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

[Martin KaFai Lau]

On Mon, Jan 24, 2022 at 05:03:20PM -0800, Alexei Starovoitov wrote:
> On Mon, Jan 24, 2022 at 4:35 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
> >
> > On Thu, Jan 20, 2022 at 09:17:27PM -0800, Alexei Starovoitov wrote:
> > > On Thu, Jan 20, 2022 at 6:18 AM Menglong Dong <menglong8.dong@xxxxxxxxx> wrote:
> > > >
> > > > On Thu, Jan 20, 2022 at 12:17 PM Alexei Starovoitov
> > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, Jan 20, 2022 at 11:02:27AM +0800, Menglong Dong wrote:
> > > > > > Hello!
> > > > > >
> > > > > > On Thu, Jan 20, 2022 at 6:03 AM Alexei Starovoitov
> > > > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > > > >
> > > > > > [...]
> > > > > > >
> > > > > > > Looks like
> > > > > > >  __sk_buff->remote_port
> > > > > > >  bpf_sock_ops->remote_port
> > > > > > >  sk_msg_md->remote_port
> > > > > > > are doing the right thing,
> > > > > > > but bpf_sock->dst_port is not correct?
> > > > > > >
> > > > > > > I think it's better to fix it,
> > > > > > > but probably need to consolidate it with
> > > > > > > convert_ctx_accesses() that deals with narrow access.
> > > > > > > I suspect reading u8 from three flavors of 'remote_port'
> > > > > > > won't be correct.
> > > > > >
> > > > > > What's the meaning of 'narrow access'? Do you mean to
> > > > > > make 'remote_port' u16? Or 'remote_port' should be made
> > > > > > accessible with u8? In fact, '*((u16 *)&skops->remote_port + 1)'
> > > > > > won't work, as it only is accessible with u32.
> > > > >
> > > > > u8 access to remote_port won't pass the verifier,
> > > > > but u8 access to dst_port will.
> > > > > Though it will return incorrect data.
> > > > > See how convert_ctx_accesses() handles narrow loads.
> > > > > I think we need to generalize it for different endian fields.
> > > >
> > > > Yeah, I understand narrower load in convert_ctx_accesses()
> > > > now. Seems u8 access to dst_port can't pass the verifier too,
> > > > which can be seen form bpf_sock_is_valid_access():
> > > >
> > > > $    switch (off) {
> > > > $    case offsetof(struct bpf_sock, state):
> > > > $    case offsetof(struct bpf_sock, family):
> > > > $    case offsetof(struct bpf_sock, type):
> > > > $    case offsetof(struct bpf_sock, protocol):
> > > > $    case offsetof(struct bpf_sock, dst_port):  // u8 access is not allowed
> > > > $    case offsetof(struct bpf_sock, src_port):
> > > > $    case offsetof(struct bpf_sock, rx_queue_mapping):
> > > > $    case bpf_ctx_range(struct bpf_sock, src_ip4):
> > > > $    case bpf_ctx_range_till(struct bpf_sock, src_ip6[0], src_ip6[3]):
> > > > $    case bpf_ctx_range(struct bpf_sock, dst_ip4):
> > > > $    case bpf_ctx_range_till(struct bpf_sock, dst_ip6[0], dst_ip6[3]):
> > > > $        bpf_ctx_record_field_size(info, size_default);
> > > > $        return bpf_ctx_narrow_access_ok(off, size, size_default);
> > > > $    }
> > > >
> > > > I'm still not sure what should we do now. Should we make all
> > > > remote_port and dst_port narrower accessable and endianness
> > > > right? For example the remote_port in struct bpf_sock_ops:
> > > >
> > > > --- a/net/core/filter.c
> > > > +++ b/net/core/filter.c
> > > > @@ -8414,6 +8414,7 @@ static bool sock_ops_is_valid_access(int off, int size,
> > > >                                 return false;
> > > >                         info->reg_type = PTR_TO_PACKET_END;
> > > >                         break;
> > > > +               case bpf_ctx_range(struct bpf_sock_ops, remote_port):
> > >
> > > Ahh. bpf_sock_ops don't have it.
> > > But bpf_sk_lookup and sk_msg_md have it.
> > >
> > > bpf_sk_lookup->remote_port
