[SDFAB-192] Report VXLAN-encaped packets to support end-host INT (#275)

* Report VXLAN-encaped packets * Add VXLAN pkt types for the INT-related PTF tests
stratum · Jun 18, 2021 · bcbb0e7 · bcbb0e7
1 parent 47c9c6b
commit bcbb0e7
Show file tree

Hide file tree

Showing 8 changed files with 125 additions and 30 deletions.
diff --git a/p4src/include/control/int.p4 b/p4src/include/control/int.p4
@@ -417,7 +417,7 @@ control IntEgress (
         fabric_md.int_report_md.eg_tstamp = eg_prsr_md.global_tstamp[31:0];
         fabric_md.int_report_md.ip_eth_type = fabric_md.bridged.base.ip_eth_type;
         fabric_md.int_report_md.flow_hash = fabric_md.bridged.base.inner_hash;
-        // fabric_md.int_report_md.gtpu_presence set by the parser
+        // fabric_md.int_report_md.encap_presence set by the parser
     }
 
     @hidden

diff --git a/p4src/include/control/int_parser.p4 b/p4src/include/control/int_parser.p4
@@ -54,7 +54,7 @@ parser IntReportParser (packet_in packet,
         fabric_md.int_report_md.ip_eth_type = ETHERTYPE_IPV4;
         fabric_md.int_report_md.report_type = IntReportType_t.DROP;
         fabric_md.int_report_md.mirror_type = FabricMirrorType_t.INVALID;
-        fabric_md.int_report_md.gtpu_presence = fabric_md.bridged.base.gtpu_presence;
+        fabric_md.int_report_md.encap_presence = fabric_md.bridged.base.encap_presence;
         fabric_md.int_report_md.flow_hash = fabric_md.bridged.base.inner_hash;
 
         /** report_fixed_header **/
@@ -174,11 +174,12 @@ parser IntReportParser (packet_in packet,
 
     state check_eth_type {
         packet.extract(hdr.eth_type);
-        transition select(hdr.eth_type.value, fabric_md.int_report_md.gtpu_presence) {
+        transition select(hdr.eth_type.value, fabric_md.int_report_md.encap_presence) {
             (ETHERTYPE_MPLS, _): strip_mpls;
-            (ETHERTYPE_IPV4, GtpuPresence.NONE): handle_ipv4;
-            (ETHERTYPE_IPV4, GtpuPresence.GTPU_ONLY): strip_ipv4_udp_gtpu;
-            (ETHERTYPE_IPV4, GtpuPresence.GTPU_WITH_PSC): strip_ipv4_udp_gtpu_psc;
+            (ETHERTYPE_IPV4, EncapPresence.NONE): handle_ipv4;
+            (ETHERTYPE_IPV4, EncapPresence.GTPU_ONLY): strip_ipv4_udp_gtpu;
+            (ETHERTYPE_IPV4, EncapPresence.GTPU_WITH_PSC): strip_ipv4_udp_gtpu_psc;
+            (ETHERTYPE_IPV4, EncapPresence.VXLAN): strip_ipv4_udp_vxlan;
             default: reject;
         }
     }
@@ -188,10 +189,11 @@ parser IntReportParser (packet_in packet,
     state strip_mpls {
         packet.advance(MPLS_HDR_BYTES * 8);
         bit<IP_VER_BITS> ip_ver = packet.lookahead<bit<IP_VER_BITS>>();
-        transition select(ip_ver, fabric_md.int_report_md.gtpu_presence) {
-            (IP_VERSION_4, GtpuPresence.NONE): handle_ipv4;
-            (IP_VERSION_4, GtpuPresence.GTPU_ONLY): strip_ipv4_udp_gtpu;
-            (IP_VERSION_4, GtpuPresence.GTPU_WITH_PSC): strip_ipv4_udp_gtpu_psc;
+        transition select(ip_ver, fabric_md.int_report_md.encap_presence) {
+            (IP_VERSION_4, EncapPresence.NONE): handle_ipv4;
+            (IP_VERSION_4, EncapPresence.GTPU_ONLY): strip_ipv4_udp_gtpu;
+            (IP_VERSION_4, EncapPresence.GTPU_WITH_PSC): strip_ipv4_udp_gtpu_psc;
+            (IP_VERSION_4, EncapPresence.VXLAN): strip_ipv4_udp_vxlan;
             default: reject;
         }
     }
@@ -207,6 +209,15 @@ parser IntReportParser (packet_in packet,
         transition handle_ipv4;
     }
 
