NAT64: multi-thread support (VPP-891)

Change-Id: Iebf859b6d86482e4465423bad598eecf87e53ec4
Signed-off-by: Matus Fabian <matfabia@cisco.com>

17 files changed, 1349 insertions, 505 deletions

diff --git a/src/plugins/nat/dslite_in2out.c b/src/plugins/nat/dslite_in2out.c
index 54568914ce0..98b3a1618f7 100644
--- a/src/plugins/nat/dslite_in2out.c
+++ b/src/plugins/nat/dslite_in2out.c
@@ -108,13 +108,13 @@ slow_path (dslite_main_t * dm, dslite_session_key_t * in2out_key,
 
       if (snat_alloc_outside_address_and_port
 	  (dm->addr_pool, 0, thread_index, &out2in_key,
-	   &s->outside_address_index, 0, dm->port_per_thread, thread_index))
+	   &s->outside_address_index, dm->port_per_thread, thread_index))
 	ASSERT (0);
     }
   else
     {
       if (snat_alloc_outside_address_and_port
-	  (dm->addr_pool, 0, thread_index, &out2in_key, &address_index, 0,
+	  (dm->addr_pool, 0, thread_index, &out2in_key, &address_index,
 	   dm->port_per_thread, thread_index))
 	{
 	  *error = DSLITE_ERROR_OUT_OF_PORTS;
diff --git a/src/plugins/nat/in2out.c b/src/plugins/nat/in2out.c
index 50d8f84f2dd..1052451ed0a 100755
--- a/src/plugins/nat/in2out.c
+++ b/src/plugins/nat/in2out.c
@@ -398,7 +398,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
 
       if (snat_alloc_outside_address_and_port (sm->addresses, rx_fib_index0,
                                                thread_index, &key1,
-                                               &address_index, sm->vrf_mode,
+                                               &address_index,
                                                sm->port_per_thread,
                                                sm->per_thread_data[thread_index].snat_thread_index))
         {
@@ -420,7 +420,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
           /* Try to create dynamic translation */
           if (snat_alloc_outside_address_and_port (sm->addresses, rx_fib_index0,
                                                    thread_index, &key1,
-                                                   &address_index, sm->vrf_mode,
+                                                   &address_index,
                                                    sm->port_per_thread,
                                                    sm->per_thread_data[thread_index].snat_thread_index))
             {
diff --git a/src/plugins/nat/nat.c b/src/plugins/nat/nat.c
index 3c891f274b5..ef26d2243fd 100644
--- a/src/plugins/nat/nat.c
+++ b/src/plugins/nat/nat.c
@@ -332,9 +332,6 @@ void snat_add_address (snat_main_t *sm, ip4_address_t *addr, u32 vrf_id)
   snat_interface_t *i;
   vlib_thread_main_t *tm = vlib_get_thread_main ();
 
-  if (vrf_id != ~0)
-    sm->vrf_mode = 1;
-
   /* Check if address already exists */
   vec_foreach (ap, sm->addresses)
     {
@@ -1384,7 +1381,6 @@ nat_alloc_addr_and_port_default (snat_address_t * addresses,
                                  u32 thread_index,
                                  snat_session_key_t * k,
                                  u32 * address_indexp,
-                                 u8 vrf_mode,
                                  u16 port_per_thread,
                                  u32 snat_thread_index);
 
@@ -1604,14 +1600,13 @@ snat_alloc_outside_address_and_port (snat_address_t * addresses,
                                      u32 thread_index,
                                      snat_session_key_t * k,
                                      u32 * address_indexp,
-                                     u8 vrf_mode,
                                      u16 port_per_thread,
                                      u32 snat_thread_index)
 {
   snat_main_t *sm = &snat_main;
 
   return sm->alloc_addr_and_port(addresses, fib_index, thread_index, k,
-                                 address_indexp, vrf_mode, port_per_thread,
+                                 address_indexp, port_per_thread,
                                  snat_thread_index);
 }
 
@@ -1621,39 +1616,44 @@ nat_alloc_addr_and_port_default (snat_address_t * addresses,
                                  u32 thread_index,
                                  snat_session_key_t * k,
                                  u32 * address_indexp,
-                                 u8 vrf_mode,
                                  u16 port_per_thread,
                                  u32 snat_thread_index)
 {
-  int i;
-  snat_address_t *a;
+  int i, gi = 0;
+  snat_address_t *a, *ga = 0;
   u32 portnum;
 
   for (i = 0; i < vec_len (addresses); i++)
     {
       a = addresses + i;
-      if (vrf_mode && a->fib_index != ~0 && a->fib_index != fib_index)
-        continue;
       switch (k->protocol)
         {
 #define _(N, j, n, s) \
         case SNAT_PROTOCOL_##N: \
           if (a->busy_##n##_ports_per_thread[thread_index] < port_per_thread) \
             { \
-              while (1) \
+              if (a->fib_index == fib_index) \
                 { \
-                  portnum = (port_per_thread * \
-                    snat_thread_index) + \
-                    snat_random_port(1, port_per_thread) + 1024; \
-                  if (clib_bitmap_get_no_check (a->busy_##n##_port_bitmap, portnum)) \
-                    continue; \
-                  clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, portnum, 1); \
-                  a->busy_##n##_ports_per_thread[thread_index]++; \
-                  a->busy_##n##_ports++; \
-                  k->addr = a->addr; \
-                  k->port = clib_host_to_net_u16(portnum); \
-                  *address_indexp = i; \
-                  return 0; \
+                  while (1) \
+                    { \
+                      portnum = (port_per_thread * \
+                        snat_thread_index) + \
+                        snat_random_port(1, port_per_thread) + 1024; \
+                      if (clib_bitmap_get_no_check (a->busy_##n##_port_bitmap, portnum)) \
+                        continue; \
+                      clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, portnum, 1); \
+                      a->busy_##n##_ports_per_thread[thread_index]++; \
+                      a->busy_##n##_ports++; \
+                      k->addr = a->addr; \
+                      k->port = clib_host_to_net_u16(portnum); \
+                      *address_indexp = i; \
+                      return 0; \
+                    } \
+                } \
+              else if (a->fib_index == ~0) \
+                { \
+                  ga = a; \
+                  gi = i; \
                 } \
             } \
           break;
@@ -1665,6 +1665,38 @@ nat_alloc_addr_and_port_default (snat_address_t * addresses,
         }
 
     }
+
+  if (ga)
+    {
+      a = ga;
+      switch (k->protocol)
+	{
+#define _(N, j, n, s) \
+        case SNAT_PROTOCOL_##N: \
+          while (1) \
+            { \
+              portnum = (port_per_thread * \
+                snat_thread_index) + \
+                snat_random_port(1, port_per_thread) + 1024; \
+              if (clib_bitmap_get_no_check (a->busy_##n##_port_bitmap, portnum)) \
+                continue; \
+              clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, portnum, 1); \
+              a->busy_##n##_ports_per_thread[thread_index]++; \
+              a->busy_##n##_ports++; \
+              k->addr = a->addr; \
+              k->port = clib_host_to_net_u16(portnum); \
+              *address_indexp = gi; \
+              return 0; \
+            }
+	  break;
+	  foreach_snat_protocol
+#undef _
+	default:
+	  clib_warning ("unknown protocol");
+	  return 1;
+	}
+    }
+
   /* Totally out of translations to use... */
   snat_ipfix_logging_addresses_exhausted(0);
   return 1;
@@ -1676,7 +1708,6 @@ nat_alloc_addr_and_port_mape (snat_address_t * addresses,
                               u32 thread_index,
                               snat_session_key_t * k,
                               u32 * address_indexp,
-                              u8 vrf_mode,
                               u16 port_per_thread,
                               u32 snat_thread_index)
 {
@@ -2462,6 +2493,25 @@ snat_get_worker_out2in_cb (ip4_header_t * ip0, u32 rx_fib_index0)
   udp = ip4_next_header (ip0);
   port = udp->dst_port;
 
+  if (PREDICT_FALSE (ip4_is_fragment (ip0)))
+    {
+      if (PREDICT_FALSE (nat_reass_is_drop_frag (0)))
+	return vlib_get_thread_index ();
+
+      if (PREDICT_TRUE (!ip4_is_first_fragment (ip0)))
+	{
+	  nat_reass_ip4_t *reass;
+
+	  reass = nat_ip4_reass_find (ip0->src_address, ip0->dst_address,
+				      ip0->fragment_id, ip0->protocol);
+
+	  if (reass && (reass->thread_index != (u32) ~ 0))
+            return reass->thread_index;
+	  else
+	    return vlib_get_thread_index ();
+	}
+    }
+
   /* unknown protocol */
   if (PREDICT_FALSE (proto == ~0))
     {
@@ -2554,6 +2604,10 @@ snat_config (vlib_main_t * vm, unformat_input_t * input)
   u32 inside_vrf_id = 0;
   u32 static_mapping_buckets = 1024;
   u32 static_mapping_memory_size = 64<<20;
+  u32 nat64_bib_buckets = 1024;
+  u32 nat64_bib_memory_size = 128 << 20;
+  u32 nat64_st_buckets = 2048;
+  u32 nat64_st_memory_size = 256 << 20;
   u8 static_mapping_only = 0;
   u8 static_mapping_connection_tracking = 0;
   snat_main_per_thread_data_t *tsm;
@@ -2588,6 +2642,17 @@ snat_config (vlib_main_t * vm, unformat_input_t * input)
         }
       else if (unformat (input, "deterministic"))
         sm->deterministic = 1;
+      else if (unformat (input, "nat64 bib hash buckets %d",
+                         &nat64_bib_buckets))
+        ;
+      else if (unformat (input, "nat64 bib hash memory %d",
+                         &nat64_bib_memory_size))
+        ;
+      else if (unformat (input, "nat64 st hash buckets %d", &nat64_st_buckets))
+        ;
+      else if (unformat (input, "nat64 st hash memory %d",
+                         &nat64_st_memory_size))
+        ;
       else
 	return clib_error_return (0, "unknown input '%U'",
 				  format_unformat_error, input);
@@ -2612,6 +2677,9 @@ snat_config (vlib_main_t * vm, unformat_input_t * input)
   sm->static_mapping_only = static_mapping_only;
   sm->static_mapping_connection_tracking = static_mapping_connection_tracking;
 
+  nat64_set_hash(nat64_bib_buckets, nat64_bib_memory_size, nat64_st_buckets,
+                 nat64_st_memory_size);
+
   if (sm->deterministic)
     {
       sm->in2out_node_index = snat_det_in2out_node.index;
@@ -3424,8 +3492,8 @@ done:
 };
 
 VLIB_CLI_COMMAND (nat44_set_alloc_addr_and_port_alg_command, static) = {
-    .path = "nat44 addr-port-assignment-alg",
-    .short_help = "nat44 addr-port-assignment-alg <alg-name> [<alg-params>]",
+    .path = "nat addr-port-assignment-alg",
+    .short_help = "nat addr-port-assignment-alg <alg-name> [<alg-params>]",
     .function = nat44_set_alloc_addr_and_port_alg_command_fn,
 };
 
diff --git a/src/plugins/nat/nat.h b/src/plugins/nat/nat.h
index 8178d76a4ac..e82c23026cb 100644
--- a/src/plugins/nat/nat.h
+++ b/src/plugins/nat/nat.h
@@ -274,7 +274,6 @@ typedef int nat_alloc_out_addr_and_port_function_t (snat_address_t * addresses,
                                                     u32 thread_index,
                                                     snat_session_key_t * k,
                                                     u32 * address_indexp,
-                                                    u8 vrf_mode,
                                                     u16 port_per_thread,
                                                     u32 snat_thread_index);
 
@@ -355,9 +354,6 @@ typedef struct snat_main_s {
   u32 inside_vrf_id;
   u32 inside_fib_index;
 
-  /* tenant VRF aware address pool activation flag */
-  u8 vrf_mode;
-
   /* values of various timeouts */
   u32 udp_timeout;
   u32 tcp_established_timeout;
@@ -399,7 +395,6 @@ int snat_alloc_outside_address_and_port (snat_address_t * addresses,
                                          u32 thread_index,
                                          snat_session_key_t * k,
                                          u32 * address_indexp,
-                                         u8 vrf_mode,
                                          u16 port_per_thread,
                                          u32 snat_thread_index);
 
diff --git a/src/plugins/nat/nat64.c b/src/plugins/nat/nat64.c
index 0054310d18b..deeb0717b7e 100644
--- a/src/plugins/nat/nat64.c
+++ b/src/plugins/nat/nat64.c
@@ -19,7 +19,9 @@
 
 #include <nat/nat64.h>
 #include <nat/nat64_db.h>
+#include <nat/nat_reass.h>
 #include <vnet/fib/ip4_fib.h>
+#include <vppinfra/crc32.h>
 
 
 nat64_main_t nat64_main;
@@ -37,6 +39,17 @@ VNET_FEATURE_INIT (nat64_out2in, static) = {
   .node_name = "nat64-out2in",
   .runs_before = VNET_FEATURES ("ip4-lookup"),
 };
+VNET_FEATURE_INIT (nat64_in2out_handoff, static) = {
+  .arc_name = "ip6-unicast",
+  .node_name = "nat64-in2out-handoff",
+  .runs_before = VNET_FEATURES ("ip6-lookup"),
+};
+VNET_FEATURE_INIT (nat64_out2in_handoff, static) = {
+  .arc_name = "ip4-unicast",
+  .node_name = "nat64-out2in-handoff",
+  .runs_before = VNET_FEATURES ("ip4-lookup"),
+};
+
 
 static u8 well_known_prefix[] = {
   0x00, 0x64, 0xff, 0x9b,
@@ -80,28 +93,137 @@ nat64_ip4_add_del_interface_address_cb (ip4_main_t * im, uword opaque,
     }
 }
 
