Re: [PATCH 3/3] csu: Move static pie self relocation later [BZ #27072]

[Szabolcs Nagy <szabolcs.nagy@arm.com>]

The 01/07/2021 13:02, Szabolcs Nagy via Libc-alpha wrote:
> this is about static pie: sorting relocs does not help, the
> problem is not that ifunc resolvers have some relocs in them,
> but that the target specific relocation processing is one
> monolithic switch in a loop that handles all relocs in one go,
> there is no api to request only a subset of relocs to be
> processed (either by type or ordering in the list of relocs).
> 
> in __libc_start_main we need to do
> 
>   do_relative_relocs();
>   setup_auxv_tunables_cpu_etc();
>   do_irelative_relocs();

i just realized we can do this because RELATIVE and IRELATIVE
happen to be in different places (one where DT_REL{A} points
to and another where DT_JUMPREL points to).

i dont know how reliable this is across targets, but we only
need this to work in static pie (which is only supported on
a small number of targets as far as i know).

so i can try to tease _dl_relocate_static_pie apart into
'normal' reloc and 'plt' reloc processing (normally this
is done so that plt relocs can be lazy evaluated, but here
we would do it so IRELATIVE is processed in a separate step).

it still sounds like a big hack that relies on where
IRELATIVE is placed with current linkers and that there
are no other relocs we need to care about in static pie.
i can create a patch to see if it works.