The vulnerability resides in xxxTrackPopupMenuEx, this function is responsible for displaying shortcut menus and tracking user selections. During this process it will try to get a reference to the GlobalMenuStateobject via a call to xxxMNAllocMenuState, if the object is in use, for example: when another pop-up menu is already active, this function will try to create a new instance.
If xxxMNAllocMenuState fails it will return False but it will also set the pGlobalMenuState thread global variable to NULL. The caller verifies the return value, and in case of failure it will try to do some cleanup in order to fail gracefully.
During this cleanup the xxxEndMenuState procedure is called. This function’s main responsibility is to free and unlock all the resources acquired and saved for the MenuState object, but it does not check that the pGlobalMenuStatevariable is not NULL before using it. As a result a bunch of kernel operations are performed on a kernel object whose address is zero and thus potentially controlled from userland memory on platforms that allow it.
Triggering the vulnerability is relatively easy by just creating and displaying two popup instances and exhausting GDI objects for the current session, as explained by Endgame. However, actually getting code execution is not trivial.
Usually a NULL dereference vulnerability in the kernel can be exploited by mapping memory at address zero in userland memory (when allowed by the OS), creating a fake object inside of this null page and then triggering the vulnerability in the kernel from the current process context of your exploit which has the null page mapped with attacker controlled data. With some luck we get a function pointer of some sort called from our controlled object data and we achieve code execution with Kernel privileges (e.g. this was the case of MS11-054). As such, NULL dereference vulnerabilities have for many years provided a simple and straightforward route to kernel exploitation and privilege escalation in scenarios where you are allowed to map at zero.
Unfortunately in the case of CVE-2013-3881 life is not that simple, even on platforms that allow the null page to be allocated.
When xxxTrackPopupMenuExcalls xxxMNAllocMenuStateand fails, it will directly jump to destroy the (non-existant) MenuState object, and after some function calls, it will inevitably try to free the memory. This means that it does not matter if we create a perfectly valid object at region zero. At some pointbefore xxxEndMenuState returns, a call to ExFreePoolWithTag(0x0, tag) will be made. This call will produce a system crash as it tries to access the pool headers which are normally located just before the poolAddress which in this case is at address 0. Thus the kernel tries to fetch at 0-minus something which is unallocated and/or uncontrolled memory and we trigger a BSOD.
This means the only viable exploitation option is to try and get code execution before this Free occurs.
At this point we try to understand the entire behavior of xxxEndMenuState, and all of the structures and objects being manipulated before we trigger any fatal crash. The main structure we have to deal with is the one that is being read from address zero, which is referenced from the pGlobalMenuState variable:
+0x000 pGlobalPopupMenu : Ptr32 tagPOPUPMENU
+0x004 fMenuStarted : Pos 0, 1 Bit
+0x004 fIsSysMenu : Pos 1, 1 Bit
+0x004 fInsideMenuLoop : Pos 2, 1 Bit
+0x004 fButtonDown : Pos 3, 1 Bit
+0x004 fInEndMenu : Pos 4, 1 Bit
+0x004 fUnderline : Pos 5, 1 Bit
+0x004 fButtonAlwaysDown : Pos 6, 1 Bit
+0x004 fDragging : Pos 7, 1 Bit
+0x004 fModelessMenu : Pos 8, 1 Bit
+0x004 fInCallHandleMenuMessages : Pos 9, 1 Bit
+0x004 fDragAndDrop : Pos 10, 1 Bit
+0x004 fAutoDismiss : Pos 11, 1 Bit
+0x004 fAboutToAutoDismiss : Pos 12, 1 Bit
+0x004 fIgnoreButtonUp : Pos 13, 1 Bit
+0x004 fMouseOffMenu : Pos 14, 1 Bit
+0x004 fInDoDragDrop : Pos 15, 1 Bit
+0x004 fActiveNoForeground : Pos 16, 1 Bit
+0x004 fNotifyByPos : Pos 17, 1 Bit
+0x004 fSetCapture : Pos 18, 1 Bit
+0x004 iAniDropDir : Pos 19, 5 Bits
+0x008 ptMouseLast : tagPOINT
+0x010 mnFocus : Int4B
+0x014 cmdLast : Int4B
+0x018 ptiMenuStateOwner : Ptr32 tagTHREADINFO
+0x01c dwLockCount : Uint4B
+0x020 pmnsPrev : Ptr32 tagMENUSTATE
+0x024 ptButtonDown : tagPOINT
+0x02c uButtonDownHitArea : Uint4B
+0x030 uButtonDownIndex : Uint4B
+0x034 vkButtonDown : Int4B
+0x038 uDraggingHitArea : Uint4B
+0x03c uDraggingIndex : Uint4B
+0x040 uDraggingFlags : Uint4B
+0x044 hdcWndAni : Ptr32 HDC__
+0x048 dwAniStartTime : Uint4B
+0x04c ixAni : Int4B
+0x050 iyAni : Int4B
+0x054 cxAni : Int4B
+0x058 cyAni : Int4B
+0x05c hbmAni : Ptr32 HBITMAP__
+0x060 hdcAni : Ptr32 HDC__
This is the main object which xxxEndMenuState will deal with, it will perform a couple of actions using the object and finally attempts to free it with the call to ExFreePoolWithTag. The interaction with the object that occurs prior to the free are the ones we have to analyze deeply as they are our only hope in getting code execution before the imminent crash.
