VPP-470: Introduce VxLAN-GPE as transport for iOAM.

This is the second of the two commits. This commit introduces VxLAN-GPE
transit functionality.

Change-Id: I12f13e00f456df0047f2b7a0737addfeb683c420
Signed-off-by: Vengada Govindan <venggovi@cisco.com>

1 files changed, 415 insertions, 38 deletions

diff --git a/plugins/ioam-plugin/ioam/lib-vxlan-gpe/vxlan_gpe_ioam.c b/plugins/ioam-plugin/ioam/lib-vxlan-gpe/vxlan_gpe_ioam.c
index 066f5821bde..6c04d9af210 100644
--- a/plugins/ioam-plugin/ioam/lib-vxlan-gpe/vxlan_gpe_ioam.c
+++ b/plugins/ioam-plugin/ioam/lib-vxlan-gpe/vxlan_gpe_ioam.c
@@ -16,6 +16,9 @@
 #include <vnet/vxlan-gpe/vxlan_gpe_packet.h>
 #include <vnet/ip/format.h>
 #include <ioam/lib-vxlan-gpe/vxlan_gpe_ioam.h>
+#include <vnet/fib/ip6_fib.h>
+#include <vnet/fib/ip4_fib.h>
+#include <vnet/fib/fib_entry.h>
 
 vxlan_gpe_ioam_main_t vxlan_gpe_ioam_main;
 
