On Sat, Jun 27, 2020 at 10:17:26AM -0700, Eric Dumazet wrote:
> On Fri, Jun 26, 2020 at 10:55 AM Martin KaFai Lau <kafai@xxxxxx> wrote:
> >
> > This patch adds logic to parse experimental kind 254 with 16 bit magic
> > 0xeB9F. The latter patch will allow bpf prog to write and parse data
> > under this experimental kind and magic.
> >
> > A one byte bpf_hdr_opt_off is added to tcp_skb_cb by using an existing
> > 4 byte hole. It is only used in rx. It stores the offset to the
> > bpf experimental option and will be made available to BPF prog
> > in a latter patch. This offset is also stored in the saved_syn.
> >
> > Signed-off-by: Martin KaFai Lau <kafai@xxxxxx>
> > ---
> > include/net/request_sock.h | 1 +
> > include/net/tcp.h | 3 +++
> > net/ipv4/tcp_input.c | 6 ++++++
> > net/ipv4/tcp_ipv4.c | 1 +
> > net/ipv6/tcp_ipv6.c | 1 +
> > 5 files changed, 12 insertions(+)
> >
> > diff --git a/include/net/request_sock.h b/include/net/request_sock.h
> > index d77237ec9fb4..55297286c066 100644
> > --- a/include/net/request_sock.h
> > +++ b/include/net/request_sock.h
> > @@ -43,6 +43,7 @@ int inet_rtx_syn_ack(const struct sock *parent, struct request_sock *req);
> >
> > struct saved_syn {
> > u32 network_hdrlen;
> > + u32 bpf_hdr_opt_off;
> > u8 data[];
> > };
> >
> > diff --git a/include/net/tcp.h b/include/net/tcp.h
> > index eab1c7d0facb..07a9dfe35242 100644
> > --- a/include/net/tcp.h
> > +++ b/include/net/tcp.h
> > @@ -191,6 +191,7 @@ void tcp_time_wait(struct sock *sk, int state, int timeo);
> > */
> > #define TCPOPT_FASTOPEN_MAGIC 0xF989
> > #define TCPOPT_SMC_MAGIC 0xE2D4C3D9
> > +#define TCPOPT_BPF_MAGIC 0xEB9F
> >
> > /*
> > * TCP option lengths
> > @@ -204,6 +205,7 @@ void tcp_time_wait(struct sock *sk, int state, int timeo);
> > #define TCPOLEN_FASTOPEN_BASE 2
> > #define TCPOLEN_EXP_FASTOPEN_BASE 4
> > #define TCPOLEN_EXP_SMC_BASE 6
> > +#define TCPOLEN_EXP_BPF_BASE 4
> >
> > /* But this is what stacks really send out. */
> > #define TCPOLEN_TSTAMP_ALIGNED 12
> > @@ -857,6 +859,7 @@ struct tcp_skb_cb {
> > has_rxtstamp:1, /* SKB has a RX timestamp */
> > unused:5;
> > __u32 ack_seq; /* Sequence number ACK'd */
> > + __u8 bpf_hdr_opt_off;/* offset to bpf hdr option. rx only. */
> > union {
> > struct {
> > /* There is space for up to 24 bytes */
> > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> > index eb0e32b2def9..640408a80b3d 100644
> > --- a/net/ipv4/tcp_input.c
> > +++ b/net/ipv4/tcp_input.c
> > @@ -3924,6 +3924,10 @@ void tcp_parse_options(const struct net *net,
> > tcp_parse_fastopen_option(opsize -
> > TCPOLEN_EXP_FASTOPEN_BASE,
> > ptr + 2, th->syn, foc, true);
> > + else if (opsize >= TCPOLEN_EXP_BPF_BASE &&
> > + get_unaligned_be16(ptr) ==
> > + TCPOPT_BPF_MAGIC)
> > + TCP_SKB_CB(skb)->bpf_hdr_opt_off = (ptr - 2) - (unsigned char *)th;
> > else
> > smc_parse_options(th, opt_rx, ptr,
> > opsize);
> > @@ -6562,6 +6566,8 @@ static void tcp_reqsk_record_syn(const struct sock *sk,
> > saved_syn = kmalloc(len + sizeof(*saved_syn), GFP_ATOMIC);
> > if (saved_syn) {
> > saved_syn->network_hdrlen = skb_network_header_len(skb);
> > + saved_syn->bpf_hdr_opt_off =