xxxEndMenuState is a destructor, and as such it will first call the destructor of all the objects contained inside the main object before actually freeing their associated allocated memory, for example:
The _MNFreePopup call is very interesting, as PopupMenu objects contain several WND objects and these have Handle references. This is relevant because if this WND object has its lock count equal to one when MNFreePopupis called, at some point it will try to destroy the object that the Handle is referencing. These objects are global to a user session. This means that we can force the deletion of any object within the current windows session, or at the very least decrement its reference count.
+0x000 fIsMenuBar : Pos 0, 1 Bit
+0x000 fHasMenuBar : Pos 1, 1 Bit
+0x000 fIsSysMenu : Pos 2, 1 Bit
+0x000 fIsTrackPopup : Pos 3, 1 Bit
+0x000 fDroppedLeft : Pos 4, 1 Bit
+0x000 fHierarchyDropped : Pos 5, 1 Bit
+0x000 fRightButton : Pos 6, 1 Bit
+0x000 fToggle : Pos 7, 1 Bit
+0x000 fSynchronous : Pos 8, 1 Bit
+0x000 fFirstClick : Pos 9, 1 Bit
+0x000 fDropNextPopup : Pos 10, 1 Bit
+0x000 fNoNotify : Pos 11, 1 Bit
+0x000 fAboutToHide : Pos 12, 1 Bit
+0x000 fShowTimer : Pos 13, 1 Bit
+0x000 fHideTimer : Pos 14, 1 Bit
+0x000 fDestroyed : Pos 15, 1 Bit
+0x000 fDelayedFree : Pos 16, 1 Bit
+0x000 fFlushDelayedFree : Pos 17, 1 Bit
+0x000 fFreed : Pos 18, 1 Bit
+0x000 fInCancel : Pos 19, 1 Bit
+0x000 fTrackMouseEvent : Pos 20, 1 Bit
+0x000 fSendUninit : Pos 21, 1 Bit
+0x000 fRtoL : Pos 22, 1 Bit
+0x000 iDropDir : Pos 23, 5 Bits
+0x000 fUseMonitorRect : Pos 28, 1 Bit
+0x004 spwndNotify : Ptr32 tagWND
+0x008 spwndPopupMenu : Ptr32 tagWND
+0x00c spwndNextPopup : Ptr32 tagWND
+0x010 spwndPrevPopup : Ptr32 tagWND
+0x014 spmenu : Ptr32 tagMENU
+0x018 spmenuAlternate : Ptr32 tagMENU
+0x01c spwndActivePopup : Ptr32 tagWND
+0x020 ppopupmenuRoot : Ptr32 tagPOPUPMENU
+0x024 ppmDelayedFree : Ptr32 tagPOPUPMENU
+0x028 posSelectedItem : Uint4B
+0x02c posDropped : Uint4B
In order to understand why this is so useful, let’s analyze what happens when a WND object is destroyed:
pWND __stdcall HMUnlockObject(pWND pWndObject)
pWND result = pWndObject;
result = HMUnlockObjectInternal(pWndObject);
The first thing done is a decrement of the cLockObj counter, and if the counter is then zero the function HMUnlockObjectInternal is called.
pWND __stdcall HMUnlockObjectInternal( pWND pWndObject)
result = pWndObject;
unsigned int entryIndex;
entryIndex = pWndObject->handle & 0xFFFF;
entry = gSharedInfo.aheList + gSharedInfo.HeEntrySize * entryIndex
if ( entry->bFlags & HANDLEF_DESTROY )
if ( !(entry->bFlags & HANDLEF_INDESTROY) )
result = 0;
Once it knows that the reference count has reached zero, it has to actually destroy the object. For this task it gets the handle value and applies a mask in order to get the index of the HandleEntry into the handle table.