+u32
+nat64_get_worker_in2out (ip6_address_t * addr)
+{
+  nat64_main_t *nm = &nat64_main;
+  snat_main_t *sm = nm->sm;
+  u32 next_worker_index = nm->sm->first_worker_index;
+  u32 hash;
+
+#ifdef clib_crc32c_uses_intrinsics
+  hash = clib_crc32c ((u8 *) addr->as_u32, 16);
+#else
+  u64 tmp = addr->as_u64[0] ^ addr->as_u64[1];
+  hash = clib_xxhash (tmp);
+#endif
+
+  if (PREDICT_TRUE (is_pow2 (_vec_len (sm->workers))))
+    next_worker_index += sm->workers[hash & (_vec_len (sm->workers) - 1)];
+  else
+    next_worker_index += sm->workers[hash % _vec_len (sm->workers)];
+
+  return next_worker_index;
+}
+
+u32
+nat64_get_worker_out2in (ip4_header_t * ip)
+{
+  nat64_main_t *nm = &nat64_main;
+  snat_main_t *sm = nm->sm;
+  udp_header_t *udp;
+  u16 port;
+  u32 proto;
+
+  proto = ip_proto_to_snat_proto (ip->protocol);
+  udp = ip4_next_header (ip);
+  port = udp->dst_port;
+
+  /* fragments */
+  if (PREDICT_FALSE (ip4_is_fragment (ip)))
+    {
+      if (PREDICT_FALSE (nat_reass_is_drop_frag (0)))
+	return vlib_get_thread_index ();
+
+      if (PREDICT_TRUE (!ip4_is_first_fragment (ip)))
+	{
+	  nat_reass_ip4_t *reass;
+
+	  reass = nat_ip4_reass_find (ip->src_address, ip->dst_address,
+				      ip->fragment_id, ip->protocol);
+
+	  if (reass && (reass->thread_index != (u32) ~ 0))
+	    return reass->thread_index;
+	  else
+	    return vlib_get_thread_index ();
+	}
+    }
+
+  /* unknown protocol */
+  if (PREDICT_FALSE (proto == ~0))
+    {
+      nat64_db_t *db;
+      ip46_address_t daddr;
+      nat64_db_bib_entry_t *bibe;
+
+      memset (&daddr, 0, sizeof (daddr));
+      daddr.ip4.as_u32 = ip->dst_address.as_u32;
+
+      /* *INDENT-OFF* */
+      vec_foreach (db, nm->db)
+        {
+          bibe = nat64_db_bib_entry_find (db, &daddr, 0, ip->protocol, 0, 0);
+          if (bibe)
+            return (u32) (db - nm->db);
+        }
+      /* *INDENT-ON* */
+      return vlib_get_thread_index ();
+    }
+
+  /* ICMP */
+  if (PREDICT_FALSE (ip->protocol == IP_PROTOCOL_ICMP))
+    {
+      icmp46_header_t *icmp = (icmp46_header_t *) udp;
+      icmp_echo_header_t *echo = (icmp_echo_header_t *) (icmp + 1);
+      if (!icmp_is_error_message (icmp))
+	port = echo->identifier;
+      else
+	{
+	  ip4_header_t *inner_ip = (ip4_header_t *) (echo + 1);
+	  proto = ip_proto_to_snat_proto (inner_ip->protocol);
+	  void *l4_header = ip4_next_header (inner_ip);
+	  switch (proto)
+	    {
+	    case SNAT_PROTOCOL_ICMP:
+	      icmp = (icmp46_header_t *) l4_header;
+	      echo = (icmp_echo_header_t *) (icmp + 1);
+	      port = echo->identifier;
+	      break;
+	    case SNAT_PROTOCOL_UDP:
+	    case SNAT_PROTOCOL_TCP:
+	      port = ((tcp_udp_header_t *) l4_header)->src_port;
+	      break;
+	    default:
+	      return vlib_get_thread_index ();
+	    }
+	}
+    }
+
+  /* worker by outside port  (TCP/UDP) */
+  port = clib_net_to_host_u16 (port);
+  if (port > 1024)
+    return (u32) ((port - 1024) / sm->port_per_thread);
+
+  return vlib_get_thread_index ();
+}
+
 clib_error_t *
 nat64_init (vlib_main_t * vm)
 {
   nat64_main_t *nm = &nat64_main;
-  clib_error_t *error = 0;
   vlib_thread_main_t *tm = vlib_get_thread_main ();
   ip4_add_del_interface_address_callback_t cb4;
   ip4_main_t *im = &ip4_main;
+  vlib_node_t *error_drop_node =
+    vlib_get_node_by_name (vm, (u8 *) "error-drop");
 
-  nm->is_disabled = 0;
+  vec_validate (nm->db, tm->n_vlib_mains - 1);
 
-  if (tm->n_vlib_mains > 1)
-    {
-      nm->is_disabled = 1;
-      goto error;
-    }
+  nm->sm = &snat_main;
 
-  if (nat64_db_init (&nm->db))
-    {
-      error = clib_error_return (0, "NAT64 DB init failed");
-      goto error;
-    }
+  nm->fq_in2out_index = ~0;
+  nm->fq_out2in_index = ~0;
+  nm->error_node_index = error_drop_node->index;
 
   /* set session timeouts to default values */
   nm->udp_timeout = SNAT_UDP_TIMEOUT;
@@ -116,8 +238,29 @@ nat64_init (vlib_main_t * vm)
   vec_add1 (im->add_del_interface_address_callbacks, cb4);
   nm->ip4_main = im;
 
-error:
-  return error;
+  return 0;
+}
+
+void
+nat64_set_hash (u32 bib_buckets, u32 bib_memory_size, u32 st_buckets,
+		u32 st_memory_size)
+{
+  nat64_main_t *nm = &nat64_main;
+  nat64_db_t *db;
+
+  nm->bib_buckets = bib_buckets;
+  nm->bib_memory_size = bib_memory_size;
+  nm->st_buckets = st_buckets;
+  nm->st_memory_size = st_memory_size;
+
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    {
+      if (nat64_db_init (db, bib_buckets, bib_memory_size, st_buckets,
+                         st_memory_size))
+        clib_warning ("NAT64 DB init failed");
+    }
+  /* *INDENT-ON* */
 }
 
 int
@@ -127,6 +270,8 @@ nat64_add_del_pool_addr (ip4_address_t * addr, u32 vrf_id, u8 is_add)
   snat_address_t *a = 0;
   snat_interface_t *interface;
   int i;
+  nat64_db_t *db;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
 
   /* Check if address already exists */
   for (i = 0; i < vec_len (nm->addr_pool); i++)
@@ -145,13 +290,15 @@ nat64_add_del_pool_addr (ip4_address_t * addr, u32 vrf_id, u8 is_add)
 
       vec_add2 (nm->addr_pool, a, 1);
       a->addr = *addr;
-      a->fib_index = 0;
+      a->fib_index = ~0;
       if (vrf_id != ~0)
 	a->fib_index =
 	  fib_table_find_or_create_and_lock (FIB_PROTOCOL_IP6, vrf_id,
 					     FIB_SOURCE_PLUGIN_HI);
-#define _(N, i, n, s) \
-      clib_bitmap_alloc (a->busy_##n##_port_bitmap, 65535);
+#define _(N, id, n, s) \
+      clib_bitmap_alloc (a->busy_##n##_port_bitmap, 65535); \
+      a->busy_##n##_ports = 0; \
+      vec_validate_init_empty (a->busy_##n##_ports_per_thread, tm->n_vlib_mains - 1, 0);
       foreach_snat_protocol
 #undef _
     }
@@ -160,17 +307,19 @@ nat64_add_del_pool_addr (ip4_address_t * addr, u32 vrf_id, u8 is_add)
       if (!a)
 	return VNET_API_ERROR_NO_SUCH_ENTRY;
 
-      if (a->fib_index)
+      if (a->fib_index != ~0)
 	fib_table_unlock (a->fib_index, FIB_PROTOCOL_IP6,
 			  FIB_SOURCE_PLUGIN_HI);
-
 #define _(N, id, n, s) \
       clib_bitmap_free (a->busy_##n##_port_bitmap);
       foreach_snat_protocol
 #undef _
 	/* Delete sessions using address */
-	nat64_db_free_out_addr (&nm->db, &a->addr);
-      vec_del1 (nm->addr_pool, i);
+        /* *INDENT-OFF* */
+        vec_foreach (db, nm->db)
+          nat64_db_free_out_addr (db, &a->addr);
+        /* *INDENT-ON* */
+	vec_del1 (nm->addr_pool, i);
     }
 
   /* Add/del external address to FIB */
@@ -300,8 +449,21 @@ nat64_add_del_interface (u32 sw_if_index, u8 is_inside, u8 is_add)
       /* *INDENT-ON* */
     }
 
+  if (nm->sm->num_workers > 1)
+    {
+      feature_name =
+	is_inside ? "nat64-in2out-handoff" : "nat64-out2in-handoff";
+      if (nm->fq_in2out_index == ~0)
+	nm->fq_in2out_index =
+	  vlib_frame_queue_main_init (nat64_in2out_node.index, 0);
+      if (nm->fq_out2in_index == ~0)
+	nm->fq_out2in_index =
+	  vlib_frame_queue_main_init (nat64_out2in_node.index, 0);
+    }
+  else
+    feature_name = is_inside ? "nat64-in2out" : "nat64-out2in";
+
   arc_name = is_inside ? "ip6-unicast" : "ip4-unicast";
-  feature_name = is_inside ? "nat64-in2out" : "nat64-out2in";
 
   return vnet_feature_enable_disable (arc_name, feature_name, sw_if_index,
 				      is_add, 0, 0);
@@ -324,93 +486,33 @@ nat64_interfaces_walk (nat64_interface_walk_fn_t fn, void *ctx)
 
 int
 nat64_alloc_out_addr_and_port (u32 fib_index, snat_protocol_t proto,
-			       ip4_address_t * addr, u16 * port)
+			       ip4_address_t * addr, u16 * port,
+			       u32 thread_index)
 {
   nat64_main_t *nm = &nat64_main;
-  snat_main_t *sm = &snat_main;
-  int i;
-  snat_address_t *a, *ga = 0;
-  u32 portnum;
+  snat_main_t *sm = nm->sm;
+  snat_session_key_t k;
+  u32 ai;
+  int rv;
 
-  for (i = 0; i < vec_len (nm->addr_pool); i++)
-    {
-      a = nm->addr_pool + i;
-      switch (proto)
-	{
-#define _(N, j, n, s) \
-        case SNAT_PROTOCOL_##N: \
-          if (a->busy_##n##_ports < (65535-1024)) \
-            { \
-              if (a->fib_index == fib_index) \
-                { \
-                  while (1) \
-                    { \
-                      portnum = random_u32 (&sm->random_seed); \
-                      portnum &= 0xFFFF; \
-                      if (portnum < 1024) \
-                        continue; \
-                      if (clib_bitmap_get_no_check (a->busy_##n##_port_bitmap, \
-                                                    portnum)) \
-                        continue; \
-                      clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, \
-                                                portnum, 1); \
-                      a->busy_##n##_ports++; \
-                      *port = portnum; \
-                      addr->as_u32 = a->addr.as_u32; \
-                      return 0; \
-                    } \
-                 } \
-               else if (a->fib_index == 0) \
-                 ga = a; \
-            } \
-          break;
-	  foreach_snat_protocol
-#undef _
-	default:
-	  clib_warning ("unknown protocol");
-	  return 1;
-	}
-    }
+  k.protocol = proto;
 
-  if (ga)
+  rv =
+    sm->alloc_addr_and_port (nm->addr_pool, fib_index, thread_index, &k, &ai,
+			     sm->port_per_thread, thread_index);
+
+  if (!rv)
     {
-      switch (proto)
-	{
-#define _(N, j, n, s) \
-        case SNAT_PROTOCOL_##N: \
-          while (1) \
-            { \
-              portnum = random_u32 (&sm->random_seed); \
-              portnum &= 0xFFFF; \
-              if (portnum < 1024) \
-                continue; \
-              if (clib_bitmap_get_no_check (a->busy_##n##_port_bitmap, \
-                                            portnum)) \
-                continue; \
-              clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, \
-                                        portnum, 1); \
-              a->busy_##n##_ports++; \
-              *port = portnum; \
-              addr->as_u32 = a->addr.as_u32; \
-              return 0; \
-            }
-	  break;
-	  foreach_snat_protocol
-#undef _
-	default:
-	  clib_warning ("unknown protocol");
-	  return 1;
-	}
+      *port = k.port;
+      addr->as_u32 = k.addr.as_u32;
     }
 
-  /* Totally out of translations to use... */
-  //TODO: IPFix
-  return 1;
+  return rv;
 }
 
 void
 nat64_free_out_addr_and_port (ip4_address_t * addr, u16 port,
-			      snat_protocol_t proto)
+			      snat_protocol_t proto, u32 thread_index)
 {
   nat64_main_t *nm = &nat64_main;
   int i;
@@ -429,6 +531,7 @@ nat64_free_out_addr_and_port (ip4_address_t * addr, u16 port,
                   port) == 1); \
           clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, port, 0); \
           a->busy_##n##_ports--; \
+          a->busy_##n##_ports_per_thread[thread_index]--; \
           break;
 	  foreach_snat_protocol
 #undef _
@@ -440,6 +543,62 @@ nat64_free_out_addr_and_port (ip4_address_t * addr, u16 port,
     }
 }
 
+/**
+ * @brief Add/delete static BIB entry in worker thread.
+ */
+static uword
+nat64_static_bib_worker_fn (vlib_main_t * vm, vlib_node_runtime_t * rt,
+			    vlib_frame_t * f)
+{
+  nat64_main_t *nm = &nat64_main;
+  u32 thread_index = vlib_get_thread_index ();
+  nat64_db_t *db = &nm->db[thread_index];
+  nat64_static_bib_to_update_t *static_bib;
+  nat64_db_bib_entry_t *bibe;
+  ip46_address_t addr;
+
+  /* *INDENT-OFF* */
+  pool_foreach (static_bib, nm->static_bibs,
+  ({
+    if ((static_bib->thread_index != thread_index) || (static_bib->done))
+      continue;
+
+    if (static_bib->is_add)
+      (void) nat64_db_bib_entry_create (db, &static_bib->in_addr,
+                                        &static_bib->out_addr,
+                                        static_bib->in_port,
+                                        static_bib->out_port,
+				        static_bib->fib_index,
+                                        static_bib->proto, 1);
+    else
+      {
+        addr.as_u64[0] = static_bib->in_addr.as_u64[0];
+        addr.as_u64[1] = static_bib->in_addr.as_u64[1];
+        bibe = nat64_db_bib_entry_find (db, &addr, static_bib->in_port,
+                                        static_bib->proto,
+                                        static_bib->fib_index, 1);
+        if (bibe)
+          nat64_db_bib_entry_free (db, bibe);
+      }
+
+      static_bib->done = 1;
+  }));
+  /* *INDENT-ON* */
+
+  return 0;
+}
+
+static vlib_node_registration_t nat64_static_bib_worker_node;
+
+/* *INDENT-OFF* */
+VLIB_REGISTER_NODE (nat64_static_bib_worker_node, static) = {
+    .function = nat64_static_bib_worker_fn,
+    .type = VLIB_NODE_TYPE_INPUT,
+    .state = VLIB_NODE_STATE_INTERRUPT,
+    .name = "nat64-static-bib-worker",
+};
+/* *INDENT-ON* */
+
 int
 nat64_add_del_static_bib_entry (ip6_address_t * in_addr,
 				ip4_address_t * out_addr, u16 in_port,
@@ -453,11 +612,24 @@ nat64_add_del_static_bib_entry (ip6_address_t * in_addr,
   ip46_address_t addr;
   int i;
   snat_address_t *a;
+  u32 thread_index = 0;
+  nat64_db_t *db;
+  nat64_static_bib_to_update_t *static_bib;
+  vlib_main_t *worker_vm;
+  u32 *to_be_free = 0, *index;
+
+  if (nm->sm->num_workers > 1)
+    {
+      thread_index = nat64_get_worker_in2out (in_addr);
+      db = &nm->db[thread_index];
+    }
+  else
+    db = &nm->db[nm->sm->num_workers];
 
   addr.as_u64[0] = in_addr->as_u64[0];
   addr.as_u64[1] = in_addr->as_u64[1];
   bibe =
-    nat64_db_bib_entry_find (&nm->db, &addr, clib_host_to_net_u16 (in_port),
+    nat64_db_bib_entry_find (db, &addr, clib_host_to_net_u16 (in_port),
 			     proto, fib_index, 1);
 
   if (is_add)
@@ -465,6 +637,13 @@ nat64_add_del_static_bib_entry (ip6_address_t * in_addr,
       if (bibe)
 	return VNET_API_ERROR_VALUE_EXIST;
 
+      /* outside port must be assigned to same thread as internall address */
+      if ((out_port > 1024) && (nm->sm->num_workers > 1))
+	{
+	  if (thread_index != ((out_port - 1024) / nm->sm->port_per_thread))
+	    return VNET_API_ERROR_INVALID_VALUE_2;
+	}
+
       for (i = 0; i < vec_len (nm->addr_pool); i++)
 	{
 	  a = nm->addr_pool + i;
@@ -480,34 +659,73 @@ nat64_add_del_static_bib_entry (ip6_address_t * in_addr,
               clib_bitmap_set_no_check (a->busy_##n##_port_bitmap, \
                                         out_port, 1); \
               if (out_port > 1024) \
-                a->busy_##n##_ports++; \
+                { \
+                  a->busy_##n##_ports++; \
+                  a->busy_##n##_ports_per_thread[thread_index]++; \
+                } \
               break;
 	      foreach_snat_protocol
 #undef _
 	    default:
 	      memset (&addr, 0, sizeof (addr));
 	      addr.ip4.as_u32 = out_addr->as_u32;
-	      if (nat64_db_bib_entry_find
-		  (&nm->db, &addr, 0, proto, fib_index, 0))
+	      if (nat64_db_bib_entry_find (db, &addr, 0, proto, fib_index, 0))
 		return VNET_API_ERROR_INVALID_VALUE;
 	    }
 	  break;
 	}
-      bibe =
-	nat64_db_bib_entry_create (&nm->db, in_addr, out_addr,
-				   clib_host_to_net_u16 (in_port),
-				   clib_host_to_net_u16 (out_port), fib_index,
-				   proto, 1);
-      if (!bibe)
-	return VNET_API_ERROR_UNSPECIFIED;
+      if (!nm->sm->num_workers)
+	{
+	  bibe =
+	    nat64_db_bib_entry_create (db, in_addr, out_addr,
+				       clib_host_to_net_u16 (in_port),
+				       clib_host_to_net_u16 (out_port),
+				       fib_index, proto, 1);
+	  if (!bibe)
+	    return VNET_API_ERROR_UNSPECIFIED;
+	}
     }
   else
     {
       if (!bibe)
 	return VNET_API_ERROR_NO_SUCH_ENTRY;
 
