Re: "arm64: vdso: Substitute gettimeofday() with C implementation" breaks clang build

[Vincenzo Frascino]

On 26/06/2019 10:38, Will Deacon wrote:
> On Tue, Jun 25, 2019 at 06:00:27PM +0100, Vincenzo Frascino wrote:
>> On 25/06/2019 17:26, Nick Desaulniers wrote:
>>> On Tue, Jun 25, 2019 at 7:54 AM Vincenzo Frascino
>>> <vincenzo.frascino@xxxxxxx> wrote:
>>>>> but clang 7.0 is still use in many distros by default, so maybe this commit can
>>>>> be fixed by adding a conditional check to use "small" if clang version < 8.0.
>>>>>
>>>>
>>>> Could you please verify that the patch below works for you?
>>>
>>> Should it be checking against CONFIG_CLANG_VERSION, or better yet be
>>> using cc-option macro?
>>>
>>
>> This is what I did in my proposed patch, but I was surprised that clang-7
>> generates relocations that clang-8 does not.
>>
>>   LD      arch/arm64/kernel/vdso/vdso.so.dbg
>>   VDSOCHK arch/arm64/kernel/vdso/vdso.so.dbg
>> 00000000000009d0 R_AARCH64_JUMP_SLOT  _mcount
> 
> Hmm. It would be nice to understand where the reference to _mcount is coming
> from, since that sounds like ftrace is getting involved where it shouldn't
> be.
> 

That's very true, it was a mistake in the Makefile change that I provided with
the original iteration of this patch, that had as a side effect of having ftrace
involved (_mcount is defined in entry-ftrace.c).

I was overriding:

CFLAGS_REMOVE_vgettimeofday.o = $(CC_FLAGS_FTRACE) -Os

with:

CFLAGS_REMOVE_vgettimeofday.o = -mcmodel=tiny (selected if clang is < 8)

that's why I said that I missed a "+" in my previous patch.

Having:

CFLAGS_REMOVE_vgettimeofday.o += -mcmodel=tiny

restores the wanted behavior.

Sorry for not being clear in my explanation.

> Will
> 

-- 
Regards,
Vincenzo