Re: [PATCH v7 0/7] Introduce support for Guest CET feature

[Jim Mattson]

On Thu, Oct 3, 2019 at 5:59 AM Yang Weijiang <weijiang.yang@xxxxxxxxx> wrote:
>
> On Wed, Oct 02, 2019 at 03:40:20PM -0700, Jim Mattson wrote:
> > On Thu, Sep 26, 2019 at 7:17 PM Yang Weijiang <weijiang.yang@xxxxxxxxx> wrote:
> > >
> > > Control-flow Enforcement Technology (CET) provides protection against
> > > Return/Jump-Oriented Programming (ROP/JOP) attack. It includes two
> > > sub-features: Shadow Stack (SHSTK) and Indirect Branch Tracking (IBT).
> > >
> > > KVM modification is required to support Guest CET feature.
> > > This patch serial implemented CET related CPUID/XSAVES enumeration, MSRs
> > > and VMEntry configuration etc.so that Guest kernel can setup CET
> > > runtime infrastructure based on them. Some MSRs and related feature
> > > flags used in the patches reference the definitions in kernel patch.
> >
> > I am still trying to make my way through the 358 page (!) spec for
> > this feature, but I already have some questions/comments about this
> > series:
> >
> > 1. Does CET "just work" with shadow paging? Shadow paging knows
> > nothing about "shadow-stack pages," and it's not clear to me how
> > shadow-stack pages will interact with dirty tracking.
> > 2. I see non-trivial changes to task switch under CET. Does
> > emulator_do_task_switch need to be updated?
> > 3. What about all of the emulator routines that emulate control
> > transfers (e.g. em_jmp_{far,abs}, em_call_(near_abs,far},
> > em_ret_{far,far_imm,near_imm}, etc)? Don't these have to be modified
> > to work correctly when CR4.CET is set?
> > 4. You don't use the new "enable supervisor shadow stack control" bit
> > in the EPTP. I assume that this is entirely optional, right?
> > 5. I think the easiest way to handle the nested issue (rather than
> > your explicit check for vmxon when setting CR4.CET when the vCPU is in
> > VMX operation) is just to leave CR4.CET out of IA32_VMX_CR4_FIXED1
> > (which is already the case).
> > 6. The function, exception_class(), in x86.c, should be updated to
> > categorize #CP as contributory.
> > 7. The function, x86_exception_has_error_code(), in x86.h, should be
> > updated to include #CP.
> > 8. There appear to be multiple changes to SMM that you haven't
> > implemented (e.g saving/restoring the SSP registers in/from SMRAM.
> >
> > CET is quite complex. Without any tests, I don't see how you can have
> > any confidence in the correctness of this patch series.
> Thanks Jim for the detailed comments.
>
> I missed adding test platform and
> result introduction in cover letter. This serial of patch has passed CET
> test in guest on Intel x86 emulator platform and develop machine.
> Some feature mentioned in the spec. has not been implemented yet. e.g.,
> "supervisor shadow stack control".

What I should have said is that I'd like to see kvm-unit-tests to
exercise the new functionality, even if no one outside Intel can run
them yet.

> CET feature itself is complex, most of the enabling work is
> inside kernel, the role of KVM is to expose CET related CPUID and MSRs
> etc. to guest, and make guest take over control of the MSRs directly so that
> CET can work efficiently for guest. There're QEMU patches for CET too.
>
> I'll review your comments carefully, thank you again!