Re: [PATCH v5 32/32] x86/mm: Add support to make use of Secure Memory Encryption

[Tom Lendacky <thomas.lendacky@amd.com>]

On 5/30/2017 9:55 AM, Borislav Petkov wrote:
> On Tue, May 30, 2017 at 09:38:36AM -0500, Tom Lendacky wrote:
>> In this case we're running identity mapped and the "on" constant ends up
>> as kernel address (0xffffffff81...) which results in a segfault.
> 
> Would
> 
> 	static const char *__on_str = "on";
> 
> 	...
> 
> 	if (!strncmp(buffer, __pa_nodebug(__on_str), 2))
> 		...
> 
> work?
> 
> __phys_addr_nodebug() seems to pay attention to phys_base and
> PAGE_OFFSET and so on...

Except that phys_base hasn't been adjusted yet so that doesn't work
either.

> 
> I'd like to avoid that rip-relative address finding in inline asm which
> looks fragile to me.

I can define the command line option and the "on" and "off" values as
character buffers in the function and initialize them on a per character
basis (using a static string causes the same issues as referencing a
string constant), i.e.:

char cmdline_arg[] = {'m', 'e', 'm', '_', 'e', 'n', 'c', 'r', 'y', 'p', 't', '\0'};
char cmdline_off[] = {'o', 'f', 'f', '\0'};
char cmdline_on[] = {'o', 'n', '\0'};

It doesn't look the greatest, but it works and removes the need for the
rip-relative addressing.

Thanks,
Tom

> 
> Thanks.
>