Re: [PATCH] ARM: initialize jump labels before setup_machine_fdt()

[Ard Biesheuvel]

On Thu, 2 Jun 2022 at 23:22, Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
>
> Stephen reported that a static key warning splat appears during early
> boot on arm64 systems that credit randomness from device trees that
> contain an "rng-seed" property, because setup_machine_fdt() is called
> before jump_label_init() during setup_arch(), which was fixed by
> 73e2d827a501 ("arm64: Initialize jump labels before
> setup_machine_fdt()").
>
> Upon cursory inspection, the same basic issue appears to apply to arm32
> as well. In this case, we reorder setup_arch() to do things in the same
> order as is now the case on arm64.
>
> Reported-by: Stephen Boyd <swboyd@xxxxxxxxxxxx>
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Ard Biesheuvel <ardb@xxxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: f5bda35fba61 ("random: use static branch for crng_ready()")

Wouldn't it be better to defer the
static_branch_enable(&crng_is_ready) call to later in the boot (e.g.,
using an initcall()), rather than going around 'fixing' fragile,
working early boot code across multiple architectures?

> Signed-off-by: Jason A. Donenfeld <Jason@xxxxxxxxx>
> ---
>  arch/arm/kernel/setup.c | 12 ++++++------
>  1 file changed, 6 insertions(+), 6 deletions(-)
>
> diff --git a/arch/arm/kernel/setup.c b/arch/arm/kernel/setup.c
> index 1e8a50a97edf..ef40d9f5d5a7 100644
> --- a/arch/arm/kernel/setup.c
> +++ b/arch/arm/kernel/setup.c
> @@ -1097,10 +1097,15 @@ void __init setup_arch(char **cmdline_p)
>         const struct machine_desc *mdesc = NULL;
>         void *atags_vaddr = NULL;
>
> +       setup_initial_init_mm(_text, _etext, _edata, _end);
> +       setup_processor();
> +       early_fixmap_init();
> +       early_ioremap_init();
> +       jump_label_init();
> +

Is it really necessary to reorder all these calls? What does
jump_label_init() actually need?

If this is related to the code patching, I wonder whether it wouldn't
be better not to rewrite all the NOPs (this is a x86-ism as every new
x86 uarch appears to have a better [faster?] NOP than the previous
one)

The issue with changes like these is that we might end up with bug
report in ~3 months' time that 'obscure platform X no longer boots or
produces any output'. In the best case, we'll have a bisect report
identifying this patch, but we won't be able to simply revert it as it
would reintroduce this issue into a kernel that is now stable.



>         if (__atags_pointer)
>                 atags_vaddr = FDT_VIRT_BASE(__atags_pointer);
>
> -       setup_processor();
>         if (atags_vaddr) {
>                 mdesc = setup_machine_fdt(atags_vaddr);
>                 if (mdesc)
> @@ -1125,15 +1130,10 @@ void __init setup_arch(char **cmdline_p)
>         if (mdesc->reboot_mode != REBOOT_HARD)
>                 reboot_mode = mdesc->reboot_mode;
>
> -       setup_initial_init_mm(_text, _etext, _edata, _end);
> -
>         /* populate cmd_line too for later use, preserving boot_command_line */
>         strlcpy(cmd_line, boot_command_line, COMMAND_LINE_SIZE);
>         *cmdline_p = cmd_line;
>
> -       early_fixmap_init();
> -       early_ioremap_init();
> -
>         parse_early_param();
>
>  #ifdef CONFIG_MMU
> --
> 2.35.1
>