Nowadays, exploitation of user-level vulnerabilities is becoming more and more difficult, because of the widespread diffusion of several protection methods, including ASLR, NX, various heap protections, stack canaries, and sandboxed execution. As a natural consequence, instead of extricating themselves with such a plethora of defensive methods, attackers prefer to take the “easy” way and started to move at the kernel-level, where sophisticated protection techniques are still not very common (indeed, things like as KASLR and SMEP are implemented only in the latest versions of the most popular OSes). This trend is also confirmed by the rising number of kernel-level vulnerabilities reported in the last few months in Windows, Linux, and OS X.
Following this trend, we recently looked at few OS X drivers (“KEXT”s) and found a integer signedness bug affecting service IOBluetoothHCIController (implemented by the IOBluetoothFamily KEXT). This vulnerability can be exploited by a local attacker to gain root privileges. The issue is present on the latest versions of OS X Mavericks (tested on 10.9.4 and 10.9.5), but has been “silently” patched by Apple in OS X Yosemite.
Vulnerability overview
In a nutshell, the bug lies in the IOBluetoothHCIUserClient::SimpleDispatchWL() function. The function eventually takes a user-supplied 32-bit signed integer value and uses it to index a global array of structures containing a function pointer. The chosen function pointer is finally called. As the reader can easily imagine, SimpleDispatchWL() fails at properly sanitizing the user-supplied index, thus bad things may happen if a malicious user is able to control the chosen function pointer.
More in detail, the vulnerable part of the function is summarized in the pseudocode below. At line 14, the user-supplied 32-bit integer is casted to a 64-bit value. Then, the "if" statement at line 16 returns an error if the casted (signed) value is greater than the number of methods available in the global _sRoutines array; obviously, due to the signed comparison, any negative value for the method_index variable will pass this test. At line 20 method_index is used to access the _sRoutines array, and the retrieved callback is finally called at line 23.
More in detail, the vulnerable part of the function is summarized in the pseudocode below. At line 14, the user-supplied 32-bit integer is casted to a 64-bit value. Then, the "if" statement at line 16 returns an error if the casted (signed) value is greater than the number of methods available in the global _sRoutines array; obviously, due to the signed comparison, any negative value for the method_index variable will pass this test. At line 20 method_index is used to access the _sRoutines array, and the retrieved callback is finally called at line 23.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 | typedef struct { void (*function_pointer)(); uint64 num_arguments; } BluetoothMethod; BluetoothMethod _sRoutines[] = { ... }; uint64 _sRoutineCount = sizeof(_sRoutines)/sizeof(BluetoothMethod); IOReturn IOBluetoothHCIUserClient::SimpleDispatchWL(IOBluetoothHCIDispatchParams *params) { // Here "user_param" is a signed 32-bit integer parameter int64 method_index = (int64) user_param; if (method_index >= _sRoutineCount) { return kIOReturnUnsupported; } BluetoothMethod method = _sRoutines[method_index]; ... if (method.num_arguments < 8) { method.function_pointer(...); } ... } |
Exploitation details
Exploitation of this vulnerability is just a matter of supplying the proper negative integer value in order to make IOBluetoothFamily index the global _sRoutines structure out of its bounds, and to fetch an attacker-controlled structure. The supplied value must be negative to index outside the _sRoutines structure while still satisfying the check at line 16.
As a foreword, consider that for our "proof-of-concept" we disabled both SMEP/SMAP and KASLR, so some additional voodoo tricks are required to get a fully weaponized exploit. Thus, our approach was actually very simple: we computed a value for the user-supplied parameter that allowed us to index a BluetoothMethod structure such that BluetoothMethod.function_ptr is a valid user-space address (where we placed our shellcode), while BluetoothMethod.num_arguments is an integer value less than 8 (to satisfy the check performed by SimpleDispatchWL() at line 22).
