https://exploitreversing.

com

ExploitingReversing (ER) series

Article 01
by Alexandre Borges
release date: APRIL/11/2023 | rev: A

0. Quote

“Success. It's got enemies. You can be successful and have enemies or you can be unsuccessful and have
friends.”. (Dominic Cattano | “American Gang er” movie - 2007)

1. Introduction

Welcome to the first article of Exploiting Reversing (ER) series, where I will review concepts, techniques
and practical steps related to binaries and, eventually, analyze vulnerabilities in general. If readers have
not read past articles about my other series (MAS – Malware Analysis Series) yet all of them are available
on the following links:
▪ MAS_1: https://exploitreversing.com/2021/12/03/malware-analysis-series-mas-article-1/
▪ MAS_2: https://exploitreversing.com/2022/02/03/malware-analysis-series-mas-article-2/
▪ MAS_3: https://exploitreversing.com/2022/05/05/malware-analysis-series-mas-article-3/
▪ MAS_4: https://exploitreversing.com/2022/05/12/malware-analysis-series-mas-article-4/
▪ MAS_5: https://exploitreversing.com/2022/09/14/malware-analysis-series-mas-article-5/
▪ MAS_6: https://exploitreversing.com/2022/11/24/malware-analysis-series-mas-article-6/
▪ MAS_7: https://exploitreversing.com/2023/01/05/malware-analysis-series-mas-article-7/
In different opportunities we have to analyze kernel drivers or mini-filter drivers to understand a
vulnerability or even a malicious driver (as known as rootkit), and this topic is usually complex and presents
many details eventually deserves to be explained. However, I still needed a better motivation to start this
new series and it came up while I was analyzing details on Microsoft Security Events Component Minifilter
(C:\Windows\system32\drivers\mssecflt.sys), which it is a required dependency that enables FltMgr
service (fltmgr.sys) to be started, and stumbled with functions from this driver that, indirectly,
remembered me about techniques used to detect different kind of evasions using NtCreateProcessEx( )
that I had read from a good article delivered by Microsoft last year:
https://www.microsoft.com/security/blog/2022/06/30/using-process-creation-properties-to-catch-
evasion-techniques/.
At that point I realized that I could really start a new series of article, covering topics as reversing
engineering and vulnerability research and, effectively, moving away from malware analysis, which it is a

1|Page
https://exploitreversing.com

tuff that I don’t work with for a long time, but also keep writing to offer information to other professionals
who need it. Somehow, this series of articles offers me this freedom and opportunity to produce
something that, eventually, could be useful for people in the area.
While I am not concerned to analyze malicious code itself in this series, I will be using a malicious driver to
illustrate a few concepts about a section that will be presented later in this article, but it will be an
exception in this series. As I mentioned previously, the main purpose of this series is being focused on
reversing engineering, vulnerability research and, eventually, something about operating system internals.
Certainly, there is nothing new here and the idea is to provide correlated information that might help
readers to understand subtle details which could go unnoticed while reading articles, books and references
on the Internet. Mainly, while doing research, we usually learn a lot, but most of the time the information
is spread over multiple sources so that it could be hard to put everything together.
Readers from my previous articles could wonder whether I have plans to continue the MAS (Malware
Analysis Series) and, definitely, I will keep writing it. The only difference is that I will alternate between
series according to inspiration and spare time, of course.
Finally, and the more important fact by far, this article will have mistakes, typos and so on, and soon I know
about them, so I will release a fixed version of this article.

2. Acknowledgments

I could not write this series and the MAS (Malware Analysis Series) without receiving the decisive help from
Ilfak Guilfanov (@ilfak), from Hex-Rays SA (@HexRaysSA), becaue I didn’t have an own IDA Pro licene,
and he kindly provided everything I needed to write this series about reversing and vulnerabilities, and
other one that are coming. However, hi help didn’t top in 2021, and he and Hex-Rays have continuously
helped until the present moment by providing immediate support for everything I need to keep these
public projects. Additionally, Ilfak is always truly kind replying to me every single time that I send a
message to him. This section, about acknowledgments, can be translated to one word: gratitude.
Personally, all messages from Ilfak and Hex-Rays expressing their trust and praises on my previous articles
are one of most motivation to keep writing as well readers who send me even a single message thanking
me. Once again: thank you for everything, Ilfak.
I have chosen a quote to start each article to subtly show my thinking about life and information security in
general, sometimes mirroring the present days and all challenges that have forced me to make a deep
reflection over. At the end of day, we should invest in the work that we really love doing, no matter our
age, because life is short, and the ahead day is our future. Enjoy the journey!

3. References

It is always a complex task to provide references and recommendations to any topic, but I want to leave
few references I have used in the last years, and which might help readers to learn about the theme,
independently whether working on vulnerability research or malware analysis:
2|Page
https://exploitreversing.com

▪ Microsoft Learn: https://learn.microsoft.com/en-us/windows-hardware/drivers/

▪ Windows drivers samples: https://github.com/Microsoft/Windows-driver-samples
▪ Windows Internals 7th edition book (Parts 1 and 2) by Pavel Yosifovich , Alex Ionescu, Mark
Russinovich and David Solomon, and Andrea Allievi, Alex Ionescu, Mark Russinovich and David
Solomon, respectively.
▪ Practical Reverse Engineering by Bruce Dang, Alexandre Gazet and Elias Bachaalany.
Mostly (over 95% of time), I have used the official Microsoft Documentation and respective Windows
drivers sample referred by the first two items above, but both Windows Internals books and Practical
Reverse Engineering book offer an excellent coverage about the topic.

4. Kernel drivers review

I don’t have an perpective to get into detail about kernel driver programming here and, certainly, it
would be impossible to touch a complex theme over a simple article, but I will try to do a minimum
revision about the topic and hopefully these words not only will help readers now, but will provide the
necessary foundation to the future ones. Actually, learning about drivers will help readers a lot while
researching for vulnerabilities in kernel drivers, as also using fuzzing tools to prospect such bugs.
To our context and concern (far away from formal WDM classification), we have distinct types of drivers:
▪ device driver: it communicates with hardware devices like printers, USB sticks and other ones.
▪ software kernel driver: this type of driver runs and establishes communication with the kernel
through resources offered by the system. Additionally, it is not the goal of this type of driver to
communicate directly with a physical device.
▪ mini-filter driver: it is a software driver that can monitor, intercept and change data transferred
between applications and/or drivers and the system (kernel or file system, for example). At the
same way, this kind of driver doen’t communicate directl with the device driver.
Certainl, we aren’t intereted in learning about device driver in thi article (although it is a fascinating
topic), but referring to device drivers is still a broad term, which could cause some confusion. In fact, a
more precise name would be function drivers, and without forgetting that we also have bus drivers that
are responsible for establishing communication between a device a PCI-X or USB bus, for example.
Anyway, in this section we will review the main concepts about kernel drivers, and in the next one we’ll
refresh concepts related to minifilter drivers.
If reader get involved in developing kernel drivers, so they will quickly learn that the development process
brings a series of challenges because as driver run on the kernel side, so any unhandled exception probably
will crash the system and, according to my experience, finding bad lines of code is not always something
trivial. One of many things that will be explained later in this article is that kernel drivers can run in
DISPATCH_LEVEL (IRQL 2), which presents a different consequence from userland applications that always
run in PASSIVE_LEVEL (IRQL 0). In fact, there is a quite extensive list of changes while programming and
writing kernel drivers than while writing user mode application, starting by the fact that most standard
libraries that help us a lot while writing userland applications are not available in kernel mode. We also
have the same concerns about security and, for example, if a driver is unloaded from memory without
3|Page
https://exploitreversing.com

doing the necessary cleaning, so there will be a memory leakage that only will be released in the next
reboot, which is also a standard issue while writing user mode programs. Unfortunately, there is an
extensive list of other programming hurdles. Of course, all of these concerns do not arise while reversing
code and understanding about internals, but they continue to be relevant aspects for differentiating user
mode and kernel mode code. Regardless of this difficulties, kernel drivers continue being an import stuff
while researching vulnerabilities and also used by criminals as an infection vector.
Another critical point is that, while writing and even analyzing a driver, we have to know that there are
different driver models that can used, which can interfere in our understanding about main characteristics:
▪ kernel drivers: Windows NT driver model and KMDF (Kernel-Mode Driver Framework).
▪ file system mini-filter drives: minidriver model.
▪ device drivers: KMDF (Kernel-Mode Driver Framework) and UMDF (User-Mode Framework Model),
and WDM (Windows Driver Model).
We need to choose a starting point, so explaining concepts related to the code, which will help while
reversing kernel drivers, could also be useful to initiate a brief discussion about the theme. Readers will
find over all kernel drivers the DriverEntry( ) routine, which is similar to the main function in C programs
that operate on the userland. This routine serves as a pivotal point to other functionalities called by the
driver. Actually, one of the main tasks performed by the DriverEntry routine is initializing structures and
resources that will be used by the driver at a later moment. In other words, it works like a midway point to
invoke other routines and prepare data structure for them.
Eventually, we also will find an unload routine that is associated with a driver object’s member named
DriverUnload, which is called automatically when the driver is unloaded and, as readers might expect, it is
responsible for performing cleaning tasks. I will be discussing about driver object, device objects and other
concepts in the next paragraphs, but for now you should know that a driver object is the parent of any
other object, and different objects such as timers, spinlocks, device objects and so on are included in this
list and, at the same way that happens for user mode application, synchronization is also a critical
component on the kernel side.
Drivers can be installed as service (sc create <driver name> type= kernel binPath= <driver path>) and, as
other services, an entry in created under HKLM\System\CurrentControlSet\Services. For sure, if Microsoft
did not sign this driver, it is necessary to setup the machine to booting in testing mode by executing bcedit
/set testsigning on followed by shutdown /r /t 0. Furthermore, whether you want to load the driver
without installing it, so there is the option to use OSR loader (available on
https://www.osronline.com/article.cfm%5Earticle=157.htm). Being honest, I haven’t ued it for a long
time, but probably it still works for legacy drivers and older versions of Windows.
We should remember that there are three main different types of memory given by POOL_TYPE
enumeration (for legacy APIs) from wdm.h (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/ne-wdm-_pool_type) or POOL_FLAGS enumeration for new APIs
(https://learn.microsoft.com/en-us/windows-hardware/drivers/kernel/pool_flags) that are used by
drivers: Paged Pool (pages can be paged out), Non-Paged Pool (pages always are kept on memory) and
NonPagedPoolNx (page alwa are kept on memor and don’t have execute permiion). Additionall, it
makes sense to mention Session Paged Pool, which can be paged but it is session independent.

4|Page
https://exploitreversing.com

Therefore, while analyzing kernel drivers, we will see routine invocations of several kernel specific memory
pool allocation functions like ExAllocatePool( ) (deprecated in Windows 10 version 2004),
ExAllocatePoolWithTag( ) (deprecated in Windows 10 version 2004), ExAllocatePool2
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-exallocatepool2),
ExAllocatePool3 (https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
exallocatepool3) and so on. It is a well-known fact that memory regions allocated with most of these
functions (deprecated and new ones) might have an associated tag, with up to four-byte value (usually in
ASCII) in reversing order, to label (tag) the allocated memory.
When a malicious drivers infects a system and allocates kernel non-paged pool memory, we might have a
chance to track these regions of memory used by the threat by looking for a specific tag if it is using one,
although it is not so common nowadays. Even without using a specific framework like Volatility, readers
can track these pools through commands such as poolmon (from WDK) and !lookaside (on WinDbg).
An essential point about kernel drivers is to understand that a single driver does not do everything alone.
Actually, when an I/O request is sent by an application, there will probably be drivers organized in a stack,
which each one is responsible for receiving the request, doing something or not, and passing the request
down to the next driver. Thus, important concepts come up from this point. After drivers are loaded, each
one is represented by a driver object, which has the following structure:

[Figure 1] _DRIVER_OBJECT structure

A driver object holds vital information, which few of them are:
▪ DeviceObject: a pointer to device objects created by the driver (IoCreateDevice( )).
▪ DriverExtension: a pointer to a driver extenion that’ ued b the driver to tore the AddDevice
routine into DriverExtension → AddDevice field.
▪ DriverInit: the entry point, configured by the I/O Manager, to the DriverEntry routine.
▪ DriverUnload: the entry point to the Unload routine.
▪ MajorFunction: a pointer to a dispatch table which contains an array of entry pointers to driver
routines.
Drivers compose a driver stack, and each one is associated with a driver object. Each driver object contains
one or more device objects represented by the _DEVICE_OBJECT structure:

5|Page
https://exploitreversing.com

[Figure 2] _DEVICE_OBJECT structure

Relevant fields in this structure follow:

▪ Type: the value 3 in this field indicates that the given object is a driver object.
▪ ReferenceCount: I/O manager uses this field to track the number of opened handles associated to
the device object.
▪ DriverObject: this field holds a pointer to the driver object (DRIVER_OBJECT), which represents the
loaded image, as explained previously.
▪ NextDevice: this field holds a pointer to the next device object.
▪ AttachedDevice: this field contains a pointer to the attached device object, which typically is
associated to a filter driver (not always).
▪ CurrentIrp: this field contains a pointer to the current IRP if the drivers are currently processing and
whether it has a StartIo routine whose entry point was set up in the driver object. StartIo and IRP
will be briefly commented later.
▪ Timer: this field contains a pointer to a timer object.
▪ Dpc: a pointer to a DPC (Deferred Procedure Call) object for the driver object. DPC will be briefly
explained later.
While there are other notable members, these mentioned above are enough for now. Anyway, a device
object (_DEVICE_OBJECT) is a key component because it works as the interface between the client and the
driver. Many functions used by user mode applications points to a device object through symbolic links
(IoCreateSymbolicLink( ) -- https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-
wdm-iocreatesymboliclink) that points to a kernel object.
A small side effect in this context is that a symbolic link (for example: \\.\ExampleDevice) usually points to
some element under \Device directory (devices as \Device\ExampleDevice are created by calling
IoCreateDevice( ): https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-

6|Page
https://exploitreversing.com

iocreatedevice), which can not be accessed from the user mode, so it is necessary to invoke
IoGetDeviceObjectPointer( ) to get the access to them (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-iogetdeviceobjectpointer).
About APIs mentioned in the last two paragraphs, we have the following one:

[Figure 3] IoCreateDevice( )
A brief summary about its parameters follows:
▪ DriverObject: it holds a pointer to driver object, which is received as parameter of DriverEntry( )
routine.
▪ DeviceExtensionSize: it represents the number of bytes reserved for the device extension of the
driver object. A device extension can be used to store private data structure associated to device,
but it is usually used with device drivers and not kernel drivers.
▪ DeviceName: optionally, it points to a buffer that holds the name of device object, as expected.
▪ DeviceType: it determines the device type, which is given by FILE_DEVICE_* constants. To add
them into IDA Pro as enumeration:
o Add the type library named ntddk64_win10 (SHIFT+11 and INS hotkeys).
o Go to Enumerations tab (SHIFT+F10), inert a new enumeration, chooe “add standard
enum by symbol name” and pick up FILE_DEVICE_DISK.

[Figure 4] _FILE_DEVICE enumeration

(truncated)

7|Page
https://exploitreversing.com

▪ DeviceCharacteristics: this parameter specifies one or more constants, but in the kernel driver’
context, it will be zero (0) or FILE_DEVICE_SECURE_OPEN in most cases. Repeating the same steps,
we have done for DeviceType, but this time add FILE_DEVICE_SECURE_OPEN.

[Figure 5] _FILE_REMOVABLE enumeration

▪ Exclusive: this parameter determines whether the device object represents an exclusive device,
which controls and determines whether more than one file object can open the device.
▪ DeviceObject: this parameter holds a pointer to the DEVICE_OBJECT structure, which is allocated in
a non-paged pool.
Based on explained concepts, we have the following scheme:
▪ driver installed → driver object (_DRIVER_OBJECT) → one or more device objects
(_DEVICE_OBJECT).
So far, the only mentioned driver routine was DriverEntry, which has the following signature:
NTSTATUS DriverEntry(
_In_ PDRIVER_OBJECT DriverObject,
_In_ PUNICODE_STRING RegistryPath
);

The first parameter is a pointer to DRIVER_OBJECT and the second parameter is a pointer to RegistryPath
structure, which is a UNICODE_STRING, and that specifies the Parameters key of the driver in the Registry:

[Figure 6] _UNICODE_STRING structure

8|Page
https://exploitreversing.com

Besides core tasks performed (actually, invoked) in DriverEntry, there is another still more relevant role
performed by the same routine that is the initialization of the Dispatch Routines, which is an array of
function pointers, and that makes part of the _DRIVER_OBJECT structure (MajorFunction member).
All indexes of this array have IRP_MJ_ prefix and, as expected, they represent the IRP major function
codes. Drivers must set entry pointers into this array, which set up associated and responsible routines for
handling and manipulating each one of planned operations and, finally, attending IRP requests.
We still have a pending list of concepts that need to be explained and cleared. An IRP (I/O Request Packet)
is a structure that represents an I/O request packet, and it is used by drivers to carry information and
communicate with other drivers. In other words, it works like a data format to be used in a well-defined
standard for communication between driver layers.
The IRP, defined in wdm.h file, is a really large structure and has many fields, but most of them are unions.
If the readers want to examine the struct using Internet, so the following reference could be interesting:
https://www.vergiliusproject.com/kernels/x64/Windows%2011/22H2%20(2022%20Update)/_IRP
Personally, I prefer retrieving the _IRP structure from IDA Pro by performing the following steps:
1. open a PE format binary in IDA Pro
2. go to Type Libraries (SHIFT+F11)
3. add ntddk64_win10 or any other similar library (ntddk_win7).
Now go to Structures tab (SHIFT+F9) and add the standard structure named _IRP, as shown below:

[Figure 7] _IRP structure: header

9|Page
https://exploitreversing.com

There are fields that provide us with important context and information about kernel driver operation,
which few of them will be explained as necessary, and need to be complemented with new concepts that
will be introduced later. Even it is not shown on the previous image, an IRP has fixed part containing the
header (caller’ rea ID evice objec ’ a re I/O a u block an o on) that is used by I/O manager
to manage the IRP and a second part that is specific to each driver (I/O stack location), which holds
parameters such as function code of the requested operation and its respective context:

IRP HEADER STATIC PART

IO_STACK_LOCATION

IO_STACK_LOCATION DYNAMIC PART

IO_STACK_LOCATION
......
[Figure 8] IRP representation

We are going to make new notes on this topic later. Focusing on the IRP major codes topic again, there is a
series of IRP major codes that are used by drivers to call the respective dispatch routine in reaction to a
specific I/O request. These IRP major codes work as indexes in an array of function pointers.
As each kernel driver offers different functionalities, so they provide different dispatch routines to handle
I/O requests passing the IRP major codes shown below:
▪ IRP_MJ_CLEANUP: this IRP major code is used for invoking a DispatchCleanup routine when the
driver needs to release resources as memory and any other object whose respective reference
counter has reached zero, so it is an appropriate and recommended routine for cleanup that is not
related to file handles.

▪ IRP_MJ_CLOSE: this IRP major code is used for invoking a DispatchClose routine when the last
handle to a file object associated with a device object has been closed and released, and any
request has been closed or cancelled.

▪ IRP_MJ_CREATE: this IRP major code is used for calling a DispatchCreate routine to open a handle
to a device or file object. A well-known example occurs when a kernel driver calls functions like
NtCreateFile | ZwCreate, and an IRP_MJ_CREATE is sent to accomplish the open operation.

▪ IRP_MJ_DEVICE_CONTROL: this IRP code, which has an associated DispatchDeviceControl routine,

is a consequence of invoking DeviceIoControl( ), which is responsible for sending a I/O control code

10 | P a g e
https://exploitreversing.com

(it could be a well-known or a private one) to the target device driver. In most situations, the
routine will pass the IRP to the next lower driver, but there are exceptions. Readers should
remember that the first two members of DeviceIoControl( ) are associated to the referred purpose:

[Figure 9] DeviceIoControl

The first two parameters of this function are:

▪ hDevice: this parameter represents a handle to a device driver, which can be easily
retrieved by using CreateFile( ) (https://learn.microsoft.com/en-
us/windows/win32/api/fileapi/nf-fileapi-createfilea).

▪ dwIoControlCode: this parameter specifies the control code for the operation. There are
multiple set of control codes organized according to the type of target device:

▪ cdrom: https://learn.microsoft.com/en-us/windows-hardware/drivers/storage/cd-
rom-io-control-codes
▪ communication: https://learn.microsoft.com/en-
us/windows/win32/devio/communications-control-codes
▪ device management: https://learn.microsoft.com/en-
us/windows/win32/devio/device-management-control-codes
▪ directory management: https://learn.microsoft.com/en-
us/windows/win32/fileio/directory-management-control-codes
▪ disk management: https://learn.microsoft.com/en-us/windows/win32/fileio/disk-
management-control-codes
▪ file management: https://learn.microsoft.com/en-us/windows/win32/fileio/file-
management-control-codes
▪ power management: https://learn.microsoft.com/en-
us/windows/win32/power/power-management-control-codes
▪ volume management: https://learn.microsoft.com/en-
us/windows/win32/fileio/volume-management-control-codes

▪ IRP_MJ_FILE_SYSTEM_CONTROL: as readers might expect, file system drivers commonly use this
IRP major code.

11 | P a g e
https://exploitreversing.com

▪ IRP_MJ_FLUSH_BUFFERS: this IRP major code means a request to the device to flush its internal
cache, and such code is used for invoking the DispatcFlushBuffers routine.

▪ IRP_MJ_INTERNAL_DEVICE_CONTROL: it is pretty similar to IRP_MJ_DEVICE_CONTROL, and

readers will see this code when another driver calls IoBuildDeviceIoControlRequest( ) or even
IoAllocateIrp( ), for example. Basically, it can be interpreted as a code used for driver-to-driver
communication while IRP_MJ_DEVICE_CONTROL is used for application to driver communication.
Finally, it is used for invoking DispatchInternalDeviceControl routine.

▪ IRP_MJ_PNP: this code is used over a request for any Plug & Play operation (enumeration or
resource balancing, for example) and used for invoking the DispatchPnP routine.

