the same
system call is a strong indicator of a serious problem in the code. For example,
such behavior could be manifested by a function assuming it is operating on
a safe, kernel pointer, while in practice it was passed an untrusted, user-mode
address. Such a scenario is illustrated in Listing 33.
1 NTSTATUS NTAPI NtGetSystemVersion(

2 PUNICODE_STRING UnicodeString,
3 PWCHAR Buffer,
4 DWORD BufferLength
5 ) {
6 UnicodeString->Length = 0;
7 UnicodeString->MaximumLength = BufferLength;
8 UnicodeString->Buffer = Buffer;
9
10 RtlAppendUnicodeToString(UnicodeString,
11 L"Microsoft Windows [Version 10.0.16299]");
12
14 }
Listing 33: Example read-after-write leading to a write-what-where condition
In lines 6-8, the syscall handler initializes the output UNICODE STRING struc-
ture as an empty string backed by a user-mode buffer. Further on, it calls the
RtlAppendUnicodeString API on that structure to fill it with a textual rep-
resentation of the system version. The problem in the code is that the latter
routine assumes that it receives a non-volatile kernel UNICODE STRING object,
while in fact it is passed a user-mode pointer whose data may change asyn-
chronously at any point of the system call execution. A malicious program could
use a concurrent thread to exploit the race condition by changing the value of
UnicodeString->Buffer, to point it into the kernel address space within the
short time window between the initialization of the pointer and its usage in the
API function.
By nature, the bug class is very similar to double fetches, with the main
difference being that the affected code trusts the contents of user-mode memory
not because it has already read it once, but because it has explicitly initialized
it to a specific value. The detection of such issues is also almost identical to the
logic implemented in the original Bochspwn project [70], with the addition of
95
instrumenting not only kernel→user memory reads, but also writes. We expect
read-after-write conditions to be mostly specific to Windows, as it seems to be
strongly tied to direct user-mode pointer manipulation and the lack of clear
distinction between ring 3 and ring 0 pointers.
A.7 Double writes

System calls are meant to be invoked synchronously; only after execution returns
to the user-mode client, should it read the output data provided by the kernel.
However, if at any point the kernel writes a piece of privileged information to
ring 3 and later overwrites it with legitimate data in the scope of the same
syscall, a concurrent thread is able to obtain the initially written value within
the time window of the race condition.
In theory, the bug class could apply to any type of sensitive data, but in
practice it most often facilitates the disclosure of kernel-mode pointers, allowing
attackers to bypass KASLR. The typical scenario for introducing a flaw of this
kind is when a common structure containing pointers is used to store data
both in kernel and user-mode. When a system call handler wishes to copy
such an internal kernel structure to a client application, it is usually achieved
by (1) copying the overall object in memory and (2) adjusting the particular
pointers to point into the corresponding userland buffers. This may seem valid at
first glance, but in fact it means that the original kernel-mode pointers reside in
ring 3 for a brief period of time before they are overwritten with the appropriate
addresses. As more effort is continuously being put in by OS vendors to protect
information about the kernel address space, leaking ring 0 pointers may prove
useful for the exploitation of other kernel memory corruption vulnerabilities.
1 typedef struct _USERNAME {

