> -----Original Message-----
> From: Xing, Beilei
> Sent: Friday, September 29, 2017 7:30 PM
> To: Sean Harte <sea...@gmail.com>
> Cc: Wu, Jingjing <jingjing...@intel.com>; Chilikin, Andrey
> <andrey.chili...@intel.com>; dev@dpdk.org
> Subject: RE: [dpdk-dev] [PATCH v6 6/8] net/i40e: add FDIR support for GTP-C
> and GTP-U
>
>
>
> > -----Original Message-----
> > From: Sean Harte [mailto:sea...@gmail.com]
> > Sent: Friday, September 29, 2017 6:10 PM
> > To: Xing, Beilei <beilei.x...@intel.com>
> > Cc: Wu, Jingjing <jingjing...@intel.com>; Chilikin, Andrey
> > <andrey.chili...@intel.com>; dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH v6 6/8] net/i40e: add FDIR support for
> > GTP-C and GTP-U
> >
> > On 29 September 2017 at 10:33, Xing, Beilei <beilei.x...@intel.com>
> wrote:
> > >
> > >
> > >> -----Original Message-----
> > >> From: Sean Harte [mailto:sea...@gmail.com]
> > >> Sent: Friday, September 29, 2017 4:15 PM
> > >> To: Xing, Beilei <beilei.x...@intel.com>
> > >> Cc: Wu, Jingjing <jingjing...@intel.com>; Chilikin, Andrey
> > >> <andrey.chili...@intel.com>; dev@dpdk.org
> > >> Subject: Re: [dpdk-dev] [PATCH v6 6/8] net/i40e: add FDIR support
> > >> for GTP-C and GTP-U
> > >>
> > >> On 29 September 2017 at 06:19, Beilei Xing <beilei.x...@intel.com>
> wrote:
> > >> > This patch adds FDIR support for GTP-C and GTP-U. The input set
> > >> > of GTP-C and GTP-U is TEID.
> > >> >
> > >> > Signed-off-by: Beilei Xing <beilei.x...@intel.com>
> > >> > ---
> > >> > drivers/net/i40e/i40e_ethdev.h | 30 +++++
> > >> > drivers/net/i40e/i40e_fdir.c | 200
> ++++++++++++++++++++++---------
> > >> > drivers/net/i40e/i40e_flow.c | 263
> > >> +++++++++++++++++++++++++++++++++++------
> > >> > 3 files changed, 396 insertions(+), 97 deletions(-)
> > >>
> > >> <snip>
> > >>
> > >> > @@ -1173,10 +1196,69 @@ i40e_flow_fdir_construct_pkt(struct
> > >> > i40e_pf
> > >> *pf,
> > >> >
> rte_cpu_to_be_16(ETHER_TYPE_ARP))
> > >> > payload += sizeof(struct arp_hdr);
> > >> > set_idx = I40E_FLXPLD_L2_IDX;
> > >> > - break;
> > >> > - default:
> > >> > - PMD_DRV_LOG(ERR, "unknown pctype %u.",
> fdir_input-
> > >pctype);
> > >> > - return -EINVAL;
> > >> > + } else if (fdir_input->flow_ext.customized_pctype) {
> > >> > + /* If customized pctype is used */
> > >> > + cus_pctype =
> > >> > + i40e_flow_fdir_find_customized_pctype(pf,
> > pctype);
> > >> > + if (cus_pctype->index == I40E_CUSTOMIZED_GTPC
> ||
> > >> > + cus_pctype->index ==
> I40E_CUSTOMIZED_GTPU_IPV4 ||
> > >> > + cus_pctype->index ==
> I40E_CUSTOMIZED_GTPU_IPV6 ||
> > >> > + cus_pctype->index ==
> I40E_CUSTOMIZED_GTPU) {
> > >> > + udp = (struct udp_hdr *)(raw_pkt + len);
> > >> > + udp->dgram_len =
> > >> > +
> > >> > + rte_cpu_to_be_16(I40E_FDIR_UDP_DEFAULT_LEN);
> > >> > +
> > >> > + gtp = (struct rte_flow_item_gtp *)
> > >> > + ((unsigned char *)udp +
> sizeof(struct udp_hdr));
> > >> > + gtp->v_pt_rsv_flags = 0x30;
> > >>
> > >> 0x30 isn't valid for GTP-C, the sequence number must be present in
> > >> GTP-C so it will be 0x32 or more. Is this byte actually matched
> > >> against by the device using the GTP pctypes?
> > >
> > > We construct packets and send the packet to HW to create flow
> > > director rule for GTP-C and GTP-U. Actually I didn’t get error info
> > > with 0x30. And in my test, GTP-C packets can hit GTP-C pctype rule.
> > > Will try
> > 0x32 later.
> > >
> > >>
> > >> > + gtp->msg_len =
> > >> > +
> rte_cpu_to_be_16(I40E_FDIR_GTP_DEFAULT_LEN);
> > >> > + gtp->teid =
> fdir_input->flow.gtp_flow.teid;
> > >> > + gtp->msg_type = 0x1;
> > >>
> > >> Why use this value?
> > >
> > > Just for constructing a GTP packet to create a fdir rule for one
> > > pctype, can
> > use other values except 0xFF.
