On Thu, 4 Feb 2021 at 22:31, Guillaume Tucker <guillaume.tuc...@collabora.com> wrote: > > On 04/02/2021 18:23, Nick Desaulniers wrote: > > On Thu, Feb 4, 2021 at 10:12 AM Nathan Chancellor <nat...@kernel.org> wrote: > >> > >> On Thu, Feb 04, 2021 at 10:06:08AM -0800, 'Nick Desaulniers' via Clang > >> Built Linux wrote: > >>> On Thu, Feb 4, 2021 at 8:02 AM Ard Biesheuvel <a...@kernel.org> wrote: > >>>> > >>>> On Thu, 4 Feb 2021 at 16:53, Guillaume Tucker > >>>> <guillaume.tuc...@collabora.com> wrote: > >>>>> > >>>>> On 04/02/2021 15:42, Ard Biesheuvel wrote: > >>>>>> On Thu, 4 Feb 2021 at 12:32, Guillaume Tucker > >>>>>> <guillaume.tuc...@collabora.com> wrote: > >>>>>>> > >>>>>>> Essentially: > >>>>>>> > >>>>>>> make -j18 ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- LLVM=1 > >>>>>>> CC="ccache clang" zImage > >>> > >>> This command should link with BFD (and assemble with GAS; it's only > >>> using clang as the compiler. > >> > >> I think you missed the 'LLVM=1' before CC="ccache clang". That should > >> use all of the LLVM utilities minus the integrated assembler while > >> wrapping clang with ccache. > > > > You're right, I missed `LLVM=1`. Adding `LD=ld.bfd` I think should > > permit fallback to BFD. > > That was close, except we're cross-compiling with GCC for arm. > So I've now built a plain next-20210203 (without Ard's fix) using > this command line: > > make LD=arm-linux-gnueabihf-ld.bfd -j18 ARCH=arm > CROSS_COMPILE=arm-linux-gnueabihf- LLVM=1 CC="ccache clang" zImage > > I'm using a modified Docker image gtucker/kernelci-build-clang-11 > with the very latest LLVM 11 and gcc-8-arm-linux-gnueabihf > packages added to be able to use the GNU linker. BTW I guess we > should enable this kind of hybrid build setup on kernelci.org as > well. > > Full build log + kernel binaries can be found here: > > > https://storage.staging.kernelci.org/gtucker/next-20210203-ard-fix/v5.10-rc4-24722-g58b6c0e507b7-gtucker_single-staging-41/arm/multi_v7_defconfig/clang-11/ > > And this booted fine, which confirms it's really down to how > ld.lld puts together the kernel image. Does it actually solve > the debate whether this is an issue to fix in the assembly code > or at link time? > > Full test job details for the record: > > https://lava.collabora.co.uk/scheduler/job/3176004 >
So the issue appears to be in the way the linker generates the _kernel_bss_size symbol, which obviously has an impact, given that the queued fix takes it into account in the cache_clean operation. On GNU ld, I see 479: 00065e14 0 NOTYPE GLOBAL DEFAULT ABS _kernel_bss_size whereas n LLVM ld.lld, I see 433: c1c86e98 0 NOTYPE GLOBAL DEFAULT ABS _kernel_bss_size and adding this value may cause the cache clean to operate on unmapped addresses, or cause the addition to wrap and not perform a cache clean at all. AFAICT, this also breaks the appended DTB case in LLVM, so this needs a separate fix in any case.