2 UNICODE_STRING String;
3 WCHAR Buffer[128];
4 } USERNAME, *PUSERNAME;
5
6 NTSTATUS NTAPI NtGetAdminUsername(PUSERNAME OutputUsername) {
7 USERNAME LocalUsername;
8
9 RtlZeroMemory(&LocalUsername, sizeof(USERNAME));
10
11 StringCchCopy(LocalUsername.Buffer, 128, "Administrator");
12 RtlInitUnicodeString(&LocalUsername.String, LocalUsername.Buffer);
13
14 RtlCopyMemory(OutputUsername, &LocalUsername, sizeof(USERNAME));
15 OutputUsername->String.Buffer = OutputUsername->Buffer;
16
18 }
Listing 34: Example double-write condition revealing a kernel-mode address
96
Let’s examine the example illustrated in Listing 34. The USERNAME structure
is a self-contained object that includes both the UNICODE STRING structure and
the corresponding textual buffer. A local object of this type is first initialized
in lines 11-12 and later copied to the client in line 14. Since the Buffer pointer
passed back to user-mode still contains a kernel address, it is overwritten with a
reference to the ring 3 buffer in line 15. The time window available for another
thread to capture the disclosed kernel pointer lasts between lines 14 and 15.
Detection of double writes is again very similar to that of double fetches – the
instrumentation should catch all kernel→userland memory writes, and signal a
bug every time a specific address is written to with non-zero data more than
once in the context of a single system call. As an additional feature, the tool can
put a special emphasis on cases where the original bytes resemble a kernel-mode
address, and the new data appears to be a user-mode pointer. This should help
highlight the reports most likely to represent actual information disclosure bugs.
To test the above idea, we implemented a simple prototype of the instru-
mentation and ran it on Windows 7 and 10 32-bit. As a result, we discovered
three double-write conditions, all leaking addresses of objects in ring 0 memory:
• A bug in nt!IopQueryNameInternal, in the copying of a UNICODE STRING
structure. The flaw is reachable through the nt!NtQueryObject and
nt!NtQueryVirtualMemory system calls, and was filed in the Project Zero
bug tracker with the corresponding proof of concept as issue #1456 [63].
• A bug in nt!PspCopyAndFixupParameters (UNICODE STRING structures
nested in RTL USER PROCESS PARAMETERS).
• A bug in win32k!NtUserfnINOUTNCCALCSIZE (NCCALCSIZE PARAMS struc-
ture).
The first of the above problems was reported to Microsoft in December 2017,
but the vendor replied that the report and all similar issues didn’t meet the bar
for a security bulletin and would be instead targeted to be fixed in the next
version of Windows. Upon publishing the details of the double-write conditions,
other researchers publicly claimed that they were also aware of the bug class
and collided with some of our findings [28, 30, 29].
A.8 Ignored function return values

Correctly checking the return values of functions and methods is essential to
maintaining a consistent program state and preventing unexpected conditions,
which can further lead to security flaws. This type of issues was categorized by
MITRE and assigned a CWE-252 weakness ID [95]. There are many examples
of real-world bugs caused by insufficient validation of return values, for instance
a buffer overflow in the FreeType font engine [48], a NULL pointer dereference
in win32k.sys exploited in the wild [110], or a lesser stack-based infoleak in
win32k.sys related to unicode strings [47]. In all of the above cases, prop-
erly handling the error conditions indicated by the called functions would have
prevented the subsequent vulnerabilities.
97
It is important to note that unchecked return values are a canonical type
of a problem that can be effectively detected using static analysis, especially
if the source code is available. As a very basic example, after finding the
aforementioned FreeType bug caused by an ignored error code, we added a
attribute ((warn unused result)) directive to the declaration of the in-
ternal FT Error type and compiled the project with the -Wno-attributes flag,
which caused gcc to warn about all instances of unchecked return values of type
FT Error [49]. The output of this experiment motivated the project’s main-
tainer to submit a series of patches to fix many potential, related bugs. While
this is a simple example, more advanced analyzers should be able to pinpoint
such behavior accurately, without the need to apply any special changes to the
tested code.
On a binary level, static analysis is more difficult, as some information is
inevitably lost during compilation. However, it is still possible to achieve by
tracking operations on the EAX or RAX registers to determine if each of the
evaluated functions has a return value, and verifying if all of their callers check
that value accordingly. A big advantage of the static approach is that it is able
to process an entire code base without the need to execute it, and hence it is not
limited by the reachable code coverage. Nonetheless, this can also be considered
a drawback, as reports regarding unchecked return values tend to be flooded
with false-positives and non-issues. In this context, the results of dynamic
analysis are easier to triage and understand, because they are supplemented
with complete information about the system state, including traces of the control
flow, the actual values returned by the functions and so on.
The execution pattern indicating potential problems in the code is straight-
forward. Excluding minor corner cases and assuming 32-bit execution mode, it
is as follows: if two instructions set the value of the EAX register and they are
separated by at least one ret instruction but no reads from EAX in between,
this suggests that the second write discards a return value that should have
been checked first. The only major problem with the above scheme is the fact
that Bochs doesn’t provide instrumentation callbacks for register operations.
On the upside, references to the emulated CPU registers are achieved through
general-purpose macros such as BX READ 32BIT REG and BX WRITE 32BIT REGZ
defined in cpu/cpu.h (Listings 35 and 36). For a demonstration of how the
macros are used in the software implementation of the mov r32, r32 instruc-
tion, see Listing 37. Thanks to this detail, we were able to introduce two
custom callbacks named bx instr genreg read and bx instr genreg write,
invoked on every access to any register; their prototypes are shown in Listing 38.
We subsequently added calls to these instrumentation-defined functions in all
register-related macros found in cpu/cpu.h.
With the capability of intercepting all references to EAX taking place in the
kernel, we implemented the high-level logic of the instrumentation. While test-
ing the tool, we learned that several instructions required special handling –
xor eax, eax had to be treated as write operations instead of the theoreti-
cal r/w, while movzx eax, ax and similar instructions are effectively no-ops in
the sense of our logic, even though they operate on various parts of EAX.
98
145 #define BX_READ_8BIT_REGL(index) (BX_CPU_THIS_PTR gen_reg[index].word.
byte.rl)
146 #define BX_READ_16BIT_REG(index) (BX_CPU_THIS_PTR gen_reg[index].word.rx)
147 #define BX_READ_32BIT_REG(index) (BX_CPU_THIS_PTR gen_reg[index].dword.
erx)
Listing 35: Definitions of macros reading from CPU registers in Bochs