-      nat64_free_out_addr_and_port (out_addr, out_port, p);
-      nat64_db_bib_entry_free (&nm->db, bibe);
+      if (!nm->sm->num_workers)
+	{
+	  nat64_free_out_addr_and_port (out_addr, out_port, p, thread_index);
+	  nat64_db_bib_entry_free (db, bibe);
+	}
+    }
+
+  if (nm->sm->num_workers)
+    {
+      /* *INDENT-OFF* */
+      pool_foreach (static_bib, nm->static_bibs,
+      ({
+        if (static_bib->done)
+          vec_add1 (to_be_free, static_bib - nm->static_bibs);
+      }));
+      vec_foreach (index, to_be_free)
+        pool_put_index (nm->static_bibs, index[0]);
+      /* *INDENT-ON* */
+      vec_free (to_be_free);
+      pool_get (nm->static_bibs, static_bib);
+      static_bib->in_addr.as_u64[0] = in_addr->as_u64[0];
+      static_bib->in_addr.as_u64[1] = in_addr->as_u64[1];
+      static_bib->in_port = clib_host_to_net_u16 (in_port);
+      static_bib->out_addr.as_u32 = out_addr->as_u32;
+      static_bib->out_port = clib_host_to_net_u16 (out_port);
+      static_bib->fib_index = fib_index;
+      static_bib->proto = proto;
+      static_bib->is_add = is_add;
+      static_bib->thread_index = thread_index;
+      static_bib->done = 0;
+      worker_vm = vlib_mains[thread_index];
+      if (worker_vm)
+	vlib_node_set_interrupt_pending (worker_vm,
+					 nat64_static_bib_worker_node.index);
+      else
+	return VNET_API_ERROR_UNSPECIFIED;
     }
 
   return 0;
@@ -911,23 +1129,65 @@ nat64_extract_ip4 (ip6_address_t * ip6, ip4_address_t * ip4, u32 fib_index)
 }
 
 /**
- * @brief The 'nat64-expire-walk' process's main loop.
- *
- * Check expire time for NAT64 sessions.
+ * @brief Per worker process checking expire time for NAT64 sessions.
+ */
+static uword
+nat64_expire_worker_walk_fn (vlib_main_t * vm, vlib_node_runtime_t * rt,
+			     vlib_frame_t * f)
+{
+  nat64_main_t *nm = &nat64_main;
+  u32 thread_index = vlib_get_thread_index ();
+  nat64_db_t *db = &nm->db[thread_index];
+  u32 now = (u32) vlib_time_now (vm);
+
+  nad64_db_st_free_expired (db, now);
+
+  return 0;
+}
+
+static vlib_node_registration_t nat64_expire_worker_walk_node;
+
+/* *INDENT-OFF* */
+VLIB_REGISTER_NODE (nat64_expire_worker_walk_node, static) = {
+    .function = nat64_expire_worker_walk_fn,
+    .type = VLIB_NODE_TYPE_INPUT,
+    .state = VLIB_NODE_STATE_INTERRUPT,
+    .name = "nat64-expire-worker-walk",
+};
+/* *INDENT-ON* */
+
+/**
+ * @brief Centralized process to drive per worker expire walk.
  */
 static uword
 nat64_expire_walk_fn (vlib_main_t * vm, vlib_node_runtime_t * rt,
 		      vlib_frame_t * f)
 {
-  nat64_main_t *nm = &nat64_main;
+  vlib_main_t **worker_vms = 0, *worker_vm;
+  int i;
 
-  while (!nm->is_disabled)
+  if (vec_len (vlib_mains) == 0)
+    vec_add1 (worker_vms, vm);
+  else
+    {
+      for (i = 0; i < vec_len (vlib_mains); i++)
+	{
+	  worker_vm = vlib_mains[i];
+	  if (worker_vm)
+	    vec_add1 (worker_vms, worker_vm);
+	}
+    }
+
+  while (1)
     {
       vlib_process_wait_for_event_or_clock (vm, 10.0);
       vlib_process_get_events (vm, NULL);
-      u32 now = (u32) vlib_time_now (vm);
-
-      nad64_db_st_free_expired (&nm->db, now);
+      for (i = 0; i < vec_len (worker_vms); i++)
+	{
+	  worker_vm = worker_vms[i];
+	  vlib_node_set_interrupt_pending (worker_vm,
+					   nat64_expire_worker_walk_node.index);
+	}
     }
 
   return 0;
diff --git a/src/plugins/nat/nat64.h b/src/plugins/nat/nat64.h
index 118076705cc..0c8fd82a730 100644
--- a/src/plugins/nat/nat64.h
+++ b/src/plugins/nat/nat64.h
@@ -49,6 +49,19 @@ typedef struct
 
 typedef struct
 {
+  ip6_address_t in_addr;
+  u16 in_port;
+  ip4_address_t out_addr;
+  u16 out_port;
+  u32 fib_index;
+  u32 thread_index;
+  u8 proto;
+  u8 is_add;
+  u8 done;
+} nat64_static_bib_to_update_t;
+
+typedef struct
+{
   /** Interface pool */
   snat_interface_t *interfaces;
 
@@ -61,18 +74,31 @@ typedef struct
   /** Pref64 vector */
   nat64_prefix_t *pref64;
 
-  /** BIB and session DB */
-  nat64_db_t db;
+  /** BIB and session DB per thread */
+  nat64_db_t *db;
+
+  /** Worker handoff */
+  u32 fq_in2out_index;
+  u32 fq_out2in_index;
+
+  /** Pool of static BIB entries to be added/deleted in worker threads */
+  nat64_static_bib_to_update_t *static_bibs;
 
-  /* values of various timeouts */
+  u32 error_node_index;
+
+  /** config parameters */
+  u32 bib_buckets;
+  u32 bib_memory_size;
+  u32 st_buckets;
+  u32 st_memory_size;
+
+  /** values of various timeouts */
   u32 udp_timeout;
   u32 icmp_timeout;
   u32 tcp_trans_timeout;
   u32 tcp_est_timeout;
   u32 tcp_incoming_syn_timeout;
 
-  u8 is_disabled;
-
   ip4_main_t *ip4_main;
   snat_main_t *sm;
 } nat64_main_t;
@@ -171,27 +197,30 @@ int nat64_add_del_static_bib_entry (ip6_address_t * in_addr,
 /**
  * @brief Alloce IPv4 address and port pair from NAT64 pool.
  *
- * @param fib_index FIB index of tenant.
- * @param proto     L4 protocol.
- * @param addr      Allocated IPv4 address.
- * @param port      Allocated port number.
+ * @param fib_index    FIB index of tenant.
+ * @param proto        L4 protocol.
+ * @param addr         Allocated IPv4 address.
+ * @param port         Allocated port number.
+ * @param thread_index Thread index.
  *
  * @returns 0 on success, non-zero value otherwise.
  */
 int nat64_alloc_out_addr_and_port (u32 fib_index, snat_protocol_t proto,
-				   ip4_address_t * addr, u16 * port);
+				   ip4_address_t * addr, u16 * port,
+				   u32 thread_index);
 
 /**
  * @brief Free IPv4 address and port pair from NAT64 pool.
  *
- * @param addr  IPv4 address to free.
- * @param port  Port number to free.
- * @param proto L4 protocol.
+ * @param addr         IPv4 address to free.
+ * @param port         Port number to free.
+ * @param proto        L4 protocol.
+ * @param thread_index Thread index.
  *
  * @returns 0 on success, non-zero value otherwise.
  */
 void nat64_free_out_addr_and_port (ip4_address_t * addr, u16 port,
-				   snat_protocol_t proto);
+				   snat_protocol_t proto, u32 thread_index);
 
 /**
  * @brief Set UDP session timeout.
@@ -322,6 +351,35 @@ void nat64_compose_ip6 (ip6_address_t * ip6, ip4_address_t * ip4,
 void nat64_extract_ip4 (ip6_address_t * ip6, ip4_address_t * ip4,
 			u32 fib_index);
 
+/**
+ * @brief Set NAT64 hash tables configuration.
+ *
+ * @param bib_buckets Number of BIB hash buckets.
+ * @param bib_memory_size Memory size of BIB hash.
+ * @param st_buckets Number of session table hash buckets.
+ * @param st_memory_size Memory size of session table hash.
+ */
+void nat64_set_hash (u32 bib_buckets, u32 bib_memory_size, u32 st_buckets,
+		     u32 st_memory_size);
+
+/**
+ * @brief Get worker thread index for NAT64 in2out.
+ *
+ * @param addr IPv6 src address.
+ *
+ * @returns worker thread index.
+ */
+u32 nat64_get_worker_in2out (ip6_address_t * addr);
+
+/**
+ * @brief Get worker thread index for NAT64 out2in.
+ *
+ * @param ip IPv4 header.
+ *
+ * @returns worker thread index.
+ */
+u32 nat64_get_worker_out2in (ip4_header_t * ip);
+
 #define u8_ptr_add(ptr, index) (((u8 *)ptr) + index)
 #define u16_net_add(u, val) clib_host_to_net_u16(clib_net_to_host_u16(u) + (val))
 
diff --git a/src/plugins/nat/nat64_cli.c b/src/plugins/nat/nat64_cli.c
index 3e15beeee3b..22bfa41eb48 100644
--- a/src/plugins/nat/nat64_cli.c
+++ b/src/plugins/nat/nat64_cli.c
@@ -26,7 +26,6 @@ nat64_add_del_pool_addr_command_fn (vlib_main_t * vm,
 				    unformat_input_t * input,
 				    vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   unformat_input_t _line_input, *line_input = &_line_input;
   ip4_address_t start_addr, end_addr, this_addr;
   u32 start_host_order, end_host_order;
@@ -35,10 +34,6 @@ nat64_add_del_pool_addr_command_fn (vlib_main_t * vm,
   u8 is_add = 1;
   clib_error_t *error = 0;
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   /* Get a line of input. */
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
@@ -129,12 +124,6 @@ nat64_show_pool_command_fn (vlib_main_t * vm,
 			    unformat_input_t * input,
 			    vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   vlib_cli_output (vm, "NAT64 pool:");
   nat64_pool_addr_walk (nat64_cli_pool_walk, vm);
 
@@ -146,7 +135,6 @@ nat64_interface_feature_command_fn (vlib_main_t * vm,
 				    unformat_input_t *
 				    input, vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   unformat_input_t _line_input, *line_input = &_line_input;
   vnet_main_t *vnm = vnet_get_main ();
   clib_error_t *error = 0;
@@ -156,10 +144,6 @@ nat64_interface_feature_command_fn (vlib_main_t * vm,
   u8 is_add = 1;
   int i, rv;
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   /* Get a line of input. */
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
@@ -279,12 +263,6 @@ nat64_show_interfaces_command_fn (vlib_main_t * vm,
 				  unformat_input_t *
 				  input, vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   vlib_cli_output (vm, "NAT64 interfaces:");
   nat64_interfaces_walk (nat64_cli_interface_walk, vm);
 
@@ -297,7 +275,6 @@ nat64_add_del_static_bib_command_fn (vlib_main_t *
 				     unformat_input_t
 				     * input, vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   unformat_input_t _line_input, *line_input = &_line_input;
   clib_error_t *error = 0;
   u8 is_add = 1;
@@ -310,10 +287,6 @@ nat64_add_del_static_bib_command_fn (vlib_main_t *
   u8 p = 0;
   int rv;
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
 
@@ -383,6 +356,8 @@ nat64_add_del_static_bib_command_fn (vlib_main_t *
 	clib_error_return (0, "Outside addres %U and port %u already in use.",
 			   format_ip4_address, &out_addr, out_port);
       goto done;
+    case VNET_API_ERROR_INVALID_VALUE_2:
+      error = clib_error_return (0, "Invalid outside port.");
     default:
       break;
     }
@@ -436,10 +411,7 @@ nat64_show_bib_command_fn (vlib_main_t * vm,
   clib_error_t *error = 0;
   u32 proto = ~0;
   u8 p = 255;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
+  nat64_db_t *db;
 
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
@@ -462,7 +434,11 @@ nat64_show_bib_command_fn (vlib_main_t * vm,
   else
     vlib_cli_output (vm, "NAT64 %U BIB entries:", format_snat_protocol,
 		     proto);
-  nat64_db_bib_walk (&nm->db, p, nat64_cli_bib_walk, vm);
+
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    nat64_db_bib_walk (db, p, nat64_cli_bib_walk, vm);
+  /* *INDENT-ON* */
 
 done:
   unformat_free (line_input);
@@ -474,7 +450,6 @@ static clib_error_t *
 nat64_set_timeouts_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			       vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   unformat_input_t _line_input, *line_input = &_line_input;
   clib_error_t *error = 0;
   u32 timeout, tcp_trans, tcp_est, tcp_incoming_syn;
@@ -483,10 +458,6 @@ nat64_set_timeouts_command_fn (vlib_main_t * vm, unformat_input_t * input,
   tcp_est = nat64_get_tcp_est_timeout ();
   tcp_incoming_syn = nat64_get_tcp_incoming_syn_timeout ();
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
 
@@ -563,12 +534,6 @@ static clib_error_t *
 nat64_show_timeouts_command_fn (vlib_main_t * vm, unformat_input_t * input,
 				vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   vlib_cli_output (vm, "NAT64 session timeouts:");
   vlib_cli_output (vm, " UDP %usec", nat64_get_udp_timeout ());
   vlib_cli_output (vm, " ICMP %usec", nat64_get_icmp_timeout ());
@@ -582,15 +547,21 @@ nat64_show_timeouts_command_fn (vlib_main_t * vm, unformat_input_t * input,
   return 0;
 }
 
+typedef struct nat64_cli_st_walk_ctx_t_
+{
+  vlib_main_t *vm;
+  nat64_db_t *db;
+} nat64_cli_st_walk_ctx_t;
+
 static int
-nat64_cli_st_walk (nat64_db_st_entry_t * ste, void *ctx)
+nat64_cli_st_walk (nat64_db_st_entry_t * ste, void *arg)
 {
-  vlib_main_t *vm = ctx;
-  nat64_main_t *nm = &nat64_main;
+  nat64_cli_st_walk_ctx_t *ctx = arg;
+  vlib_main_t *vm = ctx->vm;
   nat64_db_bib_entry_t *bibe;
   fib_table_t *fib;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, ste->proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (ctx->db, ste->proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -642,10 +613,10 @@ nat64_show_st_command_fn (vlib_main_t * vm,
   clib_error_t *error = 0;
   u32 proto = ~0;
   u8 p = 255;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
+  nat64_db_t *db;
+  nat64_cli_st_walk_ctx_t ctx = {
+    .vm = vm,
+  };
 
