snat: thread safe (VPP-443)

All traffic corresponding to a specific SANT user is handled by a CPU core.
in2out: Non-translated packets worker lookup by src address and VRF hash in
snat-in2out-worker-handoff node.
out2in: Translated packets worker lookup by dst address and port number hash in
snat-out2in-worker-handoff node.

Change-Id: Ia092a605689539469841d382588f3f486a29a769
Signed-off-by: Matus Fabian <matfabia@cisco.com>

7 files changed, 1066 insertions, 188 deletions

diff --git a/plugins/snat-plugin/snat/in2out.c b/plugins/snat-plugin/snat/in2out.c
index 9a4aeb01fd4..e1edbb81581 100644
--- a/plugins/snat-plugin/snat/in2out.c
+++ b/plugins/snat-plugin/snat/in2out.c
@@ -16,6 +16,7 @@
 #include <vlib/vlib.h>
 #include <vnet/vnet.h>
 #include <vnet/pg/pg.h>
+#include <vnet/handoff.h>
 
 #include <vnet/ip/ip.h>
 #include <vnet/ethernet/ethernet.h>
@@ -33,6 +34,11 @@ typedef struct {
   u32 is_slow_path;
 } snat_in2out_trace_t;
 
+typedef struct {
+  u32 next_worker_index;
+  u8 do_handoff;
+} snat_in2out_worker_handoff_trace_t;
+
 /* packet trace format function */
 static u8 * format_snat_in2out_trace (u8 * s, va_list * args)
 {
@@ -61,9 +67,24 @@ static u8 * format_snat_in2out_fast_trace (u8 * s, va_list * args)
   return s;
 }
 
+static u8 * format_snat_in2out_worker_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 *);
+  snat_in2out_worker_handoff_trace_t * t =
+    va_arg (*args, snat_in2out_worker_handoff_trace_t *);
+  char * m;
+
+  m = t->do_handoff ? "next worker" : "same worker";
+  s = format (s, "SNAT_IN2OUT_WORKER_HANDOFF: %s %d", m, t->next_worker_index);
+
+  return s;
+}
+
 vlib_node_registration_t snat_in2out_node;
 vlib_node_registration_t snat_in2out_slowpath_node;
 vlib_node_registration_t snat_in2out_fast_node;
+vlib_node_registration_t snat_in2out_worker_handoff_node;
 
 #define foreach_snat_in2out_error                       \
 _(UNSUPPORTED_PROTOCOL, "Unsupported protocol")         \
@@ -93,14 +114,14 @@ typedef enum {
   SNAT_IN2OUT_N_NEXT,
 } snat_in2out_next_t;
 
-
 static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
                       ip4_header_t * ip0,
                       u32 rx_fib_index0,
                       snat_session_key_t * key0,
                       snat_session_t ** sessionp,
                       vlib_node_runtime_t * node,
-                      u32 next0)
+                      u32 next0,
+                      u32 cpu_index)
 {
   snat_user_t *u;
   snat_user_key_t user_key;
@@ -115,6 +136,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
   u32 address_index = ~0;
   u32 outside_fib_index;
   uword * p;
+  snat_static_mapping_key_t worker_by_out_key;
 
   p = hash_get (sm->ip4_main->fib_index_by_table_id, sm->outside_vrf_id);
   if (! p)
@@ -132,25 +154,27 @@ 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->users, u);
+      pool_get (sm->per_thread_data[cpu_index].users, u);
       memset (u, 0, sizeof (*u));
       u->addr = ip0->src_address;
 
-      pool_get (sm->list_pool, per_user_list_head_elt);
+      pool_get (sm->per_thread_data[cpu_index].list_pool, per_user_list_head_elt);
 
       u->sessions_per_user_list_head_index = per_user_list_head_elt -
-        sm->list_pool;
+        sm->per_thread_data[cpu_index].list_pool;
 
-      clib_dlist_init (sm->list_pool, u->sessions_per_user_list_head_index);
+      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+                       u->sessions_per_user_list_head_index);
 
-      kv0.value = u - sm->users;
+      kv0.value = u - sm->per_thread_data[cpu_index].users;
 
       /* add user */
       clib_bihash_add_del_8_8 (&sm->user_hash, &kv0, 1 /* is_add */);
     }
   else
     {
-      u = pool_elt_at_index (sm->users, value0.value);
+      u = pool_elt_at_index (sm->per_thread_data[cpu_index].users,
+                             value0.value);
     }
 
   /* Over quota? Recycle the least recently used dynamic translation */
@@ -159,25 +183,26 @@ 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->list_pool, u->sessions_per_user_list_head_index);
+            clib_dlist_remove_head (sm->per_thread_data[cpu_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->list_pool,
+          clib_dlist_addtail (sm->per_thread_data[cpu_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->list_pool,
+            pool_elt_at_index (sm->per_thread_data[cpu_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->sessions, session_index);
+          s = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                 session_index);
       } while (!snat_is_session_static (s));
 
       /* Remove in2out, out2in keys */
@@ -218,7 +243,7 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
         }
 
       /* Create a new session */
-      pool_get (sm->sessions, s);
+      pool_get (sm->per_thread_data[cpu_index].sessions, s);
       memset (s, 0, sizeof (*s));
       
       s->outside_address_index = address_index;
@@ -234,16 +259,22 @@ static u32 slow_path (snat_main_t *sm, vlib_buffer_t *b0,
         }
 
       /* Create list elts */
-      pool_get (sm->list_pool, per_user_translation_list_elt);
-      clib_dlist_init (sm->list_pool, per_user_translation_list_elt -
-                       sm->list_pool);
-
-      per_user_translation_list_elt->value = s - sm->sessions;
-      s->per_user_index = per_user_translation_list_elt - sm->list_pool;
+      pool_get (sm->per_thread_data[cpu_index].list_pool,
+                per_user_translation_list_elt);
+      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+                       per_user_translation_list_elt -
+                       sm->per_thread_data[cpu_index].list_pool);
+
+      per_user_translation_list_elt->value =
+        s - sm->per_thread_data[cpu_index].sessions;
+      s->per_user_index = per_user_translation_list_elt -
+                          sm->per_thread_data[cpu_index].list_pool;
       s->per_user_list_head_index = u->sessions_per_user_list_head_index;
 
