Use thread local storage for thread index

This patch deprecates stack-based thread identification,
Also removes requirement that thread stacks are adjacent.

Finally, possibly annoying for some folks, it renames
all occurences of cpu_index and cpu_number with thread
index. Using word "cpu" is misleading here as thread can
be migrated ti different CPU, and also it is not related
to linux cpu index.

Change-Id: I68cdaf661e701d2336fc953dcb9978d10a70f7c1
Signed-off-by: Damjan Marion <damarion@cisco.com>

27 files changed, 216 insertions, 212 deletions

diff --git a/src/plugins/dpdk/buffer.c b/src/plugins/dpdk/buffer.c
index 2765c292dee..c80b3fa8875 100644
--- a/src/plugins/dpdk/buffer.c
+++ b/src/plugins/dpdk/buffer.c
@@ -132,7 +132,7 @@ dpdk_buffer_delete_free_list (vlib_main_t * vm, u32 free_list_index)
   u32 merge_index;
   int i;
 
-  ASSERT (os_get_cpu_number () == 0);
+  ASSERT (vlib_get_thread_index () == 0);
 
   f = vlib_buffer_get_free_list (vm, free_list_index);
 
diff --git a/src/plugins/dpdk/device/device.c b/src/plugins/dpdk/device/device.c
index 50b2668977f..91661246ebc 100644
--- a/src/plugins/dpdk/device/device.c
+++ b/src/plugins/dpdk/device/device.c
@@ -243,7 +243,7 @@ static_always_inline
   ASSERT (ring->tx_tail == 0);
 
   n_retry = 16;
-  queue_id = vm->cpu_index;
+  queue_id = vm->thread_index;
 
   do
     {
@@ -266,7 +266,7 @@ static_always_inline
 	{
 	  /* no wrap, transmit in one burst */
 	  dpdk_device_hqos_per_worker_thread_t *hqos =
-	    &xd->hqos_wt[vm->cpu_index];
+	    &xd->hqos_wt[vm->thread_index];
 
 	  ASSERT (hqos->swq != NULL);
 
@@ -332,7 +332,7 @@ dpdk_buffer_recycle (vlib_main_t * vm, vlib_node_runtime_t * node,
 		     vlib_buffer_t * b, u32 bi, struct rte_mbuf **mbp)
 {
   dpdk_main_t *dm = &dpdk_main;
-  u32 my_cpu = vm->cpu_index;
+  u32 my_cpu = vm->thread_index;
   struct rte_mbuf *mb_new;
 
   if (PREDICT_FALSE (b->flags & VLIB_BUFFER_RECYCLE) == 0)
@@ -376,7 +376,7 @@ dpdk_interface_tx (vlib_main_t * vm,
   tx_ring_hdr_t *ring;
   u32 n_on_ring;
 
-  my_cpu = vm->cpu_index;
+  my_cpu = vm->thread_index;
 
   queue_id = my_cpu;
 
diff --git a/src/plugins/dpdk/device/dpdk_priv.h b/src/plugins/dpdk/device/dpdk_priv.h
index dd40ff48510..52b4ca4b303 100644
--- a/src/plugins/dpdk/device/dpdk_priv.h
+++ b/src/plugins/dpdk/device/dpdk_priv.h
@@ -79,7 +79,7 @@ dpdk_update_counters (dpdk_device_t * xd, f64 now)
 {
   vlib_simple_counter_main_t *cm;
   vnet_main_t *vnm = vnet_get_main ();
-  u32 my_cpu = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   u64 rxerrors, last_rxerrors;
 
   /* only update counters for PMD interfaces */
@@ -96,7 +96,7 @@ dpdk_update_counters (dpdk_device_t * xd, f64 now)
       cm = vec_elt_at_index (vnm->interface_main.sw_if_counters,
 			     VNET_INTERFACE_COUNTER_RX_NO_BUF);
 
-      vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index,
+      vlib_increment_simple_counter (cm, thread_index, xd->vlib_sw_if_index,
 				     xd->stats.rx_nombuf -
 				     xd->last_stats.rx_nombuf);
     }
@@ -107,7 +107,7 @@ dpdk_update_counters (dpdk_device_t * xd, f64 now)
       cm = vec_elt_at_index (vnm->interface_main.sw_if_counters,
 			     VNET_INTERFACE_COUNTER_RX_MISS);
 
-      vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index,
+      vlib_increment_simple_counter (cm, thread_index, xd->vlib_sw_if_index,
 				     xd->stats.imissed -
 				     xd->last_stats.imissed);
     }
@@ -119,7 +119,7 @@ dpdk_update_counters (dpdk_device_t * xd, f64 now)
       cm = vec_elt_at_index (vnm->interface_main.sw_if_counters,
 			     VNET_INTERFACE_COUNTER_RX_ERROR);
 
-      vlib_increment_simple_counter (cm, my_cpu, xd->vlib_sw_if_index,
+      vlib_increment_simple_counter (cm, thread_index, xd->vlib_sw_if_index,
 				     rxerrors - last_rxerrors);
     }
 
diff --git a/src/plugins/dpdk/device/init.c b/src/plugins/dpdk/device/init.c
index 538db6cb94e..7eaf8da7992 100755
--- a/src/plugins/dpdk/device/init.c
+++ b/src/plugins/dpdk/device/init.c
@@ -324,7 +324,7 @@ dpdk_port_setup (dpdk_main_t * dm, dpdk_device_t * xd)
   int rv;
   int j;
 
-  ASSERT (os_get_cpu_number () == 0);
+  ASSERT (vlib_get_thread_index () == 0);
 
   if (xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP)
     {
diff --git a/src/plugins/dpdk/device/node.c b/src/plugins/dpdk/device/node.c
index e740fd18422..b10e0fadb6f 100644
--- a/src/plugins/dpdk/device/node.c
+++ b/src/plugins/dpdk/device/node.c
@@ -283,7 +283,7 @@ dpdk_buffer_init_from_template (void *d0, void *d1, void *d2, void *d3,
  */
 static_always_inline u32
 dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd,
-		   vlib_node_runtime_t * node, u32 cpu_index, u16 queue_id,
+		   vlib_node_runtime_t * node, u32 thread_index, u16 queue_id,
 		   int maybe_multiseg)
 {
   u32 n_buffers;
@@ -294,7 +294,7 @@ dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd,
   uword n_rx_bytes = 0;
   u32 n_trace, trace_cnt __attribute__ ((unused));
   vlib_buffer_free_list_t *fl;
-  vlib_buffer_t *bt = vec_elt_at_index (dm->buffer_templates, cpu_index);
+  vlib_buffer_t *bt = vec_elt_at_index (dm->buffer_templates, thread_index);
 
   if ((xd->flags & DPDK_DEVICE_FLAG_ADMIN_UP) == 0)
     return 0;
@@ -306,7 +306,7 @@ dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd,
       return 0;
     }
 
-  vec_reset_length (xd->d_trace_buffers[cpu_index]);
+  vec_reset_length (xd->d_trace_buffers[thread_index]);
   trace_cnt = n_trace = vlib_get_trace_count (vm, node);
 
   if (n_trace > 0)
@@ -318,7 +318,7 @@ dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd,
 	{
 	  struct rte_mbuf *mb = xd->rx_vectors[queue_id][mb_index++];
 	  vlib_buffer_t *b = vlib_buffer_from_rte_mbuf (mb);
-	  vec_add1 (xd->d_trace_buffers[cpu_index],
+	  vec_add1 (xd->d_trace_buffers[thread_index],
 		    vlib_get_buffer_index (vm, b));
 	}
     }
@@ -546,20 +546,22 @@ dpdk_device_input (dpdk_main_t * dm, dpdk_device_t * xd,
       vlib_put_next_frame (vm, node, next_index, n_left_to_next);
     }
 
-  if (PREDICT_FALSE (vec_len (xd->d_trace_buffers[cpu_index]) > 0))
+  if (PREDICT_FALSE (vec_len (xd->d_trace_buffers[thread_index]) > 0))
     {
-      dpdk_rx_trace (dm, node, xd, queue_id, xd->d_trace_buffers[cpu_index],
-		     vec_len (xd->d_trace_buffers[cpu_index]));
-      vlib_set_trace_count (vm, node, n_trace -
-			    vec_len (xd->d_trace_buffers[cpu_index]));
+      dpdk_rx_trace (dm, node, xd, queue_id,
+		     xd->d_trace_buffers[thread_index],
+		     vec_len (xd->d_trace_buffers[thread_index]));
+      vlib_set_trace_count (vm, node,
+			    n_trace -
+			    vec_len (xd->d_trace_buffers[thread_index]));
     }
 
   vlib_increment_combined_counter
     (vnet_get_main ()->interface_main.combined_sw_if_counters
      + VNET_INTERFACE_COUNTER_RX,
-     cpu_index, xd->vlib_sw_if_index, mb_index, n_rx_bytes);
+     thread_index, xd->vlib_sw_if_index, mb_index, n_rx_bytes);
 