@@ -163,11 +166,257 @@ vxlan_gpe_ioam_set (vxlan_gpe_tunnel_t * t,
 
 }
 
+static void
+vxlan_gpe_set_clear_output_feature_on_intf (vlib_main_t * vm,
+					    u32 sw_if_index0, u8 is_add)
+{
+
+
+
+  vnet_feature_enable_disable ("ip4-output", "vxlan-gpe-transit-ioam",
+			       sw_if_index0, is_add,
+			       0 /* void *feature_config */ ,
+			       0 /* u32 n_feature_config_bytes */ );
+  return;
+}
+
+void
+vxlan_gpe_clear_output_feature_on_all_intfs (vlib_main_t * vm)
+{
+  vnet_sw_interface_t *si = 0;
+  vnet_main_t *vnm = vnet_get_main ();
+  vnet_interface_main_t *im = &vnm->interface_main;
+
+  pool_foreach (si, im->sw_interfaces, (
+					 {
+					 vxlan_gpe_set_clear_output_feature_on_intf
+					 (vm, si->sw_if_index, 0);
+					 }));
+  return;
+}
+
+
+extern fib_forward_chain_type_t
+fib_entry_get_default_chain_type (const fib_entry_t * fib_entry);
+
+int
+vxlan_gpe_enable_disable_ioam_for_dest (vlib_main_t * vm,
+					ip46_address_t dst_addr,
+					u32 outer_fib_index,
+					u8 is_ipv4, u8 is_add)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  u32 fib_index0 = 0;
+  u32 sw_if_index0 = ~0;
+
+  fib_node_index_t fei = ~0;
+  fib_entry_t *fib_entry;
+  u32 adj_index0;
+  ip_adjacency_t *adj0;
+  fib_prefix_t fib_prefix;
+  //fib_forward_chain_type_t fct;
+  load_balance_t *lb_m, *lb_b;
+  const dpo_id_t *dpo0, *dpo1;
+  u32 i, j;
+  //vnet_hw_interface_t *hw;
+
+  if (is_ipv4)
+    {
+      memset (&fib_prefix, 0, sizeof (fib_prefix_t));
+      fib_prefix.fp_len = 32;
+      fib_prefix.fp_proto = FIB_PROTOCOL_IP4;
+      fib_prefix.fp_addr = dst_addr;
+    }
+  else
+    {
+      return 0;
+    }
+
+  fei = fib_table_lookup (fib_index0, &fib_prefix);
+  fib_entry = fib_entry_get (fei);
+
+  //fct = fib_entry_get_default_chain_type (fib_entry);
+
+  if (!dpo_id_is_valid (&fib_entry->fe_lb /*[fct] */ ))
+    {
+      return (-1);
+    }
+
+  lb_m = load_balance_get (fib_entry->fe_lb /*[fct] */ .dpoi_index);
+
+  for (i = 0; i < lb_m->lb_n_buckets; i++)
+    {
+      dpo0 = load_balance_get_bucket_i (lb_m, i);
+
+      if (dpo0->dpoi_type == DPO_LOAD_BALANCE)
+	{
+	  lb_b = load_balance_get (dpo0->dpoi_index);
+
+	  for (j = 0; j < lb_b->lb_n_buckets; j++)
+	    {
+	      dpo1 = load_balance_get_bucket_i (lb_b, j);
+
+	      if (dpo1->dpoi_type == DPO_ADJACENCY)
+		{
+		  adj_index0 = dpo1->dpoi_index;
+
+		  if (ADJ_INDEX_INVALID == adj_index0)
+		    {
+		      continue;
+		    }
+
+		  adj0 =
+		    ip_get_adjacency (&(ip4_main.lookup_main), adj_index0);
+		  sw_if_index0 = adj0->rewrite_header.sw_if_index;
+
+		  if (~0 == sw_if_index0)
+		    {
+		      continue;
+		    }
+
+
+		  if (is_add)
+		    {
+		      vnet_feature_enable_disable ("ip4-output",
+						   "vxlan-gpe-transit-ioam",
+						   sw_if_index0, is_add, 0
+						   /* void *feature_config */
+						   , 0	/* u32 n_feature_config_bytes */
+			);
+
+		      vec_validate_init_empty (hm->bool_ref_by_sw_if_index,
+					       sw_if_index0, ~0);
+		      hm->bool_ref_by_sw_if_index[sw_if_index0] = 1;
+		    }
+		  else
+		    {
+		      hm->bool_ref_by_sw_if_index[sw_if_index0] = ~0;
+		    }
+		}
+	    }
+	}
+    }
+
+  if (is_ipv4)
+    {
+
+      uword *t = NULL;
+      vxlan_gpe_ioam_dest_tunnels_t *t1;
+      fib_prefix_t key4, *key4_copy;
+      hash_pair_t *hp;
+      memset (&key4, 0, sizeof (key4));
+      key4.fp_proto = FIB_PROTOCOL_IP4;
+      key4.fp_addr.ip4.as_u32 = fib_prefix.fp_addr.ip4.as_u32;
+      t = hash_get_mem (hm->dst_by_ip4, &key4);
+      if (is_add)
+	{
+	  if (t)
+	    {
+	      return 0;
+	    }
+	  pool_get_aligned (hm->dst_tunnels, t1, CLIB_CACHE_LINE_BYTES);
+	  memset (t1, 0, sizeof (*t1));
+	  t1->fp_proto = FIB_PROTOCOL_IP4;
+	  t1->dst_addr.ip4.as_u32 = fib_prefix.fp_addr.ip4.as_u32;
+	  key4_copy = clib_mem_alloc (sizeof (*key4_copy));
+	  clib_memcpy (key4_copy, &key4, sizeof (*key4_copy));
+	  hash_set_mem (hm->dst_by_ip4, key4_copy, t1 - hm->dst_tunnels);
+	  /*
+	   * Attach to the FIB entry for the VxLAN-GPE destination
+	   * and become its child. The dest route will invoke a callback
+	   * when the fib entry changes, it can be used to
+	   * re-program the output feature on the egress interface.
+	   */
+
+	  const fib_prefix_t tun_dst_pfx = {
+	    .fp_len = 32,
+	    .fp_proto = FIB_PROTOCOL_IP4,
+	    .fp_addr = {.ip4 = t1->dst_addr.ip4,}
+	  };
+
+	  t1->fib_entry_index =
+	    fib_table_entry_special_add (outer_fib_index,
+					 &tun_dst_pfx,
+					 FIB_SOURCE_RR,
+					 FIB_ENTRY_FLAG_NONE,
+					 ADJ_INDEX_INVALID);
+	  t1->sibling_index =
+	    fib_entry_child_add (t1->fib_entry_index,
+				 hm->fib_entry_type, t1 - hm->dst_tunnels);
+	  t1->outer_fib_index = outer_fib_index;
+
+	}
+      else
+	{
+	  if (!t)
+	    {
+	      return 0;
+	    }
+	  t1 = pool_elt_at_index (hm->dst_tunnels, t[0]);
+	  hp = hash_get_pair (hm->dst_by_ip4, &key4);
+	  key4_copy = (void *) (hp->key);
+	  hash_unset_mem (hm->dst_by_ip4, &key4);
+	  clib_mem_free (key4_copy);
+	  pool_put (hm->dst_tunnels, t1);
+	}
+    }
+  else
+    {
+      // TBD for IPv6
+    }
+
+  return 0;
+}
+
+void
+vxlan_gpe_refresh_output_feature_on_all_dest (void)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  vxlan_gpe_ioam_dest_tunnels_t *t;
+  u32 i;
+  if (pool_elts (hm->dst_tunnels) == 0)
+    return;
+  vxlan_gpe_clear_output_feature_on_all_intfs (hm->vlib_main);
+  i = vec_len (hm->bool_ref_by_sw_if_index);
+  vec_free (hm->bool_ref_by_sw_if_index);
+  vec_validate_init_empty (hm->bool_ref_by_sw_if_index, i, ~0);
+  pool_foreach (t, hm->dst_tunnels, (
+				      {
+				      vxlan_gpe_enable_disable_ioam_for_dest
+				      (hm->vlib_main,
+				       t->dst_addr,
+				       t->outer_fib_index,
+				       (t->fp_proto == FIB_PROTOCOL_IP4), 1
+				       /* is_add */
+				      );
+				      }
+		));
+  return;
+}
+
+void
+vxlan_gpe_clear_output_feature_on_select_intfs (void)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  u32 sw_if_index0 = 0;
+  for (sw_if_index0 = 0;
+       sw_if_index0 < vec_len (hm->bool_ref_by_sw_if_index); sw_if_index0++)
+    {
+      if (hm->bool_ref_by_sw_if_index[sw_if_index0] == 0xFF)
+	{
+	  vxlan_gpe_set_clear_output_feature_on_intf
+	    (hm->vlib_main, sw_if_index0, 0);
+	}
+    }
+
+  return;
+}
 