+    state strip_ipv4_udp_vxlan {
+        packet.advance((IPV4_HDR_BYTES + UDP_HDR_BYTES + VXLAN_HDR_BYTES) * 8);
+        // Skip the Ethernet header.
+        // It will be removed from a packet and the outer Ethernet will be used
+        // as the inner Ethernet header.
+        packet.advance((ETH_HDR_BYTES) * 8);
+        transition handle_ipv4;
+    }
+
     state handle_ipv4 {
         // Extract only the length, required later to compute the lenght of the
         // report encap headers.

diff --git a/p4src/include/control/spgw.p4 b/p4src/include/control/spgw.p4
@@ -19,7 +19,7 @@ control DecapGtpu(inout ingress_headers_t            hdr,
         hdr.gtpu.setInvalid();
         hdr.gtpu_options.setInvalid();
         hdr.gtpu_ext_psc.setInvalid();
-        fabric_md.bridged.base.gtpu_presence = GtpuPresence.NONE;
+        fabric_md.bridged.base.encap_presence = EncapPresence.NONE;
     }
     @hidden
     action decap_inner_tcp() {
@@ -348,7 +348,7 @@ control SpgwIngress(
                 }
                 // PDRs
                 if (fabric_md.spgw.src_iface == SpgwInterface.ACCESS &&
-                        fabric_md.bridged.base.gtpu_presence != GtpuPresence.NONE) {
+                        fabric_md.bridged.base.encap_presence != EncapPresence.NONE) {
                     uplink_pdrs.apply();
                 } else if (fabric_md.spgw.src_iface == SpgwInterface.CORE ||
                             fabric_md.spgw.src_iface == SpgwInterface.FROM_DBUF) {
@@ -410,7 +410,7 @@ control SpgwEgress(
         hdr.outer_gtpu.msglen = hdr.ipv4.total_len;
         hdr.outer_gtpu.ex_flag = 0;
 #ifdef WITH_INT
-        fabric_md.int_report_md.gtpu_presence = GtpuPresence.GTPU_ONLY;
+        fabric_md.int_report_md.encap_presence = EncapPresence.GTPU_ONLY;
 #endif // WITH_INT
     }
 
@@ -432,7 +432,7 @@ control SpgwEgress(
         hdr.outer_gtpu_ext_psc.setValid();
         hdr.outer_gtpu_ext_psc.qfi = qfi;
 #ifdef WITH_INT
-        fabric_md.int_report_md.gtpu_presence = GtpuPresence.GTPU_WITH_PSC;
+        fabric_md.int_report_md.encap_presence = EncapPresence.GTPU_WITH_PSC;
 #endif // WITH_INT
     }
 

diff --git a/p4src/include/define.p4 b/p4src/include/define.p4
@@ -24,6 +24,7 @@
 #define GTPU_EXT_PSC_HDR_BYTES 4
 #define MPLS_HDR_BYTES 4
 #define VLAN_HDR_BYTES 4
+#define VXLAN_HDR_BYTES 8
 
 #define PKT_INSTANCE_TYPE_NORMAL 0
 #define PKT_INSTANCE_TYPE_INGRESS_CLONE 1
@@ -61,10 +62,15 @@ enum bit<8> SpgwInterface {
     FROM_DBUF     = 0x3
 }
 
-enum bit<2> GtpuPresence {
+// According to our design choice, we report only the inner headers to the INT collector.
+// The EncapPresence keeps track of the encapsulation protocol in use.
+// The EncapPresence is further needed by the egress INT parser to strip out the outer encapsulation headers
+// and put only inner headers in an INT report.
+enum bit<2> EncapPresence {
     NONE          = 0x0,
     GTPU_ONLY     = 0x1,
-    GTPU_WITH_PSC = 0x2
+    GTPU_WITH_PSC = 0x2,
+    VXLAN         = 0x3
 }
 
 const bit<16> GTPU_UDP_PORT = 2152;
@@ -131,6 +137,8 @@ const vlan_id_t DEFAULT_VLAN_ID = 12w4094;
 const bit<8> DEFAULT_MPLS_TTL = 64;
 const bit<8> DEFAULT_IPV4_TTL = 64;
 