-      clib_dlist_addtail (sm->list_pool, s->per_user_list_head_index,
-                          per_user_translation_list_elt - sm->list_pool);
+      clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                          s->per_user_list_head_index,
+                          per_user_translation_list_elt -
+                          sm->per_thread_data[cpu_index].list_pool);
    }
   
   s->in2out = *key0;
@@ -254,16 +285,23 @@ 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->sessions;
+  kv0.value = s - sm->per_thread_data[cpu_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->sessions;
+  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
   
   if (clib_bihash_add_del_8_8 (&sm->out2in, &kv0, 1 /* is_add */))
       clib_warning ("out2in key add failed");
 
+  /* Add to translated packets worker lookup */
+  worker_by_out_key.addr = s->out2in.addr;
+  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;
+  clib_bihash_add_del_8_8 (&sm->worker_by_out, &kv0, 1);
   return next0;
 }
                       
@@ -275,7 +313,8 @@ static inline u32 icmp_in2out_slow_path (snat_main_t *sm,
                                          u32 rx_fib_index0,
                                          vlib_node_runtime_t * node,
                                          u32 next0,
-                                         f64 now)
+                                         f64 now,
+                                         u32 cpu_index)
 {
   snat_session_key_t key0;
   icmp_echo_header_t *echo0;
@@ -320,13 +359,14 @@ static inline u32 icmp_in2out_slow_path (snat_main_t *sm,
         return next0;
       
       next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                         &s0, node, next0);
+                         &s0, node, next0, cpu_index);
       
       if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
         return next0;
     }
   else
-    s0 = pool_elt_at_index (sm->sessions, value0.value);
+    s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                            value0.value);
 
   old_addr0 = ip0->src_address.as_u32;
   ip0->src_address = s0->out2in.addr;
@@ -355,8 +395,10 @@ 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->list_pool, s0->per_user_index);
-      clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+      clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                         s0->per_user_index);
+      clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                          s0->per_user_list_head_index,
                           s0->per_user_index);
     }
 
@@ -375,6 +417,7 @@ 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 ();
 
   stats_node_index = is_slow_path ? snat_in2out_slowpath_node.index :
     snat_in2out_node.index;
@@ -445,14 +488,6 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
 
           next0 = next1 = SNAT_IN2OUT_NEXT_LOOKUP;
 
-#if 0
-          /* Formally correct, but we send to slowpath, lookup or drop */
-	  vnet_get_config_data (&cm->config_main,
-                                &b0->current_config_index,
-                                &next0,
-                                0 /* sizeof config data */);
-#endif
-
           proto0 = ~0;
           proto0 = (ip0->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto0;
@@ -471,7 +506,7 @@ 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);
+                     node, next0, now, cpu_index);
                   goto trace00;
                 }
             }
@@ -512,7 +547,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace00;
                   
                   next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                                     &s0, node, next0);
+                                     &s0, node, next0, cpu_index);
                   if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace00;
                 }
@@ -523,7 +558,8 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->sessions, value0.value);
+            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value0.value);
 
           old_addr0 = ip0->src_address.as_u32;
           ip0->src_address = s0->out2in.addr;
@@ -565,8 +601,10 @@ 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->list_pool, s0->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s0->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
         trace00:
@@ -581,7 +619,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next0;
                   t->session_index = ~0;
               if (s0)
-                  t->session_index = s0 - sm->sessions;
+                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_IN2OUT_NEXT_DROP;
@@ -595,13 +633,6 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
 	  rx_fib_index1 = vec_elt (sm->ip4_main->fib_index_by_sw_if_index, 
                                    sw_if_index1);
 
-#if 0
-	  vnet_get_config_data (&cm->config_main,
-                                &b1->current_config_index,
-                                &next1,
-                                0 /* sizeof config data */);
-#endif
-
           proto1 = ~0;
           proto1 = (ip1->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto1;
@@ -619,8 +650,8 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               if (PREDICT_FALSE (proto1 == SNAT_PROTOCOL_ICMP))
                 {
                   next1 = icmp_in2out_slow_path 
-                    (sm, b1, ip1, icmp1, sw_if_index1, rx_fib_index1, node, next1,
-                     now);
+                    (sm, b1, ip1, icmp1, sw_if_index1, rx_fib_index1, node,
+                     next1, now, cpu_index);
                   goto trace01;
                 }
             }
@@ -661,7 +692,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace01;
                   
                   next1 = slow_path (sm, b1, ip1, rx_fib_index1, &key1,
-                                     &s1, node, next1);
+                                     &s1, node, next1, cpu_index);
                   if (PREDICT_FALSE (next1 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace01;
                 }
@@ -672,7 +703,8 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s1 = pool_elt_at_index (sm->sessions, value1.value);
+            s1 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value1.value);
 
           old_addr1 = ip1->src_address.as_u32;
           ip1->src_address = s1->out2in.addr;
@@ -714,8 +746,10 @@ 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->list_pool, s1->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s1->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s1->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s1->per_user_list_head_index,
                                   s1->per_user_index);
             }
         trace01:
@@ -729,7 +763,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next1;
               t->session_index = ~0;
               if (s1)
-                t->session_index = s1 - sm->sessions;
+                t->session_index = s1 - sm->per_thread_data[cpu_index].sessions;
             }
 
           pkts_processed += next1 != SNAT_IN2OUT_NEXT_DROP;
@@ -779,14 +813,6 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
 	  rx_fib_index0 = vec_elt (sm->ip4_main->fib_index_by_sw_if_index, 
                                    sw_if_index0);
 