 static clib_error_t *
-vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
-				       unformat_input_t * input,
-				       vlib_cli_command_t * cmd)
+vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t *
+				       vm,
+				       unformat_input_t
+				       * input, vlib_cli_command_t * cmd)
 {
   vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
   ip46_address_t local, remote;
@@ -184,7 +433,6 @@ vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
   uword *p;
   vxlan_gpe_main_t *gm = &vxlan_gpe_main;
   vxlan_gpe_tunnel_t *t = 0;
-
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "local %U", unformat_ip4_address, &local.ip4))
@@ -192,8 +440,8 @@ vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
 	  local_set = 1;
 	  ipv4_set = 1;
 	}
-      else if (unformat (input, "remote %U",
-			 unformat_ip4_address, &remote.ip4))
+      else
+	if (unformat (input, "remote %U", unformat_ip4_address, &remote.ip4))
 	{
 	  remote_set = 1;
 	  ipv4_set = 1;
@@ -203,8 +451,8 @@ vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
 	  local_set = 1;
 	  ipv6_set = 1;
 	}
-      else if (unformat (input, "remote %U",
-			 unformat_ip6_address, &remote.ip6))
+      else
+	if (unformat (input, "remote %U", unformat_ip6_address, &remote.ip6))
 	{
 	  remote_set = 1;
 	  ipv6_set = 1;
@@ -219,29 +467,27 @@ vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
 
   if (local_set == 0)
     return clib_error_return (0, "tunnel local address not specified");
-
   if (remote_set == 0)
     return clib_error_return (0, "tunnel remote address not specified");
-
   if (ipv4_set && ipv6_set)
     return clib_error_return (0, "both IPv4 and IPv6 addresses specified");
-
-  if ((ipv4_set && memcmp (&local.ip4, &remote.ip4, sizeof (local.ip4)) == 0)
-      || (ipv6_set
-	  && memcmp (&local.ip6, &remote.ip6, sizeof (local.ip6)) == 0))
+  if ((ipv4_set
+       && memcmp (&local.ip4, &remote.ip4,
+		  sizeof (local.ip4)) == 0) || (ipv6_set
+						&&
+						memcmp
+						(&local.ip6,
+						 &remote.ip6,
+						 sizeof (local.ip6)) == 0))
     return clib_error_return (0, "src and dst addresses are identical");
-
   if (vni_set == 0)
     return clib_error_return (0, "vni not specified");
-
-
   if (!ipv6_set)
     {
       key4.local = local.ip4.as_u32;
       key4.remote = remote.ip4.as_u32;
       key4.vni = clib_host_to_net_u32 (vni << 8);
       key4.pad = 0;
-
       p = hash_get_mem (gm->vxlan4_gpe_tunnel_by_key, &key4);
     }
   else
@@ -251,20 +497,17 @@ vxlan_gpe_set_ioam_rewrite_command_fn (vlib_main_t * vm,
       key6.remote.as_u64[0] = remote.ip6.as_u64[0];
       key6.remote.as_u64[1] = remote.ip6.as_u64[1];
       key6.vni = clib_host_to_net_u32 (vni << 8);
-
       p = hash_get_mem (gm->vxlan6_gpe_tunnel_by_key, &key6);
     }
 