+const bit<16> VXLAN_UDP_PORT = 4789;
+
 // The recirculation port uses the same number for all HW pipes. The actual port
 // ID (DP_ID) can be obtained by prefixing the HW pipe ID (2 bits).
 const bit<7> RECIRC_PORT_NUMBER = 7w68;

diff --git a/p4src/include/header.p4 b/p4src/include/header.p4
@@ -121,6 +121,12 @@ header icmp_t {
     // Other optional fields...
 }
 
+header vxlan_t {
+    bit<8>  flags;
+    bit<24> reserved;
+    bit<24> vni;
+    bit<8>  reserved_2;
+}
 
 // GTPU v1 -- 3GPP TS 29.281 version 15.7.0
 // https://www.etsi.org/deliver/etsi_ts/129200_129299/129281/15.07.00_60/ts_129281v150700p.pdf
@@ -234,7 +240,7 @@ header local_report_header_t {
 @pa_no_overlay("egress", "fabric_md.int_report_md.ip_eth_type")
 @pa_no_overlay("egress", "fabric_md.int_report_md.report_type")
 @pa_no_overlay("egress", "fabric_md.int_report_md.flow_hash")
-@pa_no_overlay("egress", "fabric_md.int_report_md.gtpu_presence")
+@pa_no_overlay("egress", "fabric_md.int_report_md.encap_presence")
 header int_report_metadata_t {
     BridgedMdType_t       bmd_type;
     @padding bit<5>       _pad0;
@@ -252,7 +258,7 @@ header int_report_metadata_t {
     bit<8>                drop_reason;
     bit<16>               ip_eth_type;
     @padding bit<6>       _pad5;
-    GtpuPresence          gtpu_presence;
+    EncapPresence         encap_presence;
     @padding bit<6>       _pad6;
     IntReportType_t       report_type;
     flow_hash_t           flow_hash;
@@ -280,7 +286,7 @@ struct bridged_metadata_base_t {
     bool                     is_multicast;
     fwd_type_t               fwd_type;
     vlan_id_t                vlan_id;
-    GtpuPresence             gtpu_presence;
+    EncapPresence            encap_presence;
     // bit<3>                vlan_pri;
     // bit<1>                vlan_cfi;
     bit<8>                   mpls_ttl;
@@ -414,6 +420,9 @@ struct ingress_headers_t {
     gtpu_t gtpu;
     gtpu_options_t gtpu_options;
     gtpu_ext_psc_t gtpu_ext_psc;
+    vxlan_t vxlan;
+    ethernet_t inner_ethernet;
+    eth_type_t inner_eth_type;
     ipv4_t inner_ipv4;
     tcp_t inner_tcp;
     udp_t inner_udp;

diff --git a/p4src/include/parser.p4 b/p4src/include/parser.p4
@@ -33,7 +33,7 @@ parser FabricIngressParser (packet_in  packet,
 #ifdef WITH_INT
         fabric_md.bridged.int_bmd.drop_reason = IntDropReason_t.DROP_REASON_UNKNOWN;
 #endif // WITH_INT
-        fabric_md.bridged.base.gtpu_presence = GtpuPresence.NONE;
+        fabric_md.bridged.base.encap_presence = EncapPresence.NONE;
         transition check_ethernet;
     }
 
@@ -195,6 +195,7 @@ parser FabricIngressParser (packet_in  packet,
         gtpu_t gtpu = packet.lookahead<gtpu_t>();
         transition select(hdr.udp.dport, gtpu.version, gtpu.msgtype) {
             (GTPU_UDP_PORT, GTP_V1, GTPU_GPDU): parse_gtpu;
+            (VXLAN_UDP_PORT, _, _): parse_vxlan;
             // Treat GTP control traffic as payload.
             default: accept;
         }
@@ -214,7 +215,7 @@ parser FabricIngressParser (packet_in  packet,
     }
 
     state set_gtpu_only {
-        fabric_md.bridged.base.gtpu_presence = GtpuPresence.GTPU_ONLY;
+        fabric_md.bridged.base.encap_presence = EncapPresence.GTPU_ONLY;
         transition parse_inner_ipv4;
     }
 
