Re: [PATCH v2 05/10] percpu: Wire up cmpxchg128

[Arnd Bergmann]

On Thu, Feb 2, 2023, at 15:50, Peter Zijlstra wrote:
> In order to replace cmpxchg_double() with the newly minted
> cmpxchg128() family of functions, wire it up in this_cpu_cmpxchg().
>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

I commented on this in the previous version but never got any
reply from you:

https://lore.kernel.org/all/1d88ba9f-5541-4b67-9cc8-a361eef36547@xxxxxxxxxxxxxxxx/

Unless I have misunderstood what you are doing, my concerns are
still the same:

>  #define this_cpu_cmpxchg(pcp, oval, nval) \
> -	__pcpu_size_call_return2(this_cpu_cmpxchg_, pcp, oval, nval)
> +	__pcpu_size16_call_return2(this_cpu_cmpxchg_, pcp, oval, nval)
>  #define this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, 
> nval2) \
>  	__pcpu_double_call_return_bool(this_cpu_cmpxchg_double_, pcp1, pcp2, 
> oval1, oval2, nval1, nval2)

Having a variable-length this_cpu_cmpxchg() that turns into cmpxchg128()
and cmpxchg64() even on CPUs where this traps (!X86_FEATURE_CX16) seems
like a bad design to me.

I would much prefer fixed-length this_cpu_cmpxchg64()/this_cpu_cmpxchg128()
calls that never trap but fall back to the generic version on CPUs that
are lacking the atomics.

     Arnd