▪ IRP_MJ_POWER: this IRP code is used by requests, through the Power Manager, to invoke the
power callback (DispatchPower routine).

▪ IRP_MJ_QUERY_INFORMATION: this IRP code is used for invoking the DispatchQueryInformation

routine, which usually gets meta-information about a file or even a handle. For example, this event
happens when a driver call ZwQueryInformationFile( ) (https://learn.microsoft.com/en-
us/windows-hardware/drivers/ddi/ntifs/nf-ntifs-ntqueryinformationfile). Of course, the driver is
not required to handle this kind of request.

▪ IRP_MJ_SET_INFORMATION: this IRP code is sent by the operating system as a request

(ZwSetInformationFile( )) to set metadata about a file or even a handle and, as in other cases, it
invokes the DispatchSetInformation routine.

▪ IRP_MJ_SHUTDOWN: this IRP code is handled by drivers that are responsible for mass-storage
devices with internal caches, and it is used for invoking the DispatchShutdown routine. As drivers
are organized in a stack, all intermediate drivers that are associated with mass-storage devices
need to be able to manage such requests. Of course, drivers must complete any transfer of data
that is currently in cache before finishing the shutdown request.

▪ IRP_MJ_SYSTEM_CONTROL: all drivers must provide a DispatchSystemControl routine that is

invoked to handle IRP_MJ_SYSTEM_CONTROL requests, and these requests are sent by
components of WMI when a user mode data consumer requests WMI data.

▪ IRP_MJ_READ: this IRP code is used for calling DispatchRead routine, which acts when application
makes requests (ReadFile( ) and ZwReadFile( )) to transfer data from the device to the application.

▪ IRP_MJ_WRITE: this IRP code is used for invoking the DispatchWrite routine, which is used by
drivers that transfer data from the system to the associated device.
Thus, so far, we have few conclusions:
▪ a driver object (_DRIVER_OBJECT) holds one or more device objects (_DEVICE_OBJECT), which are
the main interface of communication between the application and driver.

12 | P a g e
https://exploitreversing.com

▪ APIs on user-mode refer to device objects as their parameters.

▪ To a kernel driver to become really useful it has to register Dispatch Routines to serve diverse types
of requests (user-land or kernel-land) that are done by sending one of IRP codes.
▪ In many public drivers, readers will find drivers implementing dispatch routines to handle userland
application’ call uch a ReadFile( ), DeviceIoControl( ) and WriteFile( ), for example.
▪ The IRP structure (_IRP) holds the necessary information from a request and it is used to carry
information and communicate with drivers between layers in the driver stack.
▪ The IRP’ content can hold common information for all drivers in the stack, but it also carries
private information for specific drivers over the same stack.
▪ A device object is created by drivers through IoCreateDevice( ) (exported by I/O manager).
▪ Observing Figure 2, a device object (_DEVICE_OBJECT) is linked to the next one through the
NextDevice member.
As a summary, the general execution flux established by the I/O manager is:
▪ Accepting requests from different applications.
▪ For each request it creates an IRP to represent that request.
▪ Afterwards, it sends each request to its respective drivers.
▪ It manages and tracks these IRPs until they are completed.
▪ Finally, it returns the result of the operation to the application that made the request.
However, few points are still pending to be explained so far:
▪ What are IRQLs and what are available values?
▪ What is a StartIO routine?
▪ What is DPC and which is its purpose?
▪ How are IRPs passed and stored from an upper kernel driver to a lower one?
IRQL (Interrupt Request Level) is a Windows mechanism to manage interrupts according to the respective
level of importance in the operating system context. When I mention interrupts (IRQ), readers probably
remember that there are hardware (asynchronous) and software interrupts (synchronous), and Windows
creates a map assigning a priority (IRQL) to a given interrupt source emitted by a device, although this map
is different from CPU to CPU. Thus, each CPU has an associated IRQL value, and it could be interpreted as a
particular register.
Anyway, the IRQL is represented by a number, and rule is that any code running with a lower IRQL can’t
preempt a code running with a higher IRQL, and the kernel prioritizes pieces of code such as kernel drivers
over other ones according to the higher level of priority.
We should note that IRQL (Interrupt Request Level) is not equal to IRQ (Interrupt Request), which is
related to hardware, and it is also not equal to thread priority because thread priority is an individual
thread’ propert.
The usual IRQL level are:
▪ PASSIVE LEVEL (value 0): at this level, no interrupt vectors are masked, and it is the level where
most threads usually run. It is the normal IRQL. Actually, most kernel driver routines such as
DriverEntry( ), Unload( ), AddDevice( ) as well as dispatch routines run at this level.

13 | P a g e
https://exploitreversing.com

▪ APC LEVEL (value 1): it’ the level ued b APC (Asynchronous Procedure Calls), which is a function
that executes in the context of a thread. In few words, each thread has an own APC queue and
when an application sends an APC to a thread by invoking QueueUserAPC( ) (actually, a wrapper to
NtQueueApcThread( ) -- https://learn.microsoft.com/en-
us/windows/win32/api/processthreadsapi/nf-processthreadsapi-queueuserapc), it passes the
address of the APC function as argument and an interrupt is issued by the system. Therefore,
readers can understand that queueing an APC works as a request for the thread calls/invokes the
given APC function. The application is only able to deliver an APC to a thread when this thread is in
alertable state (it called SleepEx( ), WaitForSingleObjectEx( ), WaitForMultipleObjectsEx( ) and so
on), and this APC from the thread’ queue is executed when the thread transits from alertable state
to running state. The same concept is used when malware threats do APC injection, which is only
possible when the target thread is in alertable state. At the end of day, APC is a subtle technique
that makes it possible to execute a callback method (the function passed as argument to the APC)
in an asynchronous way. APCs can be listed by using !apc extension on WinDbg.

▪ DISPATCH LEVEL (value 2): it’ the higher IRQL aociated to oftware interruption. DPC (Deferred
Procedure Call) runs at this level as well as the thread dispatcher, and it is responsible for the post-
processing of a driver after a first, critical and short job has been performed by the ISR (Interrupt
Service Routine), which is registered (IoConnectInterrupt( ) -- https://learn.microsoft.com/en-
us/windows-hardware/drivers/ddi/wdm/nf-wdm-ioconnectinterrupt) by a device driver, runs at
DIRQL (Device Interrupt Request Level), and it is responsible for a really minimal work before
queueing (KeInsertQueueDpc( ) -- https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-keinsertqueuedpc) a DPC that will be executed when the IRQL
drops to a lower level. Furthermore, in the kernel driver’ context, routine uch a StartIo( ),
IoTimer( ), Cancel( ), DpcForIsr( ), CustomDpc( ) and so on also run at this level. Finally, it is
appropriate to mention that any thread waiting on kernel object (event, emaphore, mutex…) at
this level causes a system crash.

▪ DIRQL (value 3 and higher): these levels are related to hardware interrupts.
A kernel code, which can be interrupted by other kernel code with higher IRQL, is able to change the
current IRQL (from the current CPU) by calling functions such as KeLowerIrql( )
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-kelowerirql) and
KeRaiseIrql( ) (https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
keraiseirql). In the order side, it is not possible to raise the IRQL from a user mode application.
Although the APC topic is really attractive, the only difference between PASSIVE_LEVEL and APC_LEVEL is
that a process running at APC_LEVEL cannot get interrupted by APC interrupts. While explaining about high
level drivers (not associated to devices) that process IRP, we will be focused on PASSIVE_LEVEL and
DISPATCH_LEVEL to avoid getting distracted with other topics.
Anyway, I know that professionals usually ask about the IRQL and respective thread context when one of
commented dispatch routines (callbacks) is called, so I retrieved a list from Microsoft
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ifs/dispatch-routine-irql-and-thread-
context) that could help you:

14 | P a g e
https://exploitreversing.com

[Table 1] Dispatc routines IRQL and T read’s context (credit: Microsoft)

According to experience, multiple crashes caused by drivers come from a wrong action executed at a
higher level than possible to start a given operation. Furthermore, crashes also happen because such
drivers incorrectly assume to be in a certain thread context that, actually, is not true or even possible.
Analyzing the provided table above, it is quick to realize that most dispatch routines are called from
PASSIVE_LEVEL IRQL and from a non-arbitrary context. That’ the reaon that the recommended approach
is not assuming a certain context unless you are sure about which context is invoking the thread. Of
course, as a security researcher this concern is lower because we are looking for a vulnerability or even
reversing the code of malicious drivers, but for programmers these concepts exposed here are really
important.
Returning to our main discussion, readers can check basic information on drivers according to what we
have discussed so far by using WinDbg/WinDbg Preview (that is available on Microsoft Store):

15 | P a g e
https://exploitreversing.com

[Figure 10] Listing device names under \Device (truncated output)

The output above i baed on Window 11. Jut in cae reader don’t know how to intall WinDbg, it comes
from Windows SDK installation. Actually, if readers are interested in developing kernel and minifilter
drivers, so the recommendation is to install few components in the following order:
▪ Visual Studio: https://visualstudio.microsoft.com/downloads/
▪ Windows SDK: https://developer.microsoft.com/en-us/windows/downloads/windows-sdk/.
▪ Windows WDK: https://learn.microsoft.com/en-us/windows-hardware/drivers/download-the-wdk
If readers want to use WinDbg Preview, there are two methods to install it:
▪ From Microsoft Store: https://apps.microsoft.com/store/detail/windbg-preview/9PGJGD53TN86
▪ From command line: winget install windbg
Personally, I always configure the following environment variable: _NT_SYMBOL_PATH=
srv*c:\Symbols*http://msdl.microsoft.com/download/symbols
WinDbg might take a long time to show the complete list of device names, but the idea is getting a list of
devices registered under \Device directory and, from this point, collecting additional information about a
specific driver. As we have the object address given by the output above, our next step is getting the
driver’ name and associated device objects to this driver. Remember: there can be one or more device
objects attached to a driver object. Thus, choosing vmmemctl device as example, execute:

16 | P a g e
https://exploitreversing.com

17 | P a g e
https://exploitreversing.com

[Figure 11] Getting basic information about the dispatch routines.

From these commands we got:
▪ the list of device objects associated with the driver.
▪ summarized information about the given device object.
▪ the list of the dispatch routines associated to the driver object.
If readers are wondering about how to list any pending IRPs, the WinDbg offers a command too:

[Figure 12] Listing pending IRPs (truncated output)

18 | P a g e
https://exploitreversing.com

We have learned that a basic kernel driver likely will have relevant routines, mechanisms and objects that
are critical for its perfect operation:
▪ DriverEntry( ) routine, which is called from IRQL == PASSIVE_LEVEL, and responsible for providing
an entry point to driver routines, initializing or even creating object, allocating non-paged or paged
memory using ExAllocatePoolWithTag( ) (for example) or retrieving a key-information from
Registry. Furthermore, it can also be used to call PsCreateSystemThread routine, which creates a
system thread to execute in kernel mode.

▪ Unload( ) routine, which is responsible for freeing resources, and that is a strong requirement for
WDM (Windows Driver Model) drivers. The I/O manager calls the Unload routine whether there is
not any reference or pending IRP request associated to device objects of the driver. Readers may
find a series of functions inside this routine such as ExFreePool( ), IoDeleteSymbolicLink( ),
PsTerminateSystemThread( ), IoDeleteDevice( ) and so on.

▪ An associated device object (remember: the device object is the actual interface of communication
with the driver).

▪ A symbolic link (created by IoCreateSymbolicLink( ): https://learn.microsoft.com/en-us/windows-

hardware/drivers/ddi/wdm/nf-wdm-iocreatesymboliclink) associated to the device object.

▪ We will have kernel drivers which holds one ore more dispatch routines handling function codes
such as IRP_MJ_CLOSE, IRP_MJ_READ, IRP_MJ_CREATE or IRP_MJ_DEVICE_CONTROL,
IRP_MJ_INTERNAL_DEVICE_CONTROL, IRP_MJ_SYSTEM_CONTROL, because these routines are
usually essential to most of kernel drivers, and in different cases we will have the opportunity to
work with other ones like IRP_MJ_SET_INFORMATION, IRP_MJ_CLEANUP and
IRP_MJ_SHUTDOWN, for example. If readers are programming then system functions/macros such
as ObDereferenceObject (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-obdereferenceobject), PsLookupThreadByThreadId
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/ntifs/nf-ntifs-
pslookupthreadbythreadid), and IoCompleteRequest (explained below) will be very useful.

▪ A dispatch routine might have nothing else to do with a driver, so it would complete an IRP input
with a simple STATUS_SUCCESS, but it could be suitable in contexts and scenarios. For example,
DispatchClose routine (handles IRP_MJ_CLOSE I/O function code) could be responsible for
notifying that all references to a given file were removed. Eventually, drivers that never could be
unavailable, and the DispatchClose routine wouldn’t be called. At the same way, DispatchCleanup
routine (handles IRP_MJ_CLEANUP I/O function code) is used to perform cleaning operations after
handles of a given object have been released and, for each IRP request, this routine is composed by
operations such as setting Cancel routine’ pointer to NULL, cancelling all IRP related requests (for
example, associated to the object that has been closed) that are still in the queue and, finally,
calling the IoCompleteRequest( ) routine to complete the IRP and returning STATUS_SUCCESS.
Maybe, the most important lesson is that, although few dispatch routines will be seen in most of
software drivers, it is recommended not assuming whether one of them is more important or even
critical than other one because each driver has a particular goal and different role.
19 | P a g e
https://exploitreversing.com

Of course, the list of routines mentioned above is regarding only a basic software kernel driver, which is
part of the goal of this article, but we could explain much more about them. For sure, other routines might
be relevant for readers interested in writing a device driver such as AddDevice, StartIo, ISR, DPC routines
and so on.
As happens with userland applications, the I/O manager also manages synchronous and asynchronous
operations and as expected, over an anchronou operation the kernel driver doen’t have an obligation
to process IRP requests in a specific order. In other words, a kernel can start processing the next IRP
request without having finished the previous one. From this point, the kernel driver can pass down the IRP
to the next drivers in the stack and continue the request processing.
A concept that I have not mentioned yet is completion routine, an optional feature/function, which is
called by IoCompleteRequest( ) function, and that performs an important role over the kernel processing
because a driver can register a completion routine (IoCompletion( ) routine --
https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-iocompleterequest) that
will be invoked by I/O manager soon a kernel driver has finished the processing an IRP.
The IoCompleteRoutine( ) makes the reverse path by sending back the IRP to the upper layer driver in the
driver stack. Thus, in a hypothetical asynchronous scenario, it is likely having a kernel driver processing the
next IRP while the I/O manager calls the completion routine from other driver that finished its IRP
processing.
Drivers provide the status of an operation within the I/O status block of IRP. Additionally, drivers can keep
the status of the operation inside the driver extension, which is really useful in the context with two or
more drivers that are part of the same stack. When a device object is created through IoCreateDevice
function (https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
iocreatedevice), the DriverExtensionSize parameter is used to prepare the driver for scenarios like
explained in this paragraph. A driver extension can be created or initialized by
IoAllocateDriverObjectExtension( ), which is invoked by DriverEntry( ) routine.
During the usage of the concept of driver stack, I am not assuming a specific number of drivers in this stack
to keep the explanation wide enough. However, it is suitable to explain that whether any driver, which
make part of the tack, doen’t receive a handle, or even pa down the IRP to next driver through the
right way, the system can (and probably will) crash. Additionally, and as a side note, so far we have mostly
explained and handled I/O operation as being IRP requests. Nonetheless, there is another type of
operation called Fast I/O that doen’t generate IRP and goes to specific drivers to complete the request,
but it is not the moment to discuss this kind of operations in this section.
Returning to outstanding points, it is time to provide a concise explanation about ISR and StartIo routines.
In general, hardware interrupts are associated with a priority (IRQL, as we learned), the device registers
(through IoConnectInterruptEx / WdmlibIoConnectInterruptEx routines ) one or more ISR (Interrupt
Service Routine) to handle interrupts. Drivers associated to physical devices, which generate interrupts,
need to have one ISR, at least. Once again, threads have an associated priority while CPUs have an
associated attribute named IRQL.
In other words, each time an interrupt is generated to that specific device, the system calls an ISR, which
could be InterruptService or InterruptMessageService routines. Anyway, it will be executed with the same
associated IRQL that the request arrived (masking interruptions at lower level) and, if the IRQL is zero (for
20 | P a g e
https://exploitreversing.com

example) before the ISR, then it will be raised to the same higher level of the interrupt (there in’t context
switch when IRQL is 2 or higher, and accessing paged memory causes system crash) and, after the ISR
completes, the IRQL will return to the previous level. Additionally, it is possible to enable or disable an ISR
by calling IoReportInterruptActive( ) or IoReportInterruptInactive( ) functions, whose references follow
below:
▪ https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
ioreportinterruptactive
▪ https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
ioreportinterruptinactive
ISR is short and fast. In few words, it should handle the interrupt (stop the interrupt), gather and save the
state (context), and queues a DPC (DpcForIsr or CustomDpc routines) through IoRequestDpc or
KeInsertQueueDpc routines, respectively, soon the IRQL drops below DISPATCH_LEVEL.
The DPC will be responsible for managing the I/O operation that will be conducted at a lower level than the
ISR. The ISR does only a little part of the I/O processing (the initial request), and the heavy work is left to
the DPC (Deferred Procedure Call), which has the assignment to complete the I/O operation, queue the
next IRP (ensuring the next I/O operation) and, as explained, finish the current IRP when it is possible.
The system provides a DPC object for each device object, and the first (and default) routine is DpcForIsr( ).
In case of driver to need to create additional DPC objects then CustomDpc routines are associated to these
new DPC objects. Both DpcForIsr and CustomDpc routines are called in arbitrary DPC context at
IRQL_DISPATCH_LEVEL (IRQL value 2).
The IoInitializeDpcRequest( ) routine is responsible for registering the DpcForIsr routine, receiving a
pointer to a device object represented by DEVICE_OBJECT structure (remember: a DPC object for each
device object) and also receiving a pointer to the provided DpcForIsr routine, as shown below:

[Figure 13] IoInitializeDpcRequest routine

To register a CustomDpc routine associated with a device object, the driver must call KeInitializeDpc
routine. The first parameter is a pointer to a KDPC structure, the second parameter is a pointer to the
CustomDpc routine, and the last parameter holds the context. It is timely to highlight that CustomDpc
routine is not associated with the DeviceObject, as shown below:

[Figure 14] KeInitializeDpc routine

The IoRequestDpc routine is called by ISR for queueing the DpcForIsr routine to be executed:

21 | P a g e
https://exploitreversing.com

[Figure 15] IoRequestDpc routine

The Irp parameter is a pointer to the current IRP and Context parameter is passed to the routine.
Another routine to queue a DPC for execution is KeInsertQueueDpc, which has as argument a pointer to
KDPC routine and two arguments dedicated to context, as shown below:

[Figure 16] KeInsertQueueDpc routine

According to https://www.vergiliusproject.com/ , the representation of the _KDPC structure is the
following one:
//0x40 bytes (sizeof)
struct _KDPC
{
union
{
ULONG TargetInfoAsUlong; //0x0
struct
{
UCHAR Type; //0x0
UCHAR Importance; //0x1
volatile USHORT Number; //0x2
};
};
struct _SINGLE_LIST_ENTRY DpcListEntry; //0x8
ULONGLONG ProcessorHistory; //0x10
VOID (*DeferredRoutine)(struct _KDPC* arg1, VOID* arg2, VOID* arg3, VOID* arg4); //0x18
VOID* DeferredContext; //0x20
VOID* SystemArgument1; //0x28
VOID* SystemArgument2; //0x30
VOID* DpcData; //0x38
};
[Figure 17] _KDPC structure
Although it is not the focus of this introduction about kernel drivers, there is another type of DPC named
Threaded DPC, which executes at PASSIVE_LEVEL, and that can be preempted by a normal DPC, but not
by other threads. Analyzing this feature from a strict point of view, it presents a good alternative because
as normal DPC cannot be preempted by other normal DPC, a system with multiple queued DPCs might

22 | P a g e
https://exploitreversing.com

present a big latency and, eventually, cause performance issues. Therefore, Threaded DPC, which is
enabled by default (HKLM\System\CCS\Control\SessionManager\Kernel\ThreadDpcEnable), might be
interpreted, in most cases, as a better choice than normal DPC (but it is not a rule).
Beide DPC’ uage with ISR, DPC can be alo ued with kernel timers that have a remarkably similar
behavior to other objects like semaphores, event, mutex, events and so on, as any driver can use these
objects during synchronization tasks since it happens in IRQL==PASSIVE_LEVEL and non-arbitrary context.
Independently of which of mentioned kernel objects is being taken, we can use typical waiting routines
such as:
▪ KeWaitForSingleObject (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-kewaitforsingleobject)
▪ KeWaitForMultipleObjects (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-kewaitformultipleobjects).
Getting into quite few details, kernel timer is associated and represented by a KTIMER or EX_TIMER
structure, and it is used to time out operations of kernel routines or even scheduling new operations
(other researchers and programmer might be use the term “action” or “tak”) to be executed from time
to time, so presenting well-established periodic behavior.
Kernel timers based on KTIMER structure can be set by using KeSetTimer (the timer object must have
been initialized using KeInitializeTimer/KeInitializeTimerEx routine, and its DPC also must have been
initialized by calling KeInitializeDPC routine) to set absolute or even relative interval, which after it expires
it is set to signaled state.

