Re: [PATCH/RFC v2 bpf-next 06/19] bpf: mark lo32 writes that should be zero extended into hi32

[Jakub Kicinski]

On Wed, 10 Apr 2019 20:50:20 +0100, Jiong Wang wrote:
> @@ -2169,6 +2188,12 @@ static int check_mem_access(struct bpf_verifier_env 
> *env, int insn_idx, u32 regn
>                                                   value_regno);
>                               if (reg_type_may_be_null(reg_type))
>                                       regs[value_regno].id = ++env->id_gen;
> +                             /* A load of ctx field could have different
> +                              * actual load size with the one encoded in the
> +                              * insn. When the dst is PTR, it is for sure not
> +                              * a sub-register.
> +                              */
> +                             regs[value_regno].subreg_def = DEF_NOT_SUBREG;

Can't the rewrite generate a 32bit load?  E.g. reading skb->len will be
LDX | W, we still gotta clear the top bits in that case, no?

I can't find the explanation of this case with a quick scan of the code
and cover letter..

>                       }
>                       regs[value_regno].type = reg_type;
>               }