Re: [PATCH] x86: static_cpu_has_safe: discard dynamic check after init

[Andy Lutomirski]

On Mon, Jan 18, 2016 at 8:52 AM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
> On Sun, Jan 17, 2016 at 5:33 AM, Borislav Petkov <bp@xxxxxxx> wrote:
>> On Sat, Jan 16, 2016 at 02:58:21PM -0500, Brian Gerst wrote:
>>> It saves space.  The only thing left in main text is one jump
>>> instruction (5 bytes).
>>
>> How do I measure this?
>>
>> Because with my tailored config here, the only thing I'm seeing is a growth of
>> text by 730 bytes:
>>
>> before:
>>    text    data     bss     dec     hex filename
>> 10926552        3598944 16642048        31167544        1db9438 vmlinux
>>
>> [    0.056552] Freeing SMP alternatives memory: 24K (ffffffff81dd7000 - ffffffff81ddd000)
>> [    2.883728] Freeing unused kernel memory: 2904K (ffffffff81b01000 - ffffffff81dd7000)
>>
>>
>> after:
>>    text    data     bss     dec     hex filename
>> 10927282        3598944 16642048        31168274        1db9712 vmlinux
>>
>> [    0.052559] Freeing SMP alternatives memory: 24K (ffffffff81dd7000 - ffffffff81ddd000)
>> [    3.225318] Freeing unused kernel memory: 2904K (ffffffff81b01000 - ffffffff81dd7000)
>
> It is due to page alignment padding.  It was not enough to lose a
> whole page from .text in your case.
>
> The size command includes any section that is marked executable in the
> text count, including init text.  If you use readelf -S vmlinux.o
> instead you will notice that .text is the same size or smaller, and
> .static_cpu_has (which is freed after boot) is the difference.
>

If I'm understanding this correctly, the total non-init overhead from
static_cpu_has_safe with your patch is five bytes.  I'd imagine that
the short jmp optimization in regular static_cpu_has essentially never
works, which means it will also use five bytes of text, which makes me
wonder whether we should just make static_cpu_has safe and remove the
distinction.

Also, someone should write a little script to measure kernel size
minus .init stuff.

--Andy