Hello Graeme,

On Wed, 04 Feb 2015 18:14:22 +1100, Graeme Russ gruss@tss-engineering.com wrote:

Hi Guys,

Since I did the initial 'full relocation' for x86 way back when (http://git.denx.de/?p=u-boot.git;a=commit;h=1c409bc7101a24ecd47a13a4e851845d...), there has been a lot of chatter about the ongoing need for CONFIG_NEEDS_MANUAL_RELOC. I think it's about time we looked really hard at this define and come to a definitive conclusion as to IF and HOW we can get rid of it for the remaining architectures that use it (being avr32, m68k, nds32, and sparc).

I second that.

First, a bit of a primer on why and how I finally killed off CONFIG_NEEDS_MANUAL_RELOC for x86...

Despite the best intentions of the 'manual' relocation performed by U-Boot, there were still instances where the U-Boot binary referenced the 'unrelocated' image which typically resides in flash (NOR flash in the case of the x86 board I was using). I only finally tracked this down through some arduous debugging of random crashes during the writing of a new U-Boot image to the NOR flash.

After a lot of digging around, I managed to get a good understanding of the ELF binary format for x86 and compiler/linker flags needed in order to generate a binary image that could be relocated. It turns out that for x86, it was the -pic and --emit-relocs ld flags (Oddly, I cannot find the -pic flag in the gnu ld documentation maybe it's depricated?)

When I did the ELF relocation for ARM, only -pie existed for ld; but it has an alias, --pic-executable. Maybe -pic is short for -pic-executable?

The --emit-relocs flags leaves the relocation sections (for x86 the section of interest is .rela.dyn) in the final image. These sections list all the locations in the final binary that need to be adjusted if the image is loaded at a different address than the one it is compiled for. You can see the processing in do_elf_reloc_fixups() in arch/x86/lib/relocate.c

For ARM I just specified the placement of relocation sections in the linker script; I'm surprised that x86 needs this --emit-relocs. Maybe that's needed when one does not provide a custom linker script?

Keep in mind that the relocation sections are NOT copied to RAM - they are only used once after U-Boot is copied to RAM to locate the memory locations that need to be adjusted.

So the crux of generic relocation is to generate a stripped binary with all necessary ELF relocation information located AFTER the text and data sections which are copied to RAM

So the first question that we need to answer is: Do the avr32, m68k, nds32, and sparc toolchains support the generation of the relevant sections needed to perform relocation in a manner similar to x86 (i.e generation of relocation references into sections of the final U-Boot binary)?

Those arches which use GCC should be able to use -pie / -pic-executable and possibly --emit-relocs, as none of these is arch-specific.

Then, each arch might have to look into what GCC options are needed. For the record, on ARM, I did not need any gcc option, but actually had one -fPIC option /removed/.

Regards,

Graeme

Amicalement,