> On Mar 1, 2019, at 6:38 AM, Peter Zijlstra <[email protected]> wrote:
>
>> On Fri, Mar 01, 2019 at 01:57:18PM +0100, Peter Zijlstra wrote:
>>> On Fri, Mar 01, 2019 at 01:27:45PM +0100, Peter Zijlstra wrote:
>>> arch/x86/lib/usercopy_64.o: warning: objtool: .altinstr_replacement+0x30:
>>> redundant UACCESS disable
>>
>>> The usercopy one is difficult, that's copy_user_handle_tail(), it is
>>> buggered though, because that lacks notrace and thus has a __fentry__
>>> call in.
>>>
>>> Also, afaict all exception jumps into copy_user_handle_tail() will have
>>> AC=1, but the __{get,put}_user_nocheck() things do STAC/CLAC all over
>>> again.
>>>
>>> So what do we do? Annotate that we start with AC=1 and then immediately
>>> do the clac, and then let __{get,put}_user_nocheck() do their own thing?
>>> or make it use the unsafe stuff?
>>
>> Or.. we move the thing to assembly. Of course, I suck at (writing) asm,
>> so the below is probably broken in various ways.
>
> The advantage is that it now all lives in the same .o file and objtool
> can actually follow and find the complete control flow.
>
> I've made it ENDPROC() such that it becomes STT_FUNC and objtool does
> all the normal things. I've also moved the ALIGN_DESTINATION macro into
> the .S file.
>
> Andy, do we have a sensible self-test for this path?
Not that I know of. Something like my (rejected) strncpy_from_user test could
probably be added fairly easily.
