Re: [PATCH] BaseTools/GenFw ARM: don't permit R_ARM_GOT_PREL relocations

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On Tue, 11 Dec 2018 at 12:19, Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote:
>
> On Tue, 11 Dec 2018 at 10:53, Leif Lindholm <leif.lindholm@linaro.org> wrote:
> >
> > On Tue, Dec 11, 2018 at 10:37:15AM +0100, Ard Biesheuvel wrote:
> > > We currently permit R_ARM_GOT_PREL relocations in the ELF32 conversion
> > > routines, under the assumption that relative relocations are fine as
> > > long as the section layout is the same between ELF and PE/COFF.
> > >
> > > However, as is the case with any proxy generating relocation, it is
> > > up to the linker to emit an entry in the GOT table and populate it
> > > with the correct absolute address, which should also be fixed up at
> > > PE/COFF load time. Unfortunately, the relocations covering the GOT
> > > section are not emitted into the static relocation sections processed
> > > by GenFw, but only in the dynamic relocation section as a R_ARM_RELATIVE
> > > relocation, and so GenFw fails to emit the correct PE/COFF relocation
> > > data for GOT entries.
> > >
> > > Since GOT indirection is pointless anyway for PE/COFF modules running
> > > in UEFI context, let's just drop the references to R_ARM_GOT_PREL from
> > > GenFw, resulting in a build time failure rather than a runtime failure
> > > if such relocations do occur.
> > >
> > > Cc: Bob Feng <bob.c.feng@intel.com>
> > > Cc: Liming Gao <liming.gao@intel.com>
> > > Cc: Leif Lindholm <leif.lindholm@linaro.org>
> > > Contributed-under: TianoCore Contribution Agreement 1.1
> > > Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
> >
> > Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org>
> >
> > Ouch. This sounds like the best move for now. But how do we deal with
> > builds that actually break?
> >
>
> So the only builds that are breaking due to this are ones where we run
> the linker in PIE mode (which only happens in
> ArmVirtPkg/PrePi/ArmVirtPrePiUniCoreRelocatable.inf), and using the
> GNU gold linker. The reason we need the -pie option is to force the
> linker to emit dynamic relocations into the binary so it can relocate
> itself. This is necessary because the firmware image may execute from
> a a priori unknown memory offset.
>
> I am playing around with hidden visibility and other tweaks to coerce
> the linker into emitting direct relative references instead of GOT
> based ones, and it is very tedious. The GOLD linker really doesn't
> appear to be set up for bare metal binaries.

Oh, and on AARCH64 it is even more annoying, given that the relative
GOT references are emitted as ADRP/ADD pairs, which means we have the
4 KB alignment issue as well.