(cpu/cpu.h)
166 #define BX_WRITE_32BIT_REGZ(index, val) {\

167 BX_CPU_THIS_PTR gen_reg[index].rrx = (Bit32u) val; \
168 }
169
170 #define BX_WRITE_64BIT_REG(index, val) {\
171 BX_CPU_THIS_PTR gen_reg[index].rrx = val; \
172 }
173 #define BX_CLEAR_64BIT_HIGH(index) {\
174 BX_CPU_THIS_PTR gen_reg[index].dword.hrx = 0; \
175 }
Listing 36: Definitions of macros writing to CPU registers in Bochs

(cpu/cpu.h)
60 BX_INSF_TYPE BX_CPP_AttrRegparmN(1) BX_CPU_C::MOV_GdEdR(bxInstruction_c *

i)
61 {
62 BX_WRITE_32BIT_REGZ(i->dst(), BX_READ_32BIT_REG(i->src()));
63
64 BX_NEXT_INSTR(i);
65 }
Listing 37: Implementation of the mov r32, r32 instruction in Bochs

(cpu/data xfer32.cc)
void bx_instr_genreg_read(unsigned cpu, unsigned index, unsigned size,

bool is_high);
void bx_instr_genreg_write(unsigned cpu, unsigned index, Bit64u value,

unsigned size, bool is_high);
Listing 38: Prototypes of custom register-related instrumentation callbacks
99
We evaluated the instrumentation against Windows 7 32-bit and collected
over 2700 unique reports of unchecked return values. Due to the excessive output
volume we were only able to review about 20% of the reports, which did not
manifest any high-severity bugs. Nevertheless, we believe the technique shows
great potential and can be successfully used to uncover new bugs, with more
effort put into reducing the number of flagged non-issues.
A.9 API misuse

Kernel instrumentation is not limited to detecting low-level bugs related to inter-
actions with user-mode memory, but can also prove useful in spotting problems
on a higher level, such as the way API functions are called. Because of the com-
plexity of certain kernel subsystems (e.g. the Object Manager in Windows [93]),
some of the clients may use their interface in unsafe, exploitable ways. Exam-
ples of Windows kernel vulnerabilities that could be caused by API misuse are
listed below:
• One of the arguments of the ObReferenceObjectByHandle function [87]

