On Thu, Feb 17, 2022 at 05:11 PM +01, Jakub Sitnicki wrote:
> On Thu, Feb 17, 2022 at 03:18 PM +01, Ilya Leoshkevich wrote:
>> On Wed, 2022-02-16 at 13:44 -0800, Andrii Nakryiko wrote:
>>> On Wed, Feb 9, 2022 at 10:43 AM Jakub Sitnicki <jakub@xxxxxxxxxxxxxx>
>>> wrote:
>
> [...]
>
>>> > + /* Load from remote_port field with zero padding (backward
>>> > compatibility) */
>>> > + val_u32 = *(__u32 *)&ctx->remote_port;
>>> > + if (val_u32 != bpf_htonl(bpf_ntohs(SRC_PORT) << 16))
>>> > + return SK_DROP;
>>> > +
>>>
>>> Jakub, can you please double check that your patch set doesn't break
>>> big-endian architectures? I've noticed that our s390x test runner is
>>> now failing in the sk_lookup selftest. See [0]. Also CC'ing Ilya.
>>
>> I agree that this looks like an endianness issue. The new check seems
>> to make little sense on big-endian to me, so I would just #ifdef it
>> out.
>
> We have a very similar check for a load from context in
> progs/test_sock_fields.c, which is not causing problems:
>
> static __noinline bool sk_dst_port__load_word(struct bpf_sock *sk)
> {
> __u32 *word = (__u32 *)&sk->dst_port;
> return word[0] == bpf_htonl(0xcafe0000);
> }
>
> So I think I just messed something up here. Will dig into it.
Pretty sure the source of the problem here is undefined behaviour. Can't
legally shift u16 by 16 bits like I did in the `bpf_ntohs(SRC_PORT) <<
16` expression. Will fix.
