Re: [RFC] x86: Avoid CR3 load on compatibility mode with PTI

[Nadav Amit]

Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:

> 
> 
>> On Jan 15, 2018, at 9:42 AM, Nadav Amit <namit@xxxxxxxxxx> wrote:
>> 
>> Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> 
>>>> On Jan 14, 2018, at 12:13 PM, Nadav Amit <namit@xxxxxxxxxx> wrote:
>>>> 
>>>> Currently, when page-table isolation is on to prevent the Meltdown bug
>>>> (CVE-2017-5754), CR3 is always loaded on system-call and interrupt.
>>>> 
>>>> However, it appears that this is an unnecessary measure when programs
>>>> run in compatibility mode. In this mode only 32-bit registers are
>>>> available, which means that there *should* be no way for the CPU to
>>>> access, even speculatively, memory that belongs to the kernel, which
>>>> sits in high addresses.
>>> 
>>> You're assuming that TIF_IA32 prevents the execution of 64-bit code.  It doesn't.
>>> 
>>> I've occasionally considered adding an opt-in hardening mechanism to enforce 32-bit or 64-bit execution, but we don't have this now.
>> 
>> I noticed it doesnât. I thought the removing/restoring the __USER_CS
>> descriptor on context switch, based on TIF_IA32, would be enough.
>> modify_ldt() always keeps the descriptor l-bit clear. I will review the
>> other GDT descriptors, and if needed, create two GDTs. Let me know if I
>> missed anything else.
> 
> There world need to be some opt-in control, I think, for CRIU if nothing else.
> 
> Also, on Xen PV, it's a complete nonstarter.  We don't have enough control over the GDT unless someone knows otherwise.  But there's no PTI on Xen PV either.
> 
>>> Anything like this would also need to spend on SMEP, I think -- the pseudo-SMEP granted by PTI is too valuable to give up on old boxes, I think.
>> 
>> If SMEP is not supported, compatibility mode would still require page-table
>> isolation.
>> 
>> Thanks for the feedback. I still look for an ack for the basic idea of
>> disabling page-table isolation on compatibility mode.
> 
> I'm still not really convinced this is worth it.  It will send a bad message and get people to run critical stuff compiled for 32-bit, which has its own downsides.

I can handle #GP gracefully if __USER_CS is loaded so PTI would be required
again. Doing so would eliminate the need for an opt-in, and preserve the
current semantics.