Re: [PATCH RFC v2 0/6] x86: Disabling PTI in compatibility mode

From: Andy Lutomirski
Date: Thu Feb 15 2018 - 22:06:18 EST


On Fri, Feb 16, 2018 at 12:51 AM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> wrote:
>
>> On 16/02/2018 00:25, Nadav Amit wrote:
>>> Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx> wrote:
>>>
>>>> On 02/15/2018 08:35 AM, Nadav Amit wrote:
>>>>> I removed the PTI disabling while SMEP is unsupported, although I
>>>>> must admit I did not fully understand why it is required.
>>>> Do you mean you don't fully understand how PTI gives SMEP-like behavior
>>>> on non-SMEP hardware?
>>> No. I understand how it provide SMEP-like behavior, and I understand the value
>>> of SMEP by itself.
>>>
>>> However, I do not understand why SMEP-like protection is required to protect
>>> processes that run in compatibility-mode from Meltdown/Spectre attacks. As
>>> far as I understand, the process should not be able to manipulate the kernel
>>> to execute code in the low 4GB.
>>
>> Being 32bit is itself sufficient protection against Meltdown (as long as
>> there nothing interesting of the kernels mapped below the 4G boundary).
>>
>> However, a 32bit compatibility process try to attack with Spectre/SP2 to
>> redirect speculation back into userspace, at which point (if successful)
>> the pipeline will be speculating in 64bit mode, and Meltdown is back on
>> the table. SMEP will block this attack vector, irrespective of other
>> SP2 defences the kernel may employ, but a fully SP2-defended kernel
>> doesn't require SMEP to be safe in this case.
>
> Based on Jann Hornâs description of the branch predictor, it basically only
> holds the lowest 31-bits of the target address. There might be a subtle
> problem if the prediction wrapsaround, but excluding this case, I do not see
> how Spectre v2 can be used to jump into running user code.
>

If you can make the *kernel* speculate into user code, you can create
whatever gadget you want. A 32-bit task can poison the branch
predictor and use this to attack a non-retpolined kernel.