Re: [PATCH] lds: align u-boot-nodtb with 8 bytes boundary

[Simon Glass <sjg@chromium.org>]

Hi Pali,

On Thu, 15 Dec 2022 at 16:13, Pali Rohár <pali@kernel.org> wrote:
>
> On Thursday 15 December 2022 06:24:16 Simon Glass wrote:
> > Hi Eugen,
> >
> > On Thu, 15 Dec 2022 at 03:58, Eugen Hristev <eugen.hristev@microchip.com> wrote:
> > >
> > > Newer DTC require that the DTB start address is aligned at 8 bytes.
> > > In the u-boot.bin case, the u-boot-nodtb.bin is concatenated with the
> > > DTB, but there is no alignment/padding to the next 8byte aligned address.
> > > This causes the board to fail booting, because the FDT will claim
> > > that the DTB inside u-boot.bin is not a valid DTB, it will fail with
> > > -FDT_ERR_ALIGNMENT.
> > > To solve this, have the u-boot binary `_end` aligned with 8 bytes.
> > > The objcopy in the Makefile will create the u-boot-nodtb.bin and it has to
> > > be truncated to 8 bytes to correspond to the u-boot.map file which will
> > > have the `_end` aligned to 8 bytes.
> > > The lds files which use devicetrees have been changed to align the `_end`
> > > tag with 8 bytes.
> > >
> > > This patch is also a prerequisite to have the possibility to update the
> > > dtc inside u-boot to newer versions (1.6.1+)
> > >
> > > Signed-off-by: Eugen Hristev <eugen.hristev@microchip.com>
> > > ---
> > > Hi,
> > >
> > > I could not test all affected boards, it's an impossible task.
> > > I tried this on at91 boards which I have, and ran the CI on denx.
> > > I cannot guarantee that no other boards are affected, so this patch is a bit
> > > of an RFC.
> > > If the u-boot-nodtb.bin does not have the size equal with the corresponding
> > > one in u-boot.map, the binary_size_check will fail at build time with
> > > something like this:
> > >
> > > u-boot.map shows a binary size of 502684
> > > but u-boot-nodtb.bin shows 502688
> > >
> > > Thanks,
> > > Eugen
> > >
> > >  Makefile                                    | 2 ++
> > >  arch/arm/cpu/armv8/u-boot.lds               | 4 ++--
> > >  arch/arm/cpu/u-boot-spl.lds                 | 1 +
> > >  arch/arm/cpu/u-boot.lds                     | 1 +
> > >  arch/arm/lib/elf_arm_efi.lds                | 5 +++++
> > >  arch/arm/mach-at91/arm926ejs/u-boot-spl.lds | 2 +-
> > >  arch/arm/mach-at91/armv7/u-boot-spl.lds     | 2 +-
> > >  arch/arm/mach-zynq/u-boot-spl.lds           | 2 +-
> > >  arch/mips/cpu/u-boot.lds                    | 2 +-
> > >  arch/sandbox/cpu/u-boot.lds                 | 6 ++++++
> > >  arch/sh/cpu/u-boot.lds                      | 2 ++
> > >  board/ti/am335x/u-boot.lds                  | 1 +
> > >  tools/binman/test/u_boot_binman_embed.lds   | 2 +-
> > >  13 files changed, 25 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/Makefile b/Makefile
> > > index 9d84f96481..b4d387bcce 100644
> > > --- a/Makefile
> > > +++ b/Makefile
> > > @@ -1317,6 +1317,8 @@ endif
> > >
> > >  u-boot-nodtb.bin: u-boot FORCE
> > >         $(call if_changed,objcopy_uboot)
> > > +# Make sure the size is 8 byte-aligned.
> > > +       @truncate -s %8 $@
> > >         $(BOARD_SIZE_CHECK)
> >
> > I agree this line is needed, since otherwise we will only get 4-byte
> > alignment.
>
> Hello! I do not fully agree that this line is needed.
>
> The whole issue is about DTB binary and its offset. So code/Makefile
> which construct u-boot.bin should be fixed and not u-boot-nodtb.bin.

While I agree that is true in theory, im practice we do actually need
the _image_binary_end symbol to point to the right place. It is
hopelessly confusing if u-boot-nodtb.bin is shorter than indicated by
that symbol, and this is why we have binary_size_check

So I believe that padding u-boot-nodtb.bin is the best solution.

With binman we can fairly easily pad to solve the problem, e.g. by
always adding 'align = <8>' to dtb entries, but when using 'cat' to
put images together, it is much easier if the original image has an
aligned size.

>
> > But it would be better if we could have the linker scripts
> > fill bytes out to the required alignment. Is that possible?
>
> I already investigated this. LD linker and objcopy (at least older
> version in Debian 10) drops trailing zero bytes in raw binary output.
>
> You can specify zero bytes in the linker script and they are filled in
> ELF or COFF output. But not in raw binary, which u-boot.bin is.

Is that really what is happening? If you use BYTE(0) does it actually
get dropped but with BYTE(1) it doesn't? That sounds very broken.

I thought it was actually because updating the current location
doesn't actually change what is there, e.g.:

   . = . + 4

just adds to the location pointer, whereas

   . = . + 4
   BYTE(0)

actually adds a byte there. But I do find it confusing so it would be
great to know how to do this properly.

>
> So it could be possible to extract "correct" size from ELF binary and
> call truncate on raw binary generated from objcopy.
>
>
>
> I already hit this trailing-zeros issue for powerpc and I fixed it in:
> 7696b80ec5e9 powerpc: mpc85xx: Fix loading U-Boot proper from SD card in SPL
> b898f6a6db76 powerpc: mpc85xx: Fix CONFIG_OF_SEPARATE support for NOR booting
> e8c0e0064c8a powerpc: mpc85xx: Fix CONFIG_OF_SEPARATE support
>
> All those commits re-align output to just 4 bytes, not more.
>
> So Makefile change in this proposed patch is going to break all
> powerpc/mpc85xx boards.

Regards,
Simon