Re: [PATCH bpf-next] bpf: Add document for 'dst_port' of 'struct bpf_sock'

[Jakub Sitnicki <jakub@cloudflare.com>]

On Wed, Jan 26, 2022 at 12:53 AM CET, Martin KaFai Lau wrote:
> On Tue, Jan 25, 2022 at 03:02:37PM -0800, Alexei Starovoitov wrote:
>> On Tue, Jan 25, 2022 at 2:45 PM Martin KaFai Lau <kafai@fb.com> wrote:
>> >
>> > On Tue, Jan 25, 2022 at 08:24:27PM +0100, Jakub Sitnicki wrote:
>> > > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> > > > index b0383d371b9a..891a182a749a 100644
>> > > > --- a/include/uapi/linux/bpf.h
>> > > > +++ b/include/uapi/linux/bpf.h
>> > > > @@ -5500,7 +5500,11 @@ struct bpf_sock {
>> > > >     __u32 src_ip4;
>> > > >     __u32 src_ip6[4];
>> > > >     __u32 src_port;         /* host byte order */
>> > > > -   __u32 dst_port;         /* network byte order */
>> > > > +   __u32 dst_port;         /* low 16-bits are in network byte order,
>> > > > +                            * and high 16-bits are filled by 0.
>> > > > +                            * So the real port in host byte order is
>> > > > +                            * bpf_ntohs((__u16)dst_port).
>> > > > +                            */
>> > > >     __u32 dst_ip4;
>> > > >     __u32 dst_ip6[4];
>> > > >     __u32 state;
>> > >
>> > > I'm probably missing something obvious, but is there anything stopping
>> > > us from splitting the field, so that dst_ports is 16-bit wide?
>> > >
>> > > I gave a quick check to the change below and it seems to pass verifier
>> > > checks and sock_field tests.
>> > >
>> > > IDK, just an idea. Didn't give it a deeper thought.
>> > >
>> > > --8<--
>> > >
>> > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>> > > index 4a2f7041ebae..344d62ccafba 100644
>> > > --- a/include/uapi/linux/bpf.h
>> > > +++ b/include/uapi/linux/bpf.h
>> > > @@ -5574,7 +5574,8 @@ struct bpf_sock {
>> > >       __u32 src_ip4;
>> > >       __u32 src_ip6[4];
>> > >       __u32 src_port;         /* host byte order */
>> > > -     __u32 dst_port;         /* network byte order */
>> > > +     __u16 unused;
>> > > +     __u16 dst_port;         /* network byte order */
>> > This will break the existing bpf prog.
>> 
>> I think Jakub's idea is partially expressed:
>> +       case offsetof(struct bpf_sock, dst_port):
>> +               bpf_ctx_record_field_size(info, sizeof(__u16));
>> +               return bpf_ctx_narrow_access_ok(off, size, sizeof(__u16));
>> 
>> Either 'unused' needs to be after dst_port or
>> bpf_sock_is_valid_access() needs to allow offset at 'unused'
>> and at 'dst_port'.
>> And allow u32 access though the size is actually u16.
>> Then the existing bpf progs (without recompiling) should work?
> Yes, I think that should work with the existing bpf progs.
> I suspect putting 'dst_port' first and then followed by 'unused'
> may be easier.  That will also serve as a natural doc for the
> current behavior (the value is in the lower 16 bits).

You're right. I can't count. Now fixed in [1].

>
> It can be extended to bpf_sk_lookup? bpf_sk_lookup can read at any
> offset of these 4 bytes, so may need to read 0 during
> convert_ctx_accesses?

Let's see what the feedback to [1] will be.

[1] https://lore.kernel.org/bpf/20220127172448.155686-1-jakub@cloudflare.com/T/#t