[SAIP4] Model multicast.

sai_p4/fixed/BUILD.bazel

@@ -21,7 +21,8 @@ filegroup(
     name = "fixed",
     srcs = [
         "bitwidths.p4",
-	"drop_martians.p4",
+        "bmv2_intrinsics.h",
+        "drop_martians.p4",
         "headers.p4",
         "ids.h",
         "ipv4_checksum.p4",

sai_p4/fixed/bitwidths.p4

@@ -45,4 +45,10 @@
 #define TUNNEL_ID_BITWIDTH 16
 #endif
 
+// Inherited from v1model, see `standard_metadata_t.mcast_grp`.
+#define MULTICAST_GROUP_ID_BITWIDTH 16
+
+// Inherited from v1model, see `standard_metadata_t.egress_rid`.
+#define REPLICA_INSTANCE_BITWIDTH 16
+
 #endif  // SAI_BITWIDTHS_P4_

sai_p4/fixed/bmv2_intrinsics.h

@@ -0,0 +1,15 @@
+#ifndef PINS_INFRA_SAI_P4_FIXED_BMV2_INTRINSICS_H_
+#define PINS_INFRA_SAI_P4_FIXED_BMV2_INTRINSICS_H_
+
+// Possible values of the v1model `standard_metadata_t` field `instance_type` in
+// BMv2. The semantics of these values is explained here:
+// https://github.com/p4lang/behavioral-model/blob/main/docs/simple_switch.md
+#define PKT_INSTANCE_TYPE_NORMAL 0
+#define PKT_INSTANCE_TYPE_INGRESS_CLONE 1
+#define PKT_INSTANCE_TYPE_EGRESS_CLONE 2
+#define PKT_INSTANCE_TYPE_COALESCED 3
+#define PKT_INSTANCE_TYPE_INGRESS_RECIRC 4
+#define PKT_INSTANCE_TYPE_REPLICATION 5
+#define PKT_INSTANCE_TYPE_RESUBMIT 6
+
+#endif  // PINS_INFRA_SAI_P4_FIXED_BMV2_INTRINSICS_H_

sai_p4/fixed/ids.h

@@ -19,13 +19,14 @@
 #define ROUTING_IPV4_TABLE_ID 0x02000044                // 33554500
 #define ROUTING_IPV6_TABLE_ID 0x02000045                // 33554501
 #define ROUTING_VRF_TABLE_ID 0x0200004A                 // 33554506
+#define ROUTING_MULTICAST_SRC_MAC_TABLE_ID 0x0200004C   // 33554508
 #define MIRROR_SESSION_TABLE_ID 0x02000046              // 33554502
 #define L3_ADMIT_TABLE_ID 0x02000047                    // 33554503
 #define MIRROR_PORT_TO_PRE_SESSION_TABLE_ID 0x02000048  // 33554504
 #define ECMP_HASHING_TABLE_ID 0x02000049                // 33554505
 #define ROUTING_TUNNEL_TABLE_ID 0x02000050              // 33554512
 #define IPV6_TUNNEL_TERMINATION_TABLE_ID 0x0200004B     // 33554507
-// Next available table id: 0x0200004C (33554508)
+// Next available table id: 0x0200004D (33554509)
 
 // --- Actions -----------------------------------------------------------------
 
@@ -37,10 +38,15 @@
 #define ROUTING_SET_WCMP_GROUP_ID_ACTION_ID 0x01000004               // 16777220
 #define ROUTING_SET_WCMP_GROUP_ID_AND_METADATA_ACTION_ID 0x01000011  // 16777233
 #define ROUTING_SET_NEXTHOP_ID_ACTION_ID 0x01000005                  // 16777221
+#define ROUTING_SET_IP_NEXTHOP_AND_DISABLE_REWRITES_ACTION_ID \
+  0x01000017
+// 16777239
 #define ROUTING_SET_NEXTHOP_ID_AND_METADATA_ACTION_ID 0x01000010     // 16777232
 #define ROUTING_DROP_ACTION_ID 0x01000006                            // 16777222
 #define ROUTING_SET_P2P_TUNNEL_ENCAP_NEXTHOP_ACTION_ID 0x01000012    // 16777234
 #define ROUTING_MARK_FOR_P2P_TUNNEL_ENCAP_ACTION_ID 0x01000013       // 16777235
+#define ROUTING_SET_MULTICAST_GROUP_ID_ACTION_ID 0x01000018          // 16777240
+#define ROUTING_SET_MULTICAST_SRC_MAC_ACTION_ID 0x01000019           // 16777241
 #define MIRRORING_MIRROR_AS_IPV4_ERSPAN_ACTION_ID 0x01000007         // 16777223
 #define L3_ADMIT_ACTION_ID 0x01000008                                // 16777224
 #define MIRRORING_SET_PRE_SESSION_ACTION_ID 0x01000009               // 16777225
@@ -52,7 +58,7 @@
 #define TRAP_ACTION_ID 0x0100000F                                    // 16777231
 #define ROUTING_SET_METADATA_AND_DROP_ACTION_ID 0x01000015           // 16777237
 #define MARK_FOR_TUNNEL_DECAP_AND_SET_VRF_ACTION_ID 0x01000016       // 16777238
-// Next available action id: 0x01000017 (16777239)
+// Next available action id: 0x0100001A (16777242)
 