-
-#if 0
-	  vnet_get_config_data (&cm->config_main,
-                                &b0->current_config_index,
-                                &next0,
-                                0 /* sizeof config data */);
-#endif
-
           proto0 = ~0;
           proto0 = (ip0->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto0;
@@ -804,8 +830,8 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               if (PREDICT_FALSE (proto0 == SNAT_PROTOCOL_ICMP))
                 {
                   next0 = icmp_in2out_slow_path 
-                    (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node, next0,
-                     now);
+                    (sm, b0, ip0, icmp0, sw_if_index0, rx_fib_index0, node,
+                     next0, now, cpu_index);
                   goto trace0;
                 }
             }
@@ -846,7 +872,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                     goto trace0;
                   
                   next0 = slow_path (sm, b0, ip0, rx_fib_index0, &key0,
-                                     &s0, node, next0);
+                                     &s0, node, next0, cpu_index);
                   if (PREDICT_FALSE (next0 == SNAT_IN2OUT_NEXT_DROP))
                     goto trace0;
                 }
@@ -857,7 +883,8 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
                 }
             }
           else
-            s0 = pool_elt_at_index (sm->sessions, value0.value);
+            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value0.value);
 
           old_addr0 = ip0->src_address.as_u32;
           ip0->src_address = s0->out2in.addr;
@@ -899,8 +926,10 @@ 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->list_pool, s0->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s0->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
 
@@ -915,7 +944,7 @@ snat_in2out_node_fn_inline (vlib_main_t * vm,
               t->next_index = next0;
                   t->session_index = ~0;
               if (s0)
-                  t->session_index = s0 - sm->sessions;
+                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_IN2OUT_NEXT_DROP;
@@ -999,6 +1028,183 @@ VLIB_REGISTER_NODE (snat_in2out_slowpath_node) = {
 
 VLIB_NODE_FUNCTION_MULTIARCH (snat_in2out_slowpath_node, snat_in2out_slow_path_fn);
 
+static uword
+snat_in2out_worker_handoff_fn (vlib_main_t * vm,
+                               vlib_node_runtime_t * node,
+                               vlib_frame_t * frame)
+{
+  snat_main_t *sm = &snat_main;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
+  u32 n_left_from, *from, *to_next = 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_t *f = 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 cpu_index = os_get_cpu_number ();
+
+  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,
+			       sm->first_worker_index + sm->num_workers - 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;
+      u32 sw_if_index0;
+      u32 rx_fib_index0;
+      ip4_header_t * ip0;
+      snat_user_key_t key0;
+      clib_bihash_kv_8_8_t kv0, value0;
+      u8 do_handoff;
+
+      bi0 = from[0];
+      from += 1;
+      n_left_from -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
+      rx_fib_index0 = ip4_fib_table_get_index_for_sw_if_index(sw_if_index0);
+
+      ip0 = vlib_buffer_get_current (b0);
+
+      key0.addr = ip0->src_address;
+      key0.fib_index = rx_fib_index0;
+
+      kv0.key = key0.as_u64;
+
+      /* Ever heard of of the "user" before? */
+      if (clib_bihash_search_8_8 (&sm->worker_by_in, &kv0, &value0))
+        {
+          /* No, assign next available worker (RR) */
+          next_worker_index = sm->first_worker_index +
+            sm->workers[sm->next_worker++ % vec_len (sm->workers)];
+
+          /* add non-traslated packets worker lookup */
+          kv0.value = next_worker_index;
+          clib_bihash_add_del_8_8 (&sm->worker_by_in, &kv0, 1);
+        }
+      else
+        next_worker_index = value0.value;
+
+      if (PREDICT_FALSE (next_worker_index != cpu_index))
+        {
+          do_handoff = 1;
+
+          if (next_worker_index != current_worker_index)
+            {
+              if (hf)
+                hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+              hf = vlib_get_worker_handoff_queue_elt (sm->fq_in2out_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, snat_in2out_node.index);
+              to_next = vlib_frame_vector_args (f);
+            }
+
+          to_next[0] = bi0;
+          to_next += 1;
+          f->n_vectors++;
+        }
+
+      if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)
+			 && (b0->flags & VLIB_BUFFER_IS_TRACED)))
+	{
+          snat_in2out_worker_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, snat_in2out_node.index, f);
+
+  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;
+}
+
+VLIB_REGISTER_NODE (snat_in2out_worker_handoff_node) = {
+  .function = snat_in2out_worker_handoff_fn,
+  .name = "snat-in2out-worker-handoff",
+  .vector_size = sizeof (u32),
+  .format_trace = format_snat_in2out_worker_handoff_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+  
+  .n_next_nodes = 1,
+
+  .next_nodes = {
+    [0] = "error-drop",
+  },
+};
+
+VLIB_NODE_FUNCTION_MULTIARCH (snat_in2out_worker_handoff_node, snat_in2out_worker_handoff_fn);
+
 static inline u32 icmp_in2out_static_map (snat_main_t *sm,
                                           vlib_buffer_t * b0,
                                           ip4_header_t * ip0,
diff --git a/plugins/snat-plugin/snat/out2in.c b/plugins/snat-plugin/snat/out2in.c
index 35edcc5f7b3..a4641d3a71e 100644
--- a/plugins/snat-plugin/snat/out2in.c
+++ b/plugins/snat-plugin/snat/out2in.c
@@ -16,6 +16,7 @@
 #include <vlib/vlib.h>
 #include <vnet/vnet.h>
 #include <vnet/pg/pg.h>
+#include <vnet/handoff.h>
 
 #include <vnet/ip/ip.h>
 #include <vnet/ethernet/ethernet.h>
@@ -32,6 +33,11 @@ typedef struct {
   u32 session_index;
 } snat_out2in_trace_t;
 
+typedef struct {
+  u32 next_worker_index;
+  u8 do_handoff;
+} snat_out2in_worker_handoff_trace_t;
+
 /* packet trace format function */
 static u8 * format_snat_out2in_trace (u8 * s, va_list * args)
 {
@@ -55,9 +61,23 @@ static u8 * format_snat_out2in_fast_trace (u8 * s, va_list * args)
   return s;
 }
 