is ObjectType, a pointer containing the anticipated type of the referenced
handle, or NULL, if the object may be of any type. Broadly setting the
parameter to NULL while still expecting the object to be of a specific type
can lead to type confusion and memory corruption.
• Windows kernel objects are subject to reference counting, which makes
them potentially prone to typical problems such as refcount leaks (more
references than dereferences in one self-contained interaction) or double
derefs (vice versa) [62]. Incorrect balancing of the (de)references usually
leads to use-after-free conditions.
• There are two types of Windows object handles – user-mode and kernel-
mode ones. It is essential for system security that all handles not explicitly
exported to user-mode are created with the OBJ KERNEL HANDLE flag set
in the OBJECT ATTRIBUTES.Attributes field; otherwise, rogue programs
could access system objects and abuse them for privilege escalation. Signs
of unsafe behavior include creating user-mode handles and not writing
their numerical values to ring 3 memory, or creating temporary user-mode
handles and destroying them in the scope of the same system call.
All of the above and many other types of API-related issues could be uncov-
ered by instrumenting known sensitive functions and validating their security-
related requirements.
100

Bochspwn Reloaded

Uploaded by

Copyright:

Available Formats

Bochspwn Reloaded

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bochspwn Reloaded

Uploaded by

Copyright:

Available Formats

Detecting Kernel Memory Disclosure with x86

Emulation and Taint Tracking

2 Memory disclosure in operating systems 6

3 Bochspwn Reloaded – detection with software x86 emulation 23

4 Windows bug reproduction techniques 59

6 Other data sinks 72

A Other system instrumentation schemes 88

Write input data

Invoke system call

Read input data

Write output data

Return to user space

Read output data

Figure 1: Life of a system call

• Double fetches – caused by reading a user-mode value more than once

• Unprotected memory operations – caused by the lack of exception

Write to each user-mode memory location at most once, [...], only

Due to a combination of circumstances such as the nature of the C pro-

2.1 C language-specific properties

2.1.1 Indeterminate state of uninitialized variables

7.22.3.4 The malloc function

7.22.3.5 The realloc function

1 NTSTATUS NTAPI NtMultiplyByTwo(DWORD InputValue, LPDWORD OutputPointer) {

Listing 1: Memory disclosure via a local uninitialized variable

Listing 2: Memory disclosure via a reserved structure field

2.1.2 Structure alignment and padding bytes

6.5.3.4 The sizeof and Alignof operators

6.2.8 Alignment of objects

6.7.2.1 Structure and union specifiers

1 typedef struct _SYSCALL_OUTPUT {

Listing 3: Memory disclosure via structure padding

Leaks through alignment holes are relatively common, as a variety of complex

Figure 2: Layout of a structure with padding bytes

However, Seacord R. C. argues [100] that it is a noncompliant solution,

2.1.3 Unions and diversely sized fields

6.7.2.1 Structure and union specifiers

1 typedef union _SYSCALL_OUTPUT {

Listing 4: Memory disclosure via a partially initialized union

Figure 3: Memory layout of a partially initialized union

2.1.4 sizeof considered harmful

2.2.1 Memory reuse in dynamic allocators

Note Memory that function allocates is uninitialized. A kernel-

• kmalloc – a generic function used to allocate arbitrarily sized memory

By themselves, these functions provide no assurance that the allocated re-

We see the existence of these options as an advantage for kernel security,

2.2.2 Fixed-sized arrays

1 NTSTATUS NTAPI NtGetSystemPath(PCHAR OutputPath) {

Listing 5: Memory disclosure via a partially initialized string buffer

Figure 4: Memory layout of a partially initialized string buffer

A secure implementation should only return the requested path, with no

2.2.3 Arbitrary syscall output buffer lengths

Listing 6: Memory disclosure via an arbitrary buffer output size

KernelBuffer[0] [1] [2] [3] [4] […]

Figure 5: Memory layout of a memory allocation with a user-controlled size

The scheme discussed in this section is particularly characteristic to Win-

• An optimization frequently employed in Windows system calls is to use

2.3 Further contributing factors

Lack of visible consequences As a general rule, information disclosure flaws

Leaks hidden behind system API Typical client applications implement

C:\> symchk.exe <driver> /s srv<directory>http://msdl.microsoft.com/