Re: [PATCH bpf-next 02/17] bpf: Introduce SK_LOOKUP program type with a dedicated attach point

[Jakub Sitnicki <jakub@cloudflare.com>]

On Thu, May 07, 2020 at 10:55 PM CEST, Martin KaFai Lau wrote:
> On Wed, May 06, 2020 at 03:53:35PM +0200, Jakub Sitnicki wrote:
>> On Wed, May 06, 2020 at 03:16 PM CEST, Lorenz Bauer wrote:
>> > On Wed, 6 May 2020 at 13:55, Jakub Sitnicki <jakub@cloudflare.com> wrote:
>>
>> [...]
>>
>> >> @@ -4012,4 +4051,18 @@ struct bpf_pidns_info {
>> >>         __u32 pid;
>> >>         __u32 tgid;
>> >>  };
>> >> +
>> >> +/* User accessible data for SK_LOOKUP programs. Add new fields at the end. */
>> >> +struct bpf_sk_lookup {
>> >> +       __u32 family;           /* AF_INET, AF_INET6 */
>> >> +       __u32 protocol;         /* IPPROTO_TCP, IPPROTO_UDP */
>> >> +       /* IP addresses allows 1, 2, and 4 bytes access */
>> >> +       __u32 src_ip4;
>> >> +       __u32 src_ip6[4];
>> >> +       __u32 src_port;         /* network byte order */
>> >> +       __u32 dst_ip4;
>> >> +       __u32 dst_ip6[4];
>> >> +       __u32 dst_port;         /* host byte order */
>> >
>> > Jakub and I have discussed this off-list, but we couldn't come to an
>> > agreement and decided to invite
>> > your opinion.
>> >
>> > I think that dst_port should be in network byte order, since it's one
>> > less exception to the
>> > rule to think about when writing BPF programs.
>> >
>> > Jakub's argument is that this follows __sk_buff->local_port precedent,
>> > which is also in host
>> > byte order.
>>
>> Yes, would be great to hear if there is a preference here.
>>
>> Small correction, proposed sk_lookup program doesn't have access to
>> __sk_buff, so perhaps that case matters less.
>>
>> bpf_sk_lookup->dst_port, the packet destination port, is in host byte
>> order so that it can be compared against bpf_sock->src_port, socket
>> local port, without conversion.
>>
>> But I also see how it can be a surprise for a BPF user that one field has
>> a different byte order.
> I would also prefer port and addr were all in the same byte order.
> However, it is not the cases for the other prog_type ctx.
> People has stomped on it from time to time.  May be something
> can be done at the libbpf to hide this difference.
>
> I think uapi consistency with other existing ctx is more important here.
> (i.e. keep the "local" port in host order).  Otherwise, the user will
> be slapped left and right when writting bpf_prog in different prog_type.
>
> Armed with the knowledge on skc_num, the "local" port is
> in host byte order in the current existing prog ctx.  It is
> unfortunate that the "dst"_port in this patch is the "local" port.
> The "local" port in "struct bpf_sock" is actually the "src"_port. :/
> Would "local"/"remote" be clearer than "src"/dst" in this patch?

I went and compared the field naming and byte order in existing structs:

  | struct         | field      | byte order |
  |----------------+------------+------------|
  | __sk_buff      | local_port | host       |
  | sk_msg_md      | local_port | host       |
  | bpf_sock_ops   | local_port | host       |
  | bpf_sock       | src_port   | host       |
  | bpf_fib_lookup | dport      | network    |
  | bpf_flow_keys  | dport      | network    |
  | bpf_sock_tuple | dport      | network    |
  | bpf_sock_addr  | user_port  | network    |

It does look like "local"/"remote" prefix is the sensible choice.

I got carried away trying to match the field names with bpf_sock, which
actually doesn't follow the naming convention.

Will rename fields to local_*, remote_* in v2.