-  vnet_device_increment_rx_packets (cpu_index, mb_index);
+  vnet_device_increment_rx_packets (thread_index, mb_index);
 
   return mb_index;
 }
@@ -630,19 +632,19 @@ dpdk_input (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_frame_t * f)
   dpdk_device_t *xd;
   uword n_rx_packets = 0;
   dpdk_device_and_queue_t *dq;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
 
   /*
    * Poll all devices on this cpu for input/interrupts.
    */
   /* *INDENT-OFF* */
-  vec_foreach (dq, dm->devices_by_cpu[cpu_index])
+  vec_foreach (dq, dm->devices_by_cpu[thread_index])
     {
       xd = vec_elt_at_index(dm->devices, dq->device);
       if (xd->flags & DPDK_DEVICE_FLAG_MAYBE_MULTISEG)
-        n_rx_packets += dpdk_device_input (dm, xd, node, cpu_index, dq->queue_id, /* maybe_multiseg */ 1);
+        n_rx_packets += dpdk_device_input (dm, xd, node, thread_index, dq->queue_id, /* maybe_multiseg */ 1);
       else
-        n_rx_packets += dpdk_device_input (dm, xd, node, cpu_index, dq->queue_id, /* maybe_multiseg */ 0);
+        n_rx_packets += dpdk_device_input (dm, xd, node, thread_index, dq->queue_id, /* maybe_multiseg */ 0);
     }
   /* *INDENT-ON* */
 
diff --git a/src/plugins/dpdk/hqos/hqos.c b/src/plugins/dpdk/hqos/hqos.c
index a288fca7e5e..8b251beb174 100644
--- a/src/plugins/dpdk/hqos/hqos.c
+++ b/src/plugins/dpdk/hqos/hqos.c
@@ -397,7 +397,7 @@ static_always_inline void
 dpdk_hqos_thread_internal_hqos_dbg_bypass (vlib_main_t * vm)
 {
   dpdk_main_t *dm = &dpdk_main;
-  u32 cpu_index = vm->cpu_index;
+  u32 thread_index = vm->thread_index;
   u32 dev_pos;
 
   dev_pos = 0;
@@ -405,12 +405,12 @@ dpdk_hqos_thread_internal_hqos_dbg_bypass (vlib_main_t * vm)
     {
       vlib_worker_thread_barrier_check ();
 
-      u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]);
+      u32 n_devs = vec_len (dm->devices_by_hqos_cpu[thread_index]);
       if (dev_pos >= n_devs)
 	dev_pos = 0;
 
       dpdk_device_and_queue_t *dq =
-	vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos);
+	vec_elt_at_index (dm->devices_by_hqos_cpu[thread_index], dev_pos);
       dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device);
 
       dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht;
@@ -479,7 +479,7 @@ static_always_inline void
 dpdk_hqos_thread_internal (vlib_main_t * vm)
 {
   dpdk_main_t *dm = &dpdk_main;
-  u32 cpu_index = vm->cpu_index;
+  u32 thread_index = vm->thread_index;
   u32 dev_pos;
 
   dev_pos = 0;
@@ -487,7 +487,7 @@ dpdk_hqos_thread_internal (vlib_main_t * vm)
     {
       vlib_worker_thread_barrier_check ();
 
-      u32 n_devs = vec_len (dm->devices_by_hqos_cpu[cpu_index]);
+      u32 n_devs = vec_len (dm->devices_by_hqos_cpu[thread_index]);
       if (PREDICT_FALSE (n_devs == 0))
 	{
 	  dev_pos = 0;
@@ -497,7 +497,7 @@ dpdk_hqos_thread_internal (vlib_main_t * vm)
 	dev_pos = 0;
 
       dpdk_device_and_queue_t *dq =
-	vec_elt_at_index (dm->devices_by_hqos_cpu[cpu_index], dev_pos);
+	vec_elt_at_index (dm->devices_by_hqos_cpu[thread_index], dev_pos);
       dpdk_device_t *xd = vec_elt_at_index (dm->devices, dq->device);
 
       dpdk_device_hqos_per_hqos_thread_t *hqos = xd->hqos_ht;
@@ -586,7 +586,7 @@ dpdk_hqos_thread (vlib_worker_thread_t * w)
 
   vm = vlib_get_main ();
 
-  ASSERT (vm->cpu_index == os_get_cpu_number ());
+  ASSERT (vm->thread_index == vlib_get_thread_index ());
 
   clib_time_init (&vm->clib_time);
   clib_mem_set_heap (w->thread_mheap);
@@ -595,7 +595,7 @@ dpdk_hqos_thread (vlib_worker_thread_t * w)
   while (tm->worker_thread_release == 0)
     vlib_worker_thread_barrier_check ();
 
-  if (vec_len (dm->devices_by_hqos_cpu[vm->cpu_index]) == 0)
+  if (vec_len (dm->devices_by_hqos_cpu[vm->thread_index]) == 0)
     return
       clib_error
       ("current I/O TX thread does not have any devices assigned to it");
diff --git a/src/plugins/dpdk/ipsec/cli.c b/src/plugins/dpdk/ipsec/cli.c
index cd0a6037e4e..3ae8c9b886d 100644
--- a/src/plugins/dpdk/ipsec/cli.c
+++ b/src/plugins/dpdk/ipsec/cli.c
@@ -42,8 +42,8 @@ dpdk_ipsec_show_mapping (vlib_main_t * vm, u16 detail_display)
   for (i = 0; i < tm->n_vlib_mains; i++)
     {
       uword key, data;
-      u32 cpu_index = vlib_mains[i]->cpu_index;
-      crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index];
+      u32 thread_index = vlib_mains[i]->thread_index;
+      crypto_worker_main_t *cwm = &dcm->workers_main[thread_index];
       u8 *s = 0;
 
       if (skip_master)
@@ -57,7 +57,7 @@ dpdk_ipsec_show_mapping (vlib_main_t * vm, u16 detail_display)
 	  i32 last_cdev = -1;
 	  crypto_qp_data_t *qpd;
 
-	  s = format (s, "%u\t", cpu_index);
+	  s = format (s, "%u\t", thread_index);
 
 	  /* *INDENT-OFF* */
 	  vec_foreach (qpd, cwm->qp_data)
@@ -95,7 +95,7 @@ dpdk_ipsec_show_mapping (vlib_main_t * vm, u16 detail_display)
 	    cap.sym.auth.algo = p_key->auth_algo;
 	    check_algo_is_supported (&cap, auth_str);
 	    vlib_cli_output (vm, "%u\t%10s\t%15s\t%3s\t%u\t%u\n",
-			     vlib_mains[i]->cpu_index, cipher_str, auth_str,
+			     vlib_mains[i]->thread_index, cipher_str, auth_str,
 			     p_key->is_outbound ? "out" : "in",
 			     cwm->qp_data[data].dev_id,
 			     cwm->qp_data[data].qp_id);
diff --git a/src/plugins/dpdk/ipsec/crypto_node.c b/src/plugins/dpdk/ipsec/crypto_node.c
index dc3452b286e..a3c459023a3 100644
--- a/src/plugins/dpdk/ipsec/crypto_node.c
+++ b/src/plugins/dpdk/ipsec/crypto_node.c
@@ -171,9 +171,9 @@ static uword
 dpdk_crypto_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		      vlib_frame_t * frame)
 {
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   dpdk_crypto_main_t *dcm = &dpdk_crypto_main;
-  crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index];
+  crypto_worker_main_t *cwm = &dcm->workers_main[thread_index];
   crypto_qp_data_t *qpd;
   u32 n_deq = 0;
 