+static u8 * format_snat_out2in_worker_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 *);
+  snat_out2in_worker_handoff_trace_t * t =
+    va_arg (*args, snat_out2in_worker_handoff_trace_t *);
+  char * m;
+
+  m = t->do_handoff ? "next worker" : "same worker";
+  s = format (s, "SNAT_OUT2IN_WORKER_HANDOFF: %s %d", m, t->next_worker_index);
+
+  return s;
+}
 
 vlib_node_registration_t snat_out2in_node;
 vlib_node_registration_t snat_out2in_fast_node;
+vlib_node_registration_t snat_out2in_worker_handoff_node;
 
 #define foreach_snat_out2in_error                       \
 _(UNSUPPORTED_PROTOCOL, "Unsupported protocol")         \
@@ -80,6 +100,7 @@ static char * snat_out2in_error_strings[] = {
 
 typedef enum {
   SNAT_OUT2IN_NEXT_DROP,
+  SNAT_OUT2IN_NEXT_LOOKUP,
   SNAT_OUT2IN_N_NEXT,
 } snat_out2in_next_t;
 
@@ -102,7 +123,8 @@ create_session_for_static_mapping (snat_main_t *sm,
                                    vlib_buffer_t *b0,
                                    snat_session_key_t in2out,
                                    snat_session_key_t out2in,
-                                   vlib_node_runtime_t * node)
+                                   vlib_node_runtime_t * node,
+                                   u32 cpu_index)
 {
   snat_user_t *u;
   snat_user_key_t user_key;
@@ -119,28 +141,35 @@ 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->users, u);
+      pool_get (sm->per_thread_data[cpu_index].users, u);
       memset (u, 0, sizeof (*u));
       u->addr = in2out.addr;
 
-      pool_get (sm->list_pool, per_user_list_head_elt);
+      pool_get (sm->per_thread_data[cpu_index].list_pool,
+                per_user_list_head_elt);
 
       u->sessions_per_user_list_head_index = per_user_list_head_elt -
-        sm->list_pool;
+        sm->per_thread_data[cpu_index].list_pool;
 
-      clib_dlist_init (sm->list_pool, u->sessions_per_user_list_head_index);
+      clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+                       u->sessions_per_user_list_head_index);
 
-      kv0.value = u - sm->users;
+      kv0.value = u - sm->per_thread_data[cpu_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;
+      clib_bihash_add_del_8_8 (&sm->worker_by_in, &kv0, 1);
     }
   else
     {
-      u = pool_elt_at_index (sm->users, value0.value);
+      u = pool_elt_at_index (sm->per_thread_data[cpu_index].users,
+                             value0.value);
     }
 
-  pool_get (sm->sessions, s);
+  pool_get (sm->per_thread_data[cpu_index].sessions, s);
   memset (s, 0, sizeof (*s));
 
   s->outside_address_index = ~0;
@@ -148,16 +177,22 @@ create_session_for_static_mapping (snat_main_t *sm,
   u->nstaticsessions++;
 
   /* Create list elts */
-  pool_get (sm->list_pool, per_user_translation_list_elt);
-  clib_dlist_init (sm->list_pool, per_user_translation_list_elt -
-                   sm->list_pool);
-
-  per_user_translation_list_elt->value = s - sm->sessions;
-  s->per_user_index = per_user_translation_list_elt - sm->list_pool;
+  pool_get (sm->per_thread_data[cpu_index].list_pool,
+            per_user_translation_list_elt);
+  clib_dlist_init (sm->per_thread_data[cpu_index].list_pool,
+                   per_user_translation_list_elt -
+                   sm->per_thread_data[cpu_index].list_pool);
+
+  per_user_translation_list_elt->value =
+    s - sm->per_thread_data[cpu_index].sessions;
+  s->per_user_index =
+    per_user_translation_list_elt - sm->per_thread_data[cpu_index].list_pool;
   s->per_user_list_head_index = u->sessions_per_user_list_head_index;
 
-  clib_dlist_addtail (sm->list_pool, s->per_user_list_head_index,
-                      per_user_translation_list_elt - sm->list_pool);
+  clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                      s->per_user_list_head_index,
+                      per_user_translation_list_elt -
+                      sm->per_thread_data[cpu_index].list_pool);
 
   s->in2out = in2out;
   s->out2in = out2in;
@@ -165,12 +200,12 @@ create_session_for_static_mapping (snat_main_t *sm,
 
   /* Add to translation hashes */
   kv0.key = s->in2out.as_u64;
-  kv0.value = s - sm->sessions;
+  kv0.value = s - sm->per_thread_data[cpu_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->sessions;
+  kv0.value = s - sm->per_thread_data[cpu_index].sessions;
 
   if (clib_bihash_add_del_8_8 (&sm->out2in, &kv0, 1 /* is_add */))
       clib_warning ("out2in key add failed");
@@ -185,7 +220,8 @@ static inline u32 icmp_out2in_slow_path (snat_main_t *sm,
                                          u32 sw_if_index0,
                                          u32 rx_fib_index0,
                                          vlib_node_runtime_t * node,
-                                         u32 next0, f64 now)
+                                         u32 next0, f64 now,
+                                         u32 cpu_index)
 {
   snat_session_key_t key0, sm0;
   icmp_echo_header_t *echo0;
@@ -233,12 +269,13 @@ static inline u32 icmp_out2in_slow_path (snat_main_t *sm,
 
       /* Create session initiated by host from external network */
       s0 = create_session_for_static_mapping(sm, b0, sm0, key0,
-                                             node);
+                                             node, cpu_index);
       if (!s0)
         return SNAT_OUT2IN_NEXT_DROP;
     }
   else
-    s0 = pool_elt_at_index (sm->sessions, value0.value);
+    s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                            value0.value);
 
   old_addr0 = ip0->dst_address.as_u32;
   ip0->dst_address = s0->in2out.addr;
@@ -267,8 +304,10 @@ 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->list_pool, s0->per_user_index);
-      clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+      clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                         s0->per_user_index);
+      clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                          s0->per_user_list_head_index,
                           s0->per_user_index);
     }
 
