Re: Building arm64 EFI stub with -fpie breaks build of 4.9.x (undefined reference to `__efistub__GLOBAL_OFFSET_TABLE_')

[Ard Biesheuvel]

On Wed, 5 Jun 2019 at 18:26, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Wed, Jun 05, 2019 at 05:19:40PM +0200, Rolf Eike Beer wrote:
> > I decided to dig out a toy project which uses a DragonBoard 410c. This has
> > been "running" with kernel 4.9, which I would keep this way for unrelated
> > reasons. The vanilla 4.9 kernel wasn't bootable back then, but it was
> > buildable, which was good enough.
> >
> > Upgrading the kernel to 4.9.180 caused the boot to suddenly fail:
> >
> > aarch64-unknown-linux-gnueabi-ld: ./drivers/firmware/efi/libstub/lib.a(arm64-
> > stub.stub.o): in function `handle_kernel_image':
> > /tmp/e2/build/linux-4.9.139/drivers/firmware/efi/libstub/arm64-stub.c:63:
> > undefined reference to `__efistub__GLOBAL_OFFSET_TABLE_'
> > aarch64-unknown-linux-gnueabi-ld: ./drivers/firmware/efi/libstub/lib.a(arm64-
> > stub.stub.o): relocation R_AARCH64_ADR_PREL_PG_HI21 against symbol
> > `__efistub__GLOBAL_OFFSET_TABLE_' which may bind externally can not be used
> > when making a shared object; recompile with -fPIC
> > /tmp/e2/build/linux-4.9.139/drivers/firmware/efi/libstub/arm64-stub.c:63:
> > (.init.text+0xc): dangerous relocation: unsupported relocation
> > /tmp/e2/build/linux-4.9.139/Makefile:1001: recipe for target 'vmlinux' failed
> > -make[1]: *** [vmlinux] Error 1
> >
> > This is caused by commit 27b5ebf61818749b3568354c64a8ec2d9cd5ecca from
> > linux-4.9.y (which is 91ee5b21ee026c49e4e7483de69b55b8b47042be), reverting
> > this commit fixes the build.
> >
> > This happens with vanilla binutils 2.32 and gcc 8.3.0 as well as 9.1.0. See
> > the attached .config for reference.
> >
> > If you have questions or patches just ping me.
>
> Does Linus's latest tree also fail for you (or 5.1)?
>
> Nick, do we need to add another fix that is in mainline for this to work
> properly?
>

For the record, this is an example of why I think backporting those
clang enablement patches is a bad idea. We can't actually build those
kernels with clang, can we? So what is the point? </grumpy>

It would be helpful to get a relocation dump (objdump -r) of
arm64-stub.o to figure out which symbol needs a 'hidden' annotation to
prevent GCC from emitting it as a PIC reference requiring a GOT.
Alternatively, we can just revert this patch from 4.9