diff --git a/src/plugins/dpdk/ipsec/esp.h b/src/plugins/dpdk/ipsec/esp.h
index 320295b19e6..56f0c756eec 100644
--- a/src/plugins/dpdk/ipsec/esp.h
+++ b/src/plugins/dpdk/ipsec/esp.h
@@ -170,9 +170,9 @@ static_always_inline int
 create_sym_sess (ipsec_sa_t * sa, crypto_sa_session_t * sa_sess,
 		 u8 is_outbound)
 {
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   dpdk_crypto_main_t *dcm = &dpdk_crypto_main;
-  crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index];
+  crypto_worker_main_t *cwm = &dcm->workers_main[thread_index];
   struct rte_crypto_sym_xform cipher_xform = { 0 };
   struct rte_crypto_sym_xform auth_xform = { 0 };
   struct rte_crypto_sym_xform *xfs;
diff --git a/src/plugins/dpdk/ipsec/esp_decrypt.c b/src/plugins/dpdk/ipsec/esp_decrypt.c
index 286e03f85da..bab76e3b665 100644
--- a/src/plugins/dpdk/ipsec/esp_decrypt.c
+++ b/src/plugins/dpdk/ipsec/esp_decrypt.c
@@ -88,7 +88,7 @@ dpdk_esp_decrypt_node_fn (vlib_main_t * vm,
 {
   u32 n_left_from, *from, *to_next, next_index;
   ipsec_main_t *im = &ipsec_main;
-  u32 cpu_index = os_get_cpu_number();
+  u32 thread_index = vlib_get_thread_index();
   dpdk_crypto_main_t * dcm = &dpdk_crypto_main;
   dpdk_esp_main_t * em = &dpdk_esp_main;
   u32 i;
@@ -104,7 +104,7 @@ dpdk_esp_decrypt_node_fn (vlib_main_t * vm,
       return n_left_from;
     }
 
-  crypto_worker_main_t *cwm = vec_elt_at_index(dcm->workers_main, cpu_index);
+  crypto_worker_main_t *cwm = vec_elt_at_index(dcm->workers_main, thread_index);
   u32 n_qps = vec_len(cwm->qp_data);
   struct rte_crypto_op ** cops_to_enq[n_qps];
   u32 n_cop_qp[n_qps], * bi_to_enq[n_qps];
diff --git a/src/plugins/dpdk/ipsec/esp_encrypt.c b/src/plugins/dpdk/ipsec/esp_encrypt.c
index 5b03de73bf2..f996d7df999 100644
--- a/src/plugins/dpdk/ipsec/esp_encrypt.c
+++ b/src/plugins/dpdk/ipsec/esp_encrypt.c
@@ -93,7 +93,7 @@ dpdk_esp_encrypt_node_fn (vlib_main_t * vm,
 {
   u32 n_left_from, *from, *to_next, next_index;
   ipsec_main_t *im = &ipsec_main;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   dpdk_crypto_main_t *dcm = &dpdk_crypto_main;
   dpdk_esp_main_t *em = &dpdk_esp_main;
   u32 i;
@@ -111,7 +111,8 @@ dpdk_esp_encrypt_node_fn (vlib_main_t * vm,
       return n_left_from;
     }
 
-  crypto_worker_main_t *cwm = vec_elt_at_index (dcm->workers_main, cpu_index);
+  crypto_worker_main_t *cwm =
+    vec_elt_at_index (dcm->workers_main, thread_index);
   u32 n_qps = vec_len (cwm->qp_data);
   struct rte_crypto_op **cops_to_enq[n_qps];
   u32 n_cop_qp[n_qps], *bi_to_enq[n_qps];
diff --git a/src/plugins/dpdk/ipsec/ipsec.c b/src/plugins/dpdk/ipsec/ipsec.c
index b0aaaaeca17..5d8f4fba7f3 100644
--- a/src/plugins/dpdk/ipsec/ipsec.c
+++ b/src/plugins/dpdk/ipsec/ipsec.c
@@ -289,7 +289,7 @@ dpdk_ipsec_process (vlib_main_t * vm, vlib_node_runtime_t * rt,
 	      if (!map)
 		{
 		  clib_warning ("unable to create hash table for worker %u",
-				vlib_mains[i]->cpu_index);
+				vlib_mains[i]->thread_index);
 		  goto error;
 		}
 	      cwm->algo_qp_map = map;
diff --git a/src/plugins/dpdk/ipsec/ipsec.h b/src/plugins/dpdk/ipsec/ipsec.h
index 28bffc8021e..f0f793c022b 100644
--- a/src/plugins/dpdk/ipsec/ipsec.h
+++ b/src/plugins/dpdk/ipsec/ipsec.h
@@ -95,8 +95,8 @@ static_always_inline void
 crypto_alloc_cops ()
 {
   dpdk_crypto_main_t *dcm = &dpdk_crypto_main;
-  u32 cpu_index = os_get_cpu_number ();
-  crypto_worker_main_t *cwm = &dcm->workers_main[cpu_index];
+  u32 thread_index = vlib_get_thread_index ();
+  crypto_worker_main_t *cwm = &dcm->workers_main[thread_index];
   unsigned socket_id = rte_socket_id ();
   crypto_qp_data_t *qpd;
 
diff --git a/src/plugins/dpdk/main.c b/src/plugins/dpdk/main.c
index 7ee2a785ee6..942b8b2d466 100644
--- a/src/plugins/dpdk/main.c
+++ b/src/plugins/dpdk/main.c
@@ -39,7 +39,7 @@ rte_delay_us_override (unsigned us)
    * thread then do not intercept. (Must not be called from an
    * independent pthread).
    */
-  if (os_get_cpu_number () == 0)
+  if (vlib_get_thread_index () == 0)
     {
       /*
        * We're in the vlib main thread or a vlib process. Make sure
diff --git a/src/plugins/flowperpkt/l2_node.c b/src/plugins/flowperpkt/l2_node.c
index 1c2f681e1e1..fdaf81d1a2e 100644
--- a/src/plugins/flowperpkt/l2_node.c
+++ b/src/plugins/flowperpkt/l2_node.c
@@ -102,7 +102,7 @@ add_to_flow_record_l2 (vlib_main_t * vm,
 		       u8 * src_mac, u8 * dst_mac,
 		       u16 ethertype, u64 timestamp, u16 length, int do_flush)
 {
-  u32 my_cpu_number = vm->cpu_index;
+  u32 my_cpu_number = vm->thread_index;
   flow_report_main_t *frm = &flow_report_main;
   ip4_header_t *ip;
   udp_header_t *udp;
diff --git a/src/plugins/flowperpkt/node.c b/src/plugins/flowperpkt/node.c
index f77f087dc78..0277682dc87 100644
--- a/src/plugins/flowperpkt/node.c
+++ b/src/plugins/flowperpkt/node.c
@@ -101,7 +101,7 @@ add_to_flow_record_ipv4 (vlib_main_t * vm,
 			 u32 src_address, u32 dst_address,
 			 u8 tos, u64 timestamp, u16 length, int do_flush)
 {
-  u32 my_cpu_number = vm->cpu_index;
+  u32 my_cpu_number = vm->thread_index;
   flow_report_main_t *frm = &flow_report_main;
   ip4_header_t *ip;
   udp_header_t *udp;
diff --git a/src/plugins/ioam/export-common/ioam_export.h b/src/plugins/ioam/export-common/ioam_export.h
index 2bf3fd5400b..9de0d13bd5e 100644
--- a/src/plugins/ioam/export-common/ioam_export.h
+++ b/src/plugins/ioam/export-common/ioam_export.h
@@ -477,8 +477,8 @@ do {                                                                           \
   from = vlib_frame_vector_args (F);                                           \
   n_left_from = (F)->n_vectors;                                                \
   next_index = (N)->cached_next_index;                                         \
-  while (__sync_lock_test_and_set ((EM)->lockp[(VM)->cpu_index], 1));          \
-  my_buf = ioam_export_get_my_buffer (EM, (VM)->cpu_index);                    \
+  while (__sync_lock_test_and_set ((EM)->lockp[(VM)->thread_index], 1));       \
+  my_buf = ioam_export_get_my_buffer (EM, (VM)->thread_index);                 \
   my_buf->touched_at = vlib_time_now (VM);                                     \
   while (n_left_from > 0)                                                      \
     {                                                                          \
@@ -620,7 +620,7 @@ do {                                                                           \
     }                                                                          \
   vlib_node_increment_counter (VM, export_node.index,                          \
 			       EXPORT_ERROR_RECORDED, pkts_recorded);          \
-  *(EM)->lockp[(VM)->cpu_index] = 0;                                           \
+  *(EM)->lockp[(VM)->thread_index] = 0;                                        \
 } while(0)
 
