In this guest blog from researcher Marcin Wiązowski, he details CVE-2023-21822 – a Use-After-Free (UAF) in win32kfull that could lead to a privilege escalation. The bug was reported through the ZDI program and later patched by Microsoft. Marcin has graciously provided this detailed write-up of the vulnerability, examines how it could be exploited, and a look at the patch Microsoft released to address the bug.

In the Windows kernel, there are three APIs intended for general use by device drivers for the purpose of creating bitmaps: EngCreateBitmap, EngCreateDeviceBitmap and EngCreateDeviceSurface. Each of these APIs return a bitmap handle. If the caller wants to perform some drawing operations on the bitmap, the caller must first lock the bitmap by passing its handle to EngLockSurface. EngLockSurface increases the bitmap’s reference counter and returns a pointer to a corresponding SURFOBJ record. SURFOBJ is a structure located in kernel memory containing all the information regarding the bitmap, such as the bitmap’s dimensions, pixel format, a pointer to the pixel buffer, and so forth. We’ll take a closer look at the SURFOBJ structure later. After calling EngLockSurface, the obtained SURFOBJ pointer can be passed to various drawing APIs such as EngLineTo and EngBitBlt. See winddi.h for the complete list of these drawing APIs. After the caller is finished with drawing operations, they should call EngUnlockSurface. At this point, the bitmap’s reference counter decreases to zero again, and the caller is no longer allowed to use the SURFOBJ pointer. Finally, the caller can delete the bitmap by calling EngDeleteSurface on its handle. Typical usage of these APIs is shown below:

All APIs discussed above are exported from win32k.sys kernel-mode module. Note, though, that the functions in win32k.sys are only wrappers, and the implementations are in win32kbase.sys and win32kfull.sys.

Many years ago, both display drivers and printer drivers worked in kernel mode, but since Windows Vista, printer drivers work only in user mode (hence User-Mode Printer Drivers, or UMPD). Two important facts emerge from this change:
       -- During printing operations, the kernel must now perform some callbacks to user mode to call the appropriate user-mode printer driver.
       -- To allow printer driver code to run in user mode, some kernel APIs have now been made available from user mode.

As a result, all the kernel APIs described above now have user-mode counterparts, exported from the gdi32.dll user-mode module. Let’s try to execute the same code shown above, but, this time, from user mode:

Note the reference counter values shown in the comments. The value is still zero after locking the bitmap. Why is this?

Kernel-mode code is always trusted, while user-mode code is always untrusted. So, now that printer drivers execute in user mode, they are considered untrusted and potentially malicious.

Suppose that the user-mode EngLockSurface call would increase the bitmap’s reference counter in the same way that the kernel-mode version does. An attacker, acting as a user-mode printer driver, could call EngLockSurface many times in a loop on a bitmap in order to overflow the bitmap’s reference counter, causing it to wrap around to zero. Then the bitmap could be deleted, leading to a use-after-free on the bitmap.

For this reason, the Windows kernel has implemented a different approach. The EngLockSurface API is expected to return a pointer to the bitmap’s SURFOBJ record – and it does, but, in user mode, this is a user-mode copy of the “true”, kernel-mode SURFOBJ record. We can reconstruct this user-mode data structure as follows:

The user-mode EngLockSurface implementation returns a pointer to the UMSO.so field, which is a copy of the true, kernel-mode SURFOBJ record, so that everything will work as expected. Internally, the user-mode EngLockSurface call jumps to its kernel-mode implementation win32kfull.sys!NtGdiEngLockSurface, where the user-mode UMSO record is allocated and filled in. In kernel mode, the “true”, kernel-mode EngLockSurface call is made on the bitmap, which is needed to access the bitmap’s SURFOBJ record so its data can be copied into the UMSO.so field. Afterwards, though, NtGdiEngLockSurface calls the kernel-mode EngUnlockSurface, which decreases the bitmap’s reference counter to zero again. This explains the observed reference counter values.