   if (!p)
     return clib_error_return (0, "VxLAN Tunnel not found");
-
   t = pool_elt_at_index (gm->tunnels, p[0]);
-
   if (!disable)
     {
-      rv = vxlan_gpe_ioam_set (t, hm->has_trace_option,
-			       hm->has_pot_option,
-			       hm->has_ppc_option, ipv6_set);
+      rv =
+	vxlan_gpe_ioam_set (t, hm->has_trace_option,
+			    hm->has_pot_option, hm->has_ppc_option, ipv6_set);
     }
   else
     {
@@ -285,11 +528,10 @@ VLIB_CLI_COMMAND (vxlan_gpe_set_ioam_rewrite_cmd, static) = {
 
 
 clib_error_t *
-vxlan_gpe_ioam_enable (int has_trace_option, int has_pot_option,
-		       int has_ppc_option)
+vxlan_gpe_ioam_enable (int has_trace_option,
+		       int has_pot_option, int has_ppc_option)
 {
   vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
-
   hm->has_trace_option = has_trace_option;
   hm->has_pot_option = has_pot_option;
   hm->has_ppc_option = has_ppc_option;
@@ -302,11 +544,11 @@ vxlan_gpe_ioam_enable (int has_trace_option, int has_pot_option,
 }
 
 clib_error_t *
-vxlan_gpe_ioam_disable (int has_trace_option, int has_pot_option,
-			int has_ppc_option)
+vxlan_gpe_ioam_disable (int
+			has_trace_option,
+			int has_pot_option, int has_ppc_option)
 {
   vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
-
   hm->has_trace_option = has_trace_option;
   hm->has_pot_option = has_pot_option;
   hm->has_ppc_option = has_ppc_option;
@@ -328,14 +570,13 @@ vxlan_gpe_set_next_override (uword next)
 
 static clib_error_t *
 vxlan_gpe_set_ioam_flags_command_fn (vlib_main_t * vm,
-				     unformat_input_t * input,
-				     vlib_cli_command_t * cmd)
+				     unformat_input_t
+				     * input, vlib_cli_command_t * cmd)
 {
   int has_trace_option = 0;
   int has_pot_option = 0;
   int has_ppc_option = 0;
   clib_error_t *rv = 0;
-
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (input, "trace"))
@@ -355,7 +596,6 @@ vxlan_gpe_set_ioam_flags_command_fn (vlib_main_t * vm,
 
   rv =
     vxlan_gpe_ioam_enable (has_trace_option, has_pot_option, has_ppc_option);
-
   return rv;
 }
 
@@ -368,12 +608,100 @@ VLIB_CLI_COMMAND (vxlan_gpe_set_ioam_flags_cmd, static) =
 /* *INDENT-ON* */
 
 
+int vxlan_gpe_ioam_disable_for_dest
+  (vlib_main_t * vm, ip46_address_t dst_addr, u32 outer_fib_index,
+   u8 ipv4_set)
+{
+  vxlan_gpe_ioam_dest_tunnels_t *t;
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+
+  vxlan_gpe_enable_disable_ioam_for_dest (hm->vlib_main,
+					  dst_addr, outer_fib_index, ipv4_set,
+					  0);
+  if (pool_elts (hm->dst_tunnels) == 0)
+    {
+      vxlan_gpe_clear_output_feature_on_select_intfs ();
+      return 0;
+    }
+
+  pool_foreach (t, hm->dst_tunnels, (
+				      {
+				      vxlan_gpe_enable_disable_ioam_for_dest
+				      (hm->vlib_main,
+				       t->dst_addr,
+				       t->outer_fib_index,
+				       (t->fp_proto ==
+					FIB_PROTOCOL_IP4), 1 /* is_add */ );
+				      }
+		));
+  vxlan_gpe_clear_output_feature_on_select_intfs ();
+  return (0);
 