 #endif /* __included_ioam_export_h__ */
diff --git a/src/plugins/ioam/ip6/ioam_cache_tunnel_select_node.c b/src/plugins/ioam/ip6/ioam_cache_tunnel_select_node.c
index a56dc040e82..0cf742c9be2 100644
--- a/src/plugins/ioam/ip6/ioam_cache_tunnel_select_node.c
+++ b/src/plugins/ioam/ip6/ioam_cache_tunnel_select_node.c
@@ -396,7 +396,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 					      clib_net_to_host_u32
 					      (tcp0->seq_number) + 1,
 					      no_of_responses, now,
-					      vm->cpu_index, &pool_index0))
+					      vm->thread_index, &pool_index0))
 		    {
 		      cache_ts_added++;
 		    }
@@ -419,7 +419,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 	      e2e =
 		(ioam_e2e_cache_option_t *) ((u8 *) hbh0 +
 					     cm->rewrite_pool_index_offset);
-	      e2e->pool_id = (u8) vm->cpu_index;
+	      e2e->pool_id = (u8) vm->thread_index;
 	      e2e->pool_index = pool_index0;
 	      ioam_e2e_id_rewrite_handler ((ioam_e2e_id_option_t *)
 					   ((u8 *) e2e +
@@ -455,7 +455,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 					      clib_net_to_host_u32
 					      (tcp1->seq_number) + 1,
 					      no_of_responses, now,
-					      vm->cpu_index, &pool_index1))
+					      vm->thread_index, &pool_index1))
 		    {
 		      cache_ts_added++;
 		    }
@@ -479,7 +479,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 	      e2e =
 		(ioam_e2e_cache_option_t *) ((u8 *) hbh1 +
 					     cm->rewrite_pool_index_offset);
-	      e2e->pool_id = (u8) vm->cpu_index;
+	      e2e->pool_id = (u8) vm->thread_index;
 	      e2e->pool_index = pool_index1;
 	      ioam_e2e_id_rewrite_handler ((ioam_e2e_id_option_t *)
 					   ((u8 *) e2e +
@@ -562,7 +562,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 					      clib_net_to_host_u32
 					      (tcp0->seq_number) + 1,
 					      no_of_responses, now,
-					      vm->cpu_index, &pool_index0))
+					      vm->thread_index, &pool_index0))
 		    {
 		      cache_ts_added++;
 		    }
@@ -585,7 +585,7 @@ ip6_reset_ts_hbh_node_fn (vlib_main_t * vm,
 	      e2e =
 		(ioam_e2e_cache_option_t *) ((u8 *) hbh0 +
 					     cm->rewrite_pool_index_offset);
-	      e2e->pool_id = (u8) vm->cpu_index;
+	      e2e->pool_id = (u8) vm->thread_index;
 	      e2e->pool_index = pool_index0;
 	      ioam_e2e_id_rewrite_handler ((ioam_e2e_id_option_t *)
 					   ((u8 *) e2e +
@@ -701,7 +701,7 @@ expired_cache_ts_timer_callback (u32 * expired_timers)
   ioam_cache_main_t *cm = &ioam_cache_main;
   int i;
   u32 pool_index;
-  u32 thread_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   u32 count = 0;
 
   for (i = 0; i < vec_len (expired_timers); i++)
@@ -724,7 +724,7 @@ ioam_cache_ts_timer_tick_node_fn (vlib_main_t * vm,
 				  vlib_frame_t * f)
 {
   ioam_cache_main_t *cm = &ioam_cache_main;
-  u32 my_thread_index = os_get_cpu_number ();
+  u32 my_thread_index = vlib_get_thread_index ();
   struct timespec ts, tsrem;
 
   tw_timer_expire_timers_16t_2w_512sl (&cm->timer_wheels[my_thread_index],
diff --git a/src/plugins/ixge/ixge.c b/src/plugins/ixge/ixge.c
index f3c5cc09d65..08f5b6922df 100644
--- a/src/plugins/ixge/ixge.c
+++ b/src/plugins/ixge/ixge.c
@@ -1887,7 +1887,7 @@ done:
   vlib_increment_combined_counter (vnet_main.
 				   interface_main.combined_sw_if_counters +
 				   VNET_INTERFACE_COUNTER_RX,
-				   0 /* cpu_index */ ,
+				   0 /* thread_index */ ,
 				   xd->vlib_sw_if_index, n_packets,
 				   dq->rx.n_bytes);
 
diff --git a/src/plugins/lb/lb.c b/src/plugins/lb/lb.c
index add81236ae5..addc2a42d34 100644
--- a/src/plugins/lb/lb.c
+++ b/src/plugins/lb/lb.c
@@ -63,11 +63,11 @@ u8 *format_lb_main (u8 * s, va_list * args)
   s = format(s, " #vips: %u\n", pool_elts(lbm->vips));
   s = format(s, " #ass: %u\n", pool_elts(lbm->ass) - 1);
 
-  u32 cpu_index;
-  for(cpu_index = 0; cpu_index < tm->n_vlib_mains; cpu_index++ ) {
-    lb_hash_t *h = lbm->per_cpu[cpu_index].sticky_ht;
+  u32 thread_index;
+  for(thread_index = 0; thread_index < tm->n_vlib_mains; thread_index++ ) {
+    lb_hash_t *h = lbm->per_cpu[thread_index].sticky_ht;
     if (h) {
-      s = format(s, "core %d\n", cpu_index);
+      s = format(s, "core %d\n", thread_index);
       s = format(s, "  timeout: %ds\n", h->timeout);
       s = format(s, "  usage: %d / %d\n", lb_hash_elts(h, lb_hash_time_now(vlib_get_main())),  lb_hash_size(h));
     }
diff --git a/src/plugins/lb/node.c b/src/plugins/lb/node.c
index 8b763c537d5..3171148bd3d 100644
--- a/src/plugins/lb/node.c
+++ b/src/plugins/lb/node.c
@@ -60,10 +60,10 @@ format_lb_trace (u8 * s, va_list * args)
   return s;
 }
 
-lb_hash_t *lb_get_sticky_table(u32 cpu_index)
+lb_hash_t *lb_get_sticky_table(u32 thread_index)
 {
   lb_main_t *lbm = &lb_main;
-  lb_hash_t *sticky_ht = lbm->per_cpu[cpu_index].sticky_ht;
+  lb_hash_t *sticky_ht = lbm->per_cpu[thread_index].sticky_ht;
   //Check if size changed
   if (PREDICT_FALSE(sticky_ht && (lbm->per_cpu_sticky_buckets != lb_hash_nbuckets(sticky_ht))))
     {
@@ -71,8 +71,8 @@ lb_hash_t *lb_get_sticky_table(u32 cpu_index)
       lb_hash_bucket_t *b;
       u32 i;
       lb_hash_foreach_entry(sticky_ht, b, i) {
-	vlib_refcount_add(&lbm->as_refcount, cpu_index, b->value[i], -1);
-	vlib_refcount_add(&lbm->as_refcount, cpu_index, 0, 1);
+	vlib_refcount_add(&lbm->as_refcount, thread_index, b->value[i], -1);
+	vlib_refcount_add(&lbm->as_refcount, thread_index, 0, 1);
       }
 
       lb_hash_free(sticky_ht);
@@ -81,8 +81,8 @@ lb_hash_t *lb_get_sticky_table(u32 cpu_index)
 
   //Create if necessary
   if (PREDICT_FALSE(sticky_ht == NULL)) {
-    lbm->per_cpu[cpu_index].sticky_ht = lb_hash_alloc(lbm->per_cpu_sticky_buckets, lbm->flow_timeout);
-    sticky_ht = lbm->per_cpu[cpu_index].sticky_ht;
+    lbm->per_cpu[thread_index].sticky_ht = lb_hash_alloc(lbm->per_cpu_sticky_buckets, lbm->flow_timeout);
+    sticky_ht = lbm->per_cpu[thread_index].sticky_ht;
     clib_warning("Regenerated sticky table %p", sticky_ht);
   }
 
@@ -153,10 +153,10 @@ lb_node_fn (vlib_main_t * vm,
 {
   lb_main_t *lbm = &lb_main;
   u32 n_left_from, *from, next_index, *to_next, n_left_to_next;
-  u32 cpu_index = os_get_cpu_number();
+  u32 thread_index = vlib_get_thread_index();
   u32 lb_time = lb_hash_time_now(vm);
 
-  lb_hash_t *sticky_ht = lb_get_sticky_table(cpu_index);
+  lb_hash_t *sticky_ht = lb_get_sticky_table(thread_index);
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
   next_index = node->cached_next_index;
@@ -240,9 +240,9 @@ lb_node_fn (vlib_main_t * vm,
 	  //Configuration may be changed, vectors resized, etc...
 