Once we call the user-mode EngLockSurface, we are allowed to pass its result (which is a pointer to the copied SURFOBJ data) to various drawing functions, such as EngLineTo or EngBitBlt. When corresponding calls are made from kernel mode, it works in a straightforward manner, but when calling from user mode, an additional layer is needed to translate the user-mode SURFOBJ pointers into true, kernel-mode pointers. So, for example, if the user-mode code calls gdi32.dll!EngLineTo, this will jump to the kernel-mode win32kfull.sys!NtGdiEngLineTo wrapper. The wrapper will obtain the bitmap’s true kernel-mode SURFOBJ record, so the kernel-mode win32kfull.sys!EngLineTo drawing handler ultimately can be executed.

How does the kernel obtain the needed kernel-mode SURFOBJ record? A SURFOBJ record contains sensitive data such as the bitmap’s pixel buffer pointer, so the kernel never relies on the contents of SURFOBJ records coming from user mode. Otherwise, there would be a security risk from malicious user-mode code that tampers with the contents of UMSO.so structures. Instead, inside the wrapper function (such as win32kfull.sys!NtGdiEngLineTo in the example above), the kernel verifies the UMSO.magic value, and then uses the UMSO.hsurf bitmap handle value to lock the bitmap by calling EngLockSurface. In this way, the kernel safely obtains the requested bitmap’s kernel-mode SURFOBJ record, which it can then pass to the appropriate kernel-mode win32kfull.sys!EngXXX drawing function.

The Vulnerability

The user-mode EngLockSurface function performs some validation on the supplied bitmap handle, meaning that not every kind of bitmap can be passed successfully to this call (we will discuss this in more detail later). But malicious user-mode code can now bypass this in any of these ways:

1) After making the EngLockSurface call, we can delete the already-validated bitmap and create some other bitmap with the same handle value. We could choose to create a bitmap of a kind that couldn’t be successfully passed to EngLockSurface.

2) After making the EngLockSurface call, we receive a pointer to a user-mode SURFOBJ record, which, as we already know, is a part of a UMSO record. So, we can overwrite the UMSO.hsurf field, setting it to the handle of any bitmap that we want. We can choose to set it to the handle of a bitmap that couldn’t be successfully passed to EngLockSurface.

3) Simplest of all, we could prepare a UMSO record from scratch, without making any EngLockSurface call first. All we need to do is allocate some user-mode memory, set UMSO.magic to 0x554D534F, and set UMSO.hsurf to the handle of a bitmap of our choice. The remaining part of this record (the UMSO.so field, containing the SURFOBJ record under normal circumstances) can be zeroed, as it will be disregarded by the kernel in any event.

Each of the three possibilities above will allow us to bypass the bitmap validation performed by the user-mode version of the EngLockSurface API.

Now that we have seen that the validation can be bypassed, we must ask what is the purpose of that validation, and what ramifications does it have for security? To answer this question, we must look at the SURFOBJ record definition. Some fields are publicly documented, while others have been reconstructed, as shown below:

The bitmap’s flags field is undocumented, but it is known to contain some documented HOOK_XXX flags found in the winddi.h header file. These flags tell the win32k subsystem which drawing operations should be handled by win32k itself, and which should instead be directed to a specialized device driver. The device driver is indicated by the bitmap’s hdev field.

For example, suppose we want to draw a line on some bitmap. We’ll call EngLineTo, passing a pointer to the bitmap’s SURFOBJ record. Internally, the kernel will convert the requested line into a more general drawing construct known as a “path” (which can be a sequence of lines and curves). It will then check if the bitmap’s SURFOBJ.flags field has the HOOK_STROKEPATH flag set. If this flag is not present, it will use the generic code for drawing (“stroking”) paths provided by win32kfull. If HOOK_STROKEPATH is present, though, the kernel will direct the drawing request to the device driver specified by the SURFOBJ.hdev field. The latter case, where possible, offers improved performance, as it allows individual device drivers to take advantage of accelerations offered by the specific hardware. For example, a graphics adapter may offer hardware-accelerated path stroking. Similarly, printer devices have specialized acceleration for outputting text.