[Figure 18] KeInitializeTimerEx

[Figure 19] KeSetTimer

Signaled state for timers indicates, as a flag is up, that the timer is done and any DPC object that has been
inserted in the DPC queue can execute as soon it can (during a red team operation, it would be the
moment to execute the injected code done through DPC injection).
To set a recurring time (to attribute the periodic behavior), use KeSetTimerEx routine. If the timer is based
on EX_TIMER structure (it must be allocated using ExAllocateTimer routine and can be deallocated using
ExDeleteTimer routine), then the ExSetTimer routine can be used to start a timer operation and the
expiration time. The prototype of ExAllocateTimer function is shown below:

23 | P a g e
https://exploitreversing.com

[Figure 20] ExAllocateTimer routine

Therefore, a CustomTimerDpc routine can be associated with a timer to be executed as soon as possible
when the timer is signaled. The two types of timers are notification timer (once it signaled it means the
the specified time has been reached, all threads have a green-light to proceed, and the state of the timer
stays as signaled until it is explicitly reset) and synchronization timer (once it signaled, it is kept in signaled
state until a thread waiting on it is released, and it is automatically reset to non-signaled state). If a driver
needs to disable a timer, there is the option to call KeCancelTimer routine (for timers based on KTIMER
structure) or ExCancelTimer (for timers based on EX_TIMER structure).
According to what we have reviewed so far, the DPC routine will run when the IRQL drops below
DISPATCH_LEVEL or even when a configured timer expires. No doubts, this explanation could be
extended over other kernel dispatcher objects such as mutex, events, semaphores or even other
techniques like work items and spin locks, but all these concepts can be easily learned from any resource
as Microsoft Learn (MSDN) website and books mentioned at the beginning of this article.
Returning to our planned agenda (again), we have pending items to be explained, at least, so it is time to
briefly comment about I/O stack locations as well offers a supplemental view about IRP being dynamically
passed down to other layers.
As we already know and explained previously, all I/O requests to drivers at a lower level on the driver stack
are based on IRP (I/O Request Packet). The I/O Manager allocates an array of I/O stack locations
(IO_STACK_LOCATION structure) for every configured IRP (there is a parameter named StackSize in
IoAllocateIRP function to specify the number of I/O stack locations), and each element of this array is
associated with a driver in the driver stack. In other words, the number of I/O stack locations from this
array can be translated to the number of drivers in the driver stack.

[Figure 21] IoAllocateIrp routine

Readers could use IoAllocateIrpEx function, which has three parameters, and the first one allows us to
pass a pointer to the device object. In this case, if the DeviceObject parameter is set to
DEVICE_WITH_IRP_EXTENSION, the call is intended to allocate space for IRP extension.
As each driver is the owner of the I/O stack location in the IRP, this driver can invoke
IoGetCurrentIrpStackLocation routine, which returns a pointer to t e caller’s I/O stack location in t e IRP,
to get driver specific information about the I/O operation. Actually, the I/O opera ion’ informa ion i
divided between the IRP header and the current I/O stack location.

24 | P a g e
https://exploitreversing.com

[Figure 22] IoGetCurrentIrpStackLocation routine

Each driver of the driver stack is responsible for configuring e nex lower river’ I/O ack loca ion (I/O
stack location that makes part of the IRP structure ) by calling IoGetNextIrpStackLocation routine, which
grants access to the lower I/O stack location exactly to accomplish this set up, and as readers have realized,
it is a critical task in a stack of drivers. Therefore, the I/O manager sets up the IRP header and the first I/O
stack location, and all of the next ones (for each driver) are set up by the driver immediately above.

[Figure 23] IoGetNextIrpStackLocation routine

Another possibility that should be mentioned is that a driver could be satisfied with the IRP processing
and no longer interested in making further changes. Therefore, it would call IoSkipCurrentIrpStackLocation
macro to set for the next driver in the stack exactly with the same IO_STACK_LOCATION structure that the
current driver received.
These I/O stack locations are useful for storing context about an operation such as an I/O completion
routine (registered by calling IoSetCompletionRoutine or IoSetCompletionRoutineEx functions), and it will
be called after IRP having been processed by a lower driver, allowing the I/O completion routine to
perform cleanup tasks, for example.

[Figure 24] IoSetCompletionRoutine

The CompletionRoutine argument is a pointer to an IoCompletion routine, which is called at IRQL equal or
lower than DISPATCH_LEVEL, to be invoked when the immediate lower driver to complete the IRP
processing. The second parameter is a pointer to the IO_COMPLETION_ROUTINE:

[Figure 25] IoCompletionRoutine

25 | P a g e
https://exploitreversing.com

It is really crucial to underscore that I/O completion routine can be registered and configured to any
driver in the driver stack, except the lowest one because each driver stores the completion routine from the
driver immediately above in the driver stack inside its I/O stack location.
Additionally, IoCompletion routine of a driver can be executed in two different moments or conditions: in
an arbitrary thread (thus, it is not possible to know the thread in advance) or even inside a DPC context.
Thus, after a kernel driver has completed the IRP, it invokes IoCompleteRequest routine , which is usually
called from the DpcForIsr routine) to notify that everything is done. Afterwards, the I/O manager verifies
whether the upper drivers offer an IoCompletion routine (as we described) and calls one by one, from the
immediate upper driver up to the highest driver. After everything has been done (all drivers in the stack
completed their IRP processing), so the I/O manager returns a result to the caller application.
The remaining question is: how does the driver forward the IRP to the next lower driver in the stack? It
performs this task by calling IoCallDriver, which is a macro wrapping IofCallDriver routine that accepts two
parameters such as DeviceObject (a pointer to the target device object) and Irp (a pointer to IRP):

[Figure 26] IoCallDriver routine

Now we have a very brief idea of the communication between drivers through the stack, we need to return
to the main idea in the communication between application and drivers that is the real information (data)
transferred during the communication, so it is appropriate to remember about the IRP structure again:

[Figure 27] IRP structure

26 | P a g e
https://exploitreversing.com

As I mentioned previously, I would comment some fields from IRP structure according to the need, and as
we are interested in understanding the data exchange between applications and drivers, so some of these
fields are relevant because, in general, applications can interact with a driver by writing (WriteFile:
https://learn.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-writefile), reading (ReadFile:
https://learn.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-readfile) or even controlling
(DeviceIoControl: https://learn.microsoft.com/en-us/windows/win32/api/ioapiset/nf-ioapiset-
deviceiocontrol) a device or another driver. However, it does not matter the operation, there will be some
transfer of information from application to device driver or vice-versa, and the buffer holding the
information must be pointed during the operation and, this time, other fields of IRP show their
importance:
▪ UserBuffer: this field contains a pointer (address) to a user buffer. Actually, this buffer is an address
of an output buffer, and is used in particular conditions of I/O control code (METHOD_BUFFERED or
METHOD_NEITHER) and respective major function code (IRP_MJ_DEVICE_CONTROL /
IRP_MJ_INTERNAL_DEVICE_CONTROL), as we will learn soon.

▪ SystemBuffer: this field holds a pointer to a system buffer (non-paged pool buffer), which it will be
useful for drivers using buffered I/O and the purpose of the given buffer is determined by the
associated IRP Major code such as IRP_MJ_READ (buffer will be used for reading from a device or
driver), IRP_MJ_WRITE (it will used for writing to a device or driver) and IRP_MJ_DEVICE_CONTROL
(buffer will be used for sending and receiving control data to/from a device or driver).

▪ MdlAddress: this field points to an MDL (Memory Descriptor List), which is defined by a MDL
structure, and followed by an array that describes physical page layout for a virtual memory buffer.
There is a series of functions to work with MDLs such as MmGetMdlVirtualAddress (gets the virtual
address of the I/O buffer described by the MDL), MmGetMdlByCount (retrieves the size of the I/O
buffer), IoAllocateMdl (this function allocates an MDL), IoFreeMdl (this function frees a MDL),
MmInitializeMld (this functions formats a non-paged memory block as an MDL),
MmBuildMdlForNonPagedPool (to initialize the mentioned array following the MDL structure) and
many other ones.
An important aspect to realize is that, regardless of the involvement of any field above, access to any
provided buffer is always controlled by system rules (including security aspects), and eventually a broken
rule will lead to a system crash. For example, accessing a user buffer can be done only from the context of
an application thread (IRQL==0) requesting this access. Nonetheless, associated functions such as DPC or
Start IO can execute from any thread (arbitrary context) where the provided address is meaningless
(different addresses spaces) and IRLQ == 2, which accessing user page is not allowed because part of the
buffer might have been paged out. Unfortunately, not even the dispatch routine might not to be reliable
due to the fact that, although it runs at the same context of the requesting thread and initially at IRQL == 0,
eventually it might run at IRQL == 2 (or higher), over an IRP activity between drivers in the stack.
Therefore, the I/O manager provides us two approaches to access the provided user buffer in a safe way:
▪ Buffered I/O
▪ Direct I/O

27 | P a g e
https://exploitreversing.com

Most of the time, the Buffered I/O method should be used for interactive services transferring a small
amount of data (likely 4 KB or less) between application and drivers. As most of operations are reading or
writing (IRP_MJ_READ and IRP_MJ_WRITE requests, respectively), so a driver selects this method of
operation when the Flag member of the Device Object (DEVICE_OBJECT structure – check the nineth field
of Figure 2), provided by the IoCreateDevice( ), is set as DO_BUFFERED_IO (actually Flag member works as
an OR operation). If the driver needs to handle or execute I/O device control operations through
DeviceIoControl function (IRP_MJ_DEVICE_CONTROL/IRP_MJ_INTERNAL_DEVICE_CONTROL requests), so
the IOCTL code’ value mut mirror thi method b uing METHOD_BUFFERED as its TransferType value.
Buffered I/O operations happen by allocating a buffer with the size of the user buffer inside for an
allocated non-paged pool (ExAllocatePoolWithTag / ExAllocatePool2) and this new address is stored as a
pointer into IRP (specifically, in SystemBuffer member from AssociatedIrp field). Afterwards, it allows
access to this new allocated buffer to the driver and there is no further concern because as the buffer is
stored in a non-paged pool, o driver doen’t run an rik of tring to acce paged-out data. Additionally,
as the address is in the kernel space, it is valid from any process and, better yet, the driver does not need
even to lock it before accessing it. Once the non-paged buffer has been created, data can be copied (by I/O
manager) from the user buffer into this new non-paged buffer for IRP_MJ_WRITE requests, or copied from
this new non-paged buffer to user buffer for IRP_MJ_READ requests.
Direct I/O operations, which is recommended for cases in which there is a bigger amount of data to be
transferred, presents a different approach from Buffered I/O. Instead of proposing a new buffer in the
non-paged pool as is done for Buffered I/O, this technique offers directly access to the buffers, so
improving the performance because there is not the overhead in first copying data to a new-created buffer
to be consumed afterwards. Apparently it would be a problem because, as we explained previously, the
meaning of an address is only valid to a given process address space, but the mechanism is different. When
the buffer is created by the user application, the I/O manager creates an MDL, which describes this buffer.
Actually, the content of the buffer might be scattered over different physical places in the memory, and
the created MDL represents this set of places as a one-piece in the virtual memory world. In another
words, MDL works as a kind of mapping of one virtual memory to one or more physical address ranges.
Soon after the MDL has been associated with the user buffer, the I/O manager checks whether such user
buffer is accessible and locks it (making it resident) on memory (non-paged memory) by calling
MmProbeAndLockPages (defined in wdm.h), which accepts the MDL as first argument, and make sure that
the content of the virtual memory pages will be not freed and relocated any time:

[Figure 28] MmProbeAndLockPages function

The second parameter (AccessMode) tells the mode used to check for the arguments (KernelMode or
UserMode) and the third parameter indicates the type of the planned operation (purpose) that will be
occurring while accessing the virtual memory buffer through MDL such as IoWriteAddress, IoReadAddress
or even IoModifyAddress.

28 | P a g e
https://exploitreversing.com

The user memory buffer will only be unlocked whether the I/O Manager calls the MmUnlockPages
function after the driver having completed the IRP processing.
Having created the MDL, the I/O Manager fills the IRP → MdlAddress field with the pointer to the pointer
(address) of the MDL. If the device is performing a DMA operations, so it is done because device drivers
working with DMA operations require only physical addresses. However, it is not our case because we are
interested in accessing the buffer content. Thus, we have to map the provided buffer with an associated
MDL to a non-paged system address, and this address is retrieved by calling
MmGetSystemAddressForMdlSafe( ) with the MDL’ addre a first argument. This function returns a
pointer to a non-paged virtual address for the buffer represented by MDL. Therefore, we have exactly what
we need: a non-paged system address that can be accessed from any process/thread (arbitrary context)
and any IRQL because as it is locked on memory and cannot be paged out, so a system crash will never
happen even accessing it from IRQL == 2 or higher.
There is a third option named Neither I/O, which is not managed by the I/O manager, and, in this case, the
buffer management is performed (ProbeForRead and ProbleForWrite functions), and accessed from the
same context of requesting thread because the original address of the buffer is passed into the IRP, which
will be used by the driver itself. Any broken rule likely will cause a system crash. It is not easy to manage
the necessary requirements to do all these tasks without the I/O manager and, at the end of the day, the
driver itself will have to perform manually the same tasks on his own, which would be done by the I/O
manager.
In the real world, and as I explained previously, there are writing, reading and device control operations.
The first two have been covered Buffered I/O and Direct I/O operations, but while working with I/O device
control (IRP_MJ_DEVICE_CONTROL) there is the information that is provided in the control code., which is
usually defined by driver through the CTL_CODE( ), which is a macro with the following prototype:
▪ void CTL_CODE(DeviceType, Function, Method, Access);

A fast decryption of the parameters follows:

▪ The first parameter specifies that DeviceType, but as we are interested in kernel drivers, it is zero. If
readers are looking for the possible used device types here, so they can be found on
https://learn.microsoft.com/en-us/windows-hardware/drivers/kernel/specifying-device-types.

▪ The second parameter contains the IOCTL function value, which will be used and available for user
mode applications, so it must be used with IRP_MJ_DEVICE_CONTROL requests. If it used by only
kernel-mode components, so it must be used with IRP_MJ_INTERNAL_DEVICE_CONTROL requests.

▪ The third parameter contains the method code about how the buffers are passed
(METHOD_BUFFERED, METHOD_IN_DIRECT, METHOD_OUT_DIRECT and METHOD_NEITHER).

▪ The fourth and last parameter specifies the operation: FILE_ANY_ACCESS (commonly used because
works in both directions), FILE_WRITE_ACCESS (from user application to the driver) and
FILE_READ_ACCESS (from the driver to the user application).
We finished our brief review about kernel drivers, and it is time to review filter drivers.

29 | P a g e
https://exploitreversing.com

5. Filter drivers review

Explaining concepts about kernel drivers and file system filter drivers always demands dozens of pages, but
it’ a good opportunit to touch these themes even without including too many details.
File system filter drivers are not device drivers, and the general idea of file system filter drivers is to offer
supplemental functionality to typical file system operations such as opening files, creating files, reading
and writing file, and so on, while device drivers are usually associated a hardware device (except in case of
software kernel drivers as we learned previously in this article).
No doubt, there are many common things like IRPs (I/O Request Packets) for communication, callback
methods, IOCTLs and so on, which we can also use here and, eventually, adapt concepts to explain
minifilter driver functionality. Minifilter drivers are able to filter and intercept IRPs, fast I/O (synchronous
I/O operations, where data are transferred between given user buffer and the system cache without
suffering file system or storage driver interference ) and file system callback operations.
Filter drivers are used to customize / modify operations related to the file system and, in general, file
system filter drivers are used to intercept, monitor and even modify requests to the file system, besides
eventually extending and replacing a current functionality.
Thus, as expected, you will find file system drivers and mini-filter filesystem drivers in contexts where
intercepting and monitoring are the main objective as multiple security defense products such as
antivirus, EDR, backup programs, and so on, and such fact is not a surprise, and it is pretty cool.
On Windows there are two filter system filter models that are the minifilter model, which is supported by
the Filter Manager, and the legacy file system filter model. The minifilter model is a much better choice to
be followed because it allows to unload the minifilter driver (FilterUnload( ) on user-mode,
FltUnloadFilter( ) on kernel mode and even using fltmc command, as we will learn soon) and enables
communication between a user mode application and the own minifilter driver, for example. In addition, it
also permits to lock/stick on on a specific type of operation through of the usage of callbacks (definitions
will come on the next pages) and as shown below, there is the option to control the loading order through
a concept its respective altitude (another term that will be explained).
File system filter services are available through the Filter Manager (represented by the same fltmgr.sys file
mentioned above), which are enabled when the provided minifilter is loaded, and it makes the
programming task simpler (or less complex, at least) and, as also expected , minifilter is the model used for
creating file system minifilter drivers. As kernel drivers, minifilter is also stacked, but their order of loading
(actually, positioning in stack) is determined by its respective altitude. The concept of altitude seems to be
complex, but it is not, and readers can notice it by observing the following sequence:
a. Application requests an I/O operation
b. I/O Manager receives and forwards this request to the Filter Manager (fltmgr.sys).
c. The Filter Manages receives the request from I/O manager (that is key component) and checks all
its registered minifilter drivers (mfd1, mfd2, mfd3, mfd4...) according to the registered altitude.
d. After minifilter doing its actions, the request is forwarded to the File System Filter Driver.
e. Finally, the request reaches the Storage Driver Stack.

30 | P a g e
https://exploitreversing.com

There is a list of diverse ways to represent the flux of information involving mini-filter drivers, and one of
them is through the following image, as designed by Microsoft (from MSDN):

APPLICATION
I/O Manager
Minifilter Driver 1

Minifilter Driver 2
Filter Manager
Minifilter Driver 3

Minifilter Driver 4
File System Driver

Storage Driver Stack

Hardware

[Figure 29] Filter Manager and Filter Drivers

Therefore, altitude value determines the order that minifilter drivers will be called by the Filter Manager.
In addition, there could be more than one Filter Manager loaded and each one establishes a frame for
minifilter drivers. Similar to any conventional service, mini-filter drivers can be loaded (since the user have
the due SeLoadDriverPrivilege, at least) by using information on Registry (as example: Get-Item -Path
HKLM:\SYSTEM\CurrentControlSet\Services\SysmonDrv\) , which is passed to FilterLoad( )
(https://learn.microsoft.com/en-us/windows/win32/api/fltuser/nf-fltuser-filterload) that invokes
FltLoadFilter( ) (https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/fltkernel/nf-fltkernel-
fltloadfilter). At the same way, the unloading operation must be performed by calling FilterUnload( ).
A minifilter file system driver must register itself (through FltRegisterFilter function) with the Filter
Manager and specify operations that it (minifilter driver) want to intercept and process, although
minifilter drivers do not need to set up dispatch routines themselves because they are not attached
directly in the execution flow (check image above). Callbacks (pre-operation and post-operations, which we
will talk about them soon) are specified through an array of FLT_OPERATION_REGISTRATION structures,
which also specifies major functions such as IRP_MJ_CREATE, IRP_MJ_READ, IRP_MJ_WRITE,
IRP_MJ_FILE_SYSTEM_CONTROL, IRP_MJ_DIRECTORY_CONTROL and so on. This key structure will be
appropriately used as argument of the FltRegisterFilter( ).

31 | P a g e
https://exploitreversing.com

While discussing about routines related to mini-filter drivers, there are few of them that are well-known
such as:
▪ DriverEntry( ): occurs and works as for device drivers, it is used for initialization.
▪ FltRegisterFilter( ): this function is used to register a minifilter driver (and associated callback
routines) with the filter manager.
▪ FlsStartFiltering( ): it is responsible for notifying the Filter Manager that a minifilter driver is
available and ready to attach to volumes and filter requests (IRP, fast I/O and file system callback
operations). In other words, it starts the real filtering operation.
These routines present interesting details that help to explain concepts mentioned in previous paragraphs.
The prototype of FltRegisterFilter( ), which is one the main one so far, is quite simple:

[Figure 30] FltRegisterFilter function

As readers can see, there are only three parameters:
▪ Driver: it is a pointer to the driver object representing the mini-filter driver and a expected, it’ the
same driver object pointer passed to DriverEntry( ) routine.
▪ Registration: it is a pointer to a minifilter registration structure (FLT_REGISTRATION structure).
▪ RetFilter: it is a pointer to a variable that receives a filter pointer that is returned to the caller
(baicall, it’ the function’ return).
The _FLT_REGISTRATION structure has the following members:

[Figure 31] _FLT_REGISTRATION structure

This informative structure brings information related to arrays of other structures such as
FLT_CONTEXT_REGISTRATION and FLT_OPERATION_REGISTRATION, which the former one is attributed to
each context type and the latter one is attributed for each type of I/O for which the minifilter registers
preoperation and postoperation callback routines.

32 | P a g e
https://exploitreversing.com

Anyway, there is no doubt that the most important field of this structure is OperationRegistration, which is
part of the FLT_OPERATION_REGISTRATION structure that we just mentioned, but it is not the only one.
There are other relevant fields such as FilterUnloadCallback (it holds the address of a function that is called
when a driver is about to be unloaded), InstanceSetupCallback (it is a pointer to a callback that is called by
Filter Manager when a new volume is available), InstanceSetupCallback (it points to a callback that allows
the minifilter drivers to be notified just before the be attached to a volume),
InstanceQueryTeardownStartCallback (it contains a pointer to a function that will be called by the Filter
Manager before the teardown process, making possible for minifilter to cancel pending operations and
cancel or complete I/O requests) and so on.
About the teardown process, a minifilter driver instance is torn down in the following contexts: either the
minifilter is unloaded, or there is a specific detach request to be accomplished or the volume which the
instance is attached is dismounted.
It is also suitable to highlight that, during a tearing down operation of an instance, any routine executing
preoperation and postoperation callback routines continue executing without facing any problems, but I/O
requests waiting for these preoperation and postoperation callback routines may be cancelled.
Additionally, operations initiated by the minifilter drivers proceed until they are complete.
Other valuable members of FLT_REGISTRATION structure are:
▪ ContextRegistration: it represents a pointer to an array of FLT_CONTEXT_REGISTRATION
structures, being one for each context type (formatted data to be used by the driver if it’
necessary) that the minifilter could use.

▪ OperationRegistration: it represents a pointer to an array of FLT_OPERATION structures, being one

for each type of I/O for which the minifilter registers preoperation and postoperation callback
routines. As mentioned previously, this structure has members which also specify the major
function such as IRP_MJ_CREATE, IRP_MJ_READ, IRP_MJ_WRITE,
IRP_MJ_FILE_SYSTEM_CONTROL, IRP_MJ_DIRECTORY_CONTROL, and so on.
If readers are asking about the definition of callbacks, they could interpret callbacks a a ort of “modern
hooking”. Actuall, callback method allow u to regiter routines that will be triggered and executed when
specific events occur on the system. There are a series of kernel callback functions, which will be
commented on later, and callbacks related to kernel drivers and mini-filter drivers, which some of them
will be mentioned below.
There is a list of pointers to different callbacks that can be registered, and a small amount of these most-
used callback routines are:
▪ FilterUnloadCallback: it contains a pointer to a callback routine that will be called to notify the
minifilter driver that the filter manager is going to unload the minifilter driver. This callback is
defined and viewed as optional, although without it the driver cannot be unloaded, so leaking
resources.

