On Tue, 28 Mar 2023 at 11:32, Pingfan Liu <kernelfans@xxxxxxxxx> wrote: > > On Tue, Mar 28, 2023 at 3:58 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > On Tue, 28 Mar 2023 at 09:32, Pingfan Liu <kernelfans@xxxxxxxxx> wrote: > > > > > > On Mon, Mar 27, 2023 at 6:57 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > > > > > On Mon, 27 Mar 2023 at 11:51, Pingfan Liu <kernelfans@xxxxxxxxx> wrote: > > > > > > > > > > Using objcopy to reform vmlinuz.efi.elf to vmlinuz.efi will not convey > > > > > any relocation information. That means vmlinuz.efi is expected to be > > > > > PIC. > > > > > > > > > > At present, vmlinuz.efi is PIC. But it is better to adopt the same > > > > > > > > Why is it better? > > > > > > > > > > I think except carefully coded with asm language, there is no > > > guarantee which prevents the compiler from generating position > > > dependent code. Or is there any presumption here? > > > > > > > All object files built under drivers/firmware/efi/libstub are scanned > > for absolute relocations, and if any exist, the build is aborted. > > > > Oh, I got the code piece now. Thank you very much for pointing this > out. And only if R_AARCH64_ABS64 is unavoidable, there is no need to > ask the code to fix the relocation by itself. > > But one more question, there are many other "S+A" types of relocation, > let's say R_AARCH64_ABS32 or R_AARCH64_LDST64_ABS_LO12_NC. It is also > sensitive to the delta between loading and linking. Should all of them > be excluded by the Makefile? > None of those can be used to relocate absolute addresses, so these are not the ones we should care about. There are the ones related to the GOT, which could result in GOT entries in the final executable with R_AARCH64_RELATIVE relocations. However, we use 'hidden' visibility for all codegen related to the EFI stub, so the compiler knows all symbol references can be resolved at link time. This means that under -fpic or -fpie, it will always use relative references. > > > > > solution used by the kernel to resolve the code relocation issue by > > > > > itself. That is to resolve R_AARCH64_RELATIVE at the runtime. > > > > > > > > > > > > > This breaks other architectures. > > > > > > > > > > Oops, could it be done by: > > > #if defined(_aarch64_) > > > #define R_ARCH_RELATIVE R_AARCH64_RELATIVE > > > #else > > > #define R_ARCH_RELATIVE 0 > > > #endif > > > > > > Since any r_type equaling 0 can not be ejected into the relocation > > > section, this macro will make the code dummy in essential. > > > > > > > So the assembler instructions you are using are also implemented on > > RISC-V and LoongArch? > > Oops. The whole file should be arch dependent. In those cases, they > should be a single jump instruction. > > Appreciate for your help. > > Thanks, > > Pingfan
