On Wed, 26 Aug 2026 at 15:05, Michal Simek <[email protected]> wrote:
>
> platform_id and platform_version are plain uninitialized statics,
> placing them in .bss. U-Boot's linker script overlays .bss at the
> same address as .rela.dyn (the relocation table), which is safe only
> if nothing writes to .bss before relocation completes.
Yes and we also need something the documentation explaining that since
it's hard to debug
>
> soc_detection() writes these variables during early board_init_f(),
> well before relocation, corrupting live .rela.dyn entries. When
> relocate_code() later reads the corrupted entry, it writes to an
> invalid, unaligned address. QEMU 8.x tolerated this silently, QEMU
> 10.x enforces alignment checks and traps it, causing U-Boot to hang
> right after printing "DRAM: 2 GiB", never reaching the console
> prompt.
>
> Move both variables to .data via __section(".data") so they no longer
> share an address with the relocation table.
Reviewed-by: Ilias Apalodimas <[email protected]>
>
> Fixes: 40f5046c221a ("arm64: versal2: Add support for AMD Versal Gen 2")
> Signed-off-by: Michal Simek <[email protected]>
> ---
>
> arch/arm/mach-versal2/cpu.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm/mach-versal2/cpu.c b/arch/arm/mach-versal2/cpu.c
> index d72f66f4fbac..19baa1331bdd 100644
> --- a/arch/arm/mach-versal2/cpu.c
> +++ b/arch/arm/mach-versal2/cpu.c
> @@ -241,7 +241,8 @@ void versal2_timer_setup(void)
> debug("timer 0x%llx\n", get_ticks());
> }
>
> -static u32 platform_id, platform_version;
> +static u32 platform_id __section(".data");
> +static u32 platform_version __section(".data");
>
> char *soc_name_decode(void)
> {
> ---
> base-commit: 964ad5b5c91b7be56e443e899d7f873e6aa8c9fc
> branch: xnext/data_section
>
> --
> 2.43.0
>