 	  //Dereference previously used
-	  vlib_refcount_add(&lbm->as_refcount, cpu_index,
+	  vlib_refcount_add(&lbm->as_refcount, thread_index,
 			    lb_hash_available_value(sticky_ht, hash0, available_index0), -1);
-	  vlib_refcount_add(&lbm->as_refcount, cpu_index,
+	  vlib_refcount_add(&lbm->as_refcount, thread_index,
 			    asindex0, 1);
 
 	  //Add sticky entry
@@ -260,7 +260,7 @@ lb_node_fn (vlib_main_t * vm,
 	}
 
       vlib_increment_simple_counter(&lbm->vip_counters[counter],
-				    cpu_index,
+				    thread_index,
 				    vnet_buffer (p0)->ip.adj_index[VLIB_TX],
 				    1);
 
diff --git a/src/plugins/lb/refcount.c b/src/plugins/lb/refcount.c
index 22415c8889e..6f01ab5aaf7 100644
--- a/src/plugins/lb/refcount.c
+++ b/src/plugins/lb/refcount.c
@@ -31,10 +31,10 @@ u64 vlib_refcount_get(vlib_refcount_t *r, u32 index)
 {
   u64 count = 0;
   vlib_thread_main_t *tm = vlib_get_thread_main ();
-  u32 cpu_index;
-  for (cpu_index = 0; cpu_index < tm->n_vlib_mains; cpu_index++) {
-    if (r->per_cpu[cpu_index].length > index)
-      count += r->per_cpu[cpu_index].counters[index];
+  u32 thread_index;
+  for (thread_index = 0; thread_index < tm->n_vlib_mains; thread_index++) {
+    if (r->per_cpu[thread_index].length > index)
+      count += r->per_cpu[thread_index].counters[index];
   }
   return count;
 }
diff --git a/src/plugins/lb/refcount.h b/src/plugins/lb/refcount.h
index 8c26e7be76f..dcfcb3fee58 100644
--- a/src/plugins/lb/refcount.h
+++ b/src/plugins/lb/refcount.h
@@ -45,9 +45,9 @@ typedef struct {
 void __vlib_refcount_resize(vlib_refcount_per_cpu_t *per_cpu, u32 size);
 
 static_always_inline
-void vlib_refcount_add(vlib_refcount_t *r, u32 cpu_index, u32 counter_index, i32 v)
+void vlib_refcount_add(vlib_refcount_t *r, u32 thread_index, u32 counter_index, i32 v)
 {
-  vlib_refcount_per_cpu_t *per_cpu = &r->per_cpu[cpu_index];
+  vlib_refcount_per_cpu_t *per_cpu = &r->per_cpu[thread_index];
   if (PREDICT_FALSE(counter_index >= per_cpu->length))
     __vlib_refcount_resize(per_cpu, clib_max(counter_index + 16, per_cpu->length * 2));
 
diff --git a/src/plugins/memif/node.c b/src/plugins/memif/node.c
index 659d5dfb5ee..cee1f3d1b7b 100644
--- a/src/plugins/memif/node.c
+++ b/src/plugins/memif/node.c
@@ -94,7 +94,7 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 n_rx_bytes = 0;
   u32 *to_next = 0;
   u32 n_free_bufs;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   u32 bi0, bi1;
   vlib_buffer_t *b0, *b1;
   u16 ring_size = 1 << mif->log2_ring_size;
@@ -105,14 +105,15 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   if (mif->per_interface_next_index != ~0)
     next_index = mif->per_interface_next_index;
 
-  n_free_bufs = vec_len (nm->rx_buffers[cpu_index]);
+  n_free_bufs = vec_len (nm->rx_buffers[thread_index]);
   if (PREDICT_FALSE (n_free_bufs < ring_size))
     {
-      vec_validate (nm->rx_buffers[cpu_index], ring_size + n_free_bufs - 1);
+      vec_validate (nm->rx_buffers[thread_index],
+		    ring_size + n_free_bufs - 1);
       n_free_bufs +=
-	vlib_buffer_alloc (vm, &nm->rx_buffers[cpu_index][n_free_bufs],
+	vlib_buffer_alloc (vm, &nm->rx_buffers[thread_index][n_free_bufs],
 			   ring_size);
-      _vec_len (nm->rx_buffers[cpu_index]) = n_free_bufs;
+      _vec_len (nm->rx_buffers[thread_index]) = n_free_bufs;
     }
 
   head = ring->head;
@@ -158,15 +159,15 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 			     CLIB_CACHE_LINE_BYTES, LOAD);
 	    }
 	  /* get empty buffer */
-	  u32 last_buf = vec_len (nm->rx_buffers[cpu_index]) - 1;
-	  bi0 = nm->rx_buffers[cpu_index][last_buf];
-	  bi1 = nm->rx_buffers[cpu_index][last_buf - 1];
-	  _vec_len (nm->rx_buffers[cpu_index]) -= 2;
+	  u32 last_buf = vec_len (nm->rx_buffers[thread_index]) - 1;
+	  bi0 = nm->rx_buffers[thread_index][last_buf];
+	  bi1 = nm->rx_buffers[thread_index][last_buf - 1];
+	  _vec_len (nm->rx_buffers[thread_index]) -= 2;
 
 	  if (last_buf > 4)
 	    {
-	      memif_prefetch (vm, nm->rx_buffers[cpu_index][last_buf - 2]);
-	      memif_prefetch (vm, nm->rx_buffers[cpu_index][last_buf - 3]);
+	      memif_prefetch (vm, nm->rx_buffers[thread_index][last_buf - 2]);
+	      memif_prefetch (vm, nm->rx_buffers[thread_index][last_buf - 3]);
 	    }
 
 	  /* enqueue buffer */
@@ -256,9 +257,9 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
       while (num_slots && n_left_to_next)
 	{
 	  /* get empty buffer */
-	  u32 last_buf = vec_len (nm->rx_buffers[cpu_index]) - 1;
-	  bi0 = nm->rx_buffers[cpu_index][last_buf];
-	  _vec_len (nm->rx_buffers[cpu_index]) = last_buf;
+	  u32 last_buf = vec_len (nm->rx_buffers[thread_index]) - 1;
+	  bi0 = nm->rx_buffers[thread_index][last_buf];
+	  _vec_len (nm->rx_buffers[thread_index]) = last_buf;
 
 	  /* enqueue buffer */
 	  to_next[0] = bi0;
@@ -315,7 +316,7 @@ memif_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   ring->tail = head;
 
   vlib_increment_combined_counter (vnm->interface_main.combined_sw_if_counters
-				   + VNET_INTERFACE_COUNTER_RX, cpu_index,
+				   + VNET_INTERFACE_COUNTER_RX, thread_index,
 				   mif->hw_if_index, n_rx_packets,
 				   n_rx_bytes);
 
@@ -327,7 +328,7 @@ memif_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 		vlib_frame_t * frame)
 {
   u32 n_rx_packets = 0;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
   memif_main_t *nm = &memif_main;
   memif_if_t *mif;
 
@@ -337,7 +338,7 @@ memif_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
       if (mif->flags & MEMIF_IF_FLAG_ADMIN_UP &&
 	  mif->flags & MEMIF_IF_FLAG_CONNECTED &&
 	  (mif->if_index % nm->input_cpu_count) ==
-	  (cpu_index - nm->input_cpu_first_index))
+	  (thread_index - nm->input_cpu_first_index))
 	{
 	  if (mif->flags & MEMIF_IF_FLAG_IS_SLAVE)
 	    n_rx_packets +=
diff --git a/src/plugins/snat/in2out.c b/src/plugins/snat/in2out.c
index b4961365309..e5ee965f7ae 100644
--- a/src/plugins/snat/in2out.c
+++ b/src/plugins/snat/in2out.c
@@ -212,7 +212,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
                       snat_session_t ** sessionp,
                       vlib_node_runtime_t * node,
                       u32 next0,
-                      u32 cpu_index)
+                      u32 thread_index)
 {
   snat_user_t *u;
   snat_user_key_t user_key;
@@ -246,27 +246,27 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
   if (clib_bihash_search_8_8 (&sm->user_hash, &kv0, &value0))
     {
       /* no, make a new one */
-      pool_get (sm->per_thread_data[cpu_index].users, u);
+      pool_get (sm->per_thread_data[thread_index].users, u);
       memset (u, 0, sizeof (*u));
       u->addr = ip0->src_address;
       u->fib_index = rx_fib_index0;
 
