macOS Kernel 10.12.2 (16C67) - Memory Disclosure Due to Lack of Bounds Checking in AppleIntelCapriController::getDisplayPipeCapability

EDB-ID: 41798
Author: Google Security Research
Published: 2017-04-04
CVE: CVE-2017-2489
Type: Dos
Platform: macOS
Aliases: N/A
Advisory/Source: Link
Tags: Denial of Service (DoS)
Vulnerable App: N/A

 Source: https://bugs.chromium.org/p/project-zero/issues/detail?id=1069 

MacOS kernel memory disclosure due to lack of bounds checking in AppleIntelCapriController::getDisplayPipeCapability

Selector 0x710 of IntelFBClientControl ends up in AppleIntelCapriController::getDisplayPipeCapability.

This method takes a structure input and output buffer. It reads an attacker controlled dword from the input buffer which it
uses to index an array of pointers with no bounds checking:

AppleIntelCapriController::getDisplayPipeCapability(AGDCFBGetDisplayCapability_t *, AGDCFBGetDisplayCapability_t *)
__text:000000000002A3AB mov r14, rdx ; output buffer, readable from userspace
__text:000000000002A3AE mov rbx, rsi ; input buffer, controlled from userspace
...
__text:000000000002A3B8 mov eax, [rbx] ; read dword
__text:000000000002A3BA mov rsi, [rdi+rax*8+0E40h] ; use as index for small inline buffer in this object
__text:000000000002A3C2 cmp byte ptr [rsi+1DCh], 0 ; fail if byte at +0x1dc is 0
__text:000000000002A3C9 jz short ___fail
__text:000000000002A3CB add rsi, 1E0Dh ; otherwise, memcpy from that pointer +0x1e0dh
__text:000000000002A3D2 mov edx, 1D8h ; 0x1d8 bytes
__text:000000000002A3D7 mov rdi, r14 ; to the buffer which will be sent back to userspace
__text:000000000002A3DA call _memcpy

For this PoC we try to read the pointers at 0x2000 byte boundaries after this allocation; with luck there will be a vtable
pointer there which will allow us to read back vtable contents and defeat kASLR.

With a bit more effort this could be turned into an (almost) arbitrary read by for example spraying the kernel heap with the desired read target
then using a larger offset hoping to land in one of the sprayed buffers. A kernel arbitrary read would, for example, allow you to read the sandbox.kext
HMAC key and forge sandbox extensions if it still works like that.

tested on MacOS Sierra 10.12.2 (16C67)
*/

// ianbeer

// build: clang -o capri_mem capri_mem.c -framework IOKit

#if 0
MacOS kernel memory disclosure due to lack of bounds checking in AppleIntelCapriController::getDisplayPipeCapability

Selector 0x710 of IntelFBClientControl ends up in AppleIntelCapriController::getDisplayPipeCapability.

This method takes a structure input and output buffer. It reads an attacker controlled dword from the input buffer which it
uses to index an array of pointers with no bounds checking:

AppleIntelCapriController::getDisplayPipeCapability(AGDCFBGetDisplayCapability_t *, AGDCFBGetDisplayCapability_t *)
__text:000000000002A3AB mov r14, rdx ; output buffer, readable from userspace
__text:000000000002A3AE mov rbx, rsi ; input buffer, controlled from userspace
...
__text:000000000002A3B8 mov eax, [rbx] ; read dword
__text:000000000002A3BA mov rsi, [rdi+rax*8+0E40h] ; use as index for small inline buffer in this object
__text:000000000002A3C2 cmp byte ptr [rsi+1DCh], 0 ; fail if byte at +0x1dc is 0
__text:000000000002A3C9 jz short ___fail
__text:000000000002A3CB add rsi, 1E0Dh ; otherwise, memcpy from that pointer +0x1e0dh
__text:000000000002A3D2 mov edx, 1D8h ; 0x1d8 bytes
__text:000000000002A3D7 mov rdi, r14 ; to the buffer which will be sent back to userspace
__text:000000000002A3DA call _memcpy

For this PoC we try to read the pointers at 0x2000 byte boundaries after this allocation; with luck there will be a vtable
pointer there which will allow us to read back vtable contents and defeat kASLR.

With a bit more effort this could be turned into an (almost) arbitrary read by for example spraying the kernel heap with the desired read target
then using a larger offset hoping to land in one of the sprayed buffers. A kernel arbitrary read would, for example, allow you to read the sandbox.kext
HMAC key and forge sandbox extensions if it still works like that.

tested on MacOS Sierra 10.12.2 (16C67)
Related Posts

Comments