> > >
> > >>
> > >> > +
> > >> > + if (cus_pctype->index ==
> I40E_CUSTOMIZED_GTPC)
> > >> > + udp->dst_port =
> > >> > + rte_cpu_to_be_16(2123);
> > >>
> > >> This will only match half of GTP-C messages. GTP-C messages have a
> > >> UDP port destination of 2123, or a UDP source port of 2123. To
> > >> match all GTP-C packets you need to look at both.
> > >
> > > Yes. But the GTP profile for i40e didn't support response message.
> >
> > That's not clear to a user of the rte_flow API
>
> Rte_flow is a generic API, I think it should allow users to create rule for
> response message.
> But i40e PMD does not the support response message, if user want to create
> a rule for i40e like below:
> Flow create 0 ingress pattern eth / ipv4 / udp src is 2123 / gtpc / end /
> actions queue index 4 / end It will fail. But maybe other PMD can support it.
>
> >
> > >
> > >>
> > >> > + else
> > >> > + udp->dst_port =
> > >> > + rte_cpu_to_be_16(2152);
> > >> > +
> > >> > + if (cus_pctype->index ==
> I40E_CUSTOMIZED_GTPU_IPV4) {
> > >> > + gtp->msg_type = 0xFF;
> > >> > + gtp_ipv4 = (struct ipv4_hdr *)
> > >> > + ((unsigned char *)gtp
> +
> > >> > + sizeof(struct
> > >> > + rte_flow_item_gtp));
> > >>
> > >> This is only valid if v_pt_rsv_flags is 0x30. GTP-U packets are
> > >> allowed to have a sequence number, which adds an extra 4 bytes to
> > >> the
> > GTP header.
> > >
> > > For the GTP profile, there're 4 pctypes for GTP packets: GTPC, GTPU,
> > GTPIPV4, and GTPIPV6.
> > > HW parse which pctype the GTP packets belonge to.
> > > We construct packet to create a fdir rule for one pctype, after
> > > that, all packets whose pctype matches the rule's pctype will hit the
> > > rule.
> >
> > My point is that you can only assume the inner IP header starts at an
> > offset of sizeof(struct rte_flow_item_gtp) if v_pt_rsv_flags is
> > exactly 0x30. If you match only those packets then some GTP-U packets
> will not be matched.
> > That should be clear to a user of the rte_flow API.
> >
>
> No matter if GTP-U packets has a sequence number, once message-type is
> 0xFF, the pctype of the packet parsed by i40e HW will be GTPIPV4 or
> GTPIPV6.
> So If I create a flow rule for GTP-U packets with v_pt_rsv_flags 0x30 and
> message_type is 0xFF, it means I create a rule for pctype GTPIPV4 or
> GTPIPV6.
> Then GTP-U packets with v_pt_rsv_flags 0x32 can hit the rule because its
> pctype is also GTPIPV4 or GTPIPV6.
> It's just i40e HW's behavior, doesn't mean it's effective for other NICs.
> Hope I explain it clearly:)
>
The key point is : for i40e HW, different packet types can use the same pctype.
> > >
> > >>
> > >> > + gtp_ipv4->version_ihl =
> > >> > +
> I40E_FDIR_IP_DEFAULT_VERSION_IHL;
> > >> > + gtp_ipv4->next_proto_id =
> IPPROTO_IP;
> > >> > + gtp_ipv4->total_length =
> > >> > + rte_cpu_to_be_16(
> > >> > +
> I40E_FDIR_INNER_IP_DEFAULT_LEN);
> > >> > + payload = (unsigned char
> *)gtp_ipv4 +
> > >> > + sizeof(struct
> ipv4_hdr);
> > >> > + } else if (cus_pctype->index ==
> > >> > +
> I40E_CUSTOMIZED_GTPU_IPV6) {
> > >> > + gtp->msg_type = 0xFF;
> > >> > + gtp_ipv6 = (struct ipv6_hdr *)
> > >> > + ((unsigned char *)gtp
> +
> > >> > + sizeof(struct
> > >> > + rte_flow_item_gtp));
> > >>
> > >> This is only valid if v_pt_rsv_flags is 0x30. GTP-U packets are
> > >> allowed to have a sequence number, which adds an extra 4 bytes to
> > >> the
> > GTP header.
> > >
> > > Same with above.
> > >
> > >>
> > >> > + gtp_ipv6->vtc_flow =
> > >> > + rte_cpu_to_be_32(
> > >> > +
> I40E_FDIR_IPv6_DEFAULT_VTC_FLOW |
> > >> > + (0 <<
> I40E_FDIR_IPv6_TC_OFFSET));
> > >> > + gtp_ipv6->proto =
> IPPROTO_NONE;
> > >> > + gtp_ipv6->payload_len =
> > >> > + rte_cpu_to_be_16(
> > >> > +
> I40E_FDIR_INNER_IPv6_DEFAULT_LEN);
> > >> > + gtp_ipv6->hop_limits =
> > >> > +
> I40E_FDIR_IPv6_DEFAULT_HOP_LIMITS;
> > >> > + payload = (unsigned char
> *)gtp_ipv6 +
> > >> > + sizeof(struct
> ipv6_hdr);
> > >> > + } else
> > >> > + payload = (unsigned char *)gtp
> +
> > >> > + sizeof(struct
> rte_flow_item_gtp);
> > >> > + }
> > >> > + } else {
> > >> > + PMD_DRV_LOG(ERR, "unknown pctype %u.",
> > >> > + fdir_input->pctype);
> > >> > + return -1;
> > >> > }
> > >> >
> > >> > /* fill the flexbytes to payload */
> > >>
> > >> <snip>
