On Fri, May 06, 2022 at 10:18:10AM +0800, Chen Zhongjin wrote:
> On 2022/5/5 18:56, Mark Rutland wrote:
> > On Thu, May 05, 2022 at 11:24:48AM +0200, Peter Zijlstra wrote:
> >> On Thu, May 05, 2022 at 11:36:12AM +0800, Chen Zhongjin wrote:
> >>> Hi Peter,
> >>>
> >>> IIRC now the blacklist mechanisms all run on check stage, which after
> >>> decoding, but the problem of kuser32.S happens in decoding stage. Other
> >>> than that the assembly symbols in kuser32 is STT_NOTYPE and
> >>> STACK_FRAME_NON_STANDARD will throw an error for this.
> >>>
> >>> OBJECT_FILES_NON_STANDARD works for the single file but as you said
> >>> after LTO it's invalid. However STACK_FRAME_NON_STANDARD doesn't work
> >>> for kuser32 case at all.
> >>>
> >>> Now my strategy for undecodable instructions is: show an error message
> >>> and mark insn->ignore = true, but do not stop anything so decoding work
> >>> can going on.
> >>>
> >>> To totally solve this my idea is that applying blacklist before decode.
> >>> However for this part objtool doesn't have any insn or func info, so we
> >>> should add a new blacklist just for this case...
> >>
> >> OK, so Mark explained that this is 32bit userspace (VDSO) code.
> >>
> >> And as such there's really no point in running objtool on it. Does all
> >> that live in it's own section? Should it?
> >
> > It's placed in .rodata by a linker script:
> >
> > * The 32-bit vdso + kuser code is placed in .rodata, between the `vdso32_start`
> > and `vdso32_end` symbols, as raw bytes (via .incbin).
> > See arch/arm64/kernel/vdso32-wrap.S.
> >
> > * The 64-bit vdso code is placed in .rodata, between the `vdso_start`
> > and `vdso32` symbols, as raw bytes (via .incbin).
> > See arch/arm64/kernel/vdso-wrap.S.
> >
> > The objects under arch/arm64/kernel/{vdso,vdso32}/ are all userspace objects,
> > and from userspace's PoV the existing secrtions within those objects are
> > correct, so I don't think those should change.
> >
> > How does x86 deal with its vdso objects?
> >
> > Thanks,
> > Mark.
> > .
>
> However for my build kuser32.o content is in .text
We should be able to move that into .rodata; it's never executed in kernel context.
> and there is only `vdso` symbol in .rodata without `vdso32`.
That means you're not building with CROSS_COMPILE_COMPAT, and so we can't build
the 32-bit VDSO.
> And for defconfig the CONFIG_KUSER_HELPERS=y is on.
Yes.
> According to your description, it seems something wrong here?
Sorry, I was wrong about how we linked the kuser32 code.
I believe we can move that into .rodata by adding:
.section .rodata
... to the start of that.
I think that'd be a nice cleanup to do regardless of objtool.
Thanks,
Mark.