 // --- Action Profiles and Selectors (8 most significant bits = 0x11) ----------
 // This value should ideally be 0x11000001, but we currently have this value for

sai_p4/fixed/metadata.p4

@@ -80,6 +80,8 @@ type bit<QOS_QUEUE_BITWIDTH> qos_queue_t;
 
 typedef bit<ROUTE_METADATA_BITWIDTH> route_metadata_t;
 typedef bit<ACL_METADATA_BITWIDTH> acl_metadata_t;
+typedef bit<MULTICAST_GROUP_ID_BITWIDTH> multicast_group_id_t;
+typedef bit<REPLICA_INSTANCE_BITWIDTH> replica_instance_t;
 
 // -- Meters -------------------------------------------------------------------
 
@@ -197,7 +199,13 @@ struct local_metadata_t {
   bool vlan_id_valid;  // True iff `vlan_id` is valid.
   bool admit_to_l3;
   vrf_id_t vrf_id;
+
+  // Rewrite related fields.
+  bool enable_ttl_rewrite;
+  bool enable_src_mac_rewrite;
+  bool enable_dst_mac_rewrite;
   packet_rewrites_t packet_rewrites;
+
   bit<16> l4_src_port;
   bit<16> l4_dst_port;
   bit<WCMP_SELECTOR_INPUT_BITWIDTH> wcmp_selector_input;

sai_p4/fixed/minimum_guaranteed_sizes.p4

@@ -49,6 +49,10 @@
 #define ROUTING_TUNNEL_TABLE_MINIMUM_GUARANTEED_SIZE 0
 #endif
 
+#ifndef ROUTING_MULTICAST_SOURCE_MAC_TABLE_MINIMUM_GUARANTEED_SIZE
+#define ROUTING_MULTICAST_SOURCE_MAC_TABLE_MINIMUM_GUARANTEED_SIZE 0
+#endif
+
 #ifndef L3_ADMIT_TABLE_MINIMUM_GUARANTEED_SIZE
 #define L3_ADMIT_TABLE_MINIMUM_GUARANTEED_SIZE 0
 #endif

sai_p4/fixed/packet_rewrites.p4

@@ -5,31 +5,118 @@
 #include "headers.p4"
 #include "metadata.p4"
 #include "minimum_guaranteed_sizes.p4"
+#include "bmv2_intrinsics.h"
 