@@ -285,6 +324,7 @@ 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 ();
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -301,8 +341,8 @@ snat_out2in_node_fn (vlib_main_t * vm,
 	{
           u32 bi0, bi1;
 	  vlib_buffer_t * b0, * b1;
-          u32 next0 = SNAT_OUT2IN_NEXT_DROP;
-          u32 next1 = SNAT_OUT2IN_NEXT_DROP;
+          u32 next0 = SNAT_OUT2IN_NEXT_LOOKUP;
+          u32 next1 = SNAT_OUT2IN_NEXT_LOOKUP;
           u32 sw_if_index0, sw_if_index1;
           ip4_header_t * ip0, *ip1;
           ip_csum_t sum0, sum1;
@@ -353,7 +393,6 @@ snat_out2in_node_fn (vlib_main_t * vm,
 	  rx_fib_index0 = vec_elt (sm->ip4_main->fib_index_by_sw_if_index, 
                                    sw_if_index0);
 
-	  vnet_feature_next (sw_if_index0, &next0, b0);
           proto0 = ~0;
           proto0 = (ip0->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto0;
@@ -369,7 +408,7 @@ 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);
+                 next0, now, cpu_index);
               goto trace0;
             }
 
@@ -391,12 +430,14 @@ 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);
+              s0 = create_session_for_static_mapping(sm, b0, sm0, key0, node,
+                                                     cpu_index);
               if (!s0)
                 goto trace0;
             }
           else
-            s0 = pool_elt_at_index (sm->sessions, value0.value);
+            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value0.value);
 
           old_addr0 = ip0->dst_address.as_u32;
           ip0->dst_address = s0->in2out.addr;
@@ -439,8 +480,10 @@ 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->list_pool, s0->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s0->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
         trace0:
@@ -454,7 +497,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next0;
               t->session_index = ~0;
               if (s0)
-                  t->session_index = s0 - sm->sessions;
+                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_OUT2IN_NEXT_DROP;
@@ -469,8 +512,6 @@ snat_out2in_node_fn (vlib_main_t * vm,
 	  rx_fib_index1 = vec_elt (sm->ip4_main->fib_index_by_sw_if_index, 
                                    sw_if_index1);
 
-	  vnet_feature_next (sw_if_index1, &next1, b1);
-
           proto1 = ~0;
           proto1 = (ip1->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto1;
@@ -486,7 +527,7 @@ 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);
+                 next1, now, cpu_index);
               goto trace1;
             }
 
@@ -508,12 +549,14 @@ 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);
+              s1 = create_session_for_static_mapping(sm, b1, sm1, key1, node,
+                                                     cpu_index);
               if (!s1)
                 goto trace1;
             }
           else
-            s1 = pool_elt_at_index (sm->sessions, value1.value);
+            s1 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value1.value);
 
           old_addr1 = ip1->dst_address.as_u32;
           ip1->dst_address = s1->in2out.addr;
@@ -556,8 +599,10 @@ 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->list_pool, s1->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s1->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s1->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s1->per_user_list_head_index,
                                   s1->per_user_index);
             }
         trace1:
@@ -571,10 +616,9 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next1;
               t->session_index = ~0;
               if (s1)
-                  t->session_index = s1 - sm->sessions;
+                t->session_index = s1 - sm->per_thread_data[cpu_index].sessions;
             }
 
-          pkts_processed += next0 != SNAT_OUT2IN_NEXT_DROP;
           pkts_processed += next1 != SNAT_OUT2IN_NEXT_DROP;
 
           /* verify speculative enqueues, maybe switch current next frame */
@@ -587,7 +631,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
 	{
           u32 bi0;
 	  vlib_buffer_t * b0;
-          u32 next0 = SNAT_OUT2IN_NEXT_DROP;
+          u32 next0 = SNAT_OUT2IN_NEXT_LOOKUP;
           u32 sw_if_index0;
           ip4_header_t * ip0;
           ip_csum_t sum0;
@@ -621,8 +665,6 @@ snat_out2in_node_fn (vlib_main_t * vm,
 	  rx_fib_index0 = vec_elt (sm->ip4_main->fib_index_by_sw_if_index, 
                                    sw_if_index0);
 
-	  vnet_feature_next (sw_if_index0, &next0, b0);
-
           proto0 = ~0;
           proto0 = (ip0->protocol == IP_PROTOCOL_UDP) 
             ? SNAT_PROTOCOL_UDP : proto0;
@@ -638,7 +680,7 @@ 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);
+                 next0, now, cpu_index);
               goto trace00;
             }
 
@@ -660,12 +702,14 @@ 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);
+              s0 = create_session_for_static_mapping(sm, b0, sm0, key0, node,
+                                                     cpu_index);
               if (!s0)
                 goto trace00;
             }
           else
-            s0 = pool_elt_at_index (sm->sessions, value0.value);
+            s0 = pool_elt_at_index (sm->per_thread_data[cpu_index].sessions,
+                                    value0.value);
 
           old_addr0 = ip0->dst_address.as_u32;
           ip0->dst_address = s0->in2out.addr;
@@ -708,8 +752,10 @@ 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->list_pool, s0->per_user_index);
-              clib_dlist_addtail (sm->list_pool, s0->per_user_list_head_index,
+              clib_dlist_remove (sm->per_thread_data[cpu_index].list_pool,
+                                 s0->per_user_index);
+              clib_dlist_addtail (sm->per_thread_data[cpu_index].list_pool,
+                                  s0->per_user_list_head_index,
                                   s0->per_user_index);
             }
         trace00:
@@ -723,7 +769,7 @@ snat_out2in_node_fn (vlib_main_t * vm,
               t->next_index = next0;
               t->session_index = ~0;
               if (s0)
-                  t->session_index = s0 - sm->sessions;
+                t->session_index = s0 - sm->per_thread_data[cpu_index].sessions;
             }
 
           pkts_processed += next0 != SNAT_OUT2IN_NEXT_DROP;
@@ -760,10 +806,189 @@ VLIB_REGISTER_NODE (snat_out2in_node) = {
   /* edit / add dispositions here */
   .next_nodes = {
     [SNAT_OUT2IN_NEXT_DROP] = "error-drop",
+    [SNAT_OUT2IN_NEXT_LOOKUP] = "ip4-lookup",
   },
 };
 VLIB_NODE_FUNCTION_MULTIARCH (snat_out2in_node, snat_out2in_node_fn);
 
