Re: [RFC PATCH v2 07/20] x86: Provide general kernel support for memory encryption

[Tom Lendacky]

On 09/08/2016 08:55 AM, Borislav Petkov wrote:
> On Thu, Sep 08, 2016 at 08:26:27AM -0500, Tom Lendacky wrote:
>> When does this value get initialized?  Since _PAGE_ENC is #defined to
>> sme_me_mask, which is not set until the boot process begins, I'm afraid
>> we'd end up using the initial value of sme_me_mask, which is zero.  Do
>> I have that right?
> 
> Hmm, but then that would hold true for all the other defines where you
> OR-in _PAGE_ENC, no?

As long as the #define is not a global variable like this one it's ok.

> 
> In any case, the preprocessed source looks like this:
> 
> pmdval_t early_pmd_flags = (((((((pteval_t)(1)) << 0) | (((pteval_t)(1)) << 1) | (((pteval_t)(1)) << 6) | (((pteval_t)(1)) << 5) | (((pteval_t)(1)) << 8)) | (((pteval_t)(1)) << 63)) | (((pteval_t)(1)) << 7)) | sme_me_mask) & ~((((pteval_t)(1)) << 8) | (((pteval_t)(1)) << 63));
> 
> but the problem is later, when building:
> 
> arch/x86/kernel/head64.c:39:28: error: initializer element is not constant
>  pmdval_t early_pmd_flags = (__PAGE_KERNEL_LARGE | _PAGE_ENC) & ~(_PAGE_GLOBAL | _PAGE_NX);
>                             ^
> scripts/Makefile.build:153: recipe for target 'arch/x86/kernel/head64.s' failed
> 
> so I guess not. :-\
> 
> Ok, then at least please put the early_pmd_flags init after
> sme_early_init() along with a small comment explaning what happens.

Let me verify that we won't possibly take any kind of page fault during
sme_early_init() and cause a page to be not be marked encrypted. Or... I
can add a comment indicating I need to set this as early as possible to
cover any page faults that might occur.

Thanks,
Tom

> 
> Thanks.
>