▪ InstanceSetupCallback: it is a pointer to a callback routine that will be invoked to notify the

minifilter driver that a new volume is mounted and available. In other words, the filter manager
calls this routine to notify the minifilter driver to eventually respond to an automatic or manual
33 | P a g e
https://exploitreversing.com

attachment request to the given volume. As readers can realize, there are interesting practical
usages for it.

▪ InstanceQueryTeardownCallback: it is a pointer to a callback routine that will be called to allow the

minifilter driver to respond to a manual detaching request originated from any kernel-mode
component calling FltDetachVolume or even a user-mode application calling FilterDetach function.

▪ InstanceTeardownStartCallback: it holds a pointer to a callback routine that will be called when the
filter manager starts tearing down a minifilter driver instance to allow it to complete any pending
operation such as closing opened files and stop queueing new work items, and save the
information. From a certain point of view, this callback routine can be interpreted as the first stage
preparing for a cleaning up routine.

▪ InstanceTeardownCompleteCallback: it represents a pointer to a callback routine that will be called

when the tearing down process is complete to allow the the minifilter driver to close eventual
opened files and perform any other cleanup process.

▪ GenerateFileNameCallback: it contains a pointer to a callback routine that allows the minifilter

driver to intercept file name requests by other minifilter drivers above it on the minifilter stack (it is
quite important to remember of the driver stack concept). When this callback routine is invoked,
the minifilter driver is able to generate its own file name information based on file name
information for the file that may have been retrieved through FltGetFileNameInformation( ).
The Filter Manager does its job and makes everything easier because it handles usual IRP tasks like copying
parameters to next stack location and also provide the possibility to minifilter drivers to register only for
I/O that they are really interested (it makes sense for security products, for example, and that is the main
reason that minifilter drivers | file system drivers are interpreted as optional drivers) or need to handle
through an array of FLT_OPERATION_REGISTRATION structure:

[Figure 32] _FLT_OPERATION_REGISTRATION structure

The MajorFunction parameter specifies the type of I/O operations, which are given by FLT_PARAMETERS
union and few of them are shown below:
▪ Create: IRP_MJ_CREATE
▪ CreatePipe: IRP_MJ_CREATE_NAMED_PIPE
▪ CreateMailslot: IRP_MJ_CREATE_MAILSLOT
▪ Read: IRP_MJ_READ
▪ Write: IRP_MJ_WRITE
▪ QueryFileInformation: IRP_MJ_QUERY_INFORMATION
34 | P a g e
https://exploitreversing.com

▪ SetFileInformation: IRP_MJ_SET_INFORMATION
▪ QueryEa: IRP_MJ_QUERY_EA
▪ SetEa: IRP_MJ_SET_EA
▪ QueryVolumeInformation: IRP_MJ_QUERY_VOLUME_INFORMATION
▪ SetVolumeInformation: IRP_MJ_SET_VOLUME_INFORMATION
▪ DirectoryControl: IRP_MJ_DIRECTORY_CONTROL
▪ FileSystemControl: IRP_MJ_FILE_SYSTEM_CONTROL
▪ DeviceIoControl: IRP_MJ_DEVICE_CONTROL and IRP_MJ_INTERNAL_DEVICE_CONTROL
▪ LockControl: IRP_MJ_LOCK_CONTROL
▪ QuerySecurity: IRP_MJ_QUERY_SECURITY
▪ SetSecurity: IRP_MJ_SET_SECURITY
▪ QueryQuota: IRP_MJ_QUERY_QUOTA
▪ SetQuota: IRP_MJ_SET_QUOTA
▪ Pnp: IRP_MJ_PNP
▪ AcquireForSectionSynchronization: IRP_MJ_ACQUIRE_FOR_SECTION_SYNCHRONIZATION
▪ AcquireForModifiedPageWriter: IRP_MJ_ACQUIRE_FOR_MOD_WRITE
▪ ReleaseForModifiedPageWriter: IRP_MJ_RELEASE_FOR_MOD_WRITE
▪ QueryOpen: IRP_MJ_QUERY_OPEN
▪ FastIoCheckIfPossible: IRP_MJ_FAST_IO_CHECK_IF_POSSIBLE
▪ NetworkQueryOpen: IRP_MJ_NETWORK_QUERY_OPEN
▪ MdlRead: IRP_MJ_MDL_READ
▪ MdlReadComplete: IRP_MJ_MDL_READ_COMPLETE
▪ PrepareMdlWrite: IRP_MJ_PREPARE_MDL_WRITE
▪ MdlWriteComplete: IRP_MJ_MDL_WRITE_COMPLETE
▪ MountVolume: IRP_MJ_VOLUME_MOUNT
The second parameter is Flags, which specifies when to call preoperation and postoperation callback
routines for cached I/O or paging I/O operations, but it is not quite relevant for us right now.
PreOperation and PostOperation are pointers to PFLT_PRE_OPERATION_CALLBACK and
PFLT_POST_OPERATION_CALLBACK routine that, obviously, are registered as preoperation and post-
operation callback routines, respectively.
In few and rough words, preoperation callback routines perform the processing tasks needed for complete
the I/O operation, and controls what should be done with IRP requests and post-operation routines. Post-
operation callback routines are invoked by the Filter Manager over an I/O operation when lower drivers
have already finished completion processing.
A PFLT_PRE_OPERATION_CALLBACK routine can return different values such as:
▪ FLT_PREOP_COMPLETE: this value means that the minifilter driver is completing the I/O operation,
and the filter driver does not call postoperation callbacks of any minifilter below the caller
(remember about the driver stack) and doen’t forward (pass down) any request to minifilter
drivers below the caller.

▪ FLT_PREOP_DISALLOW_FASTIO: this value means that the operation is a fast I/O operation, and
that the minifilter driver does not allow that the fast I/O path to be used for this operation. The
35 | P a g e
https://exploitreversing.com

remaining characteristics related to postoperation callbacks and forwarding requests are similar to
FLT_PREOP_COMPLETE.

▪ FLT_PREOP_PENDING: this value means that, for a provided minifilter driver, the operation is still
pending and only after FltCompletePendedPreOperation has been invoked is that the Filter
Manager will continue the I/O operation.

▪ FLT_PREOP_SUCCESS_NO_CALLBACK: this value means that the minifilter driver is returning the
I/O operation to the Filter Manager for further processing, but the the Filter Manager will not call
the postoperation callback of the minifilter drivers over the I/O completion.

▪ FLT_PREOP_SUCCESS_WITH_CALLBACK: this value means that the minifilter driver is returning the
I/O operation to the Filter Manager for further processing, which will invoke the post-operation
callback over of the minifilter driver over the I/O completion.

▪ FLT_PREOP_SYNCHRONIZE: this value indicates that the minifilter driver is returning the I/O
operation to the Filter Manager for further processing, but it will not complete the operation. In
addition, the Filter Manager will invoke the post-operation callback of the minifilter within of the
context of the current thread at IRQL <= DISPATCH_LEVEL.

▪ FLT_PREOP_DISALLOW_FSFILTER_IO: this value means that the minifilter driver is disallowing a fast
QueryOpen operation and forcing the operation proceed through the slow path.
Readers have realized the introduction of a new term in these last paragraphs: Fast I/O. In a few words,
Fast I/O is an additional mechanism, supported by minifilter drivers, to receive requests. Actually, a file
system driver filters I/O requests coming as an IRP (I/O Request Packet) or Fast I/O requests. At the same
way of IRP requests, Fast I/O requests also have callback methods.
It is fair to say that IRP requests have a kind of equivalence to Fast I/O requests, but they are not the same,
and IRP are able to handle much more I/O’ tpe than Fat I/O. Furthermore, the DriverEntry routine can
register IRP dispatch routines and also Fast I/O callback routines, but only a set of these routines can be
registered for a given filter driver.
By the way, what is the difference in the usage between IRPs and Fast I/O? The coverage of IRP is broader,
and it can be ued for nchronou/anchronou operation, and doen’t matter whether it i a cached or
non-cached I/O. In the case of Fast I/O, it is suitable for synchronous I/O operations on cached files.
Therefore, the general requisition and practical usage of filter drivers is focused on IRP requests, although
even in this scenarios filter driver mut define a Fat I/O routine returning ‘fale’ value.
Returning to the main topic, a PFLT_POS_OPERATION_CALLBACK routine can return different values such
as:
▪ FLT_POSTOP_FINISHED_PROCESSING: this value means that the minifilter driver already has
finished the completion processing and the Filter Manager will continue the completion processing
of the I/O operation.

36 | P a g e
https://exploitreversing.com

▪ FLT_POSTOP_MORE_PROCESSING_REQUIRED: this value represents that the minifilter driver has

paused the completion, will not return the control to the Filter Manager and it will not do any post-
operation task, unless that the post-operation callback has posted the I/O operation to a work
queue or the work routine to invoke FltCompletePendedPostOperation function to return the
control of the operation to the filter manager.

▪ FLT_POSTOP_DISALLOW_FSFILTER_IO: this value means that the minifilter driver is disallowed a

fast QueryOpen operation and forces the operation down the slow path.
There is a relevant fact to mention here: post-operations are called within an arbitrary thread context with
IRQL <= DISPATCH_LEVEL. Additionally, I/O completion processing with IRQL < DISPATCH_LEVEL can not
be executed in the post-operation callback routine, and must be queued to a work-queue through the
invocation of FltDoCompletionProcessingWhenSafe or FltQueueDeferredIoWorkItem routines.
Exceptions for this rule are if the pre-operation of the mini-filter driver to return
FLT_PREOP_SYNCHRONIZE or even whether there is the certainty that the post-create callback routine will
be called at IRQL_PASSIVE_LEVEL.
The registration of pre-operation and post-operation callback routines does not need a match, so a post-
operation callback routine can be registered without a respective pre-operation callback routine. Of
course, the inverse is also true.
In general, the list of possibilities provided by minifilters is quite long, and one the capability of changing
parameters such as buffer addresses, MDLs and target file objects related to I/O operations, and even
swapping buffers. These operations can be effectively done by preoperation callbacks and can be useful in
different contexts. After changing a parameter, the FltSetcallbackDataDirty is called to notify that
parameter changes have been performed. Additionally, minifilter drivers are also able to change the I/O
status for a given operation. To complete and perform the necessary cleanup, minifilter driver’ author
must free any allocated buffer.
As we have quickly discussed about the possibility of changing parameters, so readers need to know that
there is a structure named FLT_CALLBACK_DATA, that represents an I/O operation and, of course, is used
by minifilters and the own Filter Manager over I/O operations:

[Figure 33] _FLT_CALLBACK_DATA structure

37 | P a g e
https://exploitreversing.com

The main members of this structure are:

▪ Flags: this member represents a bitmask of flags that describe I/O operations and, to minifilters,
only the FLTFL_CALLBACK_DATA_DIRTY, which indicates that the content of the callback data
structure was modified, can be specified. If this structure is initialized by the Filter Manager, so
other flags can be used such as FLTFL_CALLBACK_DATA_FAST_IO_OPERATION (the callback data
structure represents a fast I/O operation), FLTFL_CALLBACK_DATA_FS_FILTER_OPERATION (the
callback data structure represents a file system minifilter callback operation),
FLTFL_CALLBACK_DATA_IRP_OPERATION (the callback data structure represents an IRP-based
operation). Readers should search for additional flags used to initialize the callback data structure
as well as during completion processing.

▪ Iobp: this member contains a pointer to an FLT_IO_PARAMETER_BLOCK structure, which contains

the parameters for the I/O operation. .

▪ IoStatus: this member contains a pointer to an IO_STATUS_BLOCK structure, which contains status
and information for an I/O operation and as mentioned previously, its content can be changed by a
preoperation callback or even a postoperation callback.
The FLT_IO_PARAMETER_BLOCK, pointed by the Iobp parameter, has the following composition:

[Figure 34] _FLT_CALLBACK_DATA structure

Certainly readers are more familiar with most the members that make part of this structure and,
eventuall, I don’t need to explain one b one, although there is an explanation on MSDN (Microsoft
Learn): https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/fltkernel/ns-fltkernel-
_flt_io_parameter_block. Additionally, note the last member is Parameters, which is given by a giant union
FLT_PARAMETERS that is described on: https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/fltkernel/ns-fltkernel-_flt_parameters.
Minifilters are involved in a quite extensive list of activities, and it also can generate and send IRP requests,
so during reverse engineering of these types of drivers we can see routines associated with opening,
reading, writing and even creating files (FltReadFile, FltWriteFile, FltCreateFile and so on).
At the same line, there is the support offered by the Filter Manager for communication between the user
mode applications and kernel mode (minifilters) through communication ports, which it is important to
control security involved in this communication through applied security descriptors.

38 | P a g e
https://exploitreversing.com

Actually, communication ports are not buffered, so they are fast, and are used by a bidirectional
communication channel. Additionally, they are created by the minifilter drivers that keep listening for any
incoming communication and, once the user mode application tries to connect to this port, so the Filter
Manager calls the ConnectNotifyCallback routine from minifilter driver to handle the connection that is
only accepted if the user mode application has the necessary and minimum rights described by the security
descriptor. Furthermore, there are many routines offered by the Filter Manager, which are involved with
communication ports such as FltSendMessage, FltCreateCommunicationPort, FltCloseClientPort, as well
as routines available for being used by the user mode application such as
FilterConnectCommunicationPort, FilterSendMessage, FilterGetMessage, FilterSendMessage and so on.
Finally, and for completeness, it is appropriate to highlight that user mode application can interact with
minifilter drivers through an extensive series of routines for loading/unloading minifilter drivers (FltLoad,
FltUnload), enumerating filters (FilterFindFirst, FilterFindNext, …), quering information
(FilterGetInformation, FilterGetInstanceInformation,…) and o on.
Unfortunately, installing a minifilter driver is not so simple as installing a kernel driver, and it is necessary
to create an INF file, which is out of the scope of this article.
On Windows system we are able to find out a series of minifilter drivers by running the following
commands:

[Figure 35] Minifilter drivers list

Of course, readers can check the altitude of a driver by checking its respective entry in the Registry. For
example, for the SysmonDrv we have:
➢ Get-ChildItem -Path HKLM:\SYSTEM\CurrentControlSet\Services\SysmonDrv\Instances
This command can do much more than only listing minifilter drivers as, for example, loading and unloading
them (as expected, unloading a minifilter driver call the FilterUnloadCallback routine):
➢ fltmc load <filter name>
➢ fltmc unload <filter name>
On WinDbg, minifilter drivers can be listed using a debugger extension (fltkd) of the WinDbg, which offers
a series of options such as listing detail information about a given minifilter, getting a list of minifilters,
listing volumes and filter manager frames, for example. Before proceeding, and a I don’t know whether

39 | P a g e
https://exploitreversing.com

readers are used to doing it, in this environment I am using two virtual machines (on VMware): the first
one running Windows 11 (host) and the second one running Windows 11 (target). In my case, both systems
have Windows SDK installed.
On target:
▪ bcdedit /set {default} DEBUG YES
▪ bcdedit /dbgsettings net hostip:<host ip> port:50100 key:1.2.3.4
▪ bcdedit /dbgsettings
▪ shutdown /r /t 0
On host:
▪ windbg -k net:port=50100,key=1.2.3.4
▪ Make sure that symbols are configured:
o File → Symbol File Path: srv*c:\symbols*https://msdl.microsoft.com/download/symbols
o set _NT_SYMBOL_PATH=srv*c:\symbols*https://msdl.microsoft.com/download/symbols
(personally, I prefer setting it at Advanced Windows Setting → Environment Variables and
creating the _NT_SYMBOL_PATH as explained above)
▪ Debug → Break
If everything is OK, you should see the WinDbg prompt, and can execute the following:

[Figure 36] Attached minifilter drivers (truncated output)

40 | P a g e
https://exploitreversing.com

We can use !fltkd.filters extension command too (it is exactly the same). As in the article from Microsoft,
which is related to Windows Defender detection that was previously mentioned at beginning of this text,
the Windows Defender Filter (WdFilter.sys) is a desirable choice. We can also list its respective
communication ports by using the same fltkd extension. Picking up it object’ addre from the output
above (FLT_FILTER: ffff880f8ae9c4d0 "WdFilter" "328010") by executing the following command:

[Figure 37] Retrieving a minifilter communication port

As listed on Figure 37, there are only five minifilter driver’ communication ports associated to the
WdFilter minifilter. If we need to collect further details about the minifilter driver itself then execute:

41 | P a g e
https://exploitreversing.com

[Figure 38] Retrieving details about a minifilter communication

The output shows us valuable information about the minifilter drivers, including the Communication Port
List. If readers have any issue with symbols, check whether the symbols path is correctly configured and
force them loading: .reload /f command.
If we pay attention to details, we will be able to realize other terms that we have not commented yet:
▪ volume: a filesystem filter driver, following the minifilter model or the legacy file system filter
model), can also perform I/O operations on one or more file system volumes as logging, I/O
filtering, modifying or monitoring (as explained previously, and based on the definition from
Microsoft MSDN). A filter device object must be created (IoCreateDevice function) and attached to
a filter driver stack by calling IoAttachDeviceToStackSafe function.

42 | P a g e
https://exploitreversing.com

▪ context: it is a structure that can be associated to the filter manager object and used to save and
pass information (the context) about an object. This structure is defined by the minifilter driver
itself, and there can be contexts associated to volumes, files, instances, transactions, stream
handles (file objects) and streams. Readers could be interested in knowing that functions such as
FltAllocateContext (to create contexts), FltRegisterFilter (registering contexts), FltSetFileContext |
FltSetInstanceContext | FltSetStreamContext | FltSetVolumeContext | FltSetTransactionContext
(setting contexts) and other ones aociated to context’ manipulation. Additionall, there i an
interesting example (code) demonstrating how to do it that is available on:
https://github.com/Microsoft/Windows-driver-samples/tree/main/filesys/miniFilter/ctx.
To get a list of volumes and their respective attached filter drivers (pay attention to WdFilter driver), you
can execute the following command:

[Figure 39] Getting a volume list

To examine information about a specific volume (FLT_VOLUME structure), execute: !fltkd.volume
ffff880f8a92f010 (it is the second volume listed previously)
43 | P a g e
https://exploitreversing.com

[Figure 40] Retrieving volume information

To list specific information about a given instance (an attachment to FLT_VOLUME structure), execute:

[Figure 41] Retrieving instance information

44 | P a g e
https://exploitreversing.com

Of course, we can get inside of structures and find out much more information. For example, we can get
information from the WdFilter driver by overlaying its address with the _FLT_FILTER structure:

[Figure 42] Getting further WdFilter details

All functions, concepts and terms we mentioned previously are present here: altitude, FilterUnload
function (called when the minifilter driver is unloaded), InstanceQueryTearDown, contexts, a pointer to
an array of FLT_OPERATION_REGISTRATION structures (contains the operation callbacks), and so on. In
the other side, the DriverObject concept we already now and, actually, we will be reviewing a typical
output using it soon.
Although I have not explained previously, each filter manager frame works like a placeholder in the I/O
driver stack, and minifilters attach to this frame. For example, there could exist two Filter Frames in the I/O
driver stack with a legacy filter driver in the middle. In this case, we could choose whether the minifilter
driver would be attached in the Filter Frame before the legacy filter driver or after the legacy filter driver.
Anyway, we can list preoperations and postoperations (routine addresses and, when it is possible,
respective names) associated to the driver. For example, we can list the first ten operations by executing
the following command:

45 | P a g e
https://exploitreversing.com

[Figure 43] Listing minifilter pre/post operations

The output’ image is small, and, in this specific case, we haven’t gotten respective names. If you try the
same command, but wit out “-c” option, you will receive a line-by-line output (longer, but better). A
similar output, but from WoF (Windows Overlay Filter) driver, is shown below to provide a case where the
routine’ name are hown (sorry for the small size):

[Figure 44] Listing minifilter pre/post operations of another driver as comparison

Returning to the WdFilter minifilter driver, we can retrieve callback information related to a given
instance:

[Figure 45] FLT_INSTANCE structure: associated callbacks

All callback nodes have an associated name such as ACQUIRE_FOR_SECTION_SYNC, CREATE, READ,
WRITE,SET_INFORMATION, QUERY_EA, SET_EA, DIRECTORY_CONTROL, FILE_SYSTEM_CONTROL and
CLEANUP.
There are multiple MUP (Multiple UNC Provider), which a MUP is a kernel component responsible for
channeling remote file system access through UNC to a network redirector, and it is associated with each
callback node (check the figure above).

46 | P a g e
https://exploitreversing.com

At the same way we did with _FLT_FILTER structure, we can pick up one of the callback nodes and getting
information by overlaying it with _CALLBACK_NODE structure as shown below:

[Figure 46] _CALLBACK_NODE structure: retrieving information to one given instance

There are multiple details to comment about the output:
▪ We have a doubly linked list of CALLBACK_NODE structures.
▪ We see a reference to PreOperation and PostOperation callbacks.
▪ All references to name are “blank”, but we alread learned that thi doen’t happen with other
minifilter drivers such WoF (Windows Overlay Filter).
As a minifilter needs to pass contexts to save and pass information about an object, so it required a
mechanism like minifilter contexts (CONTEXT_NODE) and, as expected, there is a context associated to an
instance too:

[Figure 47] _CONTEXT_NODE structure: retrieving information to one given instance

Checking the fourth line of the output, we see the reference to NonPagedPool. Except volumes contexts,
which must be allocated from NonPagedPool, all remaining contexts (instances, streams, files, transaction
and stream handles) can be allocated from PagedPool or NonPagedPool.
Anyway, if readers want, it is possible to investigate the _CONTEXT_NODE structure by using the same
technique used until now and picking up one of the context nodes, as shown on the next page:

47 | P a g e
https://exploitreversing.com

[Figure 48] _CONTEXT_NODE structure overlay wit structure’s address from last output
An organized output containing exactly the same information is given by:

[Figure 49] Context information associated to the instance

Returning to communication ports subject, it is time to examine one of those ports:

[Figure 50] _FLT_PORT_OBJECT structure

As we learned previously, a communication port (created by FltCreateCommunicationPort function) is
important to keep the communication between the minifilter driver and application and, as expected,
there is a series of functions involved with communication tasks, and few of these functions are
FilterConnectCommunicationPort, FltSendMessage, FilterSendMessage, FilterReplyMessage and so on.
Additionally, drivers uses mechanisms to exchange messages (its header is represented by
FILTER_MESSAGE_HEADER structure), to signaling that is waiting for messages (message queue,
represented by _FLT_MESSAGE_WAITER_QUEUE structure), a callback to be notified when a message is
available (MessageNotifyCallback routine, which is called at IRQL=PASSIVE_LEVEL by Filter Manager) and a
PortCookie that is used to uniquely identify the client port or server port, depending on the side of the
communication.