-      pool_get (sm->per_thread_data[cpu_index].list_pool, per_user_list_head_elt);
+      pool_get (sm->per_thread_data[thread_index].list_pool, per_user_list_head_elt);
 
       u->sessions_per_user_list_head_index = per_user_list_head_elt -
-        sm->per_thread_data[cpu_index].list_pool;
+        sm->per_thread_data[thread_index].list_pool;
 
-      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+      clib_dlist_init (sm->per_thread_data[thread_index].list_pool,
                        u->sessions_per_user_list_head_index);
 
-      kv0.value = u - sm->per_thread_data[cpu_index].users;
+      kv0.value = u - sm->per_thread_data[thread_index].users;
 
       /* add user */
       clib_bihash_add_del_8_8 (&sm->user_hash, &kv0, 1 /* is_add */);
     }
   else
     {
-      u = pool_elt_at_index (sm->per_thread_data[cpu_index].users,
+      u = pool_elt_at_index (sm->per_thread_data[thread_index].users,
                              value0.value);
     }
 
@@ -276,25 +276,25 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
       /* Remove the oldest dynamic translation */
       do {
           oldest_per_user_translation_list_index =
-            clib_dlist_remove_head (sm->per_thread_data[cpu_index].list_pool,
+            clib_dlist_remove_head (sm->per_thread_data[thread_index].list_pool,
                                     u->sessions_per_user_list_head_index);
 
           ASSERT (oldest_per_user_translation_list_index != ~0);
 
           /* add it back to the end of the LRU list */
-          clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+          clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                               u->sessions_per_user_list_head_index,
                               oldest_per_user_translation_list_index);
           /* Get the list element */
           oldest_per_user_translation_list_elt =
-            pool_elt_at_index (sm->per_thread_data[cpu_index].list_pool,
+            pool_elt_at_index (sm->per_thread_data[thread_index].list_pool,
                                oldest_per_user_translation_list_index);
 
           /* Get the session index from the list element */
           session_index = oldest_per_user_translation_list_elt->value;
 
           /* Get the session */
-          s = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+          s = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                  session_index);
       } while (snat_is_session_static (s));
 
@@ -346,7 +346,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
         }
 
       /* Create a new session */
-      pool_get (sm->per_thread_data[cpu_index].sessions, s);
+      pool_get (sm->per_thread_data[thread_index].sessions, s);
       memset (s, 0, sizeof (*s));
       
       s->outside_address_index = address_index;
@@ -362,22 +362,22 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
         }
 
       /* Create list elts */
-      pool_get (sm->per_thread_data[cpu_index].list_pool,
+      pool_get (sm->per_thread_data[thread_index].list_pool,
                 per_user_translation_list_elt);
-      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+      clib_dlist_init (sm->per_thread_data[thread_index].list_pool,
                        per_user_translation_list_elt -
-                       sm->per_thread_data[cpu_index].list_pool);
+                       sm->per_thread_data[thread_index].list_pool);
 
       per_user_translation_list_elt->value =
-        s - sm->per_thread_data[cpu_index].sessions;
+        s - sm->per_thread_data[thread_index].sessions;
       s->per_user_index = per_user_translation_list_elt -
-                          sm->per_thread_data[cpu_index].list_pool;
+                          sm->per_thread_data[thread_index].list_pool;
       s->per_user_list_head_index = u->sessions_per_user_list_head_index;
 
-      clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+      clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                           s->per_user_list_head_index,
                           per_user_translation_list_elt -
-                          sm->per_thread_data[cpu_index].list_pool);
+                          sm->per_thread_data[thread_index].list_pool);
    }
   
   s->in2out = *key0;
@@ -388,12 +388,12 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
 
   /* Add to translation hashes */
   kv0.key = s->in2out.as_u64;
-  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
+  kv0.value = s - sm->per_thread_data[thread_index].sessions;
   if (clib_bihash_add_del_8_8 (&sm->in2out, &kv0, 1 /* is_add */))
       clib_warning ("in2out key add failed");
   
   kv0.key = s->out2in.as_u64;
-  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
+  kv0.value = s - sm->per_thread_data[thread_index].sessions;
   
   if (clib_bihash_add_del_8_8 (&sm->out2in, &kv0, 1 /* is_add */))
       clib_warning ("out2in key add failed");
@@ -403,7 +403,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
   worker_by_out_key.port = s->out2in.port;
   worker_by_out_key.fib_index = s->out2in.fib_index;
   kv0.key = worker_by_out_key.as_u64;
-  kv0.value = cpu_index;
+  kv0.value = thread_index;
   clib_bihash_add_del_8_8 (&sm->worker_by_out, &kv0, 1);
 
   /* log NAT event */
@@ -465,7 +465,7 @@ snat_in2out_error_t icmp_get_key(icmp46_header_t *icmp0,
  *
  * @param[in,out] sm             SNAT main
  * @param[in,out] node           SNAT node runtime
- * @param[in] cpu_index          CPU index
+ * @param[in] thread_index       thread index
  * @param[in,out] b0             buffer containing packet to be translated
  * @param[out] p_key             address and port before NAT translation
  * @param[out] p_value           address and port after NAT translation
@@ -473,7 +473,7 @@ snat_in2out_error_t icmp_get_key(icmp46_header_t *icmp0,
  * @param d                      optional parameter
  */
 u32 icmp_match_in2out_slow(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d)
@@ -524,13 +524,13 @@ u32 icmp_match_in2out_slow(snat_main_t *sm, vlib_node_runtime_t *node,
         }
 
       next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                         &s0, node, next0, cpu_index);
+                         &s0, node, next0, thread_index);
 
       if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
         goto out;
     }
   else
-    s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+    s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                             value0.value);
 
 out:
@@ -548,7 +548,7 @@ out:
  *
  * @param[in] sm                 SNAT main
  * @param[in,out] node           SNAT node runtime
- * @param[in] cpu_index          CPU index
+ * @param[in] thread_index       thread index
  * @param[in,out] b0             buffer containing packet to be translated
  * @param[out] p_key             address and port before NAT translation
  * @param[out] p_value           address and port after NAT translation
@@ -556,7 +556,7 @@ out:
  * @param d                      optional parameter
  */
 u32 icmp_match_in2out_fast(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d)
@@ -624,7 +624,7 @@ static inline u32 icmp_in2out (snat_main_t *sm,
                                u32 rx_fib_index0,
                                vlib_node_runtime_t * node,
                                u32 next0,
-                               u32 cpu_index,
+                               u32 thread_index,
                                void *d)
 {
   snat_session_key_t key0, sm0;
@@ -641,7 +641,7 @@ static inline u32 icmp_in2out (snat_main_t *sm,
 
   echo0 = (icmp_echo_header_t *)(icmp0+1);
 
-  next0_tmp = sm->icmp_match_in2out_cb(sm, node, cpu_index, b0,
+  next0_tmp = sm->icmp_match_in2out_cb(sm, node, thread_index, b0,
                                        &key0, &sm0, &dont_translate, d);
   if (next0_tmp != ~0)
     next0 = next0_tmp;
@@ -847,11 +847,11 @@ static inline u32 icmp_in2out_slow_path (snat_main_t *sm,
                                          vlib_node_runtime_t * node,
                                          u32 next0,
                                          f64 now,
-                                         u32 cpu_index,
+                                         u32 thread_index,
                                          snat_session_t ** p_s0)
 {
   next0 = icmp_in2out(sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node,
-                      next0, cpu_index, p_s0);
+                      next0, thread_index, p_s0);
   snat_session_t * s0 = *p_s0;
   if (PREDICT_TRUE(next0 != SNAT_IN2OUT_NEXT_DROP && s0))
     {
@@ -862,9 +862,9 @@ static inline u32 icmp_in2out_slow_path (snat_main_t *sm,
       /* Per-user LRU list maintenance for dynamic translations */
       if (!snat_is_session_static (s0))
         {
-          clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+          clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                              s0->per_user_index);
-          clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+          clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                               s0->per_user_list_head_index,
                               s0->per_user_index);
         }
@@ -884,7 +884,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
   snat_runtime_t * rt = (snat_runtime_t *)node->runtime_data;
   f64 now = vlib_time_now (vm);
   u32 stats_node_index;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
 
   stats_node_index = is_slow_path ? snat_in2out_slowpath_node.index :
     snat_in2out_node.index;
@@ -977,7 +977,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 {
                   next0 = icmp_in2out_slow_path 
                     (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, 
-                     node, next0, now, cpu_index, &s0);
+                     node, next0, now, thread_index, &s0);
                   goto trace00;
                 }
             }
@@ -1006,7 +1006,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace00;
 
