There is an issue in Chakra JIT server that can be potentially exploited to compromise the JIT process from a compromised browser content process. Bugs like this could potentially be used to bypass ACG (Arbitrary Code Guard) in Microsoft Edge.
The issue has been confirmed on a ChakraCore build from the latest source.
Chakra JIT server takes bytecode as an input from the calling process. JIT server can either compile a function or a loop body. When the client asks the JIT process to compile a loop body, in addition to the bytecode buffer, the client sends a start offset and an end offset inside the buffer (`CodeGenWorkItemIDL->jitData->bodyData->loopHeaders->startOffset and CodeGenWorkItemIDL->jitData->bodyData->loopHeaders->endOffset`). These values aren't validated by the JIT server.
This can lead to out-of-bound reads in the bytecode buffer, but it can also lead to an out-of-bounds write as demonstrated below.
In `IRBuilder.cpp` in` IRBuilder::Build()` on this line
```
offsetToInstructionCount = lastOffset + 2;
```
lastOffset is user-controlled. If lastOffset is sufficiently large, an integer overflow occurs and offsetToInstructionCount wraps around to a small value. offsetToInstructionCount is then used to allocate an array:
```
m_offsetToInstruction = JitAnewArrayZ(m_tempAlloc, IR::Instr *, offsetToInstructionCount);
```
Due to the overflow, the array is goint to be too small to hold the required data. Finally, an overflow happens in `IRBuilder::AddInstr` on
```
Assert(offset < m_offsetToInstructionCount);
if (m_offsetToInstruction[offset] == nullptr)
{
m_offsetToInstruction[offset] = instr;
}
```
Note #1: While there is an assert() here it will only affect the debug build.
Note #2: Due to the if() statement, an attacker can only overwrite a null-value.
To demonstrate the issue, it is sufficient to change the value of `CodeGenWorkItemIDL->jitData->bodyData->loopHeaders->endOffset` to 0xffffffff before calling RemoteCodeGen() function on the JIT server when JITing a loop body.
暂无评论