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
@@ -667,7 +638,13 @@ nat64_show_st_command_fn (vlib_main_t * vm,
     vlib_cli_output (vm, "NAT64 sessions:");
   else
     vlib_cli_output (vm, "NAT64 %U sessions:", format_snat_protocol, proto);
-  nat64_db_st_walk (&nm->db, p, nat64_cli_st_walk, vm);
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    {
+      ctx.db = db;
+      nat64_db_st_walk (db, p, nat64_cli_st_walk, &ctx);
+    }
+  /* *INDENT-ON* */
 
 done:
   unformat_free (line_input);
@@ -679,7 +656,6 @@ static clib_error_t *
 nat64_add_del_prefix_command_fn (vlib_main_t * vm, unformat_input_t * input,
 				 vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   vnet_main_t *vnm = vnet_get_main ();
   clib_error_t *error = 0;
   unformat_input_t _line_input, *line_input = &_line_input;
@@ -689,10 +665,6 @@ nat64_add_del_prefix_command_fn (vlib_main_t * vm, unformat_input_t * input,
   u32 plen = 0;
   int rv;
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
 
@@ -748,26 +720,29 @@ nat64_add_del_prefix_command_fn (vlib_main_t * vm, unformat_input_t * input,
       fib_prefix_t fibpfx = {
 	.fp_len = plen,
 	.fp_proto = FIB_PROTOCOL_IP6,
-	.fp_addr = {.ip6 = prefix}
+	.fp_addr = {
+		    .ip6 = prefix}
       };
 
       if (is_add)
 	{
-	  fib_index = fib_table_find_or_create_and_lock (FIB_PROTOCOL_IP6,
-							 vrf_id,
-							 FIB_SOURCE_PLUGIN_HI);
+	  fib_index =
+	    fib_table_find_or_create_and_lock (FIB_PROTOCOL_IP6,
+					       vrf_id, FIB_SOURCE_PLUGIN_HI);
 	  fib_table_entry_update_one_path (fib_index, &fibpfx,
 					   FIB_SOURCE_PLUGIN_HI,
-					   FIB_ENTRY_FLAG_NONE, DPO_PROTO_IP6,
-					   NULL, sw_if_index, ~0, 0, NULL,
-					   FIB_ROUTE_PATH_INTF_RX);
+					   FIB_ENTRY_FLAG_NONE,
+					   DPO_PROTO_IP6, NULL,
+					   sw_if_index, ~0, 0,
+					   NULL, FIB_ROUTE_PATH_INTF_RX);
 	}
       else
 	{
 	  fib_index = fib_table_find (FIB_PROTOCOL_IP6, vrf_id);
 	  fib_table_entry_path_remove (fib_index, &fibpfx,
-				       FIB_SOURCE_PLUGIN_HI, DPO_PROTO_IP6,
-				       NULL, sw_if_index, ~0, 1,
+				       FIB_SOURCE_PLUGIN_HI,
+				       DPO_PROTO_IP6, NULL,
+				       sw_if_index, ~0, 1,
 				       FIB_ROUTE_PATH_INTF_RX);
 	  fib_table_unlock (fib_index, FIB_PROTOCOL_IP6,
 			    FIB_SOURCE_PLUGIN_HI);
@@ -792,15 +767,10 @@ nat64_cli_prefix_walk (nat64_prefix_t * p, void *ctx)
 }
 
 static clib_error_t *
-nat64_show_prefix_command_fn (vlib_main_t * vm, unformat_input_t * input,
+nat64_show_prefix_command_fn (vlib_main_t * vm,
+			      unformat_input_t * input,
 			      vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
-
   vlib_cli_output (vm, "NAT64 prefix:");
   nat64_prefix_walk (nat64_cli_prefix_walk, vm);
 
@@ -812,7 +782,6 @@ nat64_add_interface_address_command_fn (vlib_main_t * vm,
 					unformat_input_t * input,
 					vlib_cli_command_t * cmd)
 {
-  nat64_main_t *nm = &nat64_main;
   vnet_main_t *vnm = vnet_get_main ();
   unformat_input_t _line_input, *line_input = &_line_input;
   u32 sw_if_index;
@@ -820,9 +789,6 @@ nat64_add_interface_address_command_fn (vlib_main_t * vm,
   int is_add = 1;
   clib_error_t *error = 0;
 
-  if (nm->is_disabled)
-    return clib_error_return (0,
-			      "NAT64 disabled, multi thread not supported");
   /* Get a line of input. */
   if (!unformat_user (input, unformat_line_input, line_input))
     return 0;
@@ -830,8 +796,7 @@ nat64_add_interface_address_command_fn (vlib_main_t * vm,
   while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat
-	  (line_input, "%U", unformat_vnet_sw_interface, vnm, &sw_if_index))
-	;
+	  (line_input, "%U", unformat_vnet_sw_interface, vnm, &sw_if_index));
       else if (unformat (line_input, "del"))
 	is_add = 0;
       else
diff --git a/src/plugins/nat/nat64_db.c b/src/plugins/nat/nat64_db.c
index 008a137da10..97da672a928 100644
--- a/src/plugins/nat/nat64_db.c
+++ b/src/plugins/nat/nat64_db.c
@@ -19,13 +19,9 @@
 #include <nat/nat64_db.h>
 
 int
-nat64_db_init (nat64_db_t * db)
+nat64_db_init (nat64_db_t * db, u32 bib_buckets, u32 bib_memory_size,
+	       u32 st_buckets, u32 st_memory_size)
 {
-  u32 bib_buckets = 1024;
-  u32 bib_memory_size = 128 << 20;
-  u32 st_buckets = 2048;
-  u32 st_memory_size = 256 << 20;
-
   clib_bihash_init_24_8 (&db->bib.in2out, "bib-in2out", bib_buckets,
 			 bib_memory_size);
 
diff --git a/src/plugins/nat/nat64_db.h b/src/plugins/nat/nat64_db.h
index 94d9a8bdebf..7cda8b09a72 100644
--- a/src/plugins/nat/nat64_db.h
+++ b/src/plugins/nat/nat64_db.h
@@ -127,10 +127,15 @@ typedef struct
  * @brief Initialize NAT64 DB.
  *
  * @param db NAT64 DB.
+ * @param bib_buckets Number of BIB hash buckets.
+ * @param bib_memory_size Memory size of BIB hash.
+ * @param st_buckets Number of session table hash buckets.
+ * @param st_memory_size Memory size of session table hash.
  *
  * @returns 0 on success, non-zero value otherwise.
  */
-int nat64_db_init (nat64_db_t * db);
+int nat64_db_init (nat64_db_t * db, u32 bib_buckets, u32 bib_memory_size,
+		   u32 st_buckets, u32 st_memory_size);
 
 /**
  * @brief Create new NAT64 BIB entry.
diff --git a/src/plugins/nat/nat64_in2out.c b/src/plugins/nat/nat64_in2out.c
index 4f94575ebcc..9f77ca33fa4 100644
--- a/src/plugins/nat/nat64_in2out.c
+++ b/src/plugins/nat/nat64_in2out.c
@@ -72,6 +72,7 @@ format_nat64_in2out_reass_trace (u8 * s, va_list * args)
 vlib_node_registration_t nat64_in2out_node;
 vlib_node_registration_t nat64_in2out_slowpath_node;
 vlib_node_registration_t nat64_in2out_reass_node;
+vlib_node_registration_t nat64_in2out_handoff_node;
 
 #define foreach_nat64_in2out_error                       \
 _(UNSUPPORTED_PROTOCOL, "unsupported protocol")          \
@@ -111,6 +112,7 @@ typedef struct nat64_in2out_set_ctx_t_
 {
   vlib_buffer_t *b;
   vlib_main_t *vm;
+  u32 thread_index;
 } nat64_in2out_set_ctx_t;
 
 /**
@@ -152,6 +154,7 @@ nat64_in2out_tcp_udp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
   u8 proto = ip6->protocol;
   u16 sport = udp->src_port;
   u16 dport = udp->dst_port;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -163,19 +166,18 @@ nat64_in2out_tcp_udp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
   daddr.as_u64[1] = ip6->dst_address.as_u64[1];
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &saddr, &daddr, sport, dport, proto,
+    nat64_db_st_entry_find (db, &saddr, &daddr, sport, dport, proto,
 			    fib_index, 1);
 
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &saddr, sport, proto, fib_index, 1);
+      bibe = nat64_db_bib_entry_find (db, &saddr, sport, proto, fib_index, 1);
 
       if (!bibe)
 	{
@@ -183,11 +185,11 @@ nat64_in2out_tcp_udp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 	  ip4_address_t out_addr;
 	  if (nat64_alloc_out_addr_and_port
 	      (fib_index, ip_proto_to_snat_proto (proto), &out_addr,
-	       &out_port))
+	       &out_port, ctx->thread_index))
 	    return -1;
 
 	  bibe =
-	    nat64_db_bib_entry_create (&nm->db, &ip6->src_address, &out_addr,
+	    nat64_db_bib_entry_create (db, &ip6->src_address, &out_addr,
 				       sport, clib_host_to_net_u16 (out_port),
 				       fib_index, proto, 0);
 	  if (!bibe)
@@ -196,7 +198,7 @@ nat64_in2out_tcp_udp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 
       nat64_extract_ip4 (&ip6->dst_address, &daddr.ip4, fib_index);
       ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->dst_address,
+	nat64_db_st_entry_create (db, bibe, &ip6->dst_address,
 				  &daddr.ip4, dport);
       if (!ste)
 	return -1;
@@ -234,6 +236,7 @@ nat64_in2out_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
   ip46_address_t saddr, daddr;
   u32 sw_if_index, fib_index;
   icmp46_header_t *icmp = ip6_next_header (ip6);
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -248,13 +251,13 @@ nat64_in2out_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
     {
       u16 in_id = ((u16 *) (icmp))[2];
       ste =
-	nat64_db_st_entry_find (&nm->db, &saddr, &daddr, in_id, 0,
+	nat64_db_st_entry_find (db, &saddr, &daddr, in_id, 0,
 				IP_PROTOCOL_ICMP, fib_index, 1);
 
       if (ste)
 	{
 	  bibe =
-	    nat64_db_bib_entry_by_index (&nm->db, IP_PROTOCOL_ICMP,
+	    nat64_db_bib_entry_by_index (db, IP_PROTOCOL_ICMP,
 					 ste->bibe_index);
 	  if (!bibe)
 	    return -1;
@@ -262,7 +265,7 @@ nat64_in2out_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
       else
 	{
 	  bibe =
-	    nat64_db_bib_entry_find (&nm->db, &saddr, in_id,
+	    nat64_db_bib_entry_find (db, &saddr, in_id,
 				     IP_PROTOCOL_ICMP, fib_index, 1);
 
 	  if (!bibe)
@@ -270,11 +273,12 @@ nat64_in2out_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
 	      u16 out_id;
 	      ip4_address_t out_addr;
 	      if (nat64_alloc_out_addr_and_port
-		  (fib_index, SNAT_PROTOCOL_ICMP, &out_addr, &out_id))
+		  (fib_index, SNAT_PROTOCOL_ICMP, &out_addr, &out_id,
+		   ctx->thread_index))
 		return -1;
 
 	      bibe =
-		nat64_db_bib_entry_create (&nm->db, &ip6->src_address,
+		nat64_db_bib_entry_create (db, &ip6->src_address,
 					   &out_addr, in_id,
 					   clib_host_to_net_u16 (out_id),
 					   fib_index, IP_PROTOCOL_ICMP, 0);
@@ -284,7 +288,7 @@ nat64_in2out_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
 
 	  nat64_extract_ip4 (&ip6->dst_address, &daddr.ip4, fib_index);
 	  ste =
-	    nat64_db_st_entry_create (&nm->db, bibe, &ip6->dst_address,
+	    nat64_db_st_entry_create (db, bibe, &ip6->dst_address,
 				      &daddr.ip4, 0);
 	  if (!ste)
 	    return -1;
@@ -320,6 +324,7 @@ nat64_in2out_inner_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
   ip46_address_t saddr, daddr;
   u32 sw_if_index, fib_index;
   u8 proto = ip6->protocol;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -342,12 +347,12 @@ nat64_in2out_inner_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 	return -1;
 
       ste =
-	nat64_db_st_entry_find (&nm->db, &daddr, &saddr, in_id, 0, proto,
+	nat64_db_st_entry_find (db, &daddr, &saddr, in_id, 0, proto,
 				fib_index, 1);
       if (!ste)
 	return -1;
 
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
 
@@ -366,12 +371,12 @@ nat64_in2out_inner_icmp_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
       u16 dport = udp->dst_port;
 
       ste =
-	nat64_db_st_entry_find (&nm->db, &daddr, &saddr, dport, sport, proto,
+	nat64_db_st_entry_find (db, &daddr, &saddr, dport, sport, proto,
 				fib_index, 1);
       if (!ste)
 	return -1;
 
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
 
@@ -397,6 +402,7 @@ typedef struct unk_proto_st_walk_ctx_t_
   ip6_address_t dst_addr;
   ip4_address_t out_addr;
   u32 fib_index;
+  u32 thread_index;
   u8 proto;
 } unk_proto_st_walk_ctx_t;
 
@@ -407,11 +413,11 @@ unk_proto_st_walk (nat64_db_st_entry_t * ste, void *arg)
   unk_proto_st_walk_ctx_t *ctx = arg;
   nat64_db_bib_entry_t *bibe;
   ip46_address_t saddr, daddr;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   if (ip46_address_is_equal (&ste->in_r_addr, &ctx->dst_addr))
     {
-      bibe =
-	nat64_db_bib_entry_by_index (&nm->db, ste->proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, ste->proto, ste->bibe_index);
       if (!bibe)
 	return -1;
 
@@ -424,7 +430,7 @@ unk_proto_st_walk (nat64_db_st_entry_t * ste, void *arg)
 	  nat64_extract_ip4 (&ctx->dst_addr, &daddr.ip4, ctx->fib_index);
 
 	  if (nat64_db_st_entry_find
-	      (&nm->db, &daddr, &saddr, 0, 0, ctx->proto, ctx->fib_index, 0))
+	      (db, &daddr, &saddr, 0, 0, ctx->proto, ctx->fib_index, 0))
 	    return -1;
 
 	  ctx->out_addr.as_u32 = bibe->out_addr.as_u32;
@@ -440,15 +446,16 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 			       void *arg)
 {
   nat64_main_t *nm = &nat64_main;
-  nat64_in2out_set_ctx_t *ctx = arg;
+  nat64_in2out_set_ctx_t *s_ctx = arg;
   nat64_db_bib_entry_t *bibe;
   nat64_db_st_entry_t *ste;
   ip46_address_t saddr, daddr, addr;
   u32 sw_if_index, fib_index;
   u8 proto = ip6->protocol;
   int i;
+  nat64_db_t *db = &nm->db[s_ctx->thread_index];
 
-  sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
+  sw_if_index = vnet_buffer (s_ctx->b)->sw_if_index[VLIB_RX];
   fib_index =
     fib_table_get_index_for_sw_if_index (FIB_PROTOCOL_IP6, sw_if_index);
 
@@ -458,19 +465,17 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
   daddr.as_u64[1] = ip6->dst_address.as_u64[1];
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &saddr, &daddr, 0, 0, proto, fib_index,
-			    1);
+    nat64_db_st_entry_find (db, &saddr, &daddr, 0, 0, proto, fib_index, 1);
 
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &saddr, 0, proto, fib_index, 1);
+      bibe = nat64_db_bib_entry_find (db, &saddr, 0, proto, fib_index, 1);
 
       if (!bibe)
 	{
@@ -483,19 +488,18 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 	    .out_addr.as_u32 = 0,
 	    .fib_index = fib_index,
 	    .proto = proto,
+	    .thread_index = s_ctx->thread_index,
 	  };
 
-	  nat64_db_st_walk (&nm->db, IP_PROTOCOL_TCP, unk_proto_st_walk,
-			    &ctx);
+	  nat64_db_st_walk (db, IP_PROTOCOL_TCP, unk_proto_st_walk, &ctx);
 
 	  if (!ctx.out_addr.as_u32)
-	    nat64_db_st_walk (&nm->db, IP_PROTOCOL_UDP, unk_proto_st_walk,
-			      &ctx);
+	    nat64_db_st_walk (db, IP_PROTOCOL_UDP, unk_proto_st_walk, &ctx);
 
