Re: [PATCH 0/5] riscv: Cleanup and deduplicate unaligned access speed probe

[Nam Cao]

Michael Neuling <mikey@xxxxxxxxxxx> writes:
>> This series was applied to riscv/linux.git (for-next)
>> by Paul Walmsley <pjw@xxxxxxxxxx>:
>
>> Here is the summary with links:
>>   - [1/5] riscv: Clean up & optimize unaligned scalar access probe
>>     https://git.kernel.org/riscv/c/c202d70b2244
>
> I think this is causing a regression (SHA1 actually 6455c6c11827) .
> Fast unaligned accesses are no longer being set ever.
>
> Analysis from Claude (Opus 4.6) with Chris Masons kernel patch review skills:

I should start using these AIs..

> --
>   > diff --git a/arch/riscv/kernel/unaligned_access_speed.c
> b/arch/riscv/kernel/unaligned_access_speed.c
>   > index b36a6a56f4..1f4c128d73 100644
>   > --- a/arch/riscv/kernel/unaligned_access_speed.c
>   > +++ b/arch/riscv/kernel/unaligned_access_speed.c
>
>   [ ... ]
>
>   > -arch_initcall(check_unaligned_access_all_cpus);
>   > +late_initcall(check_unaligned_access_all_cpus);
>
>   With this change, check_unaligned_access_all_cpus() now runs at
>   late_initcall (level 7), but lock_and_set_unaligned_access_static_branch()
>   remains at arch_initcall_sync (level 3s):
...
>   Does this mean fast_unaligned_access_speed_key is never enabled at boot,
>   even on hardware with fast unaligned access? The comment in
>   set_unaligned_access_static_branches() says "This will be called after
>   check_unaligned_access_all_cpus" which is no longer true with this
>   ordering change.

Thanks, you are indeed right. This affects do_csum()'s performance.

The below patch should resolve the issue. I will send a proper patch
later today after I have tested with my hardware.

Nam

diff --git a/arch/riscv/kernel/unaligned_access_speed.c b/arch/riscv/kernel/unaligned_access_speed.c
index 485ab1d105d3..96ba80e6ea32 100644
--- a/arch/riscv/kernel/unaligned_access_speed.c
+++ b/arch/riscv/kernel/unaligned_access_speed.c
@@ -244,7 +244,7 @@ static int __init lock_and_set_unaligned_access_static_branch(void)
 	return 0;
 }
 
-arch_initcall_sync(lock_and_set_unaligned_access_static_branch);
+late_initcall_sync(lock_and_set_unaligned_access_static_branch);
 
 static int riscv_online_cpu(unsigned int cpu)
 {