@@ -229,13 +230,28 @@ parser FabricIngressParser (packet_in  packet,
 
     state parse_gtpu_ext_psc {
         packet.extract(hdr.gtpu_ext_psc);
-        fabric_md.bridged.base.gtpu_presence = GtpuPresence.GTPU_WITH_PSC;
+        fabric_md.bridged.base.encap_presence = EncapPresence.GTPU_WITH_PSC;
         transition select(hdr.gtpu_ext_psc.next_ext) {
             GTPU_NEXT_EXT_NONE: parse_inner_ipv4;
             default: accept;
         }
     }
 
+    state parse_vxlan {
+        packet.extract(hdr.vxlan);
+        fabric_md.bridged.base.encap_presence = EncapPresence.VXLAN;
+        transition parse_inner_ethernet;
+    }
+
+    state parse_inner_ethernet {
+        packet.extract(hdr.inner_ethernet);
+        packet.extract(hdr.inner_eth_type);
+        transition select(hdr.inner_eth_type.value) {
+            ETHERTYPE_IPV4: parse_inner_ipv4;
+            default: accept;
+        }
+    }
+
     state parse_inner_ipv4 {
         packet.extract(hdr.inner_ipv4);
         inner_ipv4_checksum.add(hdr.inner_ipv4);
@@ -315,6 +331,9 @@ control FabricIngressDeparser(packet_out packet,
         packet.emit(hdr.gtpu);
         packet.emit(hdr.gtpu_options);
         packet.emit(hdr.gtpu_ext_psc);
+        packet.emit(hdr.vxlan);
+        packet.emit(hdr.inner_ethernet);
+        packet.emit(hdr.inner_eth_type);
         packet.emit(hdr.inner_ipv4);
         packet.emit(hdr.inner_tcp);
         packet.emit(hdr.inner_udp);
@@ -394,7 +413,7 @@ parser FabricEgressParser (packet_in packet,
         hdr.outer_gtpu_ext_psc.next_ext  = GTPU_NEXT_EXT_NONE;
 #endif // WITH_SPGW
 #ifdef WITH_INT
-        fabric_md.int_report_md.gtpu_presence = fabric_md.bridged.base.gtpu_presence;
+        fabric_md.int_report_md.encap_presence = fabric_md.bridged.base.encap_presence;
 #endif // WITH_INT
         transition check_ethernet;
     }

diff --git a/ptf/tests/ptf/fabric.ptf/test.py b/ptf/tests/ptf/fabric.ptf/test.py
@@ -1891,7 +1891,7 @@ def runTest(self):
                     for send_report_to_spine in [False, True]:
                         if send_report_to_spine and tagged[1]:
                             continue
-                        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES:
+                        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES | VXLAN_PKT_TYPES:
                             self.doRunTest(
                                 vlan_conf,
                                 tagged,
@@ -1966,7 +1966,7 @@ def runTest(self):
                         for send_report_to_spine in [False, True]:
                             if send_report_to_spine and tagged[1]:
                                 continue
-                            for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES:
+                            for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES | VXLAN_PKT_TYPES:
                                 self.doRunTest(
                                     vlan_conf,
                                     tagged,
@@ -2037,7 +2037,7 @@ def runTest(self):
                     for send_report_to_spine in [False, True]:
                         if send_report_to_spine and tagged[1]:
                             continue
-                        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES:
+                        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES | VXLAN_PKT_TYPES:
                             self.doRunTest(
                                 vlan_conf,
                                 tagged,
@@ -2078,7 +2078,7 @@ def doRunTest(
 
     def runTest(self):
         print("")
-        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES:
+        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES | VXLAN_PKT_TYPES:
             expect_int_report = True
             # Change the IP destination to ensure we are using differnt
             # flow for diffrent test cases since the flow report filter
@@ -2199,7 +2199,7 @@ def doRunTest(
 
     def runTest(self):
         print("")
-        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES:
+        for pkt_type in BASE_PKT_TYPES | GTP_PKT_TYPES | VXLAN_PKT_TYPES:
             expect_int_report = True
             # Change the IP destination to ensure we are using differnt
             # flow for diffrent test cases since the flow report filter