48 | P a g e
https://exploitreversing.com

Just in case readers have curiosity about the stuff, there is a PowerShell module named NtObjectManager,
written by James Forshaw (https://www.powershellgallery.com/packages/NtObjectManager/1.1.33) that
provides the communication ports easily for you:
PS C:\> Install-Module -Name NtObjectManager
PS C:\> Set-ExecutionPolicy RemoteSigned
PS C:\> Import-Module NtObjectManager
PS C:\> NtObject:\ | Where-Object TypeName -eq "FilterConnectionPort"
PS C:\> ls NtObject:\ | Where-Object TypeName -eq "FilterConnectionPort"

Name TypeName
------------------------------------------------------------------ ----------------------------
UnionfsPort FilterConnectionPort
storqosfltport FilterConnectionPort
MicrosoftMalwareProtectionRemoteIoPortWD FilterConnectionPort
MicrosoftMalwareProtectionVeryLowIoPortWD FilterConnectionPort
WcifsPort FilterConnectionPort
WinSetupMonPort FilterConnectionPort
MicrosoftMalwareProtectionControlPortWD FilterConnectionPort
BindFltPort FilterConnectionPort
MicrosoftMalwareProtectionAsyncPortWD FilterConnectionPort
CLDMSGPORT FilterConnectionPort
MicrosoftMalwareProtectionPortWD FilterConnectionPort
[Figure 51] List of registered communication ports
Returning to _FLT_PORT_OBJECT structure, the MegQ member is, as we already explained, a pointer to
the _FLT_MESSAGE_WAITER_QUEUE structure, which can be applied to the address and, executing the
following sequence of commands, we have:

[Figure 52] Examining a sequence of fields since _FLT_MESSAGE_WAITER_QUEUE

As we can realize, from a given message queue structure we reached an _ETHREAD and
_IO_STACK_LOCATION structures.

49 | P a g e
https://exploitreversing.com

Investigating the fourth command, we have:

▪ dx Debugger.Utility.Collections.FromListEntry(*(nt!_LIST_ENTRY *)0xffff880f905cd1c8, "nt!_IRP",
"Tail.Overlay.ListEntry")
Readers could certainly ask from where components of this command come. This WinDbg command is
using LINQ (Language-Integrated Query), which is well-known from C# programming, and the syntax of
this command comes from WinDbg documentation on MSDN. In few words, this command parses the
nt!_LIST_ENTRY structure, and its composition is simple:
▪ 0xffff880f905cd1c8: Flink pointer
▪ nt!_IRP: structure being referenced.
▪ Tail.Overlay.ListEntry: field from _IRP structure being referenced by Flink pointer.
The remaining point is: how do I know that this list points to the nt!IRP structure and, in special, to
Tail.Overlay.ListEntry field? Open the fltmgr.sys file on the IDA Pro, and even not doing any treatment on
the code, you can easily observe that FltpAddMessageWaiter( ) receiving three arguments: a pointer to
_IO_Csq structure, a pointer to IRP structure and the third argument associated with context:

[Figure 53] FltpAddMessageWaiter function

On line 6 we have our reference to p_ListEntry = &Irp->Tail.Overlay.ListEntry and, on lines 14 and 15,
readers are able to check the doubly linked list set up. Anyway, once readers reach the _ETHREAD
structure, it is possible to retrieve the value of any field.
There are deeper details about these concepts such as filter contexts, communication ports, message
queues and so on, but it is enough for now and, hopefully, readers are forming a big picture about
minifilter drivers.
Of course, there are more details, and it is time to move on.
50 | P a g e
https://exploitreversing.com

As a summary, while examining minifilter drivers, readers will find key routines such as:
▪ DriverEntry: it is the same routine as kernel drivers and, at the same way, it is requested for all
filter drivers. Additionally, this routine serves as a starting point for key actions, and, for example, it
is where the minifilter driver can register (through FltRegisterFilter routine) one preoperation
callback and one postoperation callback (it is not necessary to be present both ones) for each of of
different I/O types been manipulated and filtered by the minifilter.
▪ FltRegisterFilter: this routine is used by minifilter drivers to register to provide a list of callback
routines to the Filter Manager and, at the same time, to register themselves to the minifilter
driver’ lit.
▪ FltStartFiltering: this routine notifies the Filter Manager that it is ready and can start to filter
requests by attaching to volumes.
▪ FltCreateCommunicationPort: this routine opens a kernel communication server port.
▪ FltCloseCommunicationPort: this routine closes a kernel communication server port.
▪ FilterUnloadCallback: it is the routine responsible for unloading the minifilter driver. It is an
optional routine.
▪ FltUnregisterFilter: this routine unregisters the minifilter driver.
It is really important to understand the concept of preoperation callback because each minifilter driver can
have its own, and every associated preoperation callback to each registered minifilter will be called from
the minifilter driver that holds the higher altitude up to the lowest one for that specific type I/O operation.
Additionally, the Register parameter from FltRegister routine is relevant because it holds a pointer to the
FLT_REGISTRATION structure. This structure holds a field/member that is actually an array of
FLT_OPERATION_REGISTRATION structures, which each one represents a type of operation being
manipulated and filtered by the minifilter driver. Certainly, it might seem confusing because there are
three levels of redirection here, but it is not so uncommon with kernel and minifilter drivers. However, it is
not the end yet and, as there are two file system filter driver models, minifilter drivers receive the I/O
operation first, and later the legacy file system filter drivers receive it for processing. Afterwards, the
associated file system receives the I/O operation for further processing. In the order side, postoperation
routines (each minifilter drivers that has registered to process that type of I/O operation can have or not a
postoperation callback) start their work in the reverse order, finish the processing of the I/O operation,
return it to the filter managers, which passes it to the next minifilter driver at the upper layer. At this point,
it is not hard to realize that a file system minifilter likely will be using many preoperation callback routines
to manipulate and filter I/O operations, and these preoperation callbacks can return values to the Filter
Manager like FLT_PREOP_SYNCHRONIZE (for IRP based operations, which can have its type confirmed by
FLT_IS_IRP_OPERATION macro, and a postoperation routine will be invoked during the I/O completion
phase), FLT_PROP_SUCCESS_NO_CALLBACK (no postoperation callback routines will be called during the
I/O completion phase) and FLT_PREOP_SUCCESS_WITH_CALLBACK (postoperation callback routines will
be invoked during the I/O completion phase), for example, as already mentioned previously in this article.
Of course, at the same way, a minifilter driver could have more than one postoperation callback routines
that can be executed at IRQL lower or equal to DISPATCH_LEVEL and, due to this fact, data structures
must be allocated in nonpaged pool. Anyway, postoperation routines are called in arbitrary context.
Minifilter drivers also transfer information (data) between applications running in user mode and other
minifilter drivers running in lower layers, which can reach device drivers and, because these data
transferring operations, they are also use some kind of buffer.

51 | P a g e
https://exploitreversing.com

There is not any news related to data buffers, and file system minifilter drivers uses the same methods
from kernel drivers to access buffers that is Buffered I/O (mainly used over IRP operations such as
IRP_MJ_CREATE and IRP_MJ_QUERY_INFORMATION, for example), Direct I/O and Neither I/O (it can used
by operations such as IRP_MJ_SYSTEM_CONTROL and IRP_MJ_QUERY_SECURITY). Additionally, important
and usual operations such as IRP_MJ_READ, IRP_MJ_WRITE, IRP_MJ_DEVICE_CONTROL and
IRP_MJ_QUERY_OPERATION (mentioned above) can be configured as Fast I/O or IRP based operations.
As readers have realized, same I/O IRP operations major codes are valid for minifilter drivers, and you can
check them by using a well-know WinDbg command:

[Figure 54] Listing IRP routines associated to the minifilter driver

52 | P a g e
https://exploitreversing.com

The Windows Cloud Files filter driver (cldflt.sys) is a file system minifilter driver that is associated to the
OneDrive, for example. The GsDriverEntry( ) is a routine generated automatically when the driver is built,
which does a short initialization and, soon after having completed the initialization, it calls the real
DriverEntry( ) that was implemented.
Moving forward, I would like to comment about ECP (Extra Create Parameters) that are structures holding
information used during file creation, and that can be attached to I/O operations by using an ECP_LIST
structure. For example, a file system filter driver can manipulate ECPs (Extra Create Parameters) to
process IRP_MJ_CREATE operations, and are exactly these ECPs that are used to distinguish between
NtCreateUserProcess( ) and NtCreateProcessEx( ) calls, which were also mentioned in the Microoft’
article at beginning of this text. ECPs can be one of two available types: System-defined ECPs that are used
by the OS to attach further information to IRP_MJ_CREATE mentioned previously, and User-Defined ECPs
that are used by kernel drivers to process and add further information to the IRP_MJ_CREATE operation.
Readers likely will recognize ECPs manipulation when find routines such as
FltAllocateExtraCreateParameterList (to allocate memory to ECP_LIST structure),
FltFreeExtraCreateParameterList (to free memory used by ECP_LIST structure),
FltAllocateExtraCreateParameter (to allocate paged-memory pool for an ECP context structure, returning
a pointer to it), FltInsertExtraCreateParameter (to insert ECP context structures into the ECP_LIST
structure), IoInitializeDriverCreateContext (to initiate an IO_DRIVER_CREATE_CONTEXT_STRUCTURE) and
finally IoCreateFileEx|FltCreateFileEx2 (to attach ECPs to a given IRP_MJ_CREATE_CONTEXT).
Of course, there is an extensive list of routines to process and manipulate ECPs such as
FltGetEcpListFromCallbackData (returns a pointer to an ECP list associated with a create operation
callback-data object), FltFindExtraCreateParameter (searches a provided ECP lit for an ECP’ context
structure) and FltIsEcpFromUserMode (checks whether the ECP is originated from the user mode). A quick
sample of usage of these routines is shown below:

[Figure 55] Routines related to ECP

53 | P a g e
https://exploitreversing.com

Returning once again to the Microsoft article, the GUID_ECP_CREATE_USER_PROCESS and respective
CREATE_USER_PROCESS_ECP_CONTEXT context, which contains the token of the process to be created,
are used by kernel while it opens the process executable file. Therefore, while the NtCreateUserProcess
adds the ECP for a process creation, the NtCreateProcessEx does not do it because it uses a section handle
already created (existing). This makes it simpler to distinguish when one or the other function is used.
Certainly, ECP is not the only interesting topic because there is a new mechanism named BypassIO that has
been introduced in Windows 11, that is requested for a file handle, and it turns the I/O access for reading
files better and quicker due to a lower overhead, and this is leveraged by minifilter drivers. The big
advantage of using BypassIO is that the I/O request does not pass through the entire driver stack, but goes
directly to NTFS file system (bypassing volume and filesystem stack, and the latter can be composed by
Volume Device Object (VDO) or Control Device Object (CDO) in addition to usual minifilter device objects)
and, from there, to the underlying volumes and disks. Furthermore, calls to functions such as
FltFsControlFile routine (or native equivalents) with FSCTL_MANAGE_BYPASS_IO control code are usual
while requesting and emitting BypassIO operations.
Readers will see FSCTL_MANAGE_BYPASS_IO and IOCTL_STORAGE_MANAGE_BYPASS_IO control codes
involved with minifilter drivers using BypassIO, which demands NTFS filesystem on NVMe storage device
on Windows 11 for while. You should also pay attention to requests such as FS_BPIO_OP_ENABLE,
FS_BPIO_OP_DISABLE, FS_BPIO_OP_QUERY, FS_BPIO_OP_GET_INFO and other similar ones, mainly
because they are involved with preoperation callbacks.
We can easily check the support for BypassIO feature by executing the following command:

[Figure 56] BypassIO: checking filesystem support

Returning to CDO (Control Device Object) and VDO (Volume Device Object) mentioned above, which are
optionally created by file system minifilter drivers (file systems must create a CDO, but it is optional to file
system minifilter driver, although it commonly used), it is suitable to highlight that CDO works like a
representation of minifilter driver to the user mode application, and besides of the system, of course.
Later, the FDO (filter driver object) will perform all related tasks of filtering on a given filesystem or
volume. This scheme and composition are independent of the driver handling IRP or Fast I/O. As explained
previously, IRPs are used in general operations (synchronous or asynchronous), while Fast I/O are used
over synchronous operations, offering advantage to make the accelerating the transfer between
application/user buffer and the system cache, so bypassing eventual filesystem and volume stack in the
middle of the way. Additionally, we should also remember that minifilter filesystem must implement Fast
I/O routines even if they do not support them (and, as recommended, returning FALSE).
So far we have explained WDM (Windows Driver Model), including a series of concepts associated with
kernel drivers and minifilter drivers because all these concepts are foundations of drivers in the current
days. However, many years ago Microsoft introduced another framework to develop drivers named
Windows Driver Frameworks (WDF), which offers a kind of abstraction that simplify the driver
development and, of course, soon or later readers will reverse and analyze a sample in their daily tasks.
54 | P a g e
https://exploitreversing.com

6. Windows Driver Frameworks (WDF) review

The first facts about WDF are that:

▪ They include two important frameworks: KMDF (Kernel-Mode Driver Framework) and UMDF
(User-Mode Driver Framework).
▪ Microsoft offers its respective source code available on: https://github.com/Microsoft/Windows-
Driver-Frameworks
▪ Microsoft Visual Studio, as expected, offers a series of templates to develop KMDF and UMDF
drivers.
These frameworks (KMDF and UMDF) offer an abstraction from WDM (readers could agree that it is really
complex) and handles important functionalities such as Plug-and-Play and Power Management, and
everything is done to offer a friendly interface to developers. We have not seen any of these details in our
previous discussions because our focus is on software driver, without interacting directly with hardware.
Anyway, although the model is different, the purpose is the same, that is to manage the communication
between user applications and devices, or other drivers. I will target KMDF in this article, but UMDF drivers
must be highlighted because they offer incredibly attractive features as handling only the memory
associated with the process, having a simpler interaction with the environment, limited access to system
files and even data from users, and a series of other advantages that, eventually, might attend
requirements of a project.
In general, WDF (Windows Driver Frameworks) is composed by a central DriverEntry routine, which is
responsible for calling the WfdDriverCreate routine (this routine creates the driver object that represents
the driver), and a series of event callback functions that finally calls object methods exported by the own
framework. In other words, the programming is oriented to events, so objects support one or more of
these possible events, which are enabled according to system’ changes or even due to new I/O requests.
The best part is that the driver framework offers default routines for all possible events. The driver is not
obliged to manage any of them and, if the driver wants to override any one of default routines to handle
the respective event, so the driver needs to register a new callback (invoked when the event happen) and
notify the driver that such event happened, which provides to driver with an opportunity to perform
further processing and tasks. If readers have any issue understanding that callback concept here, think
about it as a message to signal that something relevant happened (an event), and which the driver might
have interest in handling. The WDF model follows the proposed driver stack:
▪ application → kernel → filter device object (filter driver) → function device object (function
driver) → filter device object (filter driver) → physical device object (bus driver)
As most general concepts are similar, we have to adapt our knowledge to new function names and,
eventually, concepts. As we learned previously, drivers can implement callback methods according to
expected events, and afterwards they register these callbacks to the framework. The name convention for
callback functions is EvtObjectEvent, where the Object part represents the referred framework object and
Event part represents the provided event. The KMDF also follows a well-formed syntax to its methods,
that’ Wdf[Object][Operation], where Object refers to an object involved in the operation, and Operation
refer to the method’ goal.

55 | P a g e
https://exploitreversing.com

As I had mentioned, the own framework already offers callback implementation for events, so driver needs
to implement a callback whether it needs to perform a different processing. At end of the day, readers will
realize that KMDF drivers work similarly to minifilter drivers without imposing meaningful restrictions.
One of nomenclature aspects that readers have already realized is that most (not all) objects and routines
are prefixed with “Wdf” tring (upper cae, lower cae or mixed notation). Furthermore, you will see
names of objects like WDFDEVICE (device), WDFDPC (dpc), WDFFILEOBJECT (file), WDFINTERRUPT
(interrupt), WDFSPINLOCK (spin lock), WDFQUEUE (queue) as well as routines as WdfDriverCreate,
WdfDeviceCreate, WdmDeviceCreateSymbolicLink, WdfObjectReference,
WdfDeviceCreateDeviceInterface, WdfRequestRetrieveInputBuffer, WdfRequestRetrieveOutputBuffer,
WdfRequestRetrieveInputWdmMdl, WdfRequestRetrieveOutputWdmMdl, WdfAllocateContext
(allocated in nonpaged pool and taken as part of the object, which has an equivalent meaning of WDM
device extension), WdfIoQueueCreate and so on. Such objects have properties like ParentObject, Size,
ContextTypeInfo, and so on, that are stored into WDF_OBJECT_ATTRIBUTES structure and initialized by
WDF_OBJECT_ATTRIBUTES_INIT function. By the way, there are configuration structures associated to
objects, which hold information like pointers to the event callbacks, and nomenclature of such structures is
WDF_<object>_CONFIG, and that are usually initialized by functions/macro that also follow
WDF_<object>_CONFIG_INIT as nomenclature. Therefore, while creating a KMDF driver, readers will
follow the usual order in declaring and initializing configuration structures then initializing attributes and
finally creating an object.
Similarly, we had seen for WDM, the WDF model is composed by I/O requests, queues, memory regions
and devices, of course. Through this mechanism, when the operating system sends an I/O request to a
WDF driver, the framework is responsible for handling the dispatch operation, queueing and completion of
the request. Furthermore, as most applications will interact with drivers for reading, writing or even
controlling devices, so routines like WdfIoQueueCreate routine will be used to create a queue object that
represent the respective I/O queue (as usual, everything is about managing I/O requests and memory).
Here is appropriate to highlight that the general WDF hierarch is given by a driver object → device object
→ queue object → request object. WDF drivers also handles interrupts by calling routines like
WdfInterruptCreate routine and, as you could imagine, it will create interrupt objects to each given
interrupted and register callback functions, which I do not need to repeat the same explanation. By the
way, callbacks are usually suffixed with Evt string, so there are EvtCleanupCallback, EvtDestroyCallback,
EvtDeviceAdd, EvtIoRead, EvtIoWrite, and so on.
Certainly, KMDF is an extensive topic and has its peculiarities, but it is close to the WDM development, so
these couple of pages are enough to review basics on the KMDF.

7. Supplemental information about callbacks

Returning to callback subject, Windows offers a series of kernel callback APIs that exported by kernel
(NtosKrnl.exe + wdm.h) and which drivers can use to register their callback routines that, eventually, will
be called for specific kernel component’ event and condition.

56 | P a g e
https://exploitreversing.com

As we are discussing kernel drivers and filter drivers, leaving a few words about this topic could be useful. If
readers are writing a kernel driver, they could use a callback object from other drivers and register a
routine (InitializeObjectAttributes( ) + ExCreateCallback( ) + ExRegisterCallback( )) to be invoked when the
specific callback is triggered (a given condition happened).
The offered kernel callback functions are used mainly by security defenses to register their own callback
routines to be able to monitor the system system according to specific events and conditions, so as
expected, kernel callback functions are available to attend different purposes and goals.
The list of kernel callbacks (sometimes called as system callbacks) is really considerable, and I only will
present the definition and concepts about few of them here:
▪ CmRegisterCallbackEx( ): this function registers a RegistryCallback routine, which is a routine used
by filter driver to monitor and modif an Regitr operation uch a ke deleting, renaming, ke’
value changing, enumeration, creation and so on. For example, malware can use this callback to
restore malicious content (for example, a malicious entry used for persistence) soon after a system
administrator has removed an entry related to persistence. As we reviewed previously, the Altitude
parameter (second parameter shown below) defines the position of the minifilter driver when
compared to other minifilters in the I/O stack. Finally, we should pay attention to the fact that the
first parameter (Function) is a pointer to the RegistryCallback routine to be registered and the third
parameter (Driver) is a pointer to a traditional DRIVER_OBJECT structure, which represents the
driver itself.

[Figure 57] CmRegisterCallbackEx( )

▪ FsRtlRegisterFileSystemFilterCallbacks( ): File system drivers call this function to register
notification callback routines that will be invoked when the file system performs specific
operations. Its second parameter points to a FS_FILTER_CALLBACKS structure, which holds the
entry pointer of caller-supplied notification callback routines. At end of the execution, the usual
return value is STATUS_SUCCESS or STATUS_FSFILTER_OP_COMPLETED_SUCCESSFULLY, but this
last one means it has completed an FsFilter operation.

