Re: [PATCH] net_sched: force endianness annotation

[Edward Cree <ecree@solarflare.com>]

On 28/04/2019 06:54, Nicholas Mc Guire wrote:
> While the endiannes is being handled correctly sparse was unhappy with
> the missing annotation as be16_to_cpu()/be32_to_cpu() expects a __be16
> respectively __be32.
[...]
> diff --git a/net/sched/em_cmp.c b/net/sched/em_cmp.c
> index 1c8360a..3045ee1 100644
> --- a/net/sched/em_cmp.c
> +++ b/net/sched/em_cmp.c
> @@ -41,7 +41,7 @@ static int em_cmp_match(struct sk_buff *skb, struct tcf_ematch *em,
>  		val = get_unaligned_be16(ptr);
>  
>  		if (cmp_needs_transformation(cmp))
> -			val = be16_to_cpu(val);
> +			val = be16_to_cpu((__force __be16)val);
>  		break;
There should probably be a comment here to explain what's going on.  TBH
 it's probably a good general rule that any use of __force should have a
 comment explaining why it's needed.
AFAICT, get_unaligned_be16(ptr) is (barring alignment) equivalent to
 be16_to_cpu(*(__be16 *)ptr).  But then calling be16_to_cpu() again on
 val is bogus; it's already CPU endian.  There's a distinct lack of
 documentation around as to the intended semantics of TCF_EM_CMP_TRANS,
 but it would seem either (__force u16)cpu_to_be16(val); (which preserves
 the existing semantics, that trans is a no-op on BE) or swab16(val);
 would make more sense.

-Ed