                   next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                                     &s0, node, next0, cpu_index);
+                                     &s0, node, next0, thread_index);
                   if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace00;
                 }
@@ -1017,7 +1017,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value0.value);
 
           old_addr0 = ip0->src_address.as_u32;
@@ -1063,9 +1063,9 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s0))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s0->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
@@ -1081,7 +1081,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next0;
                   t->session_index = ~0;
               if (s0)
-                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s0 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_IN2OUT_NEXT_DROP;
@@ -1117,7 +1117,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 {
                   next1 = icmp_in2out_slow_path 
                     (sm, b1, ip1, icmp1, sw_if_index1, rx_fib_index1, node,
-                     next1, now, cpu_index, &s1);
+                     next1, now, thread_index, &s1);
                   goto trace01;
                 }
             }
@@ -1146,7 +1146,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace01;
 
                   next1 = slow_path (sm, b1, ip1, rx_fib_index1, &key1,
-                                     &s1, node, next1, cpu_index);
+                                     &s1, node, next1, thread_index);
                   if (PREDICT_FALSE (next1 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace01;
                 }
@@ -1157,7 +1157,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s1 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s1 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value1.value);
 
           old_addr1 = ip1->src_address.as_u32;
@@ -1203,9 +1203,9 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s1))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s1->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s1->per_user_list_head_index,
                                   s1->per_user_index);
             }
@@ -1220,7 +1220,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next1;
               t->session_index = ~0;
               if (s1)
-                t->session_index = s1 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s1 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next1 != SNAT_IN2OUT_NEXT_DROP;
@@ -1292,7 +1292,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 {
                   next0 = icmp_in2out_slow_path 
                     (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node,
-                     next0, now, cpu_index, &s0);
+                     next0, now, thread_index, &s0);
                   goto trace0;
                 }
             }
@@ -1321,7 +1321,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace0;
 
                   next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                                     &s0, node, next0, cpu_index);
+                                     &s0, node, next0, thread_index);
 
                   if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace0;
@@ -1333,7 +1333,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value0.value);
 
           old_addr0 = ip0->src_address.as_u32;
@@ -1379,9 +1379,9 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s0))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s0->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
@@ -1397,7 +1397,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next0;
                   t->session_index = ~0;
               if (s0)
-                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s0 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_IN2OUT_NEXT_DROP;
@@ -2010,7 +2010,7 @@ snat_in2out_worker_handoff_fn (vlib_main_t * vm,
   u32 n_left_to_next_worker = 0, *to_next_worker = 0;
   u32 next_worker_index = 0;
   u32 current_worker_index = ~0;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
 
   ASSERT (vec_len (sm->workers));
 
@@ -2048,7 +2048,7 @@ snat_in2out_worker_handoff_fn (vlib_main_t * vm,
 
       next_worker_index = sm->worker_in2out_cb(ip0, rx_fib_index0);
 
-      if (PREDICT_FALSE (next_worker_index != cpu_index))
+      if (PREDICT_FALSE (next_worker_index != thread_index))
         {
           do_handoff = 1;
 
diff --git a/src/plugins/snat/out2in.c b/src/plugins/snat/out2in.c
index 656e42dbd81..5d308d78c4f 100644
--- a/src/plugins/snat/out2in.c
+++ b/src/plugins/snat/out2in.c
@@ -129,7 +129,7 @@ create_session_for_static_mapping (snat_main_t *sm,
                                    snat_session_key_t in2out,
                                    snat_session_key_t out2in,
                                    vlib_node_runtime_t * node,
-                                   u32 cpu_index)
+                                   u32 thread_index)
 {
   snat_user_t *u;
   snat_user_key_t user_key;
@@ -146,36 +146,36 @@ create_session_for_static_mapping (snat_main_t *sm,
   if (clib_bihash_search_8_8 (&sm->user_hash, &kv0, &value0))
     {
       /* no, make a new one */
-      pool_get (sm->per_thread_data[cpu_index].users, u);
+      pool_get (sm->per_thread_data[thread_index].users, u);
       memset (u, 0, sizeof (*u));
       u->addr = in2out.addr;
       u->fib_index = in2out.fib_index;
 
-      pool_get (sm->per_thread_data[cpu_index].list_pool,
+      pool_get (sm->per_thread_data[thread_index].list_pool,
                 per_user_list_head_elt);
 
       u->sessions_per_user_list_head_index = per_user_list_head_elt -
-        sm->per_thread_data[cpu_index].list_pool;
+        sm->per_thread_data[thread_index].list_pool;
 
-      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+      clib_dlist_init (sm->per_thread_data[thread_index].list_pool,
                        u->sessions_per_user_list_head_index);
 
-      kv0.value = u - sm->per_thread_data[cpu_index].users;
+      kv0.value = u - sm->per_thread_data[thread_index].users;
 
       /* add user */
       clib_bihash_add_del_8_8 (&sm->user_hash, &kv0, 1 /* is_add */);
 
       /* add non-traslated packets worker lookup */
-      kv0.value = cpu_index;
+      kv0.value = thread_index;
       clib_bihash_add_del_8_8 (&sm->worker_by_in, &kv0, 1);
     }
   else
     {
-      u = pool_elt_at_index (sm->per_thread_data[cpu_index].users,
+      u = pool_elt_at_index (sm->per_thread_data[thread_index].users,
                              value0.value);
     }
 
-  pool_get (sm->per_thread_data[cpu_index].sessions, s);
+  pool_get (sm->per_thread_data[thread_index].sessions, s);
   memset (s, 0, sizeof (*s));
 
   s->outside_address_index = ~0;
@@ -183,22 +183,22 @@ create_session_for_static_mapping (snat_main_t *sm,
   u->nstaticsessions++;
 
   /* Create list elts */
-  pool_get (sm->per_thread_data[cpu_index].list_pool,
+  pool_get (sm->per_thread_data[thread_index].list_pool,
             per_user_translation_list_elt);
-  clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+  clib_dlist_init (sm->per_thread_data[thread_index].list_pool,
                    per_user_translation_list_elt -
-                   sm->per_thread_data[cpu_index].list_pool);
+                   sm->per_thread_data[thread_index].list_pool);
 
   per_user_translation_list_elt->value =
-    s - sm->per_thread_data[cpu_index].sessions;
+    s - sm->per_thread_data[thread_index].sessions;
   s->per_user_index =
-    per_user_translation_list_elt - sm->per_thread_data[cpu_index].list_pool;
+    per_user_translation_list_elt - sm->per_thread_data[thread_index].list_pool;
   s->per_user_list_head_index = u->sessions_per_user_list_head_index;
 
-  clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+  clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                       s->per_user_list_head_index,
                       per_user_translation_list_elt -
-                      sm->per_thread_data[cpu_index].list_pool);
+                      sm->per_thread_data[thread_index].list_pool);
 
   s->in2out = in2out;
   s->out2in = out2in;
@@ -206,12 +206,12 @@ create_session_for_static_mapping (snat_main_t *sm,
 
   /* Add to translation hashes */
   kv0.key = s->in2out.as_u64;
-  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
+  kv0.value = s - sm->per_thread_data[thread_index].sessions;
   if (clib_bihash_add_del_8_8 (&sm->in2out, &kv0, 1 /* is_add */))
       clib_warning ("in2out key add failed");
 
   kv0.key = s->out2in.as_u64;
-  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
+  kv0.value = s - sm->per_thread_data[thread_index].sessions;
 
   if (clib_bihash_add_del_8_8 (&sm->out2in, &kv0, 1 /* is_add */))
       clib_warning ("out2in key add failed");
@@ -298,7 +298,7 @@ is_interface_addr(snat_main_t *sm, vlib_node_runtime_t *node, u32 sw_if_index0,
  *
  * @param[in,out] sm             SNAT main
  * @param[in,out] node           SNAT node runtime
- * @param[in] cpu_index          CPU index
+ * @param[in] thread_index       thread index
  * @param[in,out] b0             buffer containing packet to be translated
  * @param[out] p_key             address and port before NAT translation
  * @param[out] p_value           address and port after NAT translation
@@ -306,7 +306,7 @@ is_interface_addr(snat_main_t *sm, vlib_node_runtime_t *node, u32 sw_if_index0,
  * @param d                      optional parameter
  */
 u32 icmp_match_out2in_slow(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d)
@@ -366,7 +366,7 @@ u32 icmp_match_out2in_slow(snat_main_t *sm, vlib_node_runtime_t *node,
 
