On Tue, 16 Jan 2024 at 06:21, Jan Kiszka <jan.kiszka@xxxxxxxxxxx> wrote: > > On 15.01.24 18:34, Ard Biesheuvel wrote: > > On Sat, 13 Jan 2024 at 11:35, Jan Kiszka <jan.kiszka@xxxxxxxxxxx> wrote: > >> > >> On 12.01.24 19:56, Palmer Dabbelt wrote: > >>> On Fri, 12 Jan 2024 10:51:16 PST (-0800), Ard Biesheuvel wrote: > >>>> Hi Jan, > >>>> > >>>> On Fri, 12 Jan 2024 at 19:37, Jan Kiszka <jan.kiszka@xxxxxxxxxxx> wrote: > >>>>> > >>>>> From: Jan Kiszka <jan.kiszka@xxxxxxxxxxx> > >>>>> > >>>>> The cflags for the RISC-V efistub were missing -mno-relax, thus were > >>>>> under the risk that the compiler could use GP-relative addressing. That > >>>>> happened for _edata with binutils-2.41 and kernel 6.1, causing the > >>>>> relocation to fail due to an invalid kernel_size in handle_kernel_image. > >>>>> It was not yet observed with newer versions, but that may just be luck. > >>>>> > >>>>> Signed-off-by: Jan Kiszka <jan.kiszka@xxxxxxxxxxx> > >>>>> --- > >>>>> > >>>>> Something like this should go to stable as well, but we will need > >>>>> rebased patches. > >>>>> > >>>>> drivers/firmware/efi/libstub/Makefile | 2 +- > >>>>> 1 file changed, 1 insertion(+), 1 deletion(-) > >>>>> > >>>>> diff --git a/drivers/firmware/efi/libstub/Makefile > >>>>> b/drivers/firmware/efi/libstub/Makefile > >>>>> index 06964a3c130f..d561d7de46a9 100644 > >>>>> --- a/drivers/firmware/efi/libstub/Makefile > >>>>> +++ b/drivers/firmware/efi/libstub/Makefile > >>>>> @@ -28,7 +28,7 @@ cflags-$(CONFIG_ARM) += -DEFI_HAVE_STRLEN > >>>>> -DEFI_HAVE_STRNLEN \ > >>>>> -DEFI_HAVE_MEMCHR > >>>>> -DEFI_HAVE_STRRCHR \ > >>>>> -DEFI_HAVE_STRCMP -fno-builtin > >>>>> -fpic \ > >>>>> $(call > >>>>> cc-option,-mno-single-pic-base) > >>>>> -cflags-$(CONFIG_RISCV) += -fpic -DNO_ALTERNATIVE > >>>>> +cflags-$(CONFIG_RISCV) += -fpic -DNO_ALTERNATIVE -mno-relax > >>>> > >>>> Can we detect the presence of these references (via the relocation > >>>> type)? We already do something similar for ordinary absolute > >>>> references too. > >>> > >>> If there's no `__global_pointer$` symbol then the linker won't make > >>> GP-relative relaxations (because it doesn't know where GP is). We > >>> usually define that symbol in the linker script, but I'm not entierly > >>> sure how libstub gets its linker script... > >>> > >> > >> The stub seems to be linked together with the rest of the kernel, thus > >> the regular arch/riscv/kernel/vmlinux.lds.S is used. > >> > > > > Indeed - the EFI stub is part of the same executable as vmlinux, we > > just mangle the symbol names to ensure that only code that can be > > safely called from the EFI stub can be linked to it. > > > > If the effect of -mno-relax is to stop emitting R_RISCV_RELAX > > relocations, we should perhaps add those to the STUBCOPY_RELOC-y > > Makefile variable? (in the same file). BTW R_RISCV_HI20 doesn't seem > > like the right value there to begin with: the idea of that is to > > disallow ELF relocations that evaluate to expressions that can only be > > known at runtime (like absolute addresses for global pointer > > variables) > > How to do that best? Simply replace R_RISCV_HI20 with R_RISCV_RELAX? > We'll need to keep the HI20, in fact - I got confused between HI20 and PCREL_HI20, and the former is actually used for 32-bit absolute addresses in 32-bit code. This seems to do the trick: it disallows relaxation relocations and native word sizes absolute references. AFAICT, those are the only ones we should care about. STUBCOPY_RELOC-$(CONFIG_RISCV) := -E R_RISCV_HI20\|R_RISCV_$(BITS)\|R_RISCV_RELAX