 	  /* Verify if out address is not already in use for protocol */
 	  memset (&addr, 0, sizeof (addr));
 	  addr.ip4.as_u32 = ctx.out_addr.as_u32;
-	  if (nat64_db_bib_entry_find (&nm->db, &addr, 0, proto, 0, 0))
+	  if (nat64_db_bib_entry_find (db, &addr, 0, proto, 0, 0))
 	    ctx.out_addr.as_u32 = 0;
 
 	  if (!ctx.out_addr.as_u32)
@@ -503,8 +507,7 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 	      for (i = 0; i < vec_len (nm->addr_pool); i++)
 		{
 		  addr.ip4.as_u32 = nm->addr_pool[i].addr.as_u32;
-		  if (!nat64_db_bib_entry_find
-		      (&nm->db, &addr, 0, proto, 0, 0))
+		  if (!nat64_db_bib_entry_find (db, &addr, 0, proto, 0, 0))
 		    break;
 		}
 	    }
@@ -513,7 +516,7 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 	    return -1;
 
 	  bibe =
-	    nat64_db_bib_entry_create (&nm->db, &ip6->src_address,
+	    nat64_db_bib_entry_create (db, &ip6->src_address,
 				       &ctx.out_addr, 0, 0, fib_index, proto,
 				       0);
 	  if (!bibe)
@@ -522,13 +525,12 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 
       nat64_extract_ip4 (&ip6->dst_address, &daddr.ip4, fib_index);
       ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->dst_address,
-				  &daddr.ip4, 0);
+	nat64_db_st_entry_create (db, bibe, &ip6->dst_address, &daddr.ip4, 0);
       if (!ste)
 	return -1;
     }
 
-  nat64_session_reset_timeout (ste, ctx->vm);
+  nat64_session_reset_timeout (ste, s_ctx->vm);
 
   ip4->src_address.as_u32 = bibe->out_addr.as_u32;
   ip4->dst_address.as_u32 = ste->out_r_addr.as_u32;
@@ -540,7 +542,7 @@ nat64_in2out_unk_proto_set_cb (ip6_header_t * ip6, ip4_header_t * ip4,
 
 static int
 nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
-				  ip6_header_t * ip6)
+				  ip6_header_t * ip6, u32 thread_index)
 {
   nat64_main_t *nm = &nat64_main;
   nat64_db_bib_entry_t *bibe;
@@ -554,6 +556,7 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   u16 dport = udp->dst_port;
   u16 *checksum;
   ip_csum_t csum;
+  nat64_db_t *db = &nm->db[thread_index];
 
   sw_if_index = vnet_buffer (b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -577,19 +580,18 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   csum = ip_csum_sub_even (csum, dport);
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &saddr, &daddr, sport, dport, proto,
+    nat64_db_st_entry_find (db, &saddr, &daddr, sport, dport, proto,
 			    fib_index, 1);
 
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &saddr, sport, proto, fib_index, 1);
+      bibe = nat64_db_bib_entry_find (db, &saddr, sport, proto, fib_index, 1);
 
       if (!bibe)
 	{
@@ -597,11 +599,11 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 	  ip4_address_t out_addr;
 	  if (nat64_alloc_out_addr_and_port
 	      (fib_index, ip_proto_to_snat_proto (proto), &out_addr,
-	       &out_port))
+	       &out_port, thread_index))
 	    return -1;
 
 	  bibe =
-	    nat64_db_bib_entry_create (&nm->db, &ip6->src_address, &out_addr,
+	    nat64_db_bib_entry_create (db, &ip6->src_address, &out_addr,
 				       sport, clib_host_to_net_u16 (out_port),
 				       fib_index, proto, 0);
 	  if (!bibe)
@@ -610,7 +612,7 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 
       nat64_extract_ip4 (&ip6->dst_address, &daddr.ip4, fib_index);
       ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->dst_address,
+	nat64_db_st_entry_create (db, bibe, &ip6->dst_address,
 				  &daddr.ip4, dport);
       if (!ste)
 	return -1;
@@ -621,32 +623,22 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   sport = udp->src_port = bibe->out_port;
   nat64_compose_ip6 (&ip6->src_address, &bibe->out_addr, fib_index);
 
-  memset (&saddr, 0, sizeof (saddr));
   memset (&daddr, 0, sizeof (daddr));
-  saddr.ip4.as_u32 = bibe->out_addr.as_u32;
   daddr.ip4.as_u32 = ste->out_r_addr.as_u32;
 
-  ste =
-    nat64_db_st_entry_find (&nm->db, &daddr, &saddr, dport, sport, proto, 0,
-			    0);
-
-  if (ste)
+  bibe = 0;
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
-      if (!bibe)
-	return -1;
-    }
-  else
-    {
-      bibe = nat64_db_bib_entry_find (&nm->db, &daddr, dport, proto, 0, 0);
-
-      if (!bibe)
-	return -1;
+      bibe = nat64_db_bib_entry_find (db, &daddr, dport, proto, 0, 0);
 
-      ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->src_address,
-				  &saddr.ip4, sport);
+      if (bibe)
+	break;
     }
+  /* *INDENT-ON* */
+
+  if (!bibe)
+    return -1;
 
   ip6->dst_address.as_u64[0] = bibe->in_addr.as_u64[0];
   ip6->dst_address.as_u64[1] = bibe->in_addr.as_u64[1];
@@ -665,7 +657,7 @@ nat64_in2out_tcp_udp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 
 static int
 nat64_in2out_icmp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
-			       ip6_header_t * ip6)
+			       ip6_header_t * ip6, u32 thread_index)
 {
   nat64_main_t *nm = &nat64_main;
   nat64_db_bib_entry_t *bibe;
@@ -679,6 +671,7 @@ nat64_in2out_icmp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   tcp_header_t *tcp;
   u16 *checksum, sport, dport;
   ip_csum_t csum;
+  nat64_db_t *db = &nm->db[thread_index];
 
   if (icmp->type == ICMP6_echo_request || icmp->type == ICMP6_echo_reply)
     return -1;
@@ -718,12 +711,12 @@ nat64_in2out_icmp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   csum = ip_csum_sub_even (csum, dport);
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &daddr, &saddr, dport, sport, proto,
+    nat64_db_st_entry_find (db, &daddr, &saddr, dport, sport, proto,
 			    fib_index, 1);
   if (!ste)
     return -1;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -735,13 +728,22 @@ nat64_in2out_icmp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   saddr.ip4.as_u32 = ste->out_r_addr.as_u32;
   daddr.ip4.as_u32 = bibe->out_addr.as_u32;
 
-  ste =
-    nat64_db_st_entry_find (&nm->db, &saddr, &daddr, sport, dport, proto, 0,
-			    0);
+  ste = 0;
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    {
+      ste = nat64_db_st_entry_find (db, &saddr, &daddr, sport, dport, proto,
+                                    0, 0);
+
+      if (ste)
+        break;
+    }
+  /* *INDENT-ON* */
+
   if (!ste)
     return -1;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -781,7 +783,7 @@ nat64_in2out_icmp_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 
 static int
 nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
-				    ip6_header_t * ip6)
+				    ip6_header_t * ip6, u32 thread_index)
 {
   nat64_main_t *nm = &nat64_main;
   nat64_db_bib_entry_t *bibe;
@@ -790,6 +792,7 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   u32 sw_if_index, fib_index;
   u8 proto = ip6->protocol;
   int i;
+  nat64_db_t *db = &nm->db[thread_index];
 
   sw_if_index = vnet_buffer (b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -801,19 +804,17 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
   daddr.as_u64[1] = ip6->dst_address.as_u64[1];
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &saddr, &daddr, 0, 0, proto, fib_index,
-			    1);
+    nat64_db_st_entry_find (db, &saddr, &daddr, 0, 0, proto, fib_index, 1);
 
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &saddr, 0, proto, fib_index, 1);
+      bibe = nat64_db_bib_entry_find (db, &saddr, 0, proto, fib_index, 1);
 
       if (!bibe)
 	{
@@ -826,19 +827,18 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 	    .out_addr.as_u32 = 0,
 	    .fib_index = fib_index,
 	    .proto = proto,
+	    .thread_index = thread_index,
 	  };
 
-	  nat64_db_st_walk (&nm->db, IP_PROTOCOL_TCP, unk_proto_st_walk,
-			    &ctx);
+	  nat64_db_st_walk (db, IP_PROTOCOL_TCP, unk_proto_st_walk, &ctx);
 
 	  if (!ctx.out_addr.as_u32)
-	    nat64_db_st_walk (&nm->db, IP_PROTOCOL_UDP, unk_proto_st_walk,
-			      &ctx);
+	    nat64_db_st_walk (db, IP_PROTOCOL_UDP, unk_proto_st_walk, &ctx);
 
 	  /* Verify if out address is not already in use for protocol */
 	  memset (&addr, 0, sizeof (addr));
 	  addr.ip4.as_u32 = ctx.out_addr.as_u32;
-	  if (nat64_db_bib_entry_find (&nm->db, &addr, 0, proto, 0, 0))
+	  if (nat64_db_bib_entry_find (db, &addr, 0, proto, 0, 0))
 	    ctx.out_addr.as_u32 = 0;
 
 	  if (!ctx.out_addr.as_u32)
@@ -846,8 +846,7 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 	      for (i = 0; i < vec_len (nm->addr_pool); i++)
 		{
 		  addr.ip4.as_u32 = nm->addr_pool[i].addr.as_u32;
-		  if (!nat64_db_bib_entry_find
-		      (&nm->db, &addr, 0, proto, 0, 0))
+		  if (!nat64_db_bib_entry_find (db, &addr, 0, proto, 0, 0))
 		    break;
 		}
 	    }
@@ -856,7 +855,7 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 	    return -1;
 
 	  bibe =
-	    nat64_db_bib_entry_create (&nm->db, &ip6->src_address,
+	    nat64_db_bib_entry_create (db, &ip6->src_address,
 				       &ctx.out_addr, 0, 0, fib_index, proto,
 				       0);
 	  if (!bibe)
@@ -865,8 +864,7 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 
       nat64_extract_ip4 (&ip6->dst_address, &daddr.ip4, fib_index);
       ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->dst_address,
-				  &daddr.ip4, 0);
+	nat64_db_st_entry_create (db, bibe, &ip6->dst_address, &daddr.ip4, 0);
       if (!ste)
 	return -1;
     }
@@ -875,30 +873,22 @@ nat64_in2out_unk_proto_hairpinning (vlib_main_t * vm, vlib_buffer_t * b,
 
   nat64_compose_ip6 (&ip6->src_address, &bibe->out_addr, fib_index);
 
-  memset (&saddr, 0, sizeof (saddr));
   memset (&daddr, 0, sizeof (daddr));
-  saddr.ip4.as_u32 = bibe->out_addr.as_u32;
   daddr.ip4.as_u32 = ste->out_r_addr.as_u32;
 
-  ste = nat64_db_st_entry_find (&nm->db, &daddr, &saddr, 0, 0, proto, 0, 0);
-
-  if (ste)
+  bibe = 0;
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
-      if (!bibe)
-	return -1;
-    }
-  else
-    {
-      bibe = nat64_db_bib_entry_find (&nm->db, &daddr, 0, proto, 0, 0);
+      bibe = nat64_db_bib_entry_find (db, &daddr, 0, proto, 0, 0);
 
-      if (!bibe)
-	return -1;
-
-      ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->src_address,
-				  &saddr.ip4, 0);
+      if (bibe)
+	break;
     }
+  /* *INDENT-ON* */
+
+  if (!bibe)
+    return -1;
 
   ip6->dst_address.as_u64[0] = bibe->in_addr.as_u64[0];
   ip6->dst_address.as_u64[1] = bibe->in_addr.as_u64[1];
@@ -914,6 +904,7 @@ nat64_in2out_node_fn_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   nat64_in2out_next_t next_index;
   u32 pkts_processed = 0;
   u32 stats_node_index;
+  u32 thread_index = vlib_get_thread_index ();
 
   stats_node_index =
     is_slow_path ? nat64_in2out_slowpath_node.index : nat64_in2out_node.index;
@@ -952,6 +943,7 @@ nat64_in2out_node_fn_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  ctx0.b = b0;
 	  ctx0.vm = vm;
+	  ctx0.thread_index = thread_index;
 
 	  next0 = NAT64_IN2OUT_NEXT_IP4_LOOKUP;
 
@@ -974,7 +966,8 @@ nat64_in2out_node_fn_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  if (is_hairpinning (&ip60->dst_address))
 		    {
 		      next0 = NAT64_IN2OUT_NEXT_IP6_LOOKUP;
-		      if (nat64_in2out_unk_proto_hairpinning (vm, b0, ip60))
+		      if (nat64_in2out_unk_proto_hairpinning
+			  (vm, b0, ip60, thread_index))
 			{
 			  next0 = NAT64_IN2OUT_NEXT_DROP;
 			  b0->error =
@@ -1014,7 +1007,8 @@ nat64_in2out_node_fn_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      if (is_hairpinning (&ip60->dst_address))
 		{
 		  next0 = NAT64_IN2OUT_NEXT_IP6_LOOKUP;
-		  if (nat64_in2out_icmp_hairpinning (vm, b0, ip60))
+		  if (nat64_in2out_icmp_hairpinning
+		      (vm, b0, ip60, thread_index))
 		    {
 		      next0 = NAT64_IN2OUT_NEXT_DROP;
 		      b0->error =
@@ -1037,7 +1031,8 @@ nat64_in2out_node_fn_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      if (is_hairpinning (&ip60->dst_address))
 		{
 		  next0 = NAT64_IN2OUT_NEXT_IP6_LOOKUP;
-		  if (nat64_in2out_tcp_udp_hairpinning (vm, b0, ip60))
+		  if (nat64_in2out_tcp_udp_hairpinning
+		      (vm, b0, ip60, thread_index))
 		    {
 		      next0 = NAT64_IN2OUT_NEXT_DROP;
 		      b0->error =
@@ -1145,6 +1140,7 @@ typedef struct nat64_in2out_frag_set_ctx_t_
 {
   vlib_main_t *vm;
   u32 sess_index;
+  u32 thread_index;
   u16 l4_offset;
   u8 proto;
   u8 first_frag;
@@ -1158,12 +1154,13 @@ nat64_in2out_frag_set_cb (ip6_header_t * ip6, ip4_header_t * ip4, void *arg)
   nat64_db_st_entry_t *ste;
   nat64_db_bib_entry_t *bibe;
   udp_header_t *udp;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
-  ste = nat64_db_st_entry_by_index (&nm->db, ctx->proto, ctx->sess_index);
+  ste = nat64_db_st_entry_by_index (db, ctx->proto, ctx->sess_index);
   if (!ste)
     return -1;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, ctx->proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (db, ctx->proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -1213,7 +1210,8 @@ nat64_in2out_frag_hairpinning (vlib_buffer_t * b, ip6_header_t * ip6,
   u16 dport = udp->dst_port;
   u16 *checksum;
   ip_csum_t csum;
-  ip46_address_t saddr, daddr;
+  ip46_address_t daddr;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   if (ctx->first_frag)
     {
@@ -1230,11 +1228,11 @@ nat64_in2out_frag_hairpinning (vlib_buffer_t * b, ip6_header_t * ip6,
       csum = ip_csum_sub_even (csum, dport);
     }
 
-  ste = nat64_db_st_entry_by_index (&nm->db, ctx->proto, ctx->sess_index);
+  ste = nat64_db_st_entry_by_index (db, ctx->proto, ctx->sess_index);
   if (!ste)
     return -1;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, ctx->proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (db, ctx->proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -1245,34 +1243,22 @@ nat64_in2out_frag_hairpinning (vlib_buffer_t * b, ip6_header_t * ip6,
 
   nat64_compose_ip6 (&ip6->src_address, &bibe->out_addr, bibe->fib_index);
 
-  memset (&saddr, 0, sizeof (saddr));
   memset (&daddr, 0, sizeof (daddr));
-  saddr.ip4.as_u32 = bibe->out_addr.as_u32;
   daddr.ip4.as_u32 = ste->out_r_addr.as_u32;
 