So, if we prepare a bitmap that has a screen-related SURFOBJ.hdev value, and also has the appropriate HOOK_XXX flag set, and we pass it to one of the EngXXX drawing APIs, there is the possibility of reaching an entry point of a specialized display driver, working in kernel mode. This could be cdd.dll!DrvXXX in the single-monitor case or win32kfull.sys!MulXXX in the multi-monitor case (though, there is not always a simple relationship between the requested functionality and the driver entry point ultimately called, as noted in the example above). The pointer to the bitmap’s SURFOBJ record will be passed as a parameter to the driver’s entry point.

Further note that some EngXXX APIs take not only one bitmap as a parameter, but rather two: a source bitmap and a destination bitmap. (Some optionally also take a mask bitmap, but that is not interesting for us). An example of such an API is EngBitBlt, which copies a rectangle of pixels from a source bitmap to a destination bitmap. APIs that work on two bitmaps use the SURFOBJ.flags and SURFOBJ.hdev values of the destination bitmap when determining the ultimate device driver to receive the call. Nonetheless, when the final driver’s entry point is called, both the source and destination bitmaps are passed to it.

Hence, a properly prepared, screen-related bitmap, when passed to some EngXXX API as the destination bitmap, allows us to reach a kernel-mode display driver, while also allowing an arbitrary bitmap of our choice to be passed as the source bitmap.

There is still no obvious security problem here, but let’s look at the SURFOBJ record definition once again. It contains a dhsurf field (not to be confused with the hsurf field discussed above). The win32k subsystem treats SURFOBJ.dhsurf as an opaque value. It is reserved for individual device drivers to use for their internal purposes. Setting this field on a new bitmap is easy: the EngCreateDeviceBitmap and EngCreateDeviceSurface bitmap creation APIs just take the dhsurf value as a parameter. Both the Canonical Display Driver (cdd.dll, used for single-monitor graphics output) and the multi-display driver (win32kfull.sys!MulXXX) expect to work only with their own bitmaps – bitmaps with SURFOBJ.dhsurf values set by that specific driver – rather than on arbitrary bitmaps created from user mode (or by other drivers). Internally, each of these drivers use the SURFOBJ.dhsurf value as a pointer to a block of kernel-mode memory, containing private data owned by that driver.

But we can reach a kernel-mode display driver by passing a properly prepared, destination bitmap to the EngXXX call, and we can also pass some arbitrary bitmap of our choice as the source bitmap to the same EngXXX call. This source bitmap can be an arbitrary bitmap we created, and its SURFOBJ.dhsurf value may point to arbitrary controllable memory. The kernel-mode display driver, such as the Canonical Display Driver, will work on this block of memory as if it were its own block of kernel-mode memory. This means “game over”.

For these reasons, the user-mode EngLockSurface implementation has validation to reject screen-related bitmaps that could be used to reach a kernel-mode display driver. But, thanks to the vulnerability described above, we can bypass this EngLockSurface validation easily. In fact, we can get away with not calling EngLockSurface at all, and just preparing the needed UMSO record from scratch instead, as we have explained.

Exploitation

We must first notice that user-mode EngXXX calls are intended to be used by user-mode printer drivers only, so most of these APIs will fail unless they are called during a callback from kernel to user-mode for a printing operation. But this doesn’t complicate things too much: the user-mode part of the callback is implemented as a gdi32.dll!GdiPrinterThunk function, which is a public export from gdi32.dll. It’s enough to hook or patch this function and perform our main exploitation there. This function receives four parameters (the input buffer, the input buffer size, the output buffer, and the output buffer size), but we don’t need the parameters during our exploitation at all. (However, if you are interested in more details, see Selecting Bitmaps into Mismatched Device Contexts. In particular, see sections titled “User-Mode Printer Drivers (UMPD)” and “Hooking the UMPD implementation”.)

We first need to get a callback from the kernel to our hooked gdi32.dll!GdiPrinterThunk function. To achieve this, we need to initiate some printing operation. First we must locate an installed printer. There is at least one virtual printer installed by default on every Windows machine. We can locate installed printers using a call to the user-mode winspool.drv!EnumPrintersA/W API. Then we must create a printer-related device context:

This call will go down to kernel mode, which will then perform several callbacks to user mode again – so our hooked gdi32.dll!GdiPrinterThunk function will be invoked, exactly as we need. Our main exploitation phase starts here.