+}
 
-clib_error_t *
-clear_vxlan_gpe_ioam_rewrite_command_fn (vlib_main_t * vm,
-					 unformat_input_t * input,
-					 vlib_cli_command_t * cmd)
+static clib_error_t *vxlan_gpe_set_ioam_transit_rewrite_command_fn
+  (vlib_main_t * vm, unformat_input_t * input, vlib_cli_command_t * cmd)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  ip46_address_t dst_addr;
+  u8 dst_addr_set = 0;
+  u8 ipv4_set = 0;
+  u8 ipv6_set = 0;
+  u8 disable = 0;
+  clib_error_t *rv = 0;
+  u32 outer_fib_index = 0;
+  while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (input, "dst-ip %U", unformat_ip4_address, &dst_addr.ip4))
+	{
+	  dst_addr_set = 1;
+	  ipv4_set = 1;
+	}
+      else
+	if (unformat
+	    (input, "dst-ip %U", unformat_ip6_address, &dst_addr.ip6))
+	{
+	  dst_addr_set = 1;
+	  ipv6_set = 1;
+	}
+      else if (unformat (input, "outer-fib-index %d", &outer_fib_index))
+	{
+	}
+
+      else if (unformat (input, "disable"))
+	disable = 1;
+      else
+	break;
+    }
+
+  if (dst_addr_set == 0)
+    return clib_error_return (0, "tunnel destination address not specified");
+  if (ipv4_set && ipv6_set)
+    return clib_error_return (0, "both IPv4 and IPv6 addresses specified");
+  if (!disable)
+    {
+      vxlan_gpe_enable_disable_ioam_for_dest (hm->vlib_main,
+					      dst_addr, outer_fib_index,
+					      ipv4_set, 1);
+    }
+  else
+    {
+      vxlan_gpe_ioam_disable_for_dest
+	(vm, dst_addr, outer_fib_index, ipv4_set);
+    }
+  return rv;
+}
+
+       /* *INDENT-OFF* */
+VLIB_CLI_COMMAND (vxlan_gpe_set_ioam_transit_rewrite_cmd, static) = {
+  .path = "set vxlan-gpe-ioam-transit",
+  .short_help = "set vxlan-gpe-ioam-transit dst-ip <dst_ip> [outer-fib-index <outer_fib_index>] [disable]",
+  .function = vxlan_gpe_set_ioam_transit_rewrite_command_fn,
+};
+/* *INDENT-ON* */
+
+clib_error_t *clear_vxlan_gpe_ioam_rewrite_command_fn
+  (vlib_main_t * vm, unformat_input_t * input, vlib_cli_command_t * cmd)
 {
   return (vxlan_gpe_ioam_disable (0, 0, 0));
 }
@@ -387,6 +715,55 @@ VLIB_CLI_COMMAND (vxlan_gpe_clear_ioam_flags_cmd, static) =
 };
 /* *INDENT-ON* */
 
+
+/**
+ * Function definition to backwalk a FIB node
+ */
+static fib_node_back_walk_rc_t
+vxlan_gpe_ioam_back_walk (fib_node_t * node, fib_node_back_walk_ctx_t * ctx)
+{
+  vxlan_gpe_refresh_output_feature_on_all_dest ();
+  return (FIB_NODE_BACK_WALK_CONTINUE);
+}
+
+/**
+ * Function definition to get a FIB node from its index
+ */
+static fib_node_t *
+vxlan_gpe_ioam_fib_node_get (fib_node_index_t index)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  return (&hm->node);
+}
+
+/**
+ * Function definition to inform the FIB node that its last lock has gone.
+ */
+static void
+vxlan_gpe_ioam_last_lock_gone (fib_node_t * node)
+{
+  ASSERT (0);
+}
+
+
+/*
+ * Virtual function table registered by MPLS GRE tunnels
+ * for participation in the FIB object graph.
+ */
+const static fib_node_vft_t vxlan_gpe_ioam_vft = {
+  .fnv_get = vxlan_gpe_ioam_fib_node_get,
+  .fnv_last_lock = vxlan_gpe_ioam_last_lock_gone,
+  .fnv_back_walk = vxlan_gpe_ioam_back_walk,
+};
+
+void
+vxlan_gpe_ioam_interface_init (void)
+{
+  vxlan_gpe_ioam_main_t *hm = &vxlan_gpe_ioam_main;
+  hm->fib_entry_type = fib_node_register_new_type (&vxlan_gpe_ioam_vft);
+  return;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *