I created a GitHub repository demonstrating the issue, together with instructions on how to build it.

https://github.com/Fighter19/hello-world-arm-efi

Currently, when using the old linker file, it will crash.
Personally, to test it, I built U-Boot for QEMU and loaded the EFI file via TFTP.

Am 06.11.22 um 18:03 schrieb Patrick Zacharias:
Hi Heinrich,

Thanks for the information, I'll make sure to use it for future contributions.

I encountered this issue, while building a custom EFI application using Rust by linking
the following files:

./arch/arm/lib/reloc_arm_efi.c
./arch/arm/lib/crt0_arm_efi.S
./arch/arm/lib/elf_arm_efi.lds

As well as the custom application built as a library, that exports "efi_main".

Thus, the relocations were caused by the "write_fmt" functions that Rust provides.
I encountered it, when concatenating two strings.
format!("Test {}", "Test2");

The relocation was not applied and therefor lead to a crash.

From my understanding these relocations are ELF relocations and are therefor applied by the _relocate function (inside the built application) and not by the EFI loader. This is why I assumed, that it's fine for these relocations to reside in the data section.

I noticed, that in the _relocate function in ./arch/arm/lib/reloc_arm_efi.c, that the case R_ARM_RELATIVE was never entered,
even though the generated shared object included them.
I then looked at the EFI header file (crt0_arm_efi.S) and noticed, that it specifies only up to _edata to be loaded.

I concluded that _edata needs to be extended to include the ELF relocations as well.

After I decided to place it after the relocations, I noticed, that _edata is already specified at the same (new position) in gnu-efi.
This reaffirmed me to contribute the patch as it currently is.

I assumed, that on x86, that the linker treats the ELF relocations (.rela.*) as if it was part of the data section. From my understanding the EFI relocations are in .reloc while the ELF relocations are in .rela.*.
I might be mistaken, however.

Greetings,
Patrick

Am 06.11.22 um 10:20 schrieb Heinrich Schuchardt:
On 10/31/22 21:01, Patrick Zacharias wrote:
Prior to this commit, the relocations would not get loaded by the efi
loader.

This lead to none of the relocations being applied.

Signed-off-by: Fighter19 <1475802+fighte...@users.noreply.github.com>

Thanks Patrick for your contribution.

You can use scripts/get_maintainer.pl to determine to whom a patch
should be sent.

Where did you actually see relocations?
Which code is not position independent?

---
  arch/arm/lib/elf_aarch64_efi.lds | 2 +-
  arch/arm/lib/elf_arm_efi.lds     | 2 +-
  2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/arm/lib/elf_aarch64_efi.lds
b/arch/arm/lib/elf_aarch64_efi.lds
index c0604dad46..1982864d17 100644
--- a/arch/arm/lib/elf_aarch64_efi.lds
+++ b/arch/arm/lib/elf_aarch64_efi.lds
@@ -46,12 +46,12 @@ SECTIONS
          *(COMMON)
          . = ALIGN(512);
          _bss_end = .;
-        _edata = .;
      }
      .rela.dyn : { *(.rela.dyn) }
      .rela.plt : { *(.rela.plt) }
      .rela.got : { *(.rela.got) }
      .rela.data : { *(.rela.data) *(.rela.data*) }
+    _edata = .;
      _data_size = . - _etext;

      . = ALIGN(4096);
diff --git a/arch/arm/lib/elf_arm_efi.lds b/arch/arm/lib/elf_arm_efi.lds
index 767ebda635..c1b58a8033 100644
--- a/arch/arm/lib/elf_arm_efi.lds
+++ b/arch/arm/lib/elf_arm_efi.lds
@@ -46,12 +46,12 @@ SECTIONS
          *(COMMON)
          . = ALIGN(512);
          _bss_end = .;
-        _edata = .;
      }
      .rel.dyn : { *(.rel.dyn) }
      .rel.plt : { *(.rel.plt) }
      .rel.got : { *(.rel.got) }
      .rel.data : { *(.rel.data) *(.rel.data*) }
+    _edata = .;

Relocations (if they exist) should be in the .reloc section, not in the
.data section.

If we want to create a .reloc section, we have to change
arch/arm/lib/crt0_*_efi.S too. Furthermore the relocation section must
be pointed to by field BaseRelocationTable of the Optional Header Data
Directories (see PE-COFF specification).

Please, consider the other UEFI architectures (x86 and RISC-V) too.

Best regards

Heinrich

      _data_size = . - _etext;

      /DISCARD/ : {


Reply via email to