On 26.01.2026 18:53, Andrew Cooper wrote:
> early_cpu_init() calls intel_unlock_cpuid_leaves(), which arranges for the
> boot trampoline
And, not to neglect, also the S3 wakeup entry point.
> to make the adjustment for all APs, meaning the call from
> early_init_intel() is a no-op. Drop it, allowing intel_unlock_cpuid_leaves()
> to become __init code.
>
> In turn, the adjustments in generic_identify() were a no-op too. Nothing in
> the c_early_init() hooks modifies the 1c/1d features, so their values in
> c->x86_capability[] are still good from just above.
>
> The ebx variable used to calculate c->x86_clflush_size is still good too, but
> move the logic earlier so it's more obviously correct.
This part likely can be dropped when the pending adjustment to patch 1 is
done.
> Fixes: fa4d026737a4 ("x86/Intel: unlock CPUID earlier for the BSP")
> Signed-off-by: Andrew Cooper <[email protected]>
Suitably adjusted:
Reviewed-by: Jan Beulich <[email protected]>
Jan