As shown in the C code fragment above, the user-supplied 32-bit value (user_param) is first casted to a 64-bit signed value, and then used as an index in _sRoutines. Each entry of the global _sRoutines array is 16-byte wide (two 8-byte values). These operations are implemented by the following assembly code:
As a foreword, consider that for our "proof-of-concept" we disabled both SMEP/SMAP and KASLR, so some additional voodoo tricks are required to get a fully weaponized exploit. Thus, our approach was actually very simple: we computed a value for the user-supplied parameter that allowed us to index a BluetoothMethod structure such that BluetoothMethod.function_ptr is a valid user-space address (where we placed our shellcode), while BluetoothMethod.num_arguments is an integer value less than 8 (to satisfy the check performed by SimpleDispatchWL() at line 22).
As shown in the C code fragment above, the user-supplied 32-bit value (user_param) is first casted to a 64-bit signed value, and then used as an index in _sRoutines. Each entry of the global _sRoutines array is 16-byte wide (two 8-byte values). These operations are implemented by the following assembly code:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 | ; r12+70h points to the user-supplied index value mov ecx, [r12+70h] mov r13d, kIOReturnUnsupported lea rdx, _sRoutineCount cmp ecx, [rdx] jge fail ; Go on and fetch _sRoutine[method_index] ... movsxd rax, ecx ; Sign extension to 64-bit value shl rax, 4 ; method_index *= sizeof(BluetoothMethod) lea rdx, _sRoutines mov esi, [rdx+rax+8] ; esi = _sRoutines[method_index].num_arguments cmp esi, 7 ; Check method.num_arguments < 8 ja loc_289BA ... |
At a higher-level, the address of the BluetoothMethod structure fetched when processing an index value "user_param" is computed by the following formula:
struct_addr = (ext(user_param & 0xffffffff) * 16) + _sRoutine
Where ext() is the sign-extension operation (implemented by the movsxd instruction in the assembly code snipped above).
By solving this formula for user_param and searching inside the kernel address space, we found several candidate addresses that matched our criteria (i.e., a valid user-space pointer followed by an integer value < 8). The rest of the exploit is just a matter of mmap()'ing the shellcode at the proper user-space address, connecting to the IOBluetoothHCIController service and invoking the vulnerable method.
The source code for a (very rough) proof-of-concept implementation of the aforementioned exploit is available here, while the following figure shows the exploit "in action".
A side-by-side comparison between versions 10.9.x and 10.10 of IOBluetoothFamily confirms Apple has patched the device driver by rejecting negative values for the user-supplied index. In the figure below, the user-supplied index value is compared against _sRoutineCount (orange basic block). Yosemite adds an additional check to ensure the (signed) index value is non-negative (green basic block, on the right).
By solving this formula for user_param and searching inside the kernel address space, we found several candidate addresses that matched our criteria (i.e., a valid user-space pointer followed by an integer value < 8). The rest of the exploit is just a matter of mmap()'ing the shellcode at the proper user-space address, connecting to the IOBluetoothHCIController service and invoking the vulnerable method.
The source code for a (very rough) proof-of-concept implementation of the aforementioned exploit is available here, while the following figure shows the exploit "in action".
 |
| Execution of our "proof-of-concept" exploit |
Patching
We verified the security issue both on OS X Mavericks 10.9.4 and 10.9.5 (MD5 hash values for the IOBluetoothFamily KEXT bundle on these two OS versions are 2a55b7dac51e3b546455113505b25e75 and b7411f9d80bfeab47f3eaff3c36e128f, respectively). After the release of OS X Yosemite (10.10), we noticed the vulnerability has been silently patched by Apple, with no mention about it in the security change log.A side-by-side comparison between versions 10.9.x and 10.10 of IOBluetoothFamily confirms Apple has patched the device driver by rejecting negative values for the user-supplied index. In the figure below, the user-supplied index value is compared against _sRoutineCount (orange basic block). Yosemite adds an additional check to ensure the (signed) index value is non-negative (green basic block, on the right).
 |
| Comparison of the vulnerable OS X driver (Mavericks, on the left) and patched version (Yosemite, on the right) |
Conclusions
We contacted Apple on October 20th, 2014, asking for their intention to back-port the security fix to OS X Mavericks. Unfortunately, we got no reply, so we decided to publicly disclose the details of this vulnerability: Yosemite has now been released since a while and is available for free for Apple customers; thus, we don’t think the public disclosure of this bug could endanger end-users.