-  ste =
-    nat64_db_st_entry_find (&nm->db, &daddr, &saddr, dport, sport, ctx->proto,
-			    0, 0);
-
-  if (ste)
-    {
-      bibe =
-	nat64_db_bib_entry_by_index (&nm->db, ctx->proto, ste->bibe_index);
-      if (!bibe)
-	return -1;
-    }
-  else
+  bibe = 0;
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &daddr, dport, ctx->proto, 0, 0);
+      bibe = nat64_db_bib_entry_find (db, &daddr, dport, ctx->proto, 0, 0);
 
-      if (!bibe)
-	return -1;
-
-      ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6->src_address,
-				  &saddr.ip4, sport);
+      if (bibe)
+	break;
     }
+  /* *INDENT-ON* */
+
+  if (!bibe)
+    return -1;
 
   ip6->dst_address.as_u64[0] = bibe->in_addr.as_u64[0];
   ip6->dst_address.as_u64[1] = bibe->in_addr.as_u64[1];
@@ -1303,6 +1289,7 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
   u32 *fragments_to_drop = 0;
   u32 *fragments_to_loopback = 0;
   nat64_main_t *nm = &nat64_main;
+  u32 thread_index = vlib_get_thread_index ();
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -1332,6 +1319,7 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 	  u32 sw_if_index0, fib_index0;
 	  ip46_address_t saddr0, daddr0;
 	  nat64_in2out_frag_set_ctx_t ctx0;
+	  nat64_db_t *db = &nm->db[thread_index];
 
 	  /* speculatively enqueue b0 to the current next frame */
 	  bi0 = from[0];
@@ -1349,6 +1337,8 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 	    fib_table_get_index_for_sw_if_index (FIB_PROTOCOL_IP6,
 						 sw_if_index0);
 
+	  ctx0.thread_index = thread_index;
+
 	  if (PREDICT_FALSE (nat_reass_is_drop_frag (1)))
 	    {
 	      next0 = NAT64_IN2OUT_NEXT_DROP;
@@ -1419,20 +1409,21 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 	      daddr0.as_u64[1] = ip60->dst_address.as_u64[1];
 
 	      ste0 =
-		nat64_db_st_entry_find (&nm->db, &saddr0, &daddr0,
+		nat64_db_st_entry_find (db, &saddr0, &daddr0,
 					udp0->src_port, udp0->dst_port,
 					l4_protocol0, fib_index0, 1);
 	      if (!ste0)
 		{
 		  bibe0 =
-		    nat64_db_bib_entry_find (&nm->db, &saddr0, udp0->src_port,
+		    nat64_db_bib_entry_find (db, &saddr0, udp0->src_port,
 					     l4_protocol0, fib_index0, 1);
 		  if (!bibe0)
 		    {
 		      u16 out_port0;
 		      ip4_address_t out_addr0;
 		      if (nat64_alloc_out_addr_and_port
-			  (fib_index0, proto0, &out_addr0, &out_port0))
+			  (fib_index0, proto0, &out_addr0, &out_port0,
+			   thread_index))
 			{
 			  next0 = NAT64_IN2OUT_NEXT_DROP;
 			  b0->error =
@@ -1441,7 +1432,7 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 			}
 
 		      bibe0 =
-			nat64_db_bib_entry_create (&nm->db,
+			nat64_db_bib_entry_create (db,
 						   &ip60->src_address,
 						   &out_addr0, udp0->src_port,
 						   clib_host_to_net_u16
@@ -1458,7 +1449,7 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 		  nat64_extract_ip4 (&ip60->dst_address, &daddr0.ip4,
 				     fib_index0);
 		  ste0 =
-		    nat64_db_st_entry_create (&nm->db, bibe0,
+		    nat64_db_st_entry_create (db, bibe0,
 					      &ip60->dst_address, &daddr0.ip4,
 					      udp0->dst_port);
 		  if (!ste0)
@@ -1469,8 +1460,7 @@ nat64_in2out_reass_node_fn (vlib_main_t * vm,
 		      goto trace0;
 		    }
 		}
-	      reass0->sess_index =
-		nat64_db_st_entry_get_index (&nm->db, ste0);
+	      reass0->sess_index = nat64_db_st_entry_get_index (db, ste0);
 
 	      nat_ip6_reass_get_frags (reass0, &fragments_to_loopback);
 	    }
@@ -1590,6 +1580,213 @@ VLIB_REGISTER_NODE (nat64_in2out_reass_node) = {
 VLIB_NODE_FUNCTION_MULTIARCH (nat64_in2out_reass_node,
 			      nat64_in2out_reass_node_fn);
 
+typedef struct
+{
+  u32 next_worker_index;
+  u8 do_handoff;
+} nat64_in2out_handoff_trace_t;
+
+static u8 *
+format_nat64_in2out_handoff_trace (u8 * s, va_list * args)
+{
+  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
+  CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
+  nat64_in2out_handoff_trace_t *t =
+    va_arg (*args, nat64_in2out_handoff_trace_t *);
+  char *m;
+
+  m = t->do_handoff ? "next worker" : "same worker";
+  s = format (s, "NAT64-IN2OUT-HANDOFF: %s %d", m, t->next_worker_index);
+
+  return s;
+}
+
+static inline uword
+nat64_in2out_handoff_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
+			      vlib_frame_t * frame)
+{
+  nat64_main_t *nm = &nat64_main;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
+  u32 n_left_from, *from, *to_next = 0, *to_next_drop = 0;
+  static __thread vlib_frame_queue_elt_t **handoff_queue_elt_by_worker_index;
+  static __thread vlib_frame_queue_t **congested_handoff_queue_by_worker_index
+    = 0;
+  vlib_frame_queue_elt_t *hf = 0;
+  vlib_frame_queue_t *fq;
+  vlib_frame_t *f = 0, *d = 0;
+  int i;
+  u32 n_left_to_next_worker = 0, *to_next_worker = 0;
+  u32 next_worker_index = 0;
+  u32 current_worker_index = ~0;
+  u32 thread_index = vlib_get_thread_index ();
+  u32 fq_index;
+  u32 to_node_index;
+
+  fq_index = nm->fq_in2out_index;
+  to_node_index = nat64_in2out_node.index;
+
+  if (PREDICT_FALSE (handoff_queue_elt_by_worker_index == 0))
+    {
+      vec_validate (handoff_queue_elt_by_worker_index, tm->n_vlib_mains - 1);
+
+      vec_validate_init_empty (congested_handoff_queue_by_worker_index,
+			       tm->n_vlib_mains - 1,
+			       (vlib_frame_queue_t *) (~0));
+    }
+
+  from = vlib_frame_vector_args (frame);
+  n_left_from = frame->n_vectors;
+
+  while (n_left_from > 0)
+    {
+      u32 bi0;
+      vlib_buffer_t *b0;
+      ip6_header_t *ip0;
+      u8 do_handoff;
+
+      bi0 = from[0];
+      from += 1;
+      n_left_from -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      ip0 = vlib_buffer_get_current (b0);
+
+      next_worker_index = nat64_get_worker_in2out (&ip0->src_address);
+
+      if (PREDICT_FALSE (next_worker_index != thread_index))
+	{
+	  do_handoff = 1;
+
+	  if (next_worker_index != current_worker_index)
+	    {
+	      fq =
+		is_vlib_frame_queue_congested (fq_index, next_worker_index,
+					       30,
+					       congested_handoff_queue_by_worker_index);
+
+	      if (fq)
+		{
+		  /* if this is 1st frame */
+		  if (!d)
+		    {
+		      d = vlib_get_frame_to_node (vm, nm->error_node_index);
+		      to_next_drop = vlib_frame_vector_args (d);
+		    }
+
+		  to_next_drop[0] = bi0;
+		  to_next_drop += 1;
+		  d->n_vectors++;
+		  goto trace0;
+		}
+
+	      if (hf)
+		hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+	      hf =
+		vlib_get_worker_handoff_queue_elt (fq_index,
+						   next_worker_index,
+						   handoff_queue_elt_by_worker_index);
+	      n_left_to_next_worker = VLIB_FRAME_SIZE - hf->n_vectors;
+	      to_next_worker = &hf->buffer_index[hf->n_vectors];
+	      current_worker_index = next_worker_index;
+	    }
+
+	  /* enqueue to correct worker thread */
+	  to_next_worker[0] = bi0;
+	  to_next_worker++;
+	  n_left_to_next_worker--;
+
+	  if (n_left_to_next_worker == 0)
+	    {
+	      hf->n_vectors = VLIB_FRAME_SIZE;
+	      vlib_put_frame_queue_elt (hf);
+	      current_worker_index = ~0;
+	      handoff_queue_elt_by_worker_index[next_worker_index] = 0;
+	      hf = 0;
+	    }
+	}
+      else
+	{
+	  do_handoff = 0;
+	  /* if this is 1st frame */
+	  if (!f)
+	    {
+	      f = vlib_get_frame_to_node (vm, to_node_index);
+	      to_next = vlib_frame_vector_args (f);
+	    }
+
+	  to_next[0] = bi0;
+	  to_next += 1;
+	  f->n_vectors++;
+	}
+
+    trace0:
+      if (PREDICT_FALSE
+	  ((node->flags & VLIB_NODE_FLAG_TRACE)
+	   && (b0->flags & VLIB_BUFFER_IS_TRACED)))
+	{
+	  nat64_in2out_handoff_trace_t *t =
+	    vlib_add_trace (vm, node, b0, sizeof (*t));
+	  t->next_worker_index = next_worker_index;
+	  t->do_handoff = do_handoff;
+	}
+    }
+
+  if (f)
+    vlib_put_frame_to_node (vm, to_node_index, f);
+
+  if (d)
+    vlib_put_frame_to_node (vm, nm->error_node_index, d);
+
+  if (hf)
+    hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+  /* Ship frames to the worker nodes */
+  for (i = 0; i < vec_len (handoff_queue_elt_by_worker_index); i++)
+    {
+      if (handoff_queue_elt_by_worker_index[i])
+	{
+	  hf = handoff_queue_elt_by_worker_index[i];
+	  /*
+	   * It works better to let the handoff node
+	   * rate-adapt, always ship the handoff queue element.
+	   */
+	  if (1 || hf->n_vectors == hf->last_n_vectors)
+	    {
+	      vlib_put_frame_queue_elt (hf);
+	      handoff_queue_elt_by_worker_index[i] = 0;
+	    }
+	  else
+	    hf->last_n_vectors = hf->n_vectors;
+	}
+      congested_handoff_queue_by_worker_index[i] =
+	(vlib_frame_queue_t *) (~0);
+    }
+  hf = 0;
+  current_worker_index = ~0;
+  return frame->n_vectors;
+}
+
+/* *INDENT-OFF* */
+VLIB_REGISTER_NODE (nat64_in2out_handoff_node) = {
+  .function = nat64_in2out_handoff_node_fn,
+  .name = "nat64-in2out-handoff",
+  .vector_size = sizeof (u32),
+  .format_trace = format_nat64_in2out_handoff_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+
+  .n_next_nodes = 1,
+
+  .next_nodes = {
+    [0] = "error-drop",
+  },
+};
+/* *INDENT-ON* */
+
+VLIB_NODE_FUNCTION_MULTIARCH (nat64_in2out_handoff_node,
+			      nat64_in2out_handoff_node_fn);
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/plugins/nat/nat64_out2in.c b/src/plugins/nat/nat64_out2in.c
index eb5ecb4588d..1ba73a4f84a 100644
--- a/src/plugins/nat/nat64_out2in.c
+++ b/src/plugins/nat/nat64_out2in.c
@@ -21,6 +21,7 @@
 #include <nat/nat_reass.h>
 #include <vnet/ip/ip4_to_ip6.h>
 #include <vnet/fib/ip4_fib.h>
+#include <vnet/udp/udp.h>
 
 typedef struct
 {
@@ -67,6 +68,7 @@ format_nat64_out2in_reass_trace (u8 * s, va_list * args)
 
 vlib_node_registration_t nat64_out2in_node;
 vlib_node_registration_t nat64_out2in_reass_node;
+vlib_node_registration_t nat64_out2in_handoff_node;
 
 #define foreach_nat64_out2in_error                       \
 _(UNSUPPORTED_PROTOCOL, "Unsupported protocol")          \
@@ -94,7 +96,8 @@ static char *nat64_out2in_error_strings[] = {
 
 typedef enum
 {
-  NAT64_OUT2IN_NEXT_LOOKUP,
+  NAT64_OUT2IN_NEXT_IP6_LOOKUP,
+  NAT64_OUT2IN_NEXT_IP4_LOOKUP,
   NAT64_OUT2IN_NEXT_DROP,
   NAT64_OUT2IN_NEXT_REASS,
   NAT64_OUT2IN_N_NEXT,
@@ -104,6 +107,7 @@ typedef struct nat64_out2in_set_ctx_t_
 {
   vlib_buffer_t *b;
   vlib_main_t *vm;
+  u32 thread_index;
 } nat64_out2in_set_ctx_t;
 
 static int
@@ -124,6 +128,7 @@ nat64_out2in_tcp_udp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
   u32 sw_if_index, fib_index;
   u16 *checksum;
   ip_csum_t csum;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index = ip4_fib_table_get_index_for_sw_if_index (sw_if_index);
@@ -134,26 +139,24 @@ nat64_out2in_tcp_udp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
   daddr.ip4.as_u32 = ip4->dst_address.as_u32;
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &daddr, &saddr, dport, sport, proto,
+    nat64_db_st_entry_find (db, &daddr, &saddr, dport, sport, proto,
 			    fib_index, 0);
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &daddr, dport, proto, fib_index, 0);
+      bibe = nat64_db_bib_entry_find (db, &daddr, dport, proto, fib_index, 0);
 
       if (!bibe)
 	return -1;
 
       nat64_compose_ip6 (&ip6_saddr, &ip4->src_address, bibe->fib_index);
       ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6_saddr, &saddr.ip4,
-				  sport);
+	nat64_db_st_entry_create (db, bibe, &ip6_saddr, &saddr.ip4, sport);
     }
 
   nat64_session_reset_timeout (ste, ctx->vm);
@@ -189,6 +192,7 @@ nat64_out2in_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6, void *arg)
   ip6_address_t ip6_saddr;
   u32 sw_if_index, fib_index;
   icmp46_header_t *icmp = ip4_next_header (ip4);
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index = ip4_fib_table_get_index_for_sw_if_index (sw_if_index);
@@ -202,13 +206,13 @@ nat64_out2in_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6, void *arg)
     {
       u16 out_id = ((u16 *) (icmp))[2];
       ste =
-	nat64_db_st_entry_find (&nm->db, &daddr, &saddr, out_id, 0,
+	nat64_db_st_entry_find (db, &daddr, &saddr, out_id, 0,
 				IP_PROTOCOL_ICMP, fib_index, 0);
 
       if (ste)
 	{
 	  bibe =
-	    nat64_db_bib_entry_by_index (&nm->db, IP_PROTOCOL_ICMP,
+	    nat64_db_bib_entry_by_index (db, IP_PROTOCOL_ICMP,
 					 ste->bibe_index);
 	  if (!bibe)
 	    return -1;
@@ -216,15 +220,14 @@ nat64_out2in_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6, void *arg)
       else
 	{
 	  bibe =
-	    nat64_db_bib_entry_find (&nm->db, &daddr, out_id,
+	    nat64_db_bib_entry_find (db, &daddr, out_id,
 				     IP_PROTOCOL_ICMP, fib_index, 0);
 	  if (!bibe)
 	    return -1;
 
 	  nat64_compose_ip6 (&ip6_saddr, &ip4->src_address, bibe->fib_index);
 	  ste =
-	    nat64_db_st_entry_create (&nm->db, bibe, &ip6_saddr, &saddr.ip4,
-				      0);
+	    nat64_db_st_entry_create (db, bibe, &ip6_saddr, &saddr.ip4, 0);
 	}
 
       nat64_session_reset_timeout (ste, ctx->vm);
@@ -262,6 +265,7 @@ nat64_out2in_inner_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
   ip46_address_t saddr, daddr;
   u32 sw_if_index, fib_index;
   u8 proto = ip4->protocol;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index =
@@ -284,12 +288,12 @@ nat64_out2in_inner_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
 	return -1;
 
       ste =
-	nat64_db_st_entry_find (&nm->db, &saddr, &daddr, out_id, 0, proto,
+	nat64_db_st_entry_find (db, &saddr, &daddr, out_id, 0, proto,
 				fib_index, 0);
       if (!ste)
 	return -1;
 