▪ IoRegisterBootDriverCallback( ): this function registers a BOOT_DRIVER_CALLBACK_FUNCTION

routine that will be invoked during the initialization phase of the boot-start drivers, and whose role
is to monitor boot-start events and return data to the kernel. For example, the ELAM (Early Launch
Anti-Malware) driver, which is a mechanism that can be used by defenses like antivirus programs,
is able to register callback methods using this function to verify issues due to lack of integrity of
other boot drivers or even Registry entries, that also could be monitored by using
CmRegisterCallbackEx routine as mentioned previously. Even out our focus, you can examine the
57 | P a g e
https://exploitreversing.com

WdBoot.sys (ELAM driver) using IDA Pro + WinDbg (in a remote setup configuration) if you want to
do. As a short example to help you to start:
▪ Open the WdBoot.sys driver (from C:\Windows\system32\drivers folder) from a remote
Windows system (we will debug it later) into IDA Pro.
▪ Search for DriverEntry routine (it is called by GsDriverEntry routine)
▪ Write down the DriverEntry’s address.
▪ Examine the WdBoot.sys driver on PEBear. Write down the Image Base.
▪ Through a remote WinDbg session (I explained steps previously), set up a breakpoint on the
remote (target) to stop execution when the driver gets loaded by executing sxe ld WdBoot.sys
and reboot the system. If you want to see all messages from debugger, execute ed
nt!Kd_DEFAULT_MASK 0xFFFFFFFF
▪ Once the system rebooted and stopped on WdBoot.sys loading, setup the breakpoint on
WdBoot!DriverEntry (remember that we don’t have mbol) b executing bp WdBoot +
0x1C000B000 – 0x1C0000000 (effectively is WdBoot + 0xB000).
▪ Type g to resume the system.

[Figure 58] Examining WdBoot’s Driver Entry on IDA Pro

[Figure 59] Examining WdBoot’s Driver using PE Bear 58 | P a g e

https://exploitreversing.com

[Figure 60] Setting

up a breakpoint at
WdBoot!DriverEntry

59 | P a g e
https://exploitreversing.com

From this point it is possible to perform all the usual investigations using WinDbg. Anyway, the part of the
driver using IoRegisterBootDriverCallback (and respective IoUnRegisterBootDriverCallback) routines
follows:

[Figure 61] Reversing a piece of WdBoot.sys

As MmGetSystemRoutineAddress routine is responsible for returning a pointer to the given function
specified by SystemRoutine parameter, which hold the pointer to “IoRegisterBootDriverCallback” string,
so the address of the callback is effectively resolved.
It seems that, after callbacks being resolved, Windows Defender will load its signatures according to line
133 above. Going a bit further, we will recognize another routine related to a callback that we already
mentioned previously (CmRegisterCallback) and even an API (ExFreePoolWithTag) responsible for freeing
memory pool region associated to provided tag (EBsg, in this case). Finally, we see the

60 | P a g e
https://exploitreversing.com

IoRegisterBootDriverCallback (remember that its pointer has been stored into SystemRoutineAddress
variable) being used to register a callback named MbEbBootDriverCallback, as shown on line 221:

[Figure 62] Reversing a piece of WdBoot.sys (part 2)

A BOOT_DRIVER_CALLBACK_FUNCTION routine is responsible for monitoring the startup of the a given
driver, and it matches the first parameter of IoRegisterBootDriverCallback routine as shown below:

[Figure 63] IoRegisterBootDriverCallback routine

That is enough about IoRegisterBootDriverCallback routine, and it is time to return and comment about
other system callbacks.
▪ IoRegisterFsRegistrationChangeEx( ): this routine registers a notification routine (callback routine)
of a file system filter, which is called when a file system registers or unregisters itself. Most EDRs
monitor this routine actively. The first parameter is a pointer to a driver object for the file system
filter driver, and the second parameter is a pointer to PDRIVER_FS_NOTIFICATION routine, which is
61 | P a g e
https://exploitreversing.com

called by the file system always that it registers or even unregister itself by calling functions such as
IoRegisterFileSystem( ) and IoUnregisterFileSystem( ) respectively.

▪ IoRegisterFsRegistrationChangeMountAware( ): this function aims to registers notification

routines (callback methods) of a file system filter drivers and, as expected, the second argument
points to a PSDRIVER_FS_NOTIFICATION routine, which is invoked as a file system gets mounted
(active) or unmounted (inactive). The first parameter is a pointer to a driver object for the file
system drivers, as usual.

▪ ExAllocateTimer( ): this function is responsible for allocating and initializing a timer object by using
an ExTimerCallback callback routine, which Windows calls when the time interval of a timer
(represented by EX_TIMER timer object) expires.

[Figure 64] ExAllocateTimer( )

Multiple rootkits have used this callback to create a timer object within a non-arbitrary threat
context to schedule operations that will be executed in a periodic way. For example, professionals
who are hunting timers might use WinDbg !timer extension to list all pending timers on system:

[Figure 65] WinDbg timer extension

62 | P a g e
https://exploitreversing.com

As a simple example about the usage of ExAllocateTimer routine, we could check any filter driver
as WoF.sys (Windows Overlay Filter) that initializes a timer object associated with a callback
named TlgAggregateInternalFlushTimerCallbackKernelMode. The reversing job of the routine
shown below can be improved a lot, but it is enough for now because we only want to highlight the
usage of one routine:

[Figure 66] ExAllocateTimer example

63 | P a g e
https://exploitreversing.com

▪ IoSetCompletionRoutineEx( ): Although we already have commented about this routine at a first

moment on page 25, it is valid to review that this routine registers an IoCompletion routine, which
is usually called when the next level driver (lower driver) has completed the requested operation
related to a provided IRP. The completion routine, which executes from an arbitrary thread or even
DPC ( Deferred Procedure Calls ) context, is responsible for determining whether any additional
processing is required for a given IRP. As an additional information, a DPC routine (DpcForIsr( )),
which is associated with a DPC object, is queued by the ISR (Interrupt Service Routine – its
execution must be short and fast) and executed at a later moment with a lower IRQL
(IRQL_DISPATCH_LEVEL) than the ISR’ high level and, in few words, it is responsible for performing
the heavy-work that has not been done by ISR. Any remaining work that has not been completed by
DpcForIsr( ) routine can be done by CustomDpc( ) routines, which are extra DPCs. The
DEVICE_OBJECT structure holds a KDPC structure member (Dpc field), as shown below, that is used
to request the mentioned DPC routine while within of ISR. Therefore, once we get any pending DPC
(its possible to list them by using !dpcs extension), we can get its respective address and perform an
overlay against the _KDPC structure to obtain a better comprehension on further details:

[Figure 67] WinDbg: examining DPC (part 1)

Before proceeding, just a note: eventually your test tem doen’t have anything pending at the exact
time you are performing this test because it depends on the current activity.
To get further information about a provided KDPC, execute:

64 | P a g e
https://exploitreversing.com

[Figure 68] WinDbg: examining DPC (part 2)

[Figure 69] WinDbg: examining DPC (part 3)

Note: the KPCR address (0x0xfffff806205434c0) came from !pcr extension’s output (not shown)

▪ KeInitializeDpc( ): this routine is supplemental to the topic explained above because its role is to
initialize a DPC object and register a CustomDpc routine for such object. As expected, the second
argument is a pointer to the KDEFERRED_ROUTINE callback function that is executed after the ISR
(Interrupt Service Routine). Additionally, the CustomTimerDpc routine executes after the time
interval of a given timer object expires and, of course, readers could do an association to the
timer’ tuff mentioned previoul in this article.

[Figure 70] ntoskrnl.exe: KeInitializeDPC (part 3)

65 | P a g e
https://exploitreversing.com

▪ KeInitializeApc( ): this routine is used to initialize an APC (Asynchronous Procedure Calls) object. As
readers could already know, APC is a kind of kernel mechanism that is used to queue a task that will
be performed in a context of a given thread. Additionally, APCs have been used to inject code into a
user process (in alertable state) from a kernel driver, for example. There are distinct types of APC
(UserAPC, Special User APC and Kernel APC), which the first two cases are associated with APIs
such as QueueUserAPC( ) and NtQueueApcThreadEx2( ) respectively. Kernel APC is a bit different,
runs in kernel mode at IRQL = PASSIVE_LEVEL (Special Kernel APC run at IRQL = APC_LEVEL), it is
able to prompt any user mode code running at IRQL = PASSIVE_LEVEL and one of its main
structures is the _KAPC (actually, this structure makes part of a doubly-linked structure within the
_KAPC_STATE structure, which makes part of the KTHREAD structure in the kernel) that must be
allocated from a NonPagedPool memory. At end, Kernel APC works as an interruption because it
can happen at almost any time.

▪ PsSetLoadImageNotifyRoutine( ): that is a well-known routine on Windows, and it registers a

callback routine (provided by NotifyRoutine parameter as a pointer, and typed as
PLOAD_IMAGE_NOTIFY_ROUTINE) that will be notified whenever an image is loaded. Actually, this
routine is supplemented by other similar routines such as PsSetCreateProcessNotifyRoutine (it
works at an equivalent way, but adding a callback routine that will be invoked whenever a
processes to be called or terminated) and PsSetCreateThreadNotifyRoutine (same modus
operandi, but related to thread creation and termination). About registering a callback to be
notified about process creation and termination, it is interesting to remember about
PsSetCreateProcessNotifyRoutineEx and PsSetCreateProcessNotifyRoutineEx2 too. As a simple
example, Windows drivers like mssecflt.sys (Microsoft Security Events Component file system
filter driver), which has suffered multiple fixes in last months, uses
PsSetCreateProcessNotifyRoutineEx, PsSetLoadImageNotifyRoutine,
PsSetCreateThreadNotifyRoutine actively:

[Figure 71] mssecflt.sys filter driver using callbacks

▪ KeRegisterBugCheckCallback( ): this routine is responsible for registering BugCheckCallback

routine (KBUGCHECK_CALLBACK_ROUTINE), which is executed when Windows issues a bug check.

66 | P a g e
https://exploitreversing.com

Many years ago, I could find malware threats using this callback to prevent digital forensic tools to
dump the memory image, so also preventing researchers of analyzing memory.
▪ ObRegisterCallbacks( ): this routine is one of most interesting ones because it registers a list (given
by OB_CALLBACK_REGISTRATION structure) of callback routines to thread, process and desktop
handle operation. Additionally, there is also the ObUnregisterCallbacks routine to revert all
callback’ regitration. Besides the obvious usage by malware threats (including rootkits), I have
seen it being used in anti-cheats too and, of course, Microsoft drivers also use it, of course. For
example, in the piece of code below that also comes from mssecflt.sys (it is the SecObAddCallback
function) , readers can clearly see the call for ObRegisterCallbacks routine, its parameters being
setup and even a a reference to a PreOperationCallback being setup few lines above:

[Figure 72] mssecflt.sys filter driver using ObRegisterCallbacks

There are other callbacks, and a few of them are not documented, but those ones are enough to illustrate
the idea. The advantage in using callbacks is clear because it allows to establish reactive protections and
measures (for example, enforcing a protection) that is enabled when a relevant action happens in the
system. As mentioned, these callbacks are extensively used by protective defenses as auxiliary for malware
detection.
An interesting experience is learning about callbacks that are configured to be executed as a reaction of a
system event. As expected, we have many ways to accomplish this task, and fortunately there WinDbg
extensions that makes easy to retrieve different information from system:
▪ wdkgark: https://github.com/swwwolf/wdbgark
▪ SwishDbgExt: https://github.com/comaeio/SwishDbgExt and
https://gitlab.com/opensecuritytraining/swishdbgext.git
Both extensions are old, and not all commands work as expected in recent Windows versions, but they are
still great contributions. In both cases, you must clone the project with git clone command and build them.
Personally, I always copy my extensions to the appropriate WinDbg extension folder (in this case is
C:\Program Files (x86)\Windows Kits\10\Debuggers\x64\winext), but you can store extensions wherever
you want, and afterwards passing the full path (without double quotes or spaces) while running the !load
67 | P a g e
https://exploitreversing.com

extension command. Anyway, you should make sure that you are using the right WinDbg version (x64) with
the correct extension. A simple execution retrieving callbacks using SwishDbgExt follows:

[Figure 73] Listing callbacks using SwishDbgExt.dll

68 | P a g e
https://exploitreversing.com

Of course, readers could retrieve a specified list manually. For example, get a list of
PsCreateProcessNotifyRoutines by executing the following command:
0: kd> .for (r $t0=0; $t0 < 9; r $t0=$t0+1) { r $t1=poi($t0 * 8 + nt!PspCreateProcessNotifyRoutine); .if ($t1
== 0) { .continue }; r $t1 = $t1 & 0xFFFFFFFFFFFFFFF0; dps $t1+8 L1;}

ffffb788`6fedff98 fffff804`5a5d2840
ffffb788`705fe2b8 fffff804`6115f6b0
ffffb788`705fea68 fffff804`5a74d470
ffffb788`705fea08 fffff804`6189c480
ffffb788`70c30b68 fffff804`61e00750
ffffb788`70c31a38 fffff804`605d9060
ffffb788`70c31ac8 fffff804`66bba740
ffffb788`728ac4a8 fffff804`67b90a60
ffffb788`7208b488 fffff804`695b7d00

We noticed that all addresses above do not have symbols associated, but the reason is that I tested the
command in Windows Inside Preview, and I didn’t have time to download its respective symbols.
Repeating the same procedure on a daily Windows 11 we have:
0: kd> dd nt!PspCreateProcessNotifyRoutineCount L1
fffff800`16b5377c 00000006
0: kd> .for (r $t0=0; $t0 < 6; r $t0=$t0+1) { r $t1=poi($t0 * 8 + nt!PspCreateProcessNotifyRoutine); .if ($t1
== 0) { .continue }; r $t1 = $t1 & 0xFFFFFFFFFFFFFFF0; dps $t1+8 L1;}

ffffba8a`6b49c548 fffff800`195b5500 cng!CngCreateProcessNotifyRoutine

ffffba8a`81bce5c8 fffff800`2db7f6b0 WdFilter+0x4f6b0
ffffba8a`6dff3a38 fffff800`193ec460 ksecdd!KsecCreateProcessNotifyRoutine
ffffba8a`6dff3888 fffff800`1a68fc30 tcpip!CreateProcessNotifyRoutineEx
ffffba8a`746f7408 fffff800`1abb8130 SysmonDrv+0x8130
ffffba8a`746f7bb8 fffff800`1ac7d980 iorate!IoRateProcessCreateNotify

Another way to get the same result would be executing the following sequence of commands:

[Figure 74] Retrieving PsCreateProcessNotifyRoutine callbacks

69 | P a g e
https://exploitreversing.com

As you can see, first I got the number of callback functions then I made a simple loop to retrieve the
response. Certainly, readers might ask the reason I am using PspCreateProcessNotifyRoutine (with an
extra “p” in t e name) and not PsCreateProcessNotifyRoutine (the name of the function responsible for
registering callback routines). It happens that PspCreateProcessNotifyRoutine (wit an extra “p” in t e
name) is an array the stores up to 64 callback routines.
If readers want to repeat the procedure using wdbgark, so I suggest the following commands:
▪ !load C:\Users\Administrator\Desktop\remote\wdbgark.dll (example)
▪ !wdbgark.help
▪ !wa_systemcb
The output is extensive, so I will not include it here, but readers will like it because it is very complete.
Finally, if you want to test, you can use Volatility to retrieve callbacks from Windows. To install Volatility 3
on Linux (my environment is an Ubuntu 22.10), execute the following steps:
▪ git clone https://github.com/volatilityfoundation/volatility3.git
▪ pip install -r volatility3/requirements.txt
▪ wget https://downloads.volatilityfoundation.org/volatility3/symbols/windows.zip
▪ mv windows.zip volatility3/volatility3/symbols/
Acquire the target tem’ memor b uing one of available:
▪ Surge (commercial tool): https://www.volexity.com/products-overview/surge/
▪ WinPmem: https://github.com/Velocidex/WinPmem/releases
▪ Magnet RAM Capture: https://www.magnetforensics.com/resources/magnet-ram-capture/
▪ Belkasoft RAM Capturer: https://belkasoft.com/ram-capturer
▪ Magnet DumpIt for Windows: https://www.magnetforensics.com/resources/magnet-dumpit-for-
windows/
You can list all enabled callbacks. As the output is long, so I used grep command to filter only one callback
type and I also run the command on another Windows 11 with 4 GB (and not 64 GB) to speed up the test:

[Figure 75] Retrieving PsCreateProcessNotifyRoutine callbacks using Volatility 3

Having addresses of each callback we can do further investigation. Readers can examine other callbacks
according to the context.
As I had mentioned previously, this section is only a fast review, and there are more details about the
subject, but eventually it is enough for now.

70 | P a g e
https://exploitreversing.com

8. Reversing and Windows Filtering Platform (WFP)

As I already described, programming and handling kernel events is a different approach and, as expected,
the nature of these mechanisms is also different, starting by the memory organization, where the heap is
referred by kernel pools, and these ones are presented with distinct characteristics. Actually, in recent
versions of Windows 10 and 11, the kernel is using the Segment Heap instead of being’ using the old pool
scheme, but concepts are the same. Check for the following structures:
a. _EX_POOL_HEAP_MANAGER_STATE:
https://www.vergiliusproject.com/kernels/x64/Windows%2011/22H2%20(2022%20Update)/_EX_P
OOL_HEAP_MANAGER_STATE

b. _EX_HEAP_POOL_NODE:
https://www.vergiliusproject.com/kernels/x64/Windows%2011/22H2%20(2022%20Update)/_EX_H
EAP_POOL_NODE).

The heap can be NonPagedEx (non-paged and non executable), NonPaged (non-paged), Paged, Session
and Special, although we will be using the first three types here. The non-paged heap (or pool) refers to
memory pages that can not be sent (paged out) to the disk and, of course, in the case of paged heap (or
pool) such memory pages can be sent to the disk. Modern mechanisms as Segment Heap also bring other
different concepts in terms of its organization like Low Fragmentation Heap (used for allocations lower
than 512 bytes, and now an allocation there i completel randomied in term of location’ addre),
Variable Size (for allocations between 512 bytes and 128 KB), Backend (for allocations between 128 KB
and 512 KB) and, finally, Large Block (for allocations greater than 512 KB).
Unfortunately (for researchers), many protections have been introduced or improved, and the main
protections are Kernel Mode Code Signing (KMCS), which is enforced by ci.dll and that demands that any
loaded driver to be signed, kASRL (kernel address space randomization), Hypervisor Code Integrity
(HVCI), which is VBS-based and protects the kernel against exploitation by preventing executable and
writable (W^X) privileges at same time for a page allocation on the kernel, so preventing any malware and
shellcode execution there. Additionally, any allocation must come from a signed driver and helped by the
Secure Kernel (running on VTL 1). Exploiting kernel driver’ vulnerabilitie have become harder in the lat
years. No doubt, this topic is incredibly attractive and could fill up dozens of pages, but these introductory
paragraphs are enough for us, and I recommend readers search for details on books, articles and MSDN
pages from Microsoft.
Returning to kernel drivers themselves, it could be quite complicated to know the starting point to initiate
an analysis because most drivers have dozens or hundreds of routines to examine and, of course, having
reference points are useful. Eventually an exception to this rule are malicious drivers, which might be large,
but usually are not, and sometimes it could make tasks simpler.
No doubt, all concepts I have mentioned along of this article are essential as well as all referred routines
that, almost certainly, readers will find when opening it on IDA Pro. For example, DriverEntry( ) is the first
and obvious choice because it works as a routine to invoke other important routines under certain
conditions. However, I want to comment about other aspects of the subject that will be useful for you.
71 | P a g e
https://exploitreversing.com

As we learned, applications submit requests to other drivers by calling routines like DeviceIoControl using
device I/O controls (which are also known as IOCTL), which forces the I/O Manager to create and submit
an IRP. At the same way, even other drivers can submit requests to the target driver by using well-known
functions such as IoCallDriver (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-iocalldriver) and IoBuildDeviceIoControlRequest
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-
iobuilddeviceiocontrolrequest), whose macro and routine are associated with the
IRP_MJ_INTERNAL_DEVICE_CONTROL major code. As drivers has a device object by the IoCreateDevice
routine (https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/wdm/nf-wdm-iocreatedevice),
and a link for such device object and the respective device name are given by a symbolic link created by
the IoCreateSymbolicLink routine (https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/wdm/nf-wdm-iocreatesymboliclink).
Probably readers already noticed that, at this point, the next most important piece of code is the
initialization of the dispatch routines and, in special, the array of the function pointers that is contained by
MajorFunction member field that makes part of the _DRIVER_OBJECT structure. As expected, there are
multiple dispatch routines and, sometimes, it is hard to examine all of them, so maybe a good approach
would be starting by the most used one such as DispatchRead (IRP_MJ_READ code),
DispatchWrite(IRP_MJ_WRITE code), DispatchCreate (IRP_MJ_CREATE code) and DeviceIoControl|
IoBuildDeviceIoControlRequest (IRP_MJ_DEVICE_CONTROL | IRP_MJ_INTERNAL_DEVICE_CONTROL
codes) routines. This last one is a consequence of calling DeviceIoControl |
IoBuildDeviceIoControlRequest | IoCallDriver routines (mentioned above), and it is responsible for
sending a control code (IOCTL) to a target driver. Thus, it becomes the most important for us because it
shows the meage’ flow between application and driver, or even between the current driver and other
supportive ones. While there is a list of I/O control codes defined in the SDK header files, most of these
IOCTL codes are private and defined by drivers, and it might turn analysis a bit harder. No doubt, learning
about these I/O control codes through an eventual reverse engineering task is really useful for getting a
better understanding of the kernel driver.
If readers need to a list of standard and well-known I/O control codes, so eventually some of them are
available on Internet: http://www.ioctls.net/
So far we have the following key points to be regarded at first moment of a driver analysis:
▪ Finding the DriverEntry routine.
▪ Take an initial note about key routines being invoked from DriverEntry routine as callback routines
for reading, writing and sending control codes to a device driver.
▪ Searching for the symbolic link associated with the device object.
▪ Finding the device name (DeviceName).
▪ Analyzing I/O control codes, device object and buffers used by routines such as DeviceIoControl
and IoBuildDeviceIoControlRequest.
Sure, these items are only a starting point. If readers are wondering how the IOCTL codes, which are used
with IRP_MJ_DEVICE_CONTROL requests (created by invoking DeviceIoControl( ) for communication
between user-mode application and kernel driver) or IRP_MJ_INTERNAL_DEVICE_CONTROL requests
(created by invoking IoBuildDeviceIoControlRequest for communication between two kernel drivers),
there is a macro as shown below:
72 | P a g e
https://exploitreversing.com

#define IOCTL_Device_Function CTL_CODE(DeviceType, Function, Method, Access)

IOCTL definition (it is a 32-bit value) is given by four components:
▪ DeviceType: it determines the device type.
▪ FunctionCode: it is an indicative about the function to be executed by the driver.
▪ TransferType: it determines how data will be transferred between the caller (user-mode
application or another driver) and the target driver that is responsible for handling the IRP. Possible
values are METHOD_BUFFERED, METHOD_IN_DIRECT or METHOD_OUT_DIRECT,
METHOD_NEITHER.
▪ RequiredAccess: this parameter determines the type of access requested by the caller to open the
file object that represents the device. Possible values: FILE_ANY_ACCESS, FILE_READ_DATA,
FILE_READ_DATA and FILE_WRITE_DATA.
I think I have already provided enough concepts for this article and the next ones.
It is not my intention to analyze a malicious driver (rootkit) in this article, but I will do a fast analysis of one
well known sample named Netfilter (also known as Retliften), which work as a trojan (x64) and that, at
past, was signed (at that time) by Microsoft by mistake. To download it from Malware Bazaar, execute:
malwoverview.py -b 5 -B e8e7f2f889948fd977b5941e6897921da28c8898a9ca1379816d9f3fa9bc40ff
If readers want to list and download other potential malicious drivers, this task can be done by executing
the following command:

[Figure 76] Listing malicious drivers from Malware Bazaar using Malwoverview (truncated output)
The next step is to open it on IDA Pro and observe a few facts.