+static uword
+snat_out2in_worker_handoff_fn (vlib_main_t * vm,
+                               vlib_node_runtime_t * node,
+                               vlib_frame_t * frame)
+{
+  snat_main_t *sm = &snat_main;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
+  u32 n_left_from, *from, *to_next = 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_t *f = 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 cpu_index = os_get_cpu_number ();
+
+  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,
+			       sm->first_worker_index + sm->num_workers - 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;
+      u32 sw_if_index0;
+      u32 rx_fib_index0;
+      ip4_header_t * ip0;
+      udp_header_t * udp0;
+      snat_static_mapping_key_t key0;
+      clib_bihash_kv_8_8_t kv0, value0;
+      u8 do_handoff;
+
+      bi0 = from[0];
+      from += 1;
+      n_left_from -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
+      rx_fib_index0 = ip4_fib_table_get_index_for_sw_if_index(sw_if_index0);
+
+      ip0 = vlib_buffer_get_current (b0);
+      udp0 = ip4_next_header (ip0);
+
+      key0.addr = ip0->dst_address;
+      key0.port = udp0->dst_port;
+      key0.fib_index = rx_fib_index0;
+
+      kv0.key = key0.as_u64;
+
+      /* Ever heard of of the "user" before? */
+      if (clib_bihash_search_8_8 (&sm->worker_by_out, &kv0, &value0))
+        {
+          /* No, assign next available worker (RR) */
+          next_worker_index = sm->first_worker_index +
+            sm->workers[sm->next_worker++ % vec_len (sm->workers)];
+
+          /* Add to translated packets worker lookup */
+          kv0.value = next_worker_index;
+          clib_bihash_add_del_8_8 (&sm->worker_by_out, &kv0, 1);
+        }
+      else
+        next_worker_index = value0.value;
+
+      if (PREDICT_FALSE (next_worker_index != cpu_index))
+        {
+          if (next_worker_index != current_worker_index)
+            {
+              if (hf)
+                hf->n_vectors = VLIB_FRAME_SIZE - n_left_to_next_worker;
+
+              hf = vlib_get_worker_handoff_queue_elt (sm->fq_out2in_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, snat_out2in_node.index);
+              to_next = vlib_frame_vector_args (f);
+            }
+
+          to_next[0] = bi0;
+          to_next += 1;
+          f->n_vectors++;
+        }
+
+      if (PREDICT_FALSE ((node->flags & VLIB_NODE_FLAG_TRACE)
+			 && (b0->flags & VLIB_BUFFER_IS_TRACED)))
+	{
+          snat_out2in_worker_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, snat_out2in_node.index, f);
+
+  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;
+}
+
+VLIB_REGISTER_NODE (snat_out2in_worker_handoff_node) = {
+  .function = snat_out2in_worker_handoff_fn,
+  .name = "snat-out2in-worker-handoff",
+  .vector_size = sizeof (u32),
+  .format_trace = format_snat_out2in_worker_handoff_trace,
+  .type = VLIB_NODE_TYPE_INTERNAL,
+  
+  .n_next_nodes = 1,
+
+  .next_nodes = {
+    [0] = "error-drop",
+  },
+};
+
+VLIB_NODE_FUNCTION_MULTIARCH (snat_out2in_worker_handoff_node, snat_out2in_worker_handoff_fn);
+
 static inline u32 icmp_out2in_fast (snat_main_t *sm,
                                     vlib_buffer_t * b0,
                                     ip4_header_t * ip0,
@@ -1014,6 +1239,7 @@ VLIB_REGISTER_NODE (snat_out2in_fast_node) = {
 
   /* edit / add dispositions here */
   .next_nodes = {
+    [SNAT_OUT2IN_NEXT_LOOKUP] = "ip4-lookup",
     [SNAT_OUT2IN_NEXT_DROP] = "error-drop",
   },
 };
diff --git a/plugins/snat-plugin/snat/snat.api b/plugins/snat-plugin/snat/snat.api
index d7d41f2000d..a191eed5944 100644
--- a/plugins/snat-plugin/snat/snat.api
+++ b/plugins/snat-plugin/snat/snat.api
@@ -238,3 +238,46 @@ define snat_show_config_reply
   u32 outside_vrf_id;
   u32 inside_vrf_id;
 };
+
+/** \brief Set S-NAT workers
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param worker_mask - S-NAT workers mask
+*/
+define snat_set_workers {
+  u32 client_index;
+  u32 context;
+  u64 worker_mask;
+};
+
+/** \brief Set S-NAT workers reply
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+    @param retval - return code
+*/
+define snat_set_workers_reply {
+  u32 context;
+  i32 retval;
+};
+
+/** \brief Dump S-NAT workers
+    @param client_index - opaque cookie to identify the sender
+    @param context - sender context, to match reply w/ request
+*/
+define snat_worker_dump {
+  u32 client_index;
+  u32 context;
+};
+
+/** \brief S-NAT workers details response
+    @param context - sender context, to match reply w/ request
+    @param worker_index - worker index
+    @param lcore_id - lcore ID
+    @param name - worker name
+*/
+define snat_worker_details {
+  u32 context;
+  u32 worker_index;
+  u32 lcore_id;
+  u8 name[64];
+};
diff --git a/plugins/snat-plugin/snat/snat.c b/plugins/snat-plugin/snat/snat.c
index 2956e24f054..70df44ea1d4 100644
--- a/plugins/snat-plugin/snat/snat.c
+++ b/plugins/snat-plugin/snat/snat.c
@@ -103,6 +103,16 @@ VNET_FEATURE_INIT (ip4_snat_out2in, static) = {
   .node_name = "snat-out2in",
   .runs_before = VNET_FEATURES ("ip4-lookup"),
 };
+VNET_FEATURE_INIT (ip4_snat_in2out_worker_handoff, static) = {
+  .arc_name = "ip4-unicast",
+  .node_name = "snat-in2out-worker-handoff",
+  .runs_before = VNET_FEATURES ("snat-out2in-worker-handoff"),
+};
+VNET_FEATURE_INIT (ip4_snat_out2in_worker_handoff, static) = {
+  .arc_name = "ip4-unicast",
+  .node_name = "snat-out2in-worker-handoff",
+  .runs_before = VNET_FEATURES ("ip4-lookup"),
+};
 VNET_FEATURE_INIT (ip4_snat_in2out_fast, static) = {
   .arc_name = "ip4-unicast",
   .node_name = "snat-in2out-fast",
@@ -229,13 +239,13 @@ static int is_snat_address_used_in_static_mapping (snat_main_t *sm,
 
