Re: [PATCH v3] Makefile: lld: tell clang to use lld

From: Masahiro Yamada
Date: Sat Apr 06 2019 - 22:30:12 EST


Hi Kees, Nick,


On Sat, Apr 6, 2019 at 1:52 AM Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
>
> On Fri, Apr 5, 2019 at 9:11 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > On Fri, Apr 5, 2019 at 3:17 AM Masahiro Yamada
> > <yamada.masahiro@xxxxxxxxxxxxx> wrote:
> > > I want to propose alternative solution.
> > > Please check the attached patches.
>
> ```
> My plan is to rewrite all VDSO Makefiles to use $(LD), then delete cc-ldoption.
> ```
>
> I agree with that as a possible solution as well; I think it's best to
> just invoke the linker directly rather than use the compiler as the
> driver. Just small nits with the 2 attached patches, but I think they
> will also work for us.
>
> Looks like there's 15 call sites across 12 files, plus Documentation/
> and scripts/Kbuild.include; removing it seems feasible and I'd be
> happy to help if you'd welcome the patches?


Yeah, your help is appreciated.

I just worked on two Makefiles
because I wanted to hear opinions
before investing more efforts.


Basically, you are positive to this approach,
so let's go this way.


> Let's start with that series of changes, but I suspect we will
> ultimately need to revisit -fuse-ld=. I was just thinking that even
> my proposed patch doesn't do anything for KBUILD_HOSTCFLAGS, for
> example.

Host tools are not so important, but I agree.

BTW, why does clang invoke bfd linker instead of
lld by default?

If we want to make llvm self-contained,
I believe clang should use ld.lld by default.




> I worry that bfd may still be invoked by Clang at some point
> in the build. I will try to test in an environment with no GNU
> binutils.
>
> > In the 0001 patch of arch/arm/vdso/Makefile:
> > > ...
> > > -VDSO_LDFLAGS += $(call cc-ldoption, -fuse-ld=bfd)
> > > +ldflags-y = -Bsymbolic --no-undefined -soname=linux-vdso.so.1 \
> > > + -z max-page-size=4096 -z common-page-size=4096 \
> > > + -nostdlib -shared \
> > > + $(call ld-option, --hash-style=sysv) \
>
> I think the vdso's should all REQUIRE --hash-style=sysv; the kselftest
> for vdso's will fail otherwise. I could always follow up with patches
> for that if we didn't want to conflate changes.
>
> > > + $(call ld-option, --build-id) \
> > > + -T
>
> Was it intentional to add -T standalone? The man page for `ld` says
> it needs an additional filename to follow -T.

This is because I wanted to re-use cmd_ld
defined in scripts/Makefile.lib

$(real-prereqs) contains the linker scripts and objects.


> Or rather is optional
> if a -L flag is specified?
>
> >
> > I see that "-fuse-ld=bfd" is explicitly dropped here, which could
> > reintroduce the problem fixed in d2b30cd4b722 ("ARM: 8384/1: VDSO:
> > force use of BFD linker") (and 3473f26592c1c). Does ld.gold still
> > exhibit this problem? If so, we'll need to detect gold and force bfd
> > still maybe?

I intentionally dropped -fuse-ld=bfd.

With my patch applied, users have full control
of the linker by passing LD=.

Since we cannot link vmlinux with gold,
users should definitely pass a bfd linker.



> The arm32 vdso Makefile is an oddity in this regard. I think what
> Kees described is best. Rather than always set the linker to bfd
> (which harms linkers like lld), detect if gold is being used and if so
> set the linker back to bfd.
>
> For arm64, lld has this issue with -n,
> https://github.com/ClangBuiltLinux/linux/issues/340, but I think we
> can fix that in a follow up patch. Same thoughts on --hash-style and
> -T.


If we need to add something depending on
the linker type,

I do not mind doing like this.

ldflags-$(CONFIG_LD_IS_GOLD) += ...
ldflags-$(CONFIG_LD_IS_LLD) += ...

But, it is a different issue.






--
Best Regards
Masahiro Yamada