First, we need to obtain a bitmap with a screen-related SURFOBJ.hdev value and a useful HOOK_XXX flag set in its SURFOBJ.flags field. To obtain such a bitmap, we can create a window with proper parameters, obtain the window’s device context, and grab the underlying bitmap. The obtained bitmap will act as our destination bitmap:

We also need a source bitmap, with its SURFOBJ.dhsurf field pointing to controlled user-mode memory (our FakeDhsurfBlock):

Now we can prepare two UMSO records, one for the destination bitmap and one for the source bitmap:

At this point, we have everything that we need to make a malicious EngXXX call with our bitmaps. Our screen-related, destination bitmap will have all the defined HOOK_XXX flags set, so we are free to choose any of the EngXXX APIs that accept two bitmaps:

Through reverse engineering the Canonical Display Driver or multi-display driver internals, we can learn how to prepare the user-mode FakeDhsurfBlock so that the call to the display driver yields exploitable memory primitives.

The Patch

As discussed earlier, each of the user-mode EngXXX drawing APIs (such as EngLineTo and EngBitBlt) calls its corresponding kernel-mode win32kfull.sys!NtGdiEngXXX wrapper, where, amongst other things, user-mode SURFOBJ pointers are converted to kernel-mode SURFOBJ pointers. Afterwards, a kernel-mode win32kfull.sys!EngXXX driver endpoint is called to perform the requested drawing operation.

Although it’s not related to our vulnerability, it’s worth mentioning that, for the duration of the gdi32.dll!GdiPrinterThunk user-mode callback, the kernel maintains a mapping of known user-mode SURFOBJ records to kernel-mode SURFOBJ records. When a user-mode printer driver passes a user-mode SURFOBJ pointer to some user-mode EngXXX call, the kernel will try to use the mapping to find the corresponding kernel-mode SURFOBJ pointer so it can be passed to the corresponding kernel-mode EngXXX call.

The mapping is prepared before the user-mode GdiPrinterThunk callback begins. This is because some bitmaps may be passed to the callback as parameters (though, during our exploitation, we made no use of the GdiPrinterThunk input data). However, this means that bitmaps “locked” later, that is, by calls to EngLockSurface made from inside the callback, are not present in the mapping.

Whenever some win32kfull.sys!NtGdiEngXXX receives a user-mode SURFOBJ pointer as a parameter and is not able to find it in the mapping, it assumes that the received SURFOBJ record is contained in an UMSO record (as its UMSO.so field).

Before the patch, such cases were directed to the internal win32kfull.sys!UMPDSURFOBJ::GetLockedSURFOBJ function, where the UMSO.magic value would be verified against the 0x554D534F value, and then the kernel-mode EngLockSurface call would be made on the UMSO.hsurf handle value, yielding the needed pointer to the “true”, kernel-mode SURFOBJ record, as discussed earlier.

As you may have noticed, the name GetLockedSURFOBJ is misleading, as it suggests that the bitmap is already locked. In reality, when coming from user mode, a bitmap’s reference counter is still zero. And as we saw above, a malicious user-mode printer driver may not have called EngLockSurface at all, but instead just prepared the needed UMSO record from scratch.

After the patch, the function name was changed to GetLockableSURFOBJ. A user-mode printer driver can still perform all the manipulations described above, but now GetLockableSURFOBJ considers the received bitmap handle (UMSO.hsurf) as untrusted. After using the UMSO.hsurf value to lock the bitmap in kernel mode, GetLockableSURFOBJ now once again performs the same bitmap validation that is performed when calling the user-mode EngLockSurface API. This validation is performed by calling win32kfull.sys!IsSurfaceLockable. In this way, screen-related bitmaps that could be used to reach the kernel-mode display driver from within the user-mode printer driver are now rejected by GetLockableSURFOBJ.

Thanks again to Marcin for providing this thorough write-up. He has contributed multiple bugs to the ZDI program over the last few years, and we certainly hope to see more submissions from them in the future. Until then, follow the team on Twitter, Mastodon, LinkedIn, or Instagram for the latest in exploit techniques and security patches.