 int snat_del_address (snat_main_t *sm, ip4_address_t addr)
 {
-  clib_warning("%U", format_ip4_address, &addr);
   snat_address_t *a = 0;
   snat_session_t *ses;
   u32 *ses_to_be_removed = 0, *ses_index;
   clib_bihash_kv_8_8_t kv, value;
   snat_user_key_t user_key;
   snat_user_t *u;
+  snat_main_per_thread_data_t *tsm;
 
   int i;
 
@@ -261,30 +271,33 @@ int snat_del_address (snat_main_t *sm, ip4_address_t addr)
   /* Delete sessions using address */
   if (a->busy_ports)
     {
-      pool_foreach (ses, sm->sessions, ({
-        if (ses->out2in.addr.as_u32 == addr.as_u32)
-          {
-            vec_add1 (ses_to_be_removed, ses - sm->sessions);
-            kv.key = ses->in2out.as_u64;
-            clib_bihash_add_del_8_8 (&sm->in2out, &kv, 0);
-            kv.key = ses->out2in.as_u64;
-            clib_bihash_add_del_8_8 (&sm->out2in, &kv, 0);
-            clib_dlist_remove (sm->list_pool, ses->per_user_index);
-            user_key.addr = ses->in2out.addr;
-            user_key.fib_index = ses->in2out.fib_index;
-            kv.key = user_key.as_u64;
-            if (!clib_bihash_search_8_8 (&sm->user_hash, &kv, &value))
+      vec_foreach (tsm, sm->per_thread_data)
+        {
+          pool_foreach (ses, tsm->sessions, ({
+            if (ses->out2in.addr.as_u32 == addr.as_u32)
               {
-                u = pool_elt_at_index (sm->users, value.value);
-                u->nsessions--;
+                vec_add1 (ses_to_be_removed, ses - tsm->sessions);
+                kv.key = ses->in2out.as_u64;
+                clib_bihash_add_del_8_8 (&sm->in2out, &kv, 0);
+                kv.key = ses->out2in.as_u64;
+                clib_bihash_add_del_8_8 (&sm->out2in, &kv, 0);
+                clib_dlist_remove (tsm->list_pool, ses->per_user_index);
+                user_key.addr = ses->in2out.addr;
+                user_key.fib_index = ses->in2out.fib_index;
+                kv.key = user_key.as_u64;
+                if (!clib_bihash_search_8_8 (&sm->user_hash, &kv, &value))
+                  {
+                    u = pool_elt_at_index (tsm->users, value.value);
+                    u->nsessions--;
+                  }
               }
-          }
-      }));
+          }));
 
-      vec_foreach (ses_index, ses_to_be_removed)
-        pool_put_index (sm->sessions, ses_index[0]);
+          vec_foreach (ses_index, ses_to_be_removed)
+            pool_put_index (tsm->sessions, ses_index[0]);
 
-      vec_free (ses_to_be_removed);
+          vec_free (ses_to_be_removed);
+       }
     }
 
   vec_del1 (sm->addresses, i);
@@ -469,26 +482,35 @@ int snat_add_static_mapping(ip4_address_t l_addr, ip4_address_t e_addr,
           snat_user_key_t u_key;
           snat_user_t *u;
           dlist_elt_t * head, * elt;
-          u32 elt_index, head_index;
+          u32 elt_index, head_index, del_elt_index;
           u32 ses_index;
+          u64 user_index;
           snat_session_t * s;
+          snat_main_per_thread_data_t *tsm;
 
           u_key.addr = m->local_addr;
           u_key.fib_index = m->fib_index;
           kv.key = u_key.as_u64;
           if (!clib_bihash_search_8_8 (&sm->user_hash, &kv, &value))
             {
-              u = pool_elt_at_index (sm->users, value.value);
+              user_index = value.value;
+              clib_bihash_search_8_8 (&sm->worker_by_in, &kv, &value);
+              tsm = vec_elt_at_index (sm->per_thread_data, value.value);
+              u = pool_elt_at_index (tsm->users, user_index);
               if (u->nstaticsessions)
                 {
                   head_index = u->sessions_per_user_list_head_index;
-                  head = pool_elt_at_index (sm->list_pool, head_index);
+                  head = pool_elt_at_index (tsm->list_pool, head_index);
                   elt_index = head->next;
-                  elt = pool_elt_at_index (sm->list_pool, elt_index);
+                  elt = pool_elt_at_index (tsm->list_pool, elt_index);
                   ses_index = elt->value;
                   while (ses_index != ~0)
                     {
-                      s =  pool_elt_at_index (sm->sessions, ses_index);
+                      s =  pool_elt_at_index (tsm->sessions, ses_index);
+                      del_elt_index = elt_index;
+                      elt_index = elt->next;
+                      elt = pool_elt_at_index (tsm->list_pool, elt_index);
+                      ses_index = elt->value;
 