-// This control block applies the rewrites computed during the the ingress
+// To be applied only for multicast-replicated packets, i.e. packets with
+// `standard_metadata.instance_type == PKT_INSTANCE_TYPE_REPLICATION`.
+// In P4Runtime, these are packets created by a `Replica` of a
+// `MulticastGroupEntry`.
+control multicast_rewrites(inout local_metadata_t local_metadata,
+                           in standard_metadata_t standard_metadata) {
+  // The egress port of the multicast-replicated packet.
+  // In P4Runtime, equal to the `port` value of the `Replica` of the
+  // `MulticastGroupEntry` that created this packet.
+  port_id_t multicast_replica_port = (port_id_t) standard_metadata.egress_port;
+
+  // The instance number of the multicast-replicated packet.
+  // In P4Runtime, equal to the `instance` value of the `Replica` of the
+  // `MulticastGroupEntry` that created this packet.
+  replica_instance_t multicast_replica_instance =
+      standard_metadata.egress_rid;
+
+  @id(ROUTING_SET_MULTICAST_SRC_MAC_ACTION_ID)
+  action set_multicast_src_mac(@id(1) @format(MAC_ADDRESS)
+                               ethernet_addr_t src_mac) {
+    local_metadata.enable_src_mac_rewrite = true;
+    local_metadata.packet_rewrites.src_mac = src_mac;
+  }
+
+  // Rewrites the source MAC of the multicast-replicated packet.
+  //
+  // This is a logical table that does not exist in SAI and that is managed by
+  // Orchagent. There is a many-to-one correspondence between entries in this
+  // table and SAI Router Interfaces (RIFs). Each entry corresponds to a RIF
+  // with the following attributes:
+  // * `SAI_ROUTER_INTERFACE_ATTR_PORT_ID` is equal to `multicast_replica_port`.
+  // * `SAI_ROUTER_INTERFACE_ATTR_SRC_MAC_ADDRESS` is equal to the `src_mac`
+  //   parameter of the `set_multicast_src_mac` action.
+  //
+  // Orchagent maintains a mapping from entries to RIFs, creating and destroying
+  // (possibly shared) RIFs as entries are inserted and deleted.
+  //
+  // When creating a multicast group member (`SAI_IPMC_GROUP_MEMBER`) from a
+  // P4Runtime `Replica`, Orchagent will use this table to set the value of the
+  // `SAI_IPMC_GROUP_MEMBER_ATTR_IPMC_OUTPUT_ID` attribute: it will expect to
+  // find an entry for the replica's port and instance in this table, and will
+  // use the ID of the RIF associated with that entry.
+  // This will cause the source MAC of packets generated by the group member to
+  // be rewritten to the `src_mac` of the `set_multicast_src_mac` action of the
+  // entry.
+  //
+  // TODO: Remove `@unsupported` annotation once the switch stack
+  // supports this table.
+  @unsupported
+  @p4runtime_role(P4RUNTIME_ROLE_ROUTING)
+  @id(ROUTING_MULTICAST_SRC_MAC_TABLE_ID)
+  table multicast_source_mac_table {
+    key = {
+      multicast_replica_port : exact
+        // TODO: Add this once supported by PDPI and its customers.
+        // @referenced_by(multicast_group_table, replica.port)
+        @id(1);
+      multicast_replica_instance : exact
+        // TODO: Add this once supported by PDPI and its customers.
+        // @referenced_by(multicast_group_table, replica.instance)
+        @id(2);
+    }
+    actions = {
+      @proto_id(1) set_multicast_src_mac;
+    }
+    size = ROUTING_MULTICAST_SOURCE_MAC_TABLE_MINIMUM_GUARANTEED_SIZE;
+  }
+
+  apply {
+    multicast_source_mac_table.apply();
+  }
+}  // control multicast_rewrites
+
+// This control block applies the rewrites computed during the ingress
 // stage to the actual packet.
 control packet_rewrites(inout headers_t headers,
-                        in local_metadata_t local_metadata,
+                        inout local_metadata_t local_metadata,
                         inout standard_metadata_t standard_metadata) {
   apply {
-    if (local_metadata.admit_to_l3) {
-      headers.ethernet.src_addr = local_metadata.packet_rewrites.src_mac;
-      headers.ethernet.dst_addr = local_metadata.packet_rewrites.dst_mac;
+    if (standard_metadata.instance_type == PKT_INSTANCE_TYPE_REPLICATION) {
+      // TODO: Remove guard once p4-symbolic suports assertions.
+#ifndef PLATFORM_P4SYMBOLIC
+      assert(local_metadata.admit_to_l3);
+#endif
+      local_metadata.enable_ttl_rewrite = true;
+      multicast_rewrites.apply(local_metadata, standard_metadata);
+    }
 
+    if (local_metadata.admit_to_l3){
+      if (local_metadata.enable_src_mac_rewrite) {
+        headers.ethernet.src_addr = local_metadata.packet_rewrites.src_mac;
+      }
+      if (local_metadata.enable_dst_mac_rewrite) {
+        headers.ethernet.dst_addr = local_metadata.packet_rewrites.dst_mac;
+      }
       if (headers.ipv4.isValid()) {
-        if (headers.ipv4.ttl <= 1) {
-          mark_to_drop(standard_metadata);
-        } else {
+        if (headers.ipv4.ttl > 0 && local_metadata.enable_ttl_rewrite) {
           headers.ipv4.ttl = headers.ipv4.ttl - 1;
         }
+        // TODO: Verify this is accurate when TTL rewrite is
+        // disabled and update this code if not.
+        if (headers.ipv4.ttl == 0) mark_to_drop(standard_metadata);
       }
-
       if (headers.ipv6.isValid()) {
-        if (headers.ipv6.hop_limit <= 1) {
-          mark_to_drop(standard_metadata);
-        } else {
+        if (headers.ipv6.hop_limit > 0 && local_metadata.enable_ttl_rewrite) {
           headers.ipv6.hop_limit = headers.ipv6.hop_limit - 1;
         }
+        // TODO: Verify this is accurate when TTL rewrite is
+        // disabled and update this code if not.
+        if (headers.ipv6.hop_limit == 0) mark_to_drop(standard_metadata);
       }
     }
   }

sai_p4/fixed/parser.p4

@@ -13,6 +13,9 @@ parser packet_parser(packet_in packet, out headers_t headers,
     // Initialize local metadata fields.
     local_metadata.admit_to_l3 = false;
     local_metadata.vrf_id = kDefaultVrf;
+    local_metadata.enable_ttl_rewrite = false;
+    local_metadata.enable_src_mac_rewrite = false;
+    local_metadata.enable_dst_mac_rewrite = false;
     local_metadata.packet_rewrites.src_mac = 0;
     local_metadata.packet_rewrites.dst_mac = 0;
     local_metadata.l4_src_port = 0;