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
 
@@ -311,12 +315,12 @@ nat64_out2in_inner_icmp_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
       ip_csum_t csum;
 
       ste =
-	nat64_db_st_entry_find (&nm->db, &saddr, &daddr, sport, dport, proto,
+	nat64_db_st_entry_find (db, &saddr, &daddr, sport, dport, proto,
 				fib_index, 0);
       if (!ste)
 	return -1;
 
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
 
@@ -354,6 +358,7 @@ nat64_out2in_unk_proto_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
   ip6_address_t ip6_saddr;
   u32 sw_if_index, fib_index;
   u8 proto = ip4->protocol;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
   sw_if_index = vnet_buffer (ctx->b)->sw_if_index[VLIB_RX];
   fib_index = ip4_fib_table_get_index_for_sw_if_index (sw_if_index);
@@ -364,25 +369,22 @@ nat64_out2in_unk_proto_set_cb (ip4_header_t * ip4, ip6_header_t * ip6,
   daddr.ip4.as_u32 = ip4->dst_address.as_u32;
 
   ste =
-    nat64_db_st_entry_find (&nm->db, &daddr, &saddr, 0, 0, proto, fib_index,
-			    0);
+    nat64_db_st_entry_find (db, &daddr, &saddr, 0, 0, proto, fib_index, 0);
   if (ste)
     {
-      bibe = nat64_db_bib_entry_by_index (&nm->db, proto, ste->bibe_index);
+      bibe = nat64_db_bib_entry_by_index (db, proto, ste->bibe_index);
       if (!bibe)
 	return -1;
     }
   else
     {
-      bibe =
-	nat64_db_bib_entry_find (&nm->db, &daddr, 0, proto, fib_index, 0);
+      bibe = nat64_db_bib_entry_find (db, &daddr, 0, proto, fib_index, 0);
 
       if (!bibe)
 	return -1;
 
       nat64_compose_ip6 (&ip6_saddr, &ip4->src_address, bibe->fib_index);
-      ste =
-	nat64_db_st_entry_create (&nm->db, bibe, &ip6_saddr, &saddr.ip4, 0);
+      ste = nat64_db_st_entry_create (db, bibe, &ip6_saddr, &saddr.ip4, 0);
     }
 
   nat64_session_reset_timeout (ste, ctx->vm);
@@ -405,6 +407,7 @@ nat64_out2in_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 n_left_from, *from, *to_next;
   nat64_out2in_next_t next_index;
   u32 pkts_processed = 0;
+  u32 thread_index = vlib_get_thread_index ();
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -423,6 +426,7 @@ nat64_out2in_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  ip4_header_t *ip40;
 	  u32 proto0;
 	  nat64_out2in_set_ctx_t ctx0;
+	  udp_header_t *udp0;
 
 	  /* speculatively enqueue b0 to the current next frame */
 	  bi0 = from[0];
@@ -437,8 +441,9 @@ nat64_out2in_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  ctx0.b = b0;
 	  ctx0.vm = vm;
+	  ctx0.thread_index = thread_index;
 
-	  next0 = NAT64_OUT2IN_NEXT_LOOKUP;
+	  next0 = NAT64_OUT2IN_NEXT_IP6_LOOKUP;
 
 	  proto0 = ip_proto_to_snat_proto (ip40->protocol);
 
@@ -473,6 +478,18 @@ nat64_out2in_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      if (ip4_to_ip6_tcp_udp (b0, nat64_out2in_tcp_udp_set_cb, &ctx0))
 		{
+		  udp0 = ip4_next_header (ip40);
+		  /*
+		   * Send DHCP packets to the ipv4 stack, or we won't
+		   * be able to use dhcp client on the outside interface
+		   */
+		  if ((proto0 == SNAT_PROTOCOL_UDP)
+		      && (udp0->dst_port ==
+			  clib_host_to_net_u16 (UDP_DST_PORT_dhcp_to_client)))
+		    {
+		      next0 = NAT64_OUT2IN_NEXT_IP4_LOOKUP;
+		      goto trace0;
+		    }
 		  next0 = NAT64_OUT2IN_NEXT_DROP;
 		  b0->error = node->errors[NAT64_OUT2IN_ERROR_NO_TRANSLATION];
 		  goto trace0;
@@ -516,7 +533,8 @@ VLIB_REGISTER_NODE (nat64_out2in_node) = {
   /* edit / add dispositions here */
   .next_nodes = {
     [NAT64_OUT2IN_NEXT_DROP] = "error-drop",
-    [NAT64_OUT2IN_NEXT_LOOKUP] = "ip6-lookup",
+    [NAT64_OUT2IN_NEXT_IP6_LOOKUP] = "ip6-lookup",
+    [NAT64_OUT2IN_NEXT_IP4_LOOKUP] = "ip4-lookup",
     [NAT64_OUT2IN_NEXT_REASS] = "nat64-out2in-reass",
   },
 };
@@ -529,6 +547,7 @@ typedef struct nat64_out2in_frag_set_ctx_t_
   vlib_main_t *vm;
   vlib_buffer_t *b;
   u32 sess_index;
+  u32 thread_index;
   u8 proto;
   u8 first_frag;
 } nat64_out2in_frag_set_ctx_t;
@@ -543,12 +562,13 @@ nat64_out2in_frag_set_cb (ip4_header_t * ip4, ip6_header_t * ip6, void *arg)
   udp_header_t *udp = ip4_next_header (ip4);
   ip_csum_t csum;
   u16 *checksum;
+  nat64_db_t *db = &nm->db[ctx->thread_index];
 
-  ste = nat64_db_st_entry_by_index (&nm->db, ctx->proto, ctx->sess_index);
+  ste = nat64_db_st_entry_by_index (db, ctx->proto, ctx->sess_index);
   if (!ste)
     return -1;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, ctx->proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (db, ctx->proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -629,6 +649,7 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 *fragments_to_drop = 0;
   u32 *fragments_to_loopback = 0;
   nat64_main_t *nm = &nat64_main;
+  u32 thread_index = vlib_get_thread_index ();
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -655,6 +676,7 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  nat64_db_bib_entry_t *bibe0;
 	  ip6_address_t ip6_saddr0;
 	  nat64_out2in_frag_set_ctx_t ctx0;
+	  nat64_db_t *db = &nm->db[thread_index];
 
 	  /* speculatively enqueue b0 to the current next frame */
 	  bi0 = from[0];
@@ -665,13 +687,15 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  n_left_to_next -= 1;
 
 	  b0 = vlib_get_buffer (vm, bi0);
-	  next0 = NAT64_OUT2IN_NEXT_LOOKUP;
+	  next0 = NAT64_OUT2IN_NEXT_IP6_LOOKUP;
 
 	  sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
 	  fib_index0 =
 	    fib_table_get_index_for_sw_if_index (FIB_PROTOCOL_IP4,
 						 sw_if_index0);
 
+	  ctx0.thread_index = thread_index;
+
 	  if (PREDICT_FALSE (nat_reass_is_drop_frag (1)))
 	    {
 	      next0 = NAT64_OUT2IN_NEXT_DROP;
@@ -714,13 +738,13 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      daddr0.ip4.as_u32 = ip40->dst_address.as_u32;
 
 	      ste0 =
-		nat64_db_st_entry_find (&nm->db, &daddr0, &saddr0,
+		nat64_db_st_entry_find (db, &daddr0, &saddr0,
 					udp0->dst_port, udp0->src_port,
 					ip40->protocol, fib_index0, 0);
 	      if (!ste0)
 		{
 		  bibe0 =
-		    nat64_db_bib_entry_find (&nm->db, &daddr0, udp0->dst_port,
+		    nat64_db_bib_entry_find (db, &daddr0, udp0->dst_port,
 					     ip40->protocol, fib_index0, 0);
 		  if (!bibe0)
 		    {
@@ -733,7 +757,7 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  nat64_compose_ip6 (&ip6_saddr0, &ip40->src_address,
 				     bibe0->fib_index);
 		  ste0 =
-		    nat64_db_st_entry_create (&nm->db, bibe0, &ip6_saddr0,
+		    nat64_db_st_entry_create (db, bibe0, &ip6_saddr0,
 					      &saddr0.ip4, udp0->src_port);
 
 		  if (!ste0)
@@ -744,8 +768,8 @@ nat64_out2in_reass_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		      goto trace0;
 		    }
 		}
-	      reass0->sess_index =
-		nat64_db_st_entry_get_index (&nm->db, ste0);
+	      reass0->sess_index = nat64_db_st_entry_get_index (db, ste0);
+	      reass0->thread_index = thread_index;
 
 	      nat_ip4_reass_get_frags (reass0, &fragments_to_loopback);
 	    }
@@ -857,7 +881,8 @@ VLIB_REGISTER_NODE (nat64_out2in_reass_node) = {
   /* edit / add dispositions here */
   .next_nodes = {
     [NAT64_OUT2IN_NEXT_DROP] = "error-drop",
-    [NAT64_OUT2IN_NEXT_LOOKUP] = "ip6-lookup",
+    [NAT64_OUT2IN_NEXT_IP6_LOOKUP] = "ip6-lookup",
+    [NAT64_OUT2IN_NEXT_IP4_LOOKUP] = "ip4-lookup",
     [NAT64_OUT2IN_NEXT_REASS] = "nat64-out2in-reass",
   },
 };
@@ -866,6 +891,212 @@ VLIB_REGISTER_NODE (nat64_out2in_reass_node) = {
 VLIB_NODE_FUNCTION_MULTIARCH (nat64_out2in_reass_node,
 			      nat64_out2in_reass_node_fn);
 
+typedef struct
+{
+  u32 next_worker_index;
+  u8 do_handoff;
+} nat64_out2in_handoff_trace_t;
+
+static u8 *
+format_nat64_out2in_handoff_trace (u8 * s, va_list * args)
+{
+  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
+  CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
+  nat64_out2in_handoff_trace_t *t =
+    va_arg (*args, nat64_out2in_handoff_trace_t *);
+  char *m;
+
+  m = t->do_handoff ? "next worker" : "same worker";
+  s = format (s, "NAT64-OUT2IN-HANDOFF: %s %d", m, t->next_worker_index);
+
+  return s;
+}
+
+static inline uword
+nat64_out2in_handoff_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
+			      vlib_frame_t * frame)
+{
+  nat64_main_t *nm = &nat64_main;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
+  u32 n_left_from, *from, *to_next = 0, *to_next_drop = 0;
+  static __thread vlib_frame_queue_elt_t **handoff_queue_elt_by_worker_index;
+  static __thread vlib_frame_queue_t **congested_handoff_queue_by_worker_index
+    = 0;
+  vlib_frame_queue_elt_t *hf = 0;
+  vlib_frame_queue_t *fq;
+  vlib_frame_t *f = 0, *d = 0;
+  int i;
+  u32 n_left_to_next_worker = 0, *to_next_worker = 0;
+  u32 next_worker_index = 0;
+  u32 current_worker_index = ~0;
+  u32 thread_index = vlib_get_thread_index ();
+  u32 fq_index;
+  u32 to_node_index;
+
+  fq_index = nm->fq_out2in_index;
+  to_node_index = nat64_out2in_node.index;
+
+  if (PREDICT_FALSE (handoff_queue_elt_by_worker_index == 0))
+    {
+      vec_validate (handoff_queue_elt_by_worker_index, tm->n_vlib_mains - 1);
+
+      vec_validate_init_empty (congested_handoff_queue_by_worker_index,
+			       tm->n_vlib_mains - 1,
+			       (vlib_frame_queue_t *) (~0));
+    }
+
+  from = vlib_frame_vector_args (frame);
+  n_left_from = frame->n_vectors;
+
+  while (n_left_from > 0)
+    {
+      u32 bi0;
+      vlib_buffer_t *b0;
+      ip4_header_t *ip0;
+      u8 do_handoff;
+
+      bi0 = from[0];
+      from += 1;
+      n_left_from -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      ip0 = vlib_buffer_get_current (b0);
+
+      next_worker_index = nat64_get_worker_out2in (ip0);
+
+      if (PREDICT_FALSE (next_worker_index != thread_index))
+	{
+	  do_handoff = 1;
+
+	  if (next_worker_index != current_worker_index)
+	    {
+	      fq =
+		is_vlib_frame_queue_congested (fq_index, next_worker_index,
+					       30,
+					       congested_handoff_queue_by_worker_index);
+
+	      if (fq)
+		{
+		  /* if this is 1st frame */
+		  if (!d)
+		    {
+		      d = vlib_get_frame_to_node (vm, nm->error_node_index);
+		      to_next_drop = vlib_frame_vector_args (d);
+		    }
+
+		  to_next_drop[0] = bi0;
+		  to_next_drop += 1;
+		  d->n_vectors++;
+		  goto trace0;
+		}
+
+	      if (hf)
+		hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+	      hf =
+		vlib_get_worker_handoff_queue_elt (fq_index,
+						   next_worker_index,
+						   handoff_queue_elt_by_worker_index);
+	      n_left_to_next_worker = VLIB_FRAME_SIZE - hf->n_vectors;
+	      to_next_worker = &hf->buffer_index[hf->n_vectors];
+	      current_worker_index = next_worker_index;
+	    }
+
+	  /* enqueue to correct worker thread */
+	  to_next_worker[0] = bi0;
+	  to_next_worker++;
+	  n_left_to_next_worker--;
+
+	  if (n_left_to_next_worker == 0)
+	    {
+	      hf->n_vectors = VLIB_FRAME_SIZE;
+	      vlib_put_frame_queue_elt (hf);
+	      current_worker_index = ~0;
+	      handoff_queue_elt_by_worker_index[next_worker_index] = 0;
+	      hf = 0;
+	    }
+	}
+      else
+	{
+	  do_handoff = 0;
+	  /* if this is 1st frame */
+	  if (!f)
+	    {
+	      f = vlib_get_frame_to_node (vm, to_node_index);
+	      to_next = vlib_frame_vector_args (f);
+	    }
+
+	  to_next[0] = bi0;
+	  to_next += 1;
+	  f->n_vectors++;
+	}
+
+    trace0:
+      if (PREDICT_FALSE
+	  ((node->flags & VLIB_NODE_FLAG_TRACE)
+	   && (b0->flags & VLIB_BUFFER_IS_TRACED)))
+	{
+	  nat64_out2in_handoff_trace_t *t =
+	    vlib_add_trace (vm, node, b0, sizeof (*t));
+	  t->next_worker_index = next_worker_index;
+	  t->do_handoff = do_handoff;
+	}
+    }
+
+  if (f)
+    vlib_put_frame_to_node (vm, to_node_index, f);
+
+  if (d)
+    vlib_put_frame_to_node (vm, nm->error_node_index, d);
+
+  if (hf)
+    hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+  /* Ship frames to the worker nodes */
+  for (i = 0; i < vec_len (handoff_queue_elt_by_worker_index); i++)
+    {
+      if (handoff_queue_elt_by_worker_index[i])
+	{
+	  hf = handoff_queue_elt_by_worker_index[i];
+	  /*
+	   * It works better to let the handoff node
+	   * rate-adapt, always ship the handoff queue element.
+	   */
+	  if (1 || hf->n_vectors == hf->last_n_vectors)
+	    {
+	      vlib_put_frame_queue_elt (hf);
+	      handoff_queue_elt_by_worker_index[i] = 0;
+	    }
+	  else
+	    hf->last_n_vectors = hf->n_vectors;
+	}
+      congested_handoff_queue_by_worker_index[i] =
+	(vlib_frame_queue_t *) (~0);
+    }
+  hf = 0;
+  current_worker_index = ~0;
+  return frame->n_vectors;
+}
+
+/* *INDENT-OFF* */
+VLIB_REGISTER_NODE (nat64_out2in_handoff_node) = {
+  .function = nat64_out2in_handoff_node_fn,
+  .name = "nat64-out2in-handoff",
+  .vector_size = sizeof (u32),
+  .format_trace = format_nat64_out2in_handoff_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+
+  .n_next_nodes = 1,
+
+  .next_nodes = {
+    [0] = "error-drop",
+  },
+};
+/* *INDENT-ON* */
+
+VLIB_NODE_FUNCTION_MULTIARCH (nat64_out2in_handoff_node,
+			      nat64_out2in_handoff_node_fn);
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/plugins/nat/nat_api.c b/src/plugins/nat/nat_api.c
index 3c58097e332..f1d9ec83af3 100644
--- a/src/plugins/nat/nat_api.c
+++ b/src/plugins/nat/nat_api.c
@@ -1776,7 +1776,6 @@ static void
 {
   vl_api_nat64_add_del_pool_addr_range_reply_t *rmp;
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   int rv = 0;
   ip4_address_t this_addr;
   u32 start_host_order, end_host_order;
@@ -1784,12 +1783,6 @@ static void
   int i, count;
   u32 *tmp;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   tmp = (u32 *) mp->start_addr;
   start_host_order = clib_host_to_net_u32 (tmp[0]);
   tmp = (u32 *) mp->end_addr;
@@ -1831,6 +1824,7 @@ typedef struct nat64_api_walk_ctx_t_
 {
   unix_shared_memory_queue_t *q;
   u32 context;
+  nat64_db_t *db;
 } nat64_api_walk_ctx_t;
 