> > + TCP_SKB_CB(skb)->bpf_hdr_opt_off;
> > memcpy(saved_syn->data, skb_network_header(skb), len);
> > req->saved_syn = saved_syn;
> > }
> > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> > index ea0df9fd7618..a3535b7fe002 100644
> > --- a/net/ipv4/tcp_ipv4.c
> > +++ b/net/ipv4/tcp_ipv4.c
> > @@ -1864,6 +1864,7 @@ static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
> > TCP_SKB_CB(skb)->sacked = 0;
> > TCP_SKB_CB(skb)->has_rxtstamp =
> > skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
> > + TCP_SKB_CB(skb)->bpf_hdr_opt_off = 0;
> > }
> >
> > /*
> > diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
> > index f67d45ff00b4..8356d0562279 100644
> > --- a/net/ipv6/tcp_ipv6.c
> > +++ b/net/ipv6/tcp_ipv6.c
> > @@ -1545,6 +1545,7 @@ static void tcp_v6_fill_cb(struct sk_buff *skb, const struct ipv6hdr *hdr,
> > TCP_SKB_CB(skb)->sacked = 0;
> > TCP_SKB_CB(skb)->has_rxtstamp =
> > skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
> > + TCP_SKB_CB(skb)->bpf_hdr_opt_off = 0;
> > }
> >
> > INDIRECT_CALLABLE_SCOPE int tcp_v6_rcv(struct sk_buff *skb)
> > --
> > 2.24.1
> >
>
> (Sorry for the prior empty reply, accidentally click the wrong area)
>
> It seems strange that we want to add code in TCP stack only to cover a
> limited use case (kind 254 and 0xEB9F magic)
>
> For something like the work Petar Penkov did (to be able to generate
> SYNCOOKIES from XDP), we do not go through tcp_parse_options() and BPF
> program
> would have to implement its own parsing (without having an SKB at
> hand), probably calling a helper function, with no
> TCP_SKB_CB(skb)->bpf_hdr_opt_off.
>
> This patch is hard coding a specific option and will prevent anyone
> using private option(s) from using this infrastructure in the future,
> yet paying the extra overhead.
>
> TCP_SKB_CB(skb) is tight, I would prefer keeping the space in it for
> standard TCP stack features.
>
> If an optional BPF program needs to re-parse the TCP options to find a
> specific option, maybe the extra cost is noise (especially if this is
> only for SYN & SYNACK packets) ?
Thanks for the feedback.
Re: syn & synack only
The bpf tcp hdr option infrastructure is not only limited to syn
and synack. It is available to the data/ack/fin pkt also, although
most of the use cases may be limited to syn and synack.
e.g. the latter example tests parsing the 0xeB9F option in FIN.
After a connection is established, the bpf may choose to continue hearing
for (kind 254 and 0xEB9F magic). bpf_hdr_opt_off is also used to
decide if the tcphdr has the 0xeB9F option and then call the bpf prog
to handle it.
Re: the spaces in TCP_SKB_CB(skb). I think I can avoid tapping into it.
bpf_hdr_opt_off is only needed upto calling the bpf prog. i.e. after
the bpf prog returns, the bpf_hdr_opt_off is no longer needed in TCP_SKB_CB.
Like "struct tcp_fastopen_cookie *foc", "u8 *bpf_hdr_opt_off" can be
added to tcp_parse_options() instead of saving it in TCP_SKB_CB(skb).
Then pass it all the way to the bpf prog and also save this to "saved_syn".
Does it address the concern in the spaces in TCP_SKB_CB(skb)?