73 | P a g e
https://exploitreversing.com

After launching IDA Pro and before jumping to DriverEntry routine, do not forget few basic steps:
▪ Force decompilation of the entire driver by going to File → Produce file → Create C File.
▪ Go to Edit → Plugins → Hex-Rays Decompiler → Options and change Default radix value to 16.
▪ As we are handling an x64 driver, open Type Libraries View (SHIFT+F11) and add (INSERT key) two
libraries: ntddk64_win10 and netapi64_win10.
▪ Open the Signatures View (SHIFT+F5) and check whether the following signatures are present:
ms64wdk, v64seh and vc64ucrt. If they are not, add them.
▪ Type CTRL+E to go to the Entry Point (DriverEntry).

[Figure 77] DriverEntry routine

74 | P a g e
https://exploitreversing.com

Likely readers will find common structures and routines that we have commented on in this article and,
hopefully, it will not be hard. Actually, there are references that are familiar for us:
▪ DriverEntry: driver’ entr point.
▪ DriverObject: a variable of type DRIVER_OBJECT, which represents the image of a loaded driver.
▪ DriverUnload: routine used to unload the driver.
However, there are two routine that we don’t comment about yet:
▪ RtlCopyUnicodeString: as you already realized, this routine copies a string to a destination buffer.
Remember that Rtl means Real Time Library.

▪ WdfVersionBind: this routine binds the driver to a specific WDF library version.
I could find definition of this function (and also WdfVersionUnbind) on
https://github.com/microsoft/Windows-Driver-
Frameworks/blob/main/src/framework/shared/inc/private/common/fxldr.h , which have the following
prototypes:
NTSTATUS
WdfVersionBind(
__in PDRIVER_OBJECT DriverObject,
__in PUNICODE_STRING RegistryPath,
__inout PWDF_BIND_INFO BindInfo,
__out PWDF_COMPONENT_GLOBALS* ComponentGlobals
);

NTSTATUS
WdfVersionUnbind(
__in PUNICODE_STRING RegistryPath,
__in PWDF_BIND_INFO BindInfo,
__in PWDF_COMPONENT_GLOBALS ComponentGlobals
);

Reader alread noticed that there are two tpe that we don’t do not know anthing about uch a
PWDF_BIND_INFO and PWDF_COMPONENT_GLOBALS. Usually, I have used two approaches find this
information:
▪ Cloning the repository (git clone https://github.com/microsoft/Windows-Driver-Framework) and
search recursively for the structures by using: findstr /S <string> *.

▪ Searching for structure definitions on the excellent websites such as

https://github.com/winsiderss/systeminformer and https://doxygen.reactos.org/.
Unfortunately, you will discover that these structures also mention other ones in their definitions, but
hopefully you will have all of them.
If you want to improve the WdfVersionBind definition on IDA’s idb (it is not really necessary here) then it
will be necessary to add all structure definitions into Local Types (SHIFT+F1):

75 | P a g e
https://exploitreversing.com

[Figure 78] Local types being declared and added into idb
Multiple entries will be created separately in the Local Types View, so right-click all of them and choose
Synchronize to idb option.

[Figure 79] Local types being declared and added into idb

There will not be an amazing effect in the code for this specific case, but this procedure is still valuable to
explain to readers how to proceed in similar cases. Anyway, by going to sub_140003C20 →
sub_14000395C readers will easily identify the device name associated with the driver:

[Figure 80] Device name being revealed

76 | P a g e
https://exploitreversing.com

Moving into sub_140005284 routine (not shown in the last image, but only three instructions below), we
will find the following content:

[Figure 81] sub_140005284 routine

From the last page we learned that this malicious driver named NET_FILTER is likely controlling (monitoring
or even altering) the network filtering behavior through the network communication. Although I didn’t
have explained this stuff previously, APIs to interact with network stack on Windows are offered by the
WFP (Windows Filtering Platform). In terms of nomenclature, the WFP architecture offers network stack
composed by layers (there are about a hundred of them and each one has a GUID associated), which each
layer can be composed by zero or more filters, and zero or more associated callout drivers, which are
responsible of executing by processing the data. Yes, I know that concepts here might be hard to
understand and, eventually, readers are not used to them, so a quick introduction might be useful at this
point.
A good advantage of choosing this malicious driver is that I can superficially comment about WFP
(Windows Filtering Platform), which is an amazing and powerful resource that can be used as useful
method to intercept and manipulate network data and, as everything in information security area, it can
be used to good and bad purposes. The malicious driver itself is not important or relevant for us, but
techniques and concepts definitely are. Therefore, beyond learning basic concepts about WFP, it will be
possible to provide a preview of the technology applied to a real case and even restricted to this article, to
try to correlate general concepts and details about the WFP framework with such analysis.
77 | P a g e
https://exploitreversing.com

The WFP (Windows Filtering Platform) is composed by the following large components:
▪ Filter Engine: the component is responsible for performing the filtering task, calling callouts based
on the classification and, at end, allow or not a determined traffic.
▪ Base Filtering Engine: this component is a macro component in the WFP, and it ties filters, reports,
statistics, security model and configuration together.
▪ Shims: this component represents kernel mode components that actually make the filtering
decision based on the classification.
▪ Callout: this component, as we learned so far, is a function that effectively permit, block, modify
and even reinject a network traffic. As expected, they must be registered to WFP layers.

In few words, we can directly or indirectly interact with multiple components and subcomponent of the
WFP such as:

▪ Filters: they are involved in the classification then they can be interpreted as rules to accept or
block network traffic. Filters are organized within sublayers, and the order is given by the weight,
which is similar to altitude for minifilter drivers.
▪ Layers: the work a the filter’ organiation inide the filter engine, and cannot be removed.
▪ Sublayers: they make part of layers, and generally handle exceptions in rules or a particular
scenario. They can be added or removed, and there is a set of sublayers that are inherited by
layers.
▪ Callout: they are a set of functions actively involved in the classification process as permitting or
blocking network data. Callouts can be added or removed.
▪ Shims: it is the kernel-mode component that is responsible for making classifying decisions on
filters of a specific layer. In other words, the shim component starts the classification, which is
composed by applying the filters to, at the end, decide if a network traffic should be blocked or
allowed.

The sequence of components involved in the processing is network packet → network stack → shim →
filters (from a layer) → callouts → shims (actually performing and following the filtering decision).
Decisions can be simplified as permitting (FWPM_ACTION0.type = permit) or blocking
(FWPM_ACTION0.type = block), but there are few nuances:

▪ a block decision overrides a permit decision.

▪ a block decision is a final decision, but it still depends on the flag described on the next line.
▪ there is a flag named FWPS_RIGHT_ACTION_WRITE that enables and controls whether a lower
sublayer (remember about weight concepts) can override a decision.
▪ A block decision made by a callout is a soft decision and a block decision made by a filter is a hard
decision.
Returning to the code, readers see a series of functions being called, and in few words their meaning
follow:
▪ FwpmEngineOpen0: it opens a session to the filter engine and, as expected, returns a handle to it.
▪ FwpmTransactionBegin0: starts a transaction with the current session and, to accomplish this task,
it uses the handle to the opened session returned by FwpmEngineOpen0 routine.
78 | P a g e
https://exploitreversing.com

▪ Inside of the sub_140004F2C routine, we have FwpsCalloutRegister1 function, which is responsible

for registering a callout. This function receives a pointer to Device Object, a pointer to callout
structure (typed as FWPS_CALLOUT1_) and returns a calloutId that is used to identify the callout
within the filter engine. The sub_140004F2C routine, FwpsCalloutRegister1 function and
FWPS_CALLOUT1_ structure is shown below:

[Figure 82] sub_140004F2C contains the FwpsCalloutRegister1 routine

[Figure 83] FwpsCalloutRegister1 routine

[Figure 84] FWPS_CALLOUT1_ structure

▪ The interpretation for members of callout structure (FWPS_CALLOUT1_) is direct:

▪ first member (calloutKey) contains the GUID

(0BABE0A0B870EFD9A4854F0780CF72951h);
▪ the second member represents flags (zero);
▪ the third member (classifyFn) contains a pointer to a function that works as a
notification (trigger) to invoke the callout whenever there is network data;

79 | P a g e
https://exploitreversing.com

▪ the fourth member (notifyFn) is a pointer to a function that will be called when any filter
using this callout is added or deleted, as well associated events with callout happen.
▪ the fifth parameter (flowDeleteFn) holds a pointer to a function that will be invoked
when the data flow being processed by the callout is finished.
The sub_140004FB8 is the most important routine so far:

[Figure 85] sub_140004FB8: invoking relevant calls

80 | P a g e
https://exploitreversing.com

As highlighted in the code, there are three key subroutines being called:
▪ FwpmCalloutAdd0: this routine is responsible for adding a new callout to the system and its
prototype is DWORD FwpmCalloutAdd0([in] HANDLE engineHandle, const FWPM_CALLOUT0
*callout, PSECURITY_DESCRIPTOR sd,[out, optional] UINT32 *id). The first parameter is a handle
to the open session to the filter engine, the second parameter is a pointer to the callout object
(FWPM_CALLOUT0 structure) and the last parameter represents the output, which is a runtime
identifier.

[Figure 86] FWPM_CALLOUT0 structure

▪ FwpmSubLayerAdd0: this routine adds a sublayer to the system, and its prototype is given is
DWORD FwpmSubLayerAdd0([in] HANDLE engineHandle, [in] const FWPM_SUBLAYER0
*subLayer, [in, optional] PSECURITY_DESCRIPTOR sd). The second argument represents the
sublayer to be added.

[Figure 87] FWPM_SUBLAYER0 structure

▪ FwpmFilterAdd0: this routine adds a new filter object to the system, and its prototype is DWORD
FwpmFilterAdd0([in] HANDLEengineHandle, [in] const FWPM_FILTER0 *filter, [in, optional]
PSECURITY_DESCRIPTOR sd, [out, optional] UINT64 *id), whose second parameter is a pointer to
the filter object to be added and the fourth parameter, similar to the FwpmCalloutAdd0,
represents the output as a runtime identifier.
Line 7 from the last figure has a reference to xmmword_140007680. Actually, if we follow this data
reference, we will ee a big hexadecimal number. Preing “U hotkey” (or even “A hotkey”), we will ee a
Unicode string, but without an appropriate representation (actually, it is not necessary to press U or A hot
keys, and I show it to prove that is a Unicode string). Selecting all lines containing characters and going to
Edit → Strings → Unicode, and the “redirectCalloutV4” tring will pop up. There are other Unicode strings
being used by the pseudo code within this routine, so readers can repeat the same approach for them.
After handling strings and renaming variables, we have the following pseudo code:

81 | P a g e
https://exploitreversing.com

[Figure 88] sub_140004FB8: improved code

82 | P a g e
https://exploitreversing.com

[Figure 89] FWPM_FILTER0 structure (from FwpmFilterAdd0 routine)

From the pseudo code, we have that:
▪ The callout i diplaed a “redirectCalloutV4”.
▪ The callout’ decription i “IPv4 callout for redirect”.
▪ Remember that a callout object is represented by FWPM_CALLOUT0 structure.
▪ The displayData field from FWPM_CALLOUT0_ structure is represented by the
FWPM_DISPLAY_DATA0 structure, which is composed by wchar_t pointers that are name and
description fields (check for lines 11 and 12).
▪ On line 6, flags (from FWPM_CALLOUT0_ structure) are zero, but it could be
FWPM_CALLOUT_FLAG_PERSISTENT (0x00010000), FWPM_CALLOUT_FLAG_PERSISTENT
(0x00020000) and FWPM_CALLOUT_FLAG_REGISTERED (0x00040000) values.
▪ The calloutKey identifies a session and applicableLayer indicates which layer such callout will be
used, so this field forces that only filters from this provided layer are allowed to invoke the callout.
▪ The ublaer’ decription i “Sublayer for redirect” and its displayName is “redirect for Sublayer
(lines 27 and 30).
▪ The sublayer, which has a FWPM_SUBLAYER0 structure associated, is also identified by a GUID in
the subLayerKey. Sure, there is a list of built-in sublayers, but in this specific case there is a
provided key (check for line 24). If we follow the key reference we will find the following
information:

[Figure 90] Sublayer’s key (from FWPM_SUBLAYER0 structure)

83 | P a g e
https://exploitreversing.com

▪ To format this GUID I used the following simple IDC script:

[Figure 91] IDC script to format GUID

▪ On the IDA Pro command line, run this macro proving the address of the start of the GUID:
Guid(0x00000001400084F8) == {AE1E820A-C60A-42A8-B4A2-9ACFB050387F}.
▪ The weight of the sublayer is 0xFFFF (line 29), which means that it is the first to be invoked.
▪ The number of filter conditions (numFilterConditions) is zero. Thus, there is not any established
condition to invoke the filter.
▪ The dipla’ name of the filter is redirectFilterV4 and its respective description is “IPv4 fil er for
re irec ” (lines 42 and 45).
▪ The filter’ action tpe i FWP_ACTION_CALLOUT_TERMINATING, which basically forces
invoking a callout that always returns block or permit. To show this string representation, I
searched for a macro (M hotkey).
▪ The FWPM_FILTER0_.weight.type equal to FWP_UINT64 (line 48) means that the Base Filtering
Engine will use the provided value as weight, which is 0xFFFFFFFFFFFFFFFF (lines 37 and 50).
▪ On line 53, calloutKey is the GUID for a callout that is valid in the layer (line 16) and layerKey
(line 64) holds the GUID which the filter is hosted, and it matches against the line 17.
▪ On line 55, finally the code adds a filter object into the system by calling FwpmFilterAdd0
routine, which used the filter object constructed in previous lines.
Readers already noticed that WFP is basically a set of hooks inside the network stack and also filtering
engine, which allow us interacting, monitoring and eventually controlling the network data information. By
the way if you are wondering about the meaning of FWPM, it is Filtering Windows Platform Management,
which is an appropriate name for the framework. Therefore, apparently the malware is adding a new
sublayer, filter and associated callout to handle the IPv4 communication that, in this case, it is working as
an IPv4 redirector to another IP address, but it early to conclusions. We alo have mentioned an “arbitrar
GUID” and there i nothing new here becaue a a callout i a common kernel driver, any GUID can be
generated by Visual Studio and likel the malware’ author did it.

84 | P a g e
https://exploitreversing.com

On purpose I quickly commented about the the sub_140004F2C routine (Figure 82), but we must
remember that is this routine which is responsible for registering the callout with the filter engine.
Additionally, its members like classifyFn (points to a function that will be called whenever there is data to
be processed) and notifyFn (points to a function that is called whenever data flow that is being processed
is terminated) from the FWPS_CALLOUT1_ structure are relevant.
The classifyFn is actually a callout of the callout, and its prototype is given the following:

[Figure 92] FwpsCalloutClassifyFn1

This callback has the following parameters:
▪ inFixedValues: it contains a pointer to an FWPS_INCOMING_VALUES0 structure, which holds the
values for each of data fields in the layer being filtered.
▪ inMetaValues: it contains a pointer to an FWPS_INCOMING_METADATA_VALUES0 structure,
which holds the values of each metadata field being in the layer being filtered.
▪ layerData: it contains a pointer to a structure describing the data being filtered.
▪ classifyContext: it contains a pointer to context data.
▪ filter: it holds a pointer to an FWPS_FILTER1 structure.
▪ flowContext: it holds the context associated with data flow.
▪ classifyOut: it is a pointer to an FWPS_CLASSIFY_OUT0 structure, which receives the return that
will be returned by classifyFn1 function to the caller.
From Figure 82, we know that:
▪ sub_1400053A0 is the classifyFn callout.
▪ sub_140005520 is the notifyFn callout.
Moving inside the sub_1400053A0 subroutine (classifyFn callout), we will not see a friendly aspect,
unfortunately (check Figure 93 ahead). Thus, I performed the following steps:
▪ I renamed (N hotkey) all its parameters according to prototype described on
https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/fwpsk/nc-fwpsk-
fwps_callout_classify_fn1.
▪ I added (SHIFT + F9 → INS → Add standard structure) all missing structures:
FWPS_INCOMING_VALUES0, FWPS_INCOMING_METADATA_VALUES0, FWPS_FILTER1,
FWPS_CLASSIFY_OUT0_ and FWPS_INCOMING_METADATA_VALUES0_.

85 | P a g e
https://exploitreversing.com

▪ I changed all argument’ tpe (Y hotkey) according to function’ ignature.

▪ I renamed variables over the code and applied two macros (M hotkey).
The result on Figure 94 is far from being perfect, but it is already possible to have a better idea view:

[Figure 93] Original sub_1400053A0

86 | P a g e
https://exploitreversing.com

[Figure 94] Improved sub_1400053A0 (classifyFn)

87 | P a g e
https://exploitreversing.com

[Figure 95] Improved sub_1400053A0 (second part): classifyFn

Analyzing the resulting function, we can do the following observations:
▪ The FWPS_CALLOUT_ structure (as shown on Figure 86 and applied on Figure 82), which is used
and associated to the FwpsCalloutRegister routine, was our starting point to get at this point of
analysis because it involves three relevant callouts such as classifyFn, notitfyFN and flowDeleteFn
and, at this moment, we are analyzing classifyFn. The route up to this point is sub_140004F2C →
sub_1400053A0.
▪ Therefore, on line 18 (Figure 94), the layerId field, which determines the runtime filtering layer, is
tested and verified whether is equal to FWPS_LAYER_ALE_CONNECT_REDIRECT_V4 (TCP traffic – a
sender|client component). This filtering layer allows any modification of remote address and port
of outgoing connections, so it is involved with redirecting.
▪ The “ALE” tring mean Application Later Enforcement and, as expected, is composed of multiple
filtering layers and also matching discard layers, which are involved in logging.
▪ Sometimes readers will find FWPM (Filtering Windows Platform Management) data types, which
are related to management tasks (callouts and adding filters) and other times will see FWPS data,
which is associated to callout data types (the actual filtering). There are counterparts on both sides,
although FWPS data types are usually smaller than FWPM data types. That is the reason we see a
layerId field being compared to FWPS_LAYER_ALE_CONNECT_REDIRECT_V4 (0x42 – represented
by 16 bits) while for FWPM filtering layers that GUIDs have 16-bytes. Furthermore, there are other
subtle differences that will not be commented on here.
▪ On line 24, if the layerId is not FWPS_LAYER_ALE_CONNECT_REDIRECT_V4, the decision is
FWP_ACTION_PERMIT (loaded into actionType field), which means that the network filter allows
the network data to be transmitted or received. It could be suitable to know that classifyOut, which
is a member of FwpsCalloutClassifyFn1 callout, is a pointer to FWPS_CLASSIFY_OUT0 structure,
and it receives a decision returned by the classifyFn callout function. Possible values are
FWP_ACTION_PERMIT (our case), FWP_ACTION_BLOCK and FWP_ACTION_CONTINUE.
FWP_ACTION_NONE. Thus, at the end, the final decision is taken by the classifyFn callout function.
▪ The FwpsAcquireClassifyHandle0 routine is responsible for generating a classification handle that
will be used for asynchronous classification and, most importantly, data modification in other

88 | P a g e
https://exploitreversing.com

functions such as FwpsApplyModifiedLayerData0, FwpsAcquireWritableLayerDataPointer0,

FwpsAcquireWritableLayerDataPointer0 and FwpsReleaseClassifyHandle0 functions. All of these
routines are present within sub_140005524 routine (line 74).
Before proceeding, remember: FWPM refers to WFP user mode objects identified by GUIDs and FPWS
refers to WFP kernel mode objects identified by LUIDs (locally unique identifier). Once again, the
execution flows take to another routine, sub_140005524, which is composed of a series of calls related
directly or indirectly to callouts. As usual, it is interesting to show the code before any treatment as
presented on the next page:

[Figure 96] sub_140005524: original code

There few WFP routines being called, so a summary about them follows:
▪ FwpsAcquireWritableLayerDataPointer0: this function returns layer-specific data that can be
inspected or even changed. The second parameter (filterId) is the same from classifyFn routine’s filter
parameter, and its internal organization is given by FWPS_FILTER1_ structure, which establishes
subLayerWeight, numFilterConditions, action and filterCondition, among other fields.
89 | P a g e
https://exploitreversing.com

▪ FwpsReleaseClassifyHandle0: this routine releases the previously acquired classification handle by

FwpsAcquireClassifyHandle0 routine (check page 87).
▪ FwpsApplyModifiedLayerData0: this function applies changes produced by the
FwpsAcquireWritableLayerDataPointer0 routine.
▪ FwpsCompleteClassify0: this routine completes a pending classify request.
Thus, after performing a quick analysis and a bit of reversing, the improved version of sub_140005524
follows below:

[Figure 97] sub_140005524: improved code view

No doubts, the presentation of the code is better than the original version, and I did the following:
▪ I renamed a1 to arg_1 and a2 to arg_2 (N hotkey).
▪ As arg_1 apparently was clearly a structure, so I created one by right-clicking it and choosing Create
a new structure type.
▪ I used the prototype of FwpsAcquireWritableLayerDataPointer0 routine to rename the arguments.
90 | P a g e
https://exploitreversing.com