 static int
@@ -1864,10 +1858,6 @@ static void
 vl_api_nat64_pool_addr_dump_t_handler (vl_api_nat64_pool_addr_dump_t * mp)
 {
   unix_shared_memory_queue_t *q;
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return;
 
   q = vl_api_client_index_to_input_queue (mp->client_index);
   if (q == 0)
@@ -1897,16 +1887,9 @@ vl_api_nat64_add_del_interface_t_handler (vl_api_nat64_add_del_interface_t *
 					  mp)
 {
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   vl_api_nat64_add_del_interface_reply_t *rmp;
   int rv = 0;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   VALIDATE_SW_IF_INDEX (mp);
 
   rv =
@@ -1915,7 +1898,6 @@ vl_api_nat64_add_del_interface_t_handler (vl_api_nat64_add_del_interface_t *
 
   BAD_SW_IF_INDEX_LABEL;
 
-send_reply:
   REPLY_MACRO (VL_API_NAT64_ADD_DEL_INTERFACE_REPLY);
 }
 
@@ -1958,10 +1940,6 @@ static void
 vl_api_nat64_interface_dump_t_handler (vl_api_nat64_interface_dump_t * mp)
 {
   unix_shared_memory_queue_t *q;
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return;
 
   q = vl_api_client_index_to_input_queue (mp->client_index);
   if (q == 0)
@@ -1991,18 +1969,11 @@ static void
   (vl_api_nat64_add_del_static_bib_t * mp)
 {
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   vl_api_nat64_add_del_static_bib_reply_t *rmp;
   ip6_address_t in_addr;
   ip4_address_t out_addr;
   int rv = 0;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   memcpy (&in_addr.as_u8, mp->i_addr, 16);
   memcpy (&out_addr.as_u8, mp->o_addr, 4);
 
@@ -2014,7 +1985,6 @@ static void
 				    clib_net_to_host_u32 (mp->vrf_id),
 				    mp->is_add);
 
-send_reply:
   REPLY_MACRO (VL_API_NAT64_ADD_DEL_STATIC_BIB_REPLY);
 }
 
@@ -2069,9 +2039,7 @@ vl_api_nat64_bib_dump_t_handler (vl_api_nat64_bib_dump_t * mp)
 {
   unix_shared_memory_queue_t *q;
   nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return;
+  nat64_db_t *db;
 
   q = vl_api_client_index_to_input_queue (mp->client_index);
   if (q == 0)
@@ -2082,7 +2050,10 @@ vl_api_nat64_bib_dump_t_handler (vl_api_nat64_bib_dump_t * mp)
     .context = mp->context,
   };
 
-  nat64_db_bib_walk (&nm->db, mp->proto, nat64_api_bib_walk, &ctx);
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    nat64_db_bib_walk (db, mp->proto, nat64_api_bib_walk, &ctx);
+  /* *INDENT-ON* */
 }
 
 static void *
@@ -2099,16 +2070,9 @@ static void
 vl_api_nat64_set_timeouts_t_handler (vl_api_nat64_set_timeouts_t * mp)
 {
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   vl_api_nat64_set_timeouts_reply_t *rmp;
   int rv = 0;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   rv = nat64_set_icmp_timeout (ntohl (mp->icmp));
   if (rv)
     goto send_reply;
@@ -2142,13 +2106,9 @@ static void
 vl_api_nat64_get_timeouts_t_handler (vl_api_nat64_get_timeouts_t * mp)
 {
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   vl_api_nat64_get_timeouts_reply_t *rmp;
   int rv = 0;
 
-  if (nm->is_disabled)
-    return;
-
   /* *INDENT-OFF* */
   REPLY_MACRO2 (VL_API_NAT64_GET_TIMEOUTS_REPLY,
   ({
@@ -2177,11 +2137,10 @@ nat64_api_st_walk (nat64_db_st_entry_t * ste, void *arg)
   vl_api_nat64_st_details_t *rmp;
   snat_main_t *sm = &snat_main;
   nat64_api_walk_ctx_t *ctx = arg;
-  nat64_main_t *nm = &nat64_main;
   nat64_db_bib_entry_t *bibe;
   fib_table_t *fib;
 
-  bibe = nat64_db_bib_entry_by_index (&nm->db, ste->proto, ste->bibe_index);
+  bibe = nat64_db_bib_entry_by_index (ctx->db, ste->proto, ste->bibe_index);
   if (!bibe)
     return -1;
 
@@ -2213,9 +2172,7 @@ vl_api_nat64_st_dump_t_handler (vl_api_nat64_st_dump_t * mp)
 {
   unix_shared_memory_queue_t *q;
   nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return;
+  nat64_db_t *db;
 
   q = vl_api_client_index_to_input_queue (mp->client_index);
   if (q == 0)
@@ -2226,7 +2183,13 @@ vl_api_nat64_st_dump_t_handler (vl_api_nat64_st_dump_t * mp)
     .context = mp->context,
   };
 
-  nat64_db_st_walk (&nm->db, mp->proto, nat64_api_st_walk, &ctx);
+  /* *INDENT-OFF* */
+  vec_foreach (db, nm->db)
+    {
+      ctx.db = db;
+      nat64_db_st_walk (db, mp->proto, nat64_api_st_walk, &ctx);
+    }
+  /* *INDENT-ON* */
 }
 
 static void *
@@ -2244,22 +2207,14 @@ vl_api_nat64_add_del_prefix_t_handler (vl_api_nat64_add_del_prefix_t * mp)
 {
   vl_api_nat64_add_del_prefix_reply_t *rmp;
   snat_main_t *sm = &snat_main;
-  nat64_main_t *nm = &nat64_main;
   ip6_address_t prefix;
   int rv = 0;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   memcpy (&prefix.as_u8, mp->prefix, 16);
 
   rv =
     nat64_add_del_prefix (&prefix, mp->prefix_len,
 			  clib_net_to_host_u32 (mp->vrf_id), mp->is_add);
-send_reply:
   REPLY_MACRO (VL_API_NAT64_ADD_DEL_PREFIX_REPLY);
 }
 
@@ -2300,10 +2255,6 @@ static void
 vl_api_nat64_prefix_dump_t_handler (vl_api_nat64_prefix_dump_t * mp)
 {
   unix_shared_memory_queue_t *q;
-  nat64_main_t *nm = &nat64_main;
-
-  if (nm->is_disabled)
-    return;
 
   q = vl_api_client_index_to_input_queue (mp->client_index);
   if (q == 0)
@@ -2332,24 +2283,17 @@ static void
   vl_api_nat64_add_del_interface_addr_t_handler
   (vl_api_nat64_add_del_interface_addr_t * mp)
 {
-  nat64_main_t *nm = &nat64_main;
   snat_main_t *sm = &snat_main;
   vl_api_nat64_add_del_interface_addr_reply_t *rmp;
   u32 sw_if_index = ntohl (mp->sw_if_index);
   int rv = 0;
 
-  if (nm->is_disabled)
-    {
-      rv = VNET_API_ERROR_FEATURE_DISABLED;
-      goto send_reply;
-    }
-
   VALIDATE_SW_IF_INDEX (mp);
 
   rv = nat64_add_interface_address (sw_if_index, mp->is_add);
 
   BAD_SW_IF_INDEX_LABEL;
-send_reply:
+
   REPLY_MACRO (VL_API_NAT64_ADD_DEL_INTERFACE_ADDR_REPLY);
 }
 
diff --git a/src/plugins/nat/nat_reass.c b/src/plugins/nat/nat_reass.c
index 239bc70d836..a97d8f017e9 100644
--- a/src/plugins/nat/nat_reass.c
+++ b/src/plugins/nat/nat_reass.c
@@ -195,6 +195,27 @@ nat_ip4_reass_lookup (nat_reass_ip4_key_t * k, f64 now)
 }
 
 nat_reass_ip4_t *
+nat_ip4_reass_find (ip4_address_t src, ip4_address_t dst, u16 frag_id,
+		    u8 proto)
+{
+  nat_reass_main_t *srm = &nat_reass_main;
+  nat_reass_ip4_t *reass = 0;
+  nat_reass_ip4_key_t k;
+  f64 now = vlib_time_now (srm->vlib_main);
+
+  k.src.as_u32 = src.as_u32;
+  k.dst.as_u32 = dst.as_u32;
+  k.frag_id = frag_id;
+  k.proto = proto;
+
+  clib_spinlock_lock_if_init (&srm->ip4_reass_lock);
+  reass = nat_ip4_reass_lookup (&k, now);
+  clib_spinlock_unlock_if_init (&srm->ip4_reass_lock);
+
+  return reass;
+}
+
+nat_reass_ip4_t *
 nat_ip4_reass_find_or_create (ip4_address_t src, ip4_address_t dst,
 			      u16 frag_id, u8 proto, u8 reset_timeout,
 			      u32 ** bi_to_drop)
@@ -282,6 +303,7 @@ nat_ip4_reass_find_or_create (ip4_address_t src, ip4_address_t dst,
   reass->key.as_u64[1] = kv.key[1] = k.as_u64[1];
   kv.value = reass - srm->ip4_reass_pool;
   reass->sess_index = (u32) ~ 0;
+  reass->thread_index = (u32) ~ 0;
   reass->last_heard = now;
 
   if (clib_bihash_add_del_16_8 (&srm->ip4_reass_hash, &kv, 1))
diff --git a/src/plugins/nat/nat_reass.h b/src/plugins/nat/nat_reass.h
index ae14a9604aa..4a9137e3255 100644
--- a/src/plugins/nat/nat_reass.h
+++ b/src/plugins/nat/nat_reass.h
@@ -53,6 +53,7 @@ typedef CLIB_PACKED(struct
   nat_reass_ip4_key_t key;
   u32 lru_list_index;
   u32 sess_index;
+  u32 thread_index;
   f64 last_heard;
   u32 frags_per_reass_list_head_index;
   u8 frag_n;
@@ -185,6 +186,20 @@ u8 nat_reass_is_drop_frag (u8 is_ip6);
 clib_error_t *nat_reass_init (vlib_main_t * vm);
 
 /**
+ * @brief Find reassembly.
+ *
+ * @param src Source IPv4 address.
+ * @param dst Destination IPv4 address.
+ * @param frag_id Fragment ID.
+ * @param proto L4 protocol.
+ *
+ * @returns Reassembly data or 0 if not found.
+ */
+nat_reass_ip4_t *nat_ip4_reass_find (ip4_address_t src,
+				     ip4_address_t dst,
+				     u16 frag_id, u8 proto);
+
+/**
  * @brief Find or create reassembly.
  *
  * @param src Source IPv4 address.
@@ -201,6 +216,7 @@ nat_reass_ip4_t *nat_ip4_reass_find_or_create (ip4_address_t src,
 					       u16 frag_id, u8 proto,
 					       u8 reset_timeout,
 					       u32 ** bi_to_drop);
+
 /**
  * @brief Cache fragment.
  *
diff --git a/src/plugins/nat/out2in.c b/src/plugins/nat/out2in.c
index 489afadb4a2..15f6f556d73 100755
--- a/src/plugins/nat/out2in.c
+++ b/src/plugins/nat/out2in.c
@@ -1701,6 +1701,7 @@ nat44_out2in_reass_node_fn (vlib_main_t * vm,
                       goto trace0;
                     }
                   reass0->sess_index = s0 - per_thread_data->sessions;
+                  reass0->thread_index = thread_index;
                 }
               else
                 {
diff --git a/src/scripts/vnet/nat64 b/src/scripts/vnet/nat64
new file mode 100644
index 00000000000..f0d98ea8c69
--- /dev/null
+++ b/src/scripts/vnet/nat64
@@ -0,0 +1,42 @@
+create packet-generator interface pg0
+create packet-generator interface pg1
+
+packet-generator new {
+  name f1
+  limit 10
+  node ip6-input
+  size 64-64
+  no-recycle
+  worker 0
+  interface pg0
+  data {
+    UDP: 2000::3 -> 64:ff9b::ac10:102
+    UDP: 3000 -> 3001
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+
+packet-generator new {
+  name f2
+  limit 10
+  node ip6-input
+  size 64-64
+  no-recycle
+  interface pg0
+  worker 1
+  data {
+    UDP: 2000::3 -> 64:ff9b::ac10:102
+    UDP: 3005 -> 3006
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+nat64 add pool address 172.16.1.3
+set int ip address pg0 2000::1/64
+set int ip address pg1 172.16.1.1/24
+set int state pg0 up
+set int state pg1 up
+set ip6 neighbor pg0 2000::3 abcd.abcd.abcd static
+set ip arp static pg1 172.16.1.2 cdef.abcd.abcd
+set interface nat64 in pg0 out pg1
diff --git a/src/scripts/vnet/nat64_static b/src/scripts/vnet/nat64_static
new file mode 100644
index 00000000000..ff5e8dbb2c7
--- /dev/null
+++ b/src/scripts/vnet/nat64_static
@@ -0,0 +1,44 @@
+create packet-generator interface pg0
+create packet-generator interface pg1
+
+packet-generator new {
+  name f1
+  limit 10
+  node ip4-input
+  size 64-64
+  no-recycle
+  worker 0
+  interface pg1
+  data {
+    UDP: 172.16.1.2 -> 172.16.1.3
+    UDP: 3000 -> 37678
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+
+packet-generator new {
+  name f2
+  limit 10
+  node ip4-input
+  size 64-64
+  no-recycle
+  interface pg1
+  worker 1
+  data {
+    UDP: 172.16.1.2 -> 172.16.1.3
+    UDP: 3005 -> 38678
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+nat64 add pool address 172.16.1.3
+set int ip address pg0 2000::1/64
+set int ip address pg1 172.16.1.1/24
+set int state pg0 up
+set int state pg1 up
+set ip6 neighbor pg0 2000::3 abcd.abcd.abcd static
+set ip arp static pg1 172.16.1.2 cdef.abcd.abcd
+set interface nat64 in pg0 out pg1
+nat64 add static bib 2000::3 3001 172.16.1.3 37678 udp
+nat64 add static bib 2000::3 3006 172.16.1.3 38678 udp