[PATCH v1] efi_loader: Handle GD_FLG_SKIP_RELOC

[Ilias Apalodimas]

+CC Heinrich who has looked the relocation stuff a bit more

On Thu, 27 Mar 2025 at 23:22, Ilias Apalodimas
<ilias.apalodimas at linaro.org> wrote:
>
> On Thu, 27 Mar 2025 at 08:12, Varadarajan Narayanan
> <quic_varada at quicinc.com> wrote:
> >
> > On Wed, Mar 26, 2025 at 09:28:04AM +0200, Ilias Apalodimas wrote:
> > > Hi Varadarajan
> > >
> > > On Wed, 26 Mar 2025 at 07:47, Varadarajan Narayanan
> > > <quic_varada at quicinc.com> wrote:
> > > >
> > > > If the EFI runtime services pointers are relocated even though
> > > > relocation is skipped, it corrupts some other data resulting in some
> > > > unexpected behaviour.
> > > >
> > > > In this specific case, it overwrote some page table entries resulting in
> > > > the device memory address range's mappings getting removed. Eventually,
> > > > after the completion of efi_runtime_relocate(), when a driver tries to
> > > > access its device's registers it crashes since the mappings are absent.
> > >
> > > How were those page table entries added?
> >
> > initr_caches ->
> >         enable_caches ->
> >                 setup_pgtables (arch/arm/mach-snapdragon/board.c)
> >
> > > The runtime service relocation might have overwritten those now, but
> > > since they are not in protected memory this can happen arbitrarily.
> >
> > To avoid relocating U-Boot, I had OR'ed GD_FLG_SKIP_RELOC to gd->flags
> > in init_sequence_f -> mach_cpu_init(). Hope that is the right place and
> > way to do it.

Looking at it a bit more, doing it there seems correct.  This will
skip relocation in board_f functions

>
> I'll have a look at the relocation stuff, but off the top of my head,
> I don't remember.
>
> >
> > > > Signed-off-by: Varadarajan Narayanan <quic_varada at quicinc.com>
> > > > ---
> > > >  common/board_r.c | 3 ++-
> > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/common/board_r.c b/common/board_r.c
> > > > index 179259b00de..1dd3b96c2de 100644
> > > > --- a/common/board_r.c
> > > > +++ b/common/board_r.c
> > > > @@ -169,7 +169,8 @@ static int initr_reloc_global_data(void)
> > > >          */
> > > >         efi_save_gd();
> > > >
> > > > -       efi_runtime_relocate(gd->relocaddr, NULL);
> > > > +       if (!(gd->flags & GD_FLG_SKIP_RELOC))
> > > > +               efi_runtime_relocate(gd->relocaddr, NULL);
> > >
> > > Have you tested booting with EFI with this change?
> >
> > Yes, tested this with both GD_FLG_SKIP_RELOC set and not set.
> >
> > Used the command 'bootefi bootmgr' and Linux is able to boot.
>
> Have you tried to call any runtime services -- e.g efival -l will dump
> all the EFI variables

The efi runtime relocation data are not in rela.dyn, that data is in
.efi_runtime_rel. I don't know why we relocate runtime services later
than the initial relocation code, but skipping them here seems
correct.

Heinrich?

Thanks
/Ilias

>
> Thanks
> /Ilias
> >
> > Thanks
> > Varada