> > > supports narrow access.
> > >
> > > When it accesses sport from bpf_sk_lookup_kern.
> > >
> > > and we have tests that do u8 access from remote_port.
> > > See verifier/ctx_sk_lookup.c
> > >
> > > >                 case offsetof(struct bpf_sock_ops, skb_tcp_flags):
> > > >                         bpf_ctx_record_field_size(info, size_default);
> > > >                         return bpf_ctx_narrow_access_ok(off, size,
> > > >
> > > > If remote_port/dst_port are made narrower accessable, the
> > > > result will be right. Therefore, *((u16*)&sk->remote_port) will
> > > > be the port with network byte order. And the port in host byte
> > > > order can be get with:
> > > > bpf_ntohs(*((u16*)&sk->remote_port))
> > > > or
> > > > bpf_htonl(sk->remote_port)
> > >
> > > So u8, u16, u32 will work if we make them narrow-accessible, right?
> > >
> > > The summary if I understood it:
> > > . only bpf_sk_lookup->remote_port is doing it correctly for u8,u16,u32 ?
> > > . bpf_sock->dst_port is not correct for u32,
> > >   since it's missing bpf_ctx_range() ?
> > > . __sk_buff->remote_port
> > >  bpf_sock_ops->remote_port
> > >  sk_msg_md->remote_port
> > >  correct for u32 access only. They don't support narrow access.
> > >
> > > but wait
> > > we have a test for bpf_sock->dst_port in progs/test_sock_fields.c.
> > > How does it work then?
> > >
> > > I think we need more eyes on the problem.
> > > cc-ing more experts.
> > iiuc,  I think both bpf_sk_lookup and bpf_sock allow narrow access.
> > bpf_sock only allows ((__u8 *)&bpf_sock->dst_port)[0] but
> > not ((__u8 *)&bpf_sock->dst_port)[1].  bpf_sk_lookup allows reading
> > a byte at [0], [1], [2], and [3].
> >
> > The test_sock_fields.c currently works because it is comparing
> > with another __u16: "sk->dst_port == srv_sa6.sin6_port".
> > It should also work with bpf_ntohS() which usually is what the
> > userspace program expects when dealing with port instead of using bpf_ntohl()?
> > Thus, I think we can keep the lower 16 bits way that bpf_sock->dst_port
> > and bpf_sk_lookup->remote_port (and also bpf_sock_addr->user_port ?) are
> > using.  Also, changing it to the upper 16 bits will break existing
> > bpf progs.
> >
> > For narrow access with any number of bytes at any offset may be useful
> > for IP[6] addr.  Not sure about the port though.  Ideally it should only
> > allow sizeof(__u16) read at offset 0.  However, I think at this point it makes
> > sense to make them consistent with how bpf_sk_lookup does it also,
> > i.e. allow byte [0], [1], [2], and [3] access.
> 
> Sounds like the proposal is to do:
> diff --git a/net/core/filter.c b/net/core/filter.c
> index a06931c27eeb..1a8c97bc1927 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -8276,9 +8276,9 @@ bool bpf_sock_is_valid_access(int off, int size,
> enum bpf_access_type type,
>         case offsetof(struct bpf_sock, family):
>         case offsetof(struct bpf_sock, type):
>         case offsetof(struct bpf_sock, protocol):
> -       case offsetof(struct bpf_sock, dst_port):
>         case offsetof(struct bpf_sock, src_port):
>         case offsetof(struct bpf_sock, rx_queue_mapping):
> +       case bpf_ctx_range(struct bpf_sock, dst_port):
>         case bpf_ctx_range(struct bpf_sock, src_ip4):
> 
> and then document bpf_sock->dst_port and bpf_sk_lookup->remote_port
also bpf_sock_addr->user_port

> behavior and their difference vs
>   __sk_buff->remote_port
>   bpf_sock_ops->remote_port
>   sk_msg_md->remote_port
> ?
Yes, agree on the code change and adding doc.

> I suspect we cannot remove lshift_16 from them either,
> since it might break some prog as well.
Right, I believe the existing lshift_16 has to stay.