U-boot version selection

[Joel Stanley]

On Thu, 8 Jul 2021 at 19:19, Zev Weiss <zweiss at equinix.com> wrote:
>
> On Wed, Jun 30, 2021 at 11:51:49PM CDT, Joel Stanley wrote:
> >egacOn Thu, 1 Jul 2021 at 02:48, Zev Weiss <zweiss at equinix.com> wrote:
> >>
> >> Hi,
> >>
> >> I recently found myself needing to make some tweaks to u-boot to
> >> accommodate a new board I'm targeting with a larger flash part, but in
> >> going to do so I remembered that I'm currently using u-boot v2016.7,
> >> whereas new development is strongly encouraged to use v2019.04 [1].
> >>
> >> As far as I know that happened entirely by default (i.e. I didn't go out
> >> of my way to use the older version), so I hunted around a bit for how to
> >> override that to use the newer one, but wasn't able to find anything
> >> obvious.  What's the recommended way to go about switching that for my
> >> board?
> >
> >You can see Lei's change to use the newer tree here:
> >
> > https://github.com/openbmc/openbmc/commit/1aa72efd0f54
> >
> >UBOOT_DEVICETREE = "ast2500-evb"
> >UBOOT_MACHINE = "evb-ast2500_defconfig"
> >
> >PREFERRED_PROVIDER_u-boot = "u-boot-aspeed-sdk"
> >PREFERRED_PROVIDER_u-boot-fw-utils = "u-boot-fw-utils-aspeed-sdk"
> >PREFERRED_PROVIDER_virtual/bootloader = "u-boot-aspeed-sdk"
> >
> >The important change is to point it to a valid defconfig for the new
> >tree, to specify the u-boot device tree to use, and to change some
> >yocto PROVIDER variables to use the "u-boot-aspeed-sdk" variant.
> >
>
> Great, thanks for the pointers there -- that worked smoothly on the 64M
> board I'm currently working on a port for.  Unfortunately when I tried
> it out on one of the 32M platforms I've got I realized that the u-boot
> image with the new branch comes out to around 430K, which in addition to
> being quite a bit larger than the old branch (~225K), is too big to fit
> in the u-boot partition of the 32MB static flash layout (384K).
>
> With a bit of haphazard experimentation, I found that disabling
> CONFIG_EFI_LOADER got it down to 370K, and turning off
> CONFIG_SYS_LONGHELP reduced that to 357K, leaving a bit of breathing
> room, though perhaps still less than would be ideal.  Since a quick poll
> of FLASH_SIZE settings seems to indicate that most existing OpenBMC
> platforms are 32M, would those be appropriate candidates to add to the
> evb-ast2500_defconfig?

Yes, that would make sense. Confirming that this set got it down to 345 KB:

+# CONFIG_SYS_LONGHELP is not set
+# CONFIG_CMD_ELF is not set
+# CONFIG_CMD_IMI is not set
+# CONFIG_CMD_XIMG is not set
+# CONFIG_REGEX is not set
+# CONFIG_EFI_LOADER is not set

Are there any ast2500 boards that use MMC? If not, disabling MMC and
block devices gets it down to 321 KB:

-CONFIG_CMD_MMC=y
-CONFIG_DM_MMC=y
-# CONFIG_MMC_HW_PARTITIONING is not set
-CONFIG_MMC_SDHCI=y
-CONFIG_MMC_SDHCI_ASPEED=y
+# CONFIG_MMC is not set

If we disable MMC, we should probably create our own defconfig
(ast2500_openbmc_defconifg). If we just want to disable the unused
options, I think patching the evb defconfig is fine. I'll leave it to
you; please send a patch to the list and I'll merge it.

>
> (I also encountered an ftgmac100 phy-related null pointer dereference
> that leads to it spewing a bunch of garbage to the console, which I
> band-aided temporarily with a dts patch to disable the second mac.)

Thanks for the report.