       /* Create session initiated by host from external network */
       s0 = create_session_for_static_mapping(sm, b0, sm0, key0,
-                                             node, cpu_index);
+                                             node, thread_index);
 
       if (!s0)
         {
@@ -375,7 +375,7 @@ u32 icmp_match_out2in_slow(snat_main_t *sm, vlib_node_runtime_t *node,
         }
     }
   else
-    s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+    s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                             value0.value);
 
 out:
@@ -393,7 +393,7 @@ out:
  *
  * @param[in] sm                 SNAT main
  * @param[in,out] node           SNAT node runtime
- * @param[in] cpu_index          CPU index
+ * @param[in] thread_index       thread index
  * @param[in,out] b0             buffer containing packet to be translated
  * @param[out] p_key             address and port before NAT translation
  * @param[out] p_value           address and port after NAT translation
@@ -401,7 +401,7 @@ out:
  * @param d                      optional parameter
  */
 u32 icmp_match_out2in_fast(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d)
@@ -460,7 +460,7 @@ static inline u32 icmp_out2in (snat_main_t *sm,
                                u32 rx_fib_index0,
                                vlib_node_runtime_t * node,
                                u32 next0,
-                               u32 cpu_index,
+                               u32 thread_index,
                                void *d)
 {
   snat_session_key_t key0, sm0;
@@ -477,7 +477,7 @@ static inline u32 icmp_out2in (snat_main_t *sm,
 
   echo0 = (icmp_echo_header_t *)(icmp0+1);
 
-  next0_tmp = sm->icmp_match_out2in_cb(sm, node, cpu_index, b0,
+  next0_tmp = sm->icmp_match_out2in_cb(sm, node, thread_index, b0,
                                        &key0, &sm0, &dont_translate, d);
   if (next0_tmp != ~0)
     next0 = next0_tmp;
@@ -589,11 +589,11 @@ static inline u32 icmp_out2in_slow_path (snat_main_t *sm,
                                          u32 rx_fib_index0,
                                          vlib_node_runtime_t * node,
                                          u32 next0, f64 now,
-                                         u32 cpu_index,
+                                         u32 thread_index,
                                          snat_session_t ** p_s0)
 {
   next0 = icmp_out2in(sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node,
-                      next0, cpu_index, p_s0);
+                      next0, thread_index, p_s0);
   snat_session_t * s0 = *p_s0;
   if (PREDICT_TRUE(next0 != SNAT_OUT2IN_NEXT_DROP && s0))
     {
@@ -604,9 +604,9 @@ static inline u32 icmp_out2in_slow_path (snat_main_t *sm,
       /* Per-user LRU list maintenance for dynamic translation */
       if (!snat_is_session_static (s0))
         {
-          clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+          clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                              s0->per_user_index);
-          clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+          clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                               s0->per_user_list_head_index,
                               s0->per_user_index);
         }
@@ -624,7 +624,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
   u32 pkts_processed = 0;
   snat_main_t * sm = &snat_main;
   f64 now = vlib_time_now (vm);
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -712,7 +712,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
             {
               next0 = icmp_out2in_slow_path 
                 (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node, 
-                 next0, now, cpu_index, &s0);
+                 next0, now, thread_index, &s0);
               goto trace0;
             }
 
@@ -743,7 +743,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
 
               /* Create session initiated by host from external network */
               s0 = create_session_for_static_mapping(sm, b0, sm0, key0, node,
-                                                     cpu_index);
+                                                     thread_index);
               if (!s0)
                 {
                   b0->error = node->errors[SNAT_OUT2IN_ERROR_NO_TRANSLATION];
@@ -752,7 +752,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value0.value);
 
           old_addr0 = ip0->dst_address.as_u32;
@@ -796,9 +796,9 @@ snat_out2in_node_fn (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s0))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s0->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
@@ -813,7 +813,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next0;
               t->session_index = ~0;
               if (s0)
-                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s0 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_OUT2IN_NEXT_DROP;
@@ -847,7 +847,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
             {
               next1 = icmp_out2in_slow_path 
                 (sm, b1, ip1, icmp1, sw_if_index1, rx_fib_index1, node, 
-                 next1, now, cpu_index, &s1);
+                 next1, now, thread_index, &s1);
               goto trace1;
             }
 
@@ -878,7 +878,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
 
               /* Create session initiated by host from external network */
               s1 = create_session_for_static_mapping(sm, b1, sm1, key1, node,
-                                                     cpu_index);
+                                                     thread_index);
               if (!s1)
                 {
                   b1->error = node->errors[SNAT_OUT2IN_ERROR_NO_TRANSLATION];
@@ -887,7 +887,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
                 }
             }
           else
-            s1 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s1 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value1.value);
 
           old_addr1 = ip1->dst_address.as_u32;
@@ -931,9 +931,9 @@ snat_out2in_node_fn (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s1))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s1->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s1->per_user_list_head_index,
                                   s1->per_user_index);
             }
@@ -948,7 +948,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next1;
               t->session_index = ~0;
               if (s1)
-                t->session_index = s1 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s1 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next1 != SNAT_OUT2IN_NEXT_DROP;
@@ -1016,7 +1016,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
             {
               next0 = icmp_out2in_slow_path 
                 (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node, 
-                 next0, now, cpu_index, &s0);
+                 next0, now, thread_index, &s0);
               goto trace00;
             }
 
@@ -1048,7 +1048,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
 
               /* Create session initiated by host from external network */
               s0 = create_session_for_static_mapping(sm, b0, sm0, key0, node,
-                                                     cpu_index);
+                                                     thread_index);
               if (!s0)
                 {
                   b0->error = node->errors[SNAT_OUT2IN_ERROR_NO_TRANSLATION];
@@ -1057,7 +1057,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+            s0 = pool_elt_at_index (sm->per_thread_data[thread_index].sessions,
                                     value0.value);
 
           old_addr0 = ip0->dst_address.as_u32;
@@ -1101,9 +1101,9 @@ snat_out2in_node_fn (vlib_main_t * vm,
           /* Per-user LRU list maintenance for dynamic translation */
           if (!snat_is_session_static (s0))
             {
-              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_remove (sm->per_thread_data[thread_index].list_pool,
                                  s0->per_user_index);
-              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+              clib_dlist_addtail (sm->per_thread_data[thread_index].list_pool,
                                   s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
@@ -1118,7 +1118,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next0;
               t->session_index = ~0;
               if (s0)
-                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
+                t->session_index = s0 - sm->per_thread_data[thread_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_OUT2IN_NEXT_DROP;
@@ -1599,7 +1599,7 @@ snat_out2in_worker_handoff_fn (vlib_main_t * vm,
   u32 n_left_to_next_worker = 0, *to_next_worker = 0;
   u32 next_worker_index = 0;
   u32 current_worker_index = ~0;
-  u32 cpu_index = os_get_cpu_number ();
+  u32 thread_index = vlib_get_thread_index ();
 
   ASSERT (vec_len (sm->workers));
 
@@ -1637,7 +1637,7 @@ snat_out2in_worker_handoff_fn (vlib_main_t * vm,
 
       next_worker_index = sm->worker_out2in_cb(ip0, rx_fib_index0);
 
-      if (PREDICT_FALSE (next_worker_index != cpu_index))
+      if (PREDICT_FALSE (next_worker_index != thread_index))
         {
           do_handoff = 1;
 
diff --git a/src/plugins/snat/snat.h b/src/plugins/snat/snat.h
index 017825c0904..f4e1c5c0007 100644
--- a/src/plugins/snat/snat.h
+++ b/src/plugins/snat/snat.h
@@ -221,7 +221,7 @@ struct snat_main_s;
 
 typedef u32 snat_icmp_match_function_t (struct snat_main_s *sm,
                                         vlib_node_runtime_t *node,
-                                        u32 cpu_index,
+                                        u32 thread_index,
                                         vlib_buffer_t *b0,
                                         snat_session_key_t *p_key,
                                         snat_session_key_t *p_value,
@@ -402,22 +402,22 @@ typedef struct {
 } tcp_udp_header_t;
 
 u32 icmp_match_in2out_fast(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d);
 u32 icmp_match_in2out_slow(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d);
 u32 icmp_match_out2in_fast(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d);
 u32 icmp_match_out2in_slow(snat_main_t *sm, vlib_node_runtime_t *node,
-                           u32 cpu_index, vlib_buffer_t *b0,
+                           u32 thread_index, vlib_buffer_t *b0,
                            snat_session_key_t *p_key,
                            snat_session_key_t *p_value,
                            u8 *p_dont_translate, void *d);