Then it validates some state flags, and calls HMDestroyUnlockedObject.
The HandleEntry contains information about the object type and state. This information will be used to select between the different destructor functions.
int __stdcall HMDestroyUnlockedObject(pHandleEntry handleEntry)
index = 0xC * handleEntry->bType
handleEntry->bFlags |= HANDLEF_INDESTROY;
The handle type information table (gahti) holds properties specific to each object type, as well as their Destroy functions. So this function will use the bType value from the handleEntry in order to determine which Destroy function to call.
At this point it is important to remember that we have full control over the MenuState object, and that means we can create and fully control its inner PopupMenu object, and in turn the WND objects inside this PopupMenu. This implies that we have control over the handle value in the WND object.
Another important fact is that entry zero on the gahti table is always zero, and it represents the FREE object type.
So our strategy in order to get code execution here is to, by some means, create an object whose HandleEntry in the HandleEntry table has a bType = 0x0, and bFlags = 0x1. If we can manage to do so we can then create a fake WND object with a handle that makes reference to this object of bType=0x0. When the HMDestroyUnlockedObject is called it will end up in a call gahti[0x0]. As the first element in gahti table is zero, this ends up as a “call 0”. In other words we can force a path that will execute our controlled data at address zero.
What we need
We need to create a user object of bType=FREE (0x0) and bFlags= HANDLEF_DESTROY (0x1).
This is not possible directly, so we first focus on getting an object with the bFlag value equal to 0x1. For this purpose we create a Menu object, set it to a window, and then Destroy it. The internal reference count for the object did not reach zero because it is still being accessed by the window object, so it is not actually deleted but instead flagged as HANDLEF_DESTROY on the HandleEntry. This means the bFlag will equal to 0x1.
The bType value is directly associated to the Object Type. In the case of a menu object the value is 0x2 and there is no way of creating an object of type 0x0. So we focus on what ways we have to alter this value using some of the functions being called before destroying the WND object.
As you can probably remember from the PopupMenu structure shown before, it contains several WND objects, and one of the first actions performed when HMUnlockObject(pWnd) is called is to decrement the lockCount. So we simply set-up two fake WND objects in such a way that the lockCount field will be pointing to the HandleEntry->bType field. When each of those fake WND objects is destroyed it will actually perform a “dec” operation over the bType of our menu object, thus decrementing it from 0x2 to 0x0. We now have a bFlag of 0x1 and a bType of 0x0.
Using this little trick we are able to create a User object with the needed values on the HandleEntry.
First we will create a MenuObject and force it to be flagged as HANDLEF_DESTROY.
Then we will trigger the vulnerability, where xxxEndMenuState will get a reference to the menuState structure from a global thread pointer, and its value will be zero. So we map this address and create a fake MenuState structure at zero.
XxxEndMenuStatewill call FreePopup(..) on a popup object instance we created, and will in turn try to destroy its internal objects. Three of these objects will be fake WND objects which we also create. The first two will serve the purpose of decrementing the bType value of our menu object, and the third one will trigger a HMDestroyUnlockedObjecton this same object. This will result on code execution being redirected to address 0x0 as previously discussed.
We have to remember that when we redirect execution to address 0, this memory also servers as a MenuState object. In particular the first field is a pointer to the PopupMenu object that we need to use. So what we do is to choose the address of this popup menu object in such a way that the least significant bytes of the address also represent a valid X86 jump opcode (e.g. 0x04eb translates to eb 04 in little endian memory ordering which represents a jump 4).
Once we achieve execution at ring 0 we patch the Enabled field on the _SEP_TOKEN_PRIVILEGES structure from the MOSDEF callback process in order to enable all the privileges for the process. We fix up the HandleEntry we modified before, and restore the stack in order to return after the PoolFree thus skipping the BSOD.
Once all of this is done we return to user-land, but now our MOSDEF process has all the privileges, this allows us to for example migrate to LSASS and get System privileges.
— Matias Soler