On Wed, Jan 13, 2021 at 09:40:47AM +0800, Jia He wrote:
> There is a 10s stall in idmap_kpti_install_ng_mappings when kernel boots
> on a Ampere EMAG server.
>
> Commit f992b4dfd58b ("arm64: kpti: Add ->enable callback to remap
> swapper using nG mappings") updates the nG bit runtime if kpti is required.
> But things get worse if rodata=full in map_mem(). NO_BLOCK_MAPPINGS |
> NO_CONT_MAPPINGS is required when creating pagetable mapping. Hence all
> ptes are fully mapped in this case. On a Ampere EMAG server with 256G
> memory(pagesize=4k), it causes the 10s stall.
>
> After previous commit moving init_cpu_features(), we can use
> cpu_have_const_cap earlier than before. Hence we can avoid this stall
> by updating arm64_use_ng_mappings.
>
> Signed-off-by: Jia He
> ---
> arch/arm64/kernel/setup.c | 4
> 1 file changed, 4 insertions(+)
>
> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c
> index e078ab068f3b..51098ceb7159 100644
> --- a/arch/arm64/kernel/setup.c
> +++ b/arch/arm64/kernel/setup.c
> @@ -306,6 +306,10 @@ void __init __no_sanitize_address setup_arch(char
> **cmdline_p)
> /* Init the cpu feature codes for boot cpu */
> cpuinfo_store_boot_cpu();
>
> + /* ARM64_UNMAP_KERNEL_AT_EL0 cap can be updated in
> cpuinfo_store_boot_cpu() */
> + if (!arm64_use_ng_mappings)
> + arm64_use_ng_mappings =
> cpus_have_const_cap(ARM64_UNMAP_KERNEL_AT_EL0);
Are you sure it's safe to run the cpu feature initialisation code this
early? For example, we haven't even parsed the command-line yet, so I think
a fair amount of stuff will break.
Of course, you could also just pass "mitigations=off" if you want your
performance back.
Will