▪ I applied macros such as FWP_ACTION_PERMIT and FWPS_RIGHT_ACTION_WRITE. Having the

right FWPS_RIGHT_ACTION_WRITE allows the callout driver to write the actionType member of
this structure, and changing as intended. If there was not this right here, it could write to
actionType if it needed to block a previous FWP_ACTION_PERMIT decision took by a filter with
higher weight (remember: weight presents the same idea of altitude in mini-filter drivers).
▪ I added the enum MACRO_FWPS to be able to apply
FWPS_CLASSIFY_FLAG_REAUTHORIZE_IF_MODIFIED_BY_OTHERS. The information provided by
FwpsApplyModifiedLayerData0 on MSDN about its prototype was essential to do it
(https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/fwpsk/nf-fwpsk-
fwpsapplymodifiedlayerdata0).
▪ The prototype of FwpsAcquireWritableLayerDataPointer0 (https://learn.microsoft.com/en-
us/windows-hardware/drivers/ddi/fwpsk/nf-fwpsk-fwpsacquirewritablelayerdatapointer0)
provided another useful hint. When describing writableLayerData, which is an output argument,
the description says that it a void pointer to be cast later to the appropriate structure type.
However, under the Remarks section, the MSDN tells us that it could be only two possible
structures: FWPS_BIND_REQUEST0 and FWPS_CONNECT_REQUEST0. Examining them, so it
became clear that the code i referring to the econd one becaue “ defines modifiable data for
the FWPM_LAYER_ALE_AUTH_CONNECT_REDIRECT_V4 and FWPM_LAYER_ALE_AUTH_CONNEC
T_REDIRECT_V6 layers.” (check: https://learn.microsoft.com/en-us/windows-
hardware/drivers/ddi/fwpsk/ns-fwpsk-_fwps_connect_request0). The same applies to
writableLayerData_1 because they are the same.
▪ The _FWPS_CONNECT_REQUEST0 structure has few interesting fields, but the first two of them are
more attractive at this time. As they are of SOCKADDR_STORAGE type, I changed their types (Y
hotkey) to sockaddr_in based on my previous experience. Once fields become clearer, I just
renamed other fields of arg_1 according to the context.
▪ I added at least one enumeration starting with ‘AF_’ ‘SOCK_’ and ‘IPPROTO’ (remember: adding
one enumeration value forces the IDA Pro to insert the whole enumeration associated) by going to
Enum tab, pressing INS key and choosing Add standard enum by symbol name. Afterwards, I used
these values to apply the missing macros.
▪ Other variables also have been renamed (N hotkey) according to the context.
Certainly, it could seem difficult to get an improvement of the prior code, but once readers can understand
my explanations above then the process becomes easier than expected. So far, our anali’ paths have
been the following:
▪ sub_14000395C → sub_140005284 → sub_140004F2C → sub_1400053A0 → sub_140005524
▪ sub_14000395C → sub_140005284 → sub_140004FB8
Returning to sub_140005284 we have the remaining functions:
▪ FwpmTransactionCommit0: this function commits the opened transaction.
▪ FwpsCalloutUnregisterById0: this function unregisters a callout.
We can now return to sub_14000395C routine (figure 80), and try to draw conclusions and get further
details from other routines that we left behind. It is important to highlight that I am focusing only on a
small part of the code that is related to device object and Windows Filtering Platform (WFP) as an
opportunity to explain new concepts and not due to the malicious driver itself.
91 | P a g e
https://exploitreversing.com

The whole subroutine sub_14000395C is shown below:

[Figure 97] sub_14000395C routine

Moving inside sub_140002BBC → sub_1400031F8 routine, we find the following code:

[Figure 98] sub_1400031F8 routine

92 | P a g e
https://exploitreversing.com

The WSK_CLIENT_NPI structure is used when Network Programming Interface (NPI) is being
implemented. In a few words, NPI defines an interface between network modules, which implements a
function in the network stack, which can be attached and integrated one with other. Thus, the
WSK_CLIENT_NPI structure is described and defined as shown below:

[Figure 99] WSK_CLIENT_NPI structure

The ClientContext member is a pointer to the context of the WSK (Winsock Kernel) application’ binding
and the Dispatch member is a pointer to another structure named WSK_CLIENT_DISPATCH, which
provides a dispatch table for callback functions associated with events that are not related to a specific
socket, and that will be available to be called when necessary. Its composition is given by the following:

[Figure 100] WSK_CLIENT_DISPATCH structure

Its members are:
▪ Version: it indicates the version of WSK NPI.
▪ Reserved: it must be zero.
▪ WskClientEvent: a pointer to the WskClientEvent event callback function, which will notify the
WSK application about events not related to a specific socket.
The WskClientEvent callback function is defined as PFN_WSK_CLIENT_EVENT type as shown below:

[Figure 101] WskClientEvent callback definition

The ClientContext argument is a pointer to the context value coming from WskRegister routine;
EventType argument would be a specific event to notify the WSK application; Information argument that
is used to pass additional information to WSK application is most of times NULL; InformationLength
parameter provides the size of information. Therefore, returning to the sub_1400031F8 routine, we see
two routines being invoked: WskRegister( ) and WskCaptureProvideNPI( ).
93 | P a g e
https://exploitreversing.com

WskRegister routine registers a WSK application that is provided and implemented by WSK application
(WskClientNpi) and a pointer to a memory location identifying the registration instance of the WSK
Application (WskRegistration), which is actually initialized by WskRegister routine as the result from its
processing. Once the return is success then the WskCaptureProviderNPI routine, which is running at IRQL
<= DISPATCH LEVEL in this case because its second argument is 0xFFFFFFFF (WSK_INFINITE_WAIT), is
invoked and it captures a provider NPI when it becomes available. The first parameter (WskRegistration)
has been initialized by WskRegister routine and the third parameter contains a pointer to the WSK
provider dispatch table, which provides callbacks that the WSK application will be able to call.
Return to the sub_14000395C routine, it is time to quickly examine the sub_140004A10 routine, as shown
below:

[Figure 102] sub_140004A10 routine

Previously in this article, I commented about the CmRegisterCallbackEx( ) API, which is responsible for
registering a routine that will be used by kernel and filter drivers to monitor and, eventually, modify any
Registry operation such as renaming, enumeration, key deleting, key creation and so on. Now we have a
real example being used here and, as we also already learned, the first parameter is a callback function
(given by Function in this case), the second parameter is the altitude (320000, as readers can see on line
8), a pointer to the DRIVER_OBJECT structure and a Cookie reference, which is a pointer to
LARGE_INTEGER structure that receives a defined value that identifies the callback routine.
I will not show the content of the Function callback (provided as first argument to CmRegisterCallbackEx( )
API), but the most interesting information there are two calls to CmCallbackGetKeyObjectID routine,
which retrieves an identifier and respective object name associated with the provided Registry key object.
Note that the second parameter of CmCallbackGetKeyObjectID routine is exactly a pointer that
RegistryCallback routine of the driver receives as being a reference to the REG_XXX_KEY_INFORMATION
structure.
Returning once again to sub_14000395C routine, there are two other routines that there is something
useful inside them. The first one is the sub_140006548 routine, which has only one function being called
that is PsCreateSystemThread( ), which creates a system thread, as shown below:

94 | P a g e
https://exploitreversing.com

[Figure 103] sub_140006548 routine

The most important parameter here is StartRoutine (sixth parameter), which is a pointer to a routine
(KSTART_ROUTINE callback) to be executed. We can see that it is the second argument of this
sub_140006548 routine, and according to Figure 97 (line 33), it is the routine sub_140003A70, which is
shown below:

[Figure 104] sub_140003A70 routine

95 | P a g e
https://exploitreversing.com

The code starts calling KeEnterCriticalRegion routine on line 05, which disables the execution of normal
kernel APCs. This is a usual action when is expected that the threat performs an I/O operation. The kernel
APCs will only be re-enabled again when the code call KeLeaveCriticalRegion( ) on line 34.
On line 06, the KeSetBasePriorityThread routine is called to set the run-time priority of the current threat
by adding 5 to the base priority of the process holding the thread.
From this point at the code, the number of functions explodes, and there are too many to analyze in this
article, so I will offer only a few insights and readers can investigate by themselves if it is necessary.
The routine sub_140005678, which is called five times using different arguments, has as its main content
non-paged pool allocation using ExAllocatePoolWithTag routine (go to sub_140005678 →
sub_1400044FC). The tag used by ExAllocatePoolWithTag routine is “TLXE”. Of course, we already know
that this routine has been deprecated and replaced by ExAllocatePool2( ), but malware’ author continue
using it. Additionally, sub_140005678 routine receive a function’ pointer a firt argument, and a
mentioned, it is provided one different function by each call.
The sub_1400069A4 routine (sub_140003BF0 → sub_140004A7C → sub_140004B5C → sub_1400069A4)
has intereting function’ invocation as shown below:

[Figure 105] sub_1400069A4 routine

96 | P a g e
https://exploitreversing.com

On line 11 the sub_140006B74 is called, which has the following code:

[Figure 106] sub_140006B74 routine

In the code from sub_140006B74 routine, the sub_140006B2C routine is invoked on line 13:

[Figure 107] sub_140006B2C routine

97 | P a g e
https://exploitreversing.com

We should do an analysis in reverse order to get an overview of the code. The sub_140006B2C routine
(Figure 107) is being called with ProcessID == 4 (check line 9 in sub_140006B74 routine), which know that
is the System process. Inside sub_140006B2C routine, this processes are searched by
PsLookProcessByProcessId function, and a handle to the EPROCESS structure of the provided process is
returned. Using this handle, the PsGetProcessImageFileName function is called, and a pointer to the image
file (executable file) backing up the process in the disk is returned. Finally, the ObDereferenceObject
function is called to decrease the reference count to the EPROCESS structure and, at end of the routine,
the same pointer to the image file is returned to sub_140006B74 routine.
Returning to sub_140006B74 routine, there is a while(true) condition parsing each process until a
provided PID limit (0x10000) and searching for the first occurrence of the tring “explorer.exe”. Once it i
found, it returned through by invoking PsLookProcessByProcessId function the pointer to its respective
EPROCESS structure.
Now going up to sub_1400069A4 routine (Figure 105), which is the caller of sub_140006B74 routine, we
know that ObOpenObjectByPointer function opens an object referenced by the returned pointer from
sub_140006B74 routine, and returns a pointer to the object. In other words, it is returning a pointer to the
process represented by the EPROCESS structure that, in this case, it is the explorer.exe. Pay attention to
line 20, which confirms our interpretation that it is a pointer to a process because the fifth parameter
(ObjectType) is exactly PsProcessType, and the AccessMode given by the sixth parameter is KernelMode
(zero).
Having thi proce’ handle, it is opened by ZwOpenProcessTokenEx function, which returns the
respective TokenHandle into its fifth parameter. On the next line ExAllocatePoolWithTag is called to
allocating a PagedPool (so its content can be paged out) with the tag “WENE” and size 0x1000 bytes, and
the validity of this allocated pool is checked by invoking MmIsAddressValid function (although Microsoft
doen’t recommend using this function).
On line 41, the NtQueryInformationToken is invoked to retrieve information about the provided access
token (first parameter: TokenHandle), with second parameter equal to TokenUser which is a
TOKEN_INFORMATION_CLASS value that determines that the allocated buffer receives a TOKEN_USER
structure with the user account of the token , the third parameter is a pointer to the allocated paged pool,
the fourth parameter indicating the size of the TokenInformationBuffer (0x1000) and finally the last
parameter (ReturnLength) as being the length of the returned information.
At the end, the SID_AND_ATTRIBUTES structure, which is the only member of TOKEN_USER structure and
represents the user related to the access token, is used as argument of RtlConvertSidToUnicodeString
function (line 53) to convert it to a Unicode string representation of the SID. In other words, we have the
SID of the account associated with the explorer.exe process, which is returned within a UNICODE_STRING
structure:

[Figure 108] _UNICODE_STRING structure

98 | P a g e
https://exploitreversing.com

Returning to sub_140004CB8 routine (sub_140003BF0 → sub_140003BF0 → sub_140004CB8), there is a

call the sub_140006684 routine, which basically handles ACL, ACEs and ownership related to SIDs.
The sub_140006C90 routine (sub_140003BF0 → sub_140004A7C → sub_140006C90) is quite similar to
sub_140005678, using ExAllocatePoolWithTag function, but it allocates Paged Pool instead of NonPaged
Pool, and the tag is different: “WENE”. In this same routine, there are other Registry key manipulations
involving OBJECT_ATTRIBUTES structure.
Readers can easily realize that the following routines handle with Registry key configuration related to
Internet access (proxy) and also SID/ACL manipulation (in these specific cases, it happens in subroutines
inside the following ones):
▪ sub_140004B5C: (sub_140003BF0 → sub_140004A7C → sub_140004B5C)
▪ sub_140004E30: (sub_140003BF0 → sub_140004E30)
▪ sub_140004CB8: (sub_140003BF0 → sub_140004CB8)
Few Registry entries being manipulated:
▪ \\Registry\\User\\
▪ \\Software\Microsoft\Windows\CurrentVersion\Internet Settings\Connections
▪ EnableLegacyAutoProxyFeatures
▪ AutoConfigURL
▪ DefaultConnectionSettings
Surprisingly, we just finished reviewing one (sub_140003BF0) of five routines referred to sub_140005678
routine (Figure 104), inside sub_140003A70 routine. The next two routines, sub_140003C10 and
sub_140003B90, are simpler and similar to the sub_140003BF0, and allocate memory pool, manipulate
strings and Registry keys.
The other two routines (sub_140003B80 and sub_140003BD0) are more interesting, but they call multiple
other subroutines, and it would become our analysis an endless procedure. Of course, readers could get
interested in analyzing them because there is the presence of routines interacting with
IO_STACK_LOCATION and Completion Routines, for example.
We cannot ignore the clear proxy reference on line 19 (Figure 104), suggesting a network redirection via
proxy configuration: http://110.42.4.180:2080/u. Furthermore, readers might get interested in a
Certificate Store handling inside the sub_140005D5C routine
(“\\Registry\\Machine\\SOFTWARE\\Microsoft\\SystemCertificates\\ROOT\\Certificates\\”). Finally, if
we returned a level upper of sub_14000395C (Figure 97), we are going finding multiple routines undoing
and freeing everything: releasing pools, unregistering callbacks (CmUnRegisterCallback routine),
releasingWSK application’s registration instance releasingNetwork ProgrammingInterface (NPI)
removing filter object, removing callout and, at end, closing the session to the filter engine.
Anyway, I already had said that would be only a fast overview about few pieces of code of this malicious
binary, but after having analyzed those few routines, the malicious drivers apparently try to open a kind
of exception in the filtering rule and redirecting the network data to a determined remote address and
IP port. Actually, its global plan is to manage to accomplish this task in kernel and user mode sides.
Certainly, readers can continue examining other routines by themselves.

99 | P a g e
https://exploitreversing.com

9. Further details about driver reversing

Analyzing drivers demands a good effort because they can contain multiple routines and, as expected, it
demands time. No doubts, when analyzing a system driver on Windows we have the offered public symbol
b Microoft and the function’ name are alread provided. The goal here is not analyze a driver, but only
interact with the first routines to show that everything we learned so far in this article is present and
readers can move forward by themselves without any serious issues.
I picked up the srv2.sys driver, which is the Smb2.0 Server driver (a network driver), which has been
updated very often in the last months, and a few of them due to security issues. Opening it on IDA Pro and
making a complete decompilation (File → Produce File → Create C File), the routine shown as entry point
will be GsDriverEntry, which is automatically generated when the driver was compiled and initialize the
security cookie, calls the DriverEntry at its end:

[Figure 109] srv2.sys: GsDriverEntry( )

Going inside DriverEntry( ), we have the following:

100 | P a g e
https://exploitreversing.com

[Figure 110] srv2.sys: DriverEntry( ) (truncated)

There is nothing really new in the DriverEntry routine above, but considerations follow below:
▪ From lines 11 to 30, the driver handles with WPP (Windows software trace preprocessor) aspects
aiming to establish a tracing (a logging capability that is similar to Windows event logging services )
of the operation, which is really useful during debugging sessions and, additionally, it offers the
possibility to publish events to ETW (Event Tracing for Windows). We are not interested in this
part of the driver, so we can skip it.
▪ From line 31 onwards, variables have been renamed.

101 | P a g e
https://exploitreversing.com

▪ Macros (M hotkey) have been applied to IoCreateDevice routine and also to major functions from
lines 65 to 69.
▪ A device object (network device) has been created by IoCreateDevice routine, and its name is
\Device\Srv2.
▪ The IoGetCurrentProcess function is called, and it returns a pointer to the current process.
▪ The DriverObject’s dispatc table contains pointers to four dispatch routines: cleanup
(Srv2Cleanup), close (Srv2Close), create (Srv2Create) and device control (Srv2DeviceControl).
▪ As usual and recommended, there is a DriverUnload routine to unload the driver.
We could examine the drivers and, as usual, the DispatchDeviceControl dispatch routine
(Srv2DeviceControl) is always a good starting point. I will not do it here because it is not the purpose of the
article analyze any kernel or filesystem driver in particular, but helping readers to learn about them and
respective techniques involved in the procedure.
Unfortunately, when reversing drivers that we do not have their symbols in hands, the task is harder and,
as a consequence, it might take an extended time to be finished. Readers can pick up any non-Microsoft
driver from their system during this example exercise. There are multiple applications to list drivers and
respective details from a running system, and readers could use applications such as driverquery (from
Windows: https://learn.microsoft.com/en-us/windows-server/administration/windows-
commands/driverquery) and DriverView (from Nirsoft: https://www.nirsoft.net/utils/driverview.html) that
are very simple. In my case I picked up the veracrypt.sys driver just to show the meaningful difference
between both examples (with and without debugging symbols):

102 | P a g e
https://exploitreversing.com

103 | P a g e
https://exploitreversing.com

[Figure 112] veracrypt.sys: DriverEntry( )

As readers already noticed, it will need to interpret the code and apply macros to improve it a bit, but it
was already expected. Anyway, everything we have learned in the previous section will be useful to get a
better understanding of the code.

104 | P a g e
https://exploitreversing.com

To avoid extending this article, I will be using an IDA Pro plugin named DriverBuddyReloaded
(https://github.com/VoidSec/DriverBuddyReloaded) to decode the IOCTL:

105 | P a g e
https://exploitreversing.com

106 | P a g e
https://exploitreversing.com

107 | P a g e
https://exploitreversing.com

[Figure 113] veracrypt.sys: improved DriverEntry( )

The output from DriverBuddyReloaded shows the decoding of every IOCTL found over the code:

[Figure 114] DriverBuddyReloaded’s output

Pay attention to hotkeys such as CTRL+ALT+F to decode all IOCTLs within a function; CTRL+ALT+A to start
auto-analysis and CTRL+ALT+D to decode a single IOCTL code. They can help you a lot.
I have done a quick marking up on the first routine (DriverEntry), created a structure (line 93), applied
macros (M hotkey) and created an enumeration containing all IOCTL names and their respective values.
At this point, all function invocations could be normally analyzed because that is legit driver, non-malicious,
and it follow and ue the ame concept I’ve hown over thi article. Nonethele, it would not be very
productive and would only make the article bigger.
I tried to provide the necessary basic foundation to the kernel drivers, minifilter drivers and WFP
(Windows Filtering Platform), without delving into too many programming details. It will be useful for
readers in my next articles.

108 | P a g e
https://exploitreversing.com

10. Recommended Blogs and Websites

There are excellent cyber security researchers and companies keeping blogs and writing really good articles
about operating system internals, reverse engineering, vulnerability research and exploit development. A
list of interesting websites and respective Twitter handles, in alphabetical order, follows below:

▪ https://hasherezade.github.io/articles.html (by Aleksandra Doniec: @hasherezade)

▪ https://malwareunicorn.org/#/workshops (by Amanda Rousseau: @malwareunicorn)
▪ https://captmeelo.com/ (by Capt. Meelo: @CaptMeelo)
▪ https://csandker.io/ (by Carsten Sandker: @0xcsandker)
▪ https://chuongdong.com/ (by Chuong Dong: @cPeterr)
▪ https://doar-e.github.io/ (Diary of a reverse-engineer)
▪ https://elis531989.medium.com/ (by Eli Salem: @elisalem9)
▪ http://0xeb.net/ (by Elias Bachaalany: @0xeb)
▪ https://googleprojectzero.blogspot.com/ (Google Project Zero)
▪ https://www.hexacorn.com/index.html (@Hexacorn)
▪ https://hex-rays.com/blog/ (by Hex-Rays: @HexRaysSA)
▪ https://github.com/Dump-GUY/Malware-analysis-and-Reverse-engineering (by Jiří Vinopal
@vinopaljiri)
▪ https://kienmanowar.wordpress.com/ (by Kien Tran Trung: @kienbigmummy)
▪ https://www.inversecos.com/ (by Lina Lau: @inversecos)
▪ https://maldroid.github.io/ (Łukas Siewierski @maldr0id)
▪ https://github.com/mnrkbys (by Minoru Kobayashi: @unkn0wnbit)
▪ https://voidsec.com/member/voidsec/ (by Paolo Stagno: @Void_Sec)
▪ https://www.youtube.com/@OffByOneSecurity (by Stephen Sims: @Steph3nSims)
▪ https://windows-internals.com/author/yarden/ (by Yarden Shafir @yarden_shafir)

11. Conclusion
This article, as I said at its beginning, is really an introduction to a complex topic that are kernel drivers and
minifilter drivers. The objective is to help professionals to get a minimal knowledge about involved
concepts, and provide the necessary foundation for the next articles.
Nowadays I have been working in a different area today (reversing + exploit development), but I always like
to remember closer researchers that each person has a unique perspective of the information ecurit’
world, and none of them are wrong. Follow your heart. :)
Just in case you want to stay connected:
▪ Twitter: @ale_sp_brazil
▪ Blog: https://exploitreversing.com
Keep reversing and I see you at next time!
Alexandre Borges

109 | P a g e