                       if (!addr_only)
                         {
@@ -496,35 +518,25 @@ int snat_add_static_mapping(ip4_address_t l_addr, ip4_address_t e_addr,
                               (clib_net_to_host_u16 (s->out2in.port) != e_port))
                             continue;
                         }
+
                       value.key = s->in2out.as_u64;
                       clib_bihash_add_del_8_8 (&sm->in2out, &value, 0);
                       value.key = s->out2in.as_u64;
                       clib_bihash_add_del_8_8 (&sm->out2in, &value, 0);
-                      pool_put (sm->sessions, s);
+                      pool_put (tsm->sessions, s);
+
+                      clib_dlist_remove (tsm->list_pool, del_elt_index);
+                      pool_put_index (tsm->list_pool, del_elt_index);
+                      u->nstaticsessions--;
 
                       if (!addr_only)
                         break;
-
-                      elt_index = elt->next;
-                      elt = pool_elt_at_index (sm->list_pool, elt_index);
-                      ses_index = elt->value;
                     }
                   if (addr_only)
                     {
-                      while ((elt_index = clib_dlist_remove_head(sm->list_pool, head_index)) != ~0)
-                        pool_put_index (sm->list_pool, elt_index);
-                      pool_put (sm->users, u);
+                      pool_put (tsm->users, u);
                       clib_bihash_add_del_8_8 (&sm->user_hash, &kv, 0);
                     }
-                  else
-                    {
-                      if (ses_index != ~0)
-                        {
-                          clib_dlist_remove (sm->list_pool, elt_index);
-                          pool_put (sm->list_pool, elt);
-                          u->nstaticsessions--;
-                        }
-                    }
                 }
             }
         }
@@ -545,11 +557,22 @@ static int snat_interface_add_del (u32 sw_if_index, u8 is_inside, int is_del)
   if (sm->static_mapping_only && !(sm->static_mapping_connection_tracking))
     feature_name = is_inside ?  "snat-in2out-fast" : "snat-out2in-fast";
   else
-    feature_name = is_inside ?  "snat-in2out" : "snat-out2in";
+    {
+      if (sm->num_workers > 1)
+        feature_name = is_inside ?  "snat-in2out-worker-handoff" : "snat-out2in-worker-handoff";
+      else
+        feature_name = is_inside ?  "snat-in2out" : "snat-out2in";
+    }
 
   vnet_feature_enable_disable ("ip4-unicast", feature_name, sw_if_index,
 			       !is_del, 0, 0);
 
+  if (sm->fq_in2out_index == ~0)
+    sm->fq_in2out_index = vlib_frame_queue_main_init (snat_in2out_node.index, 0);
+
+  if (sm->fq_out2in_index == ~0)
+    sm->fq_out2in_index = vlib_frame_queue_main_init (snat_out2in_node.index, 0);
+
   pool_foreach (i, sm->interfaces,
   ({
     if (i->sw_if_index == sw_if_index)
@@ -573,6 +596,26 @@ static int snat_interface_add_del (u32 sw_if_index, u8 is_inside, int is_del)
   return 0;
 }
 
+static int snat_set_workers (uword * bitmap)
+{
+  snat_main_t *sm = &snat_main;
+  int i;
+
+  if (sm->num_workers < 2)
+    return VNET_API_ERROR_FEATURE_DISABLED;
+
+  if (clib_bitmap_last_set (bitmap) >= sm->num_workers)
+    return VNET_API_ERROR_INVALID_WORKER;
+
+  vec_free (sm->workers);
+  clib_bitmap_foreach (i, bitmap,
+    ({
+      vec_add1(sm->workers, i);
+    }));
+
+  return 0;
+}
+
 static void 
 vl_api_snat_add_address_range_t_handler
 (vl_api_snat_add_address_range_t * mp)
@@ -925,6 +968,101 @@ static void *vl_api_snat_show_config_t_print
   FINISH;
 }
 
+static void 
+vl_api_snat_set_workers_t_handler
+(vl_api_snat_set_workers_t * mp)
+{
+  snat_main_t * sm = &snat_main;
+  vl_api_snat_set_workers_reply_t * rmp;
+  int rv = 0;
+  uword *bitmap = 0;
+  u64 mask = clib_net_to_host_u64 (mp->worker_mask);
+
+  if (sm->num_workers < 2)
+    {
+      rv = VNET_API_ERROR_FEATURE_DISABLED;
+      goto send_reply;
+    }
+
+  bitmap = clib_bitmap_set_multiple (bitmap, 0, mask, BITS (mask)); 
+  rv = snat_set_workers(bitmap);
+  clib_bitmap_free (bitmap);
+
+ send_reply:
+  REPLY_MACRO (VL_API_SNAT_SET_WORKERS_REPLY);
+}
+
+static void *vl_api_snat_set_workers_t_print
+(vl_api_snat_set_workers_t *mp, void * handle)
+{
+  u8 * s;
+  uword *bitmap = 0;
+  u8 first = 1;
+  int i;
+  u64 mask = clib_net_to_host_u64 (mp->worker_mask);
+
+  s = format (0, "SCRIPT: snat_set_workers ");
+  bitmap = clib_bitmap_set_multiple (bitmap, 0, mask, BITS (mask)); 
+  clib_bitmap_foreach (i, bitmap,
+    ({
+      if (first)
+        s = format (s, "%d", i);
+      else
+        s = format (s, ",%d", i);
+      first = 0;
+    }));
+  clib_bitmap_free (bitmap);
+  FINISH;
+}
+
+static void
+send_snat_worker_details
+(u32 worker_index, unix_shared_memory_queue_t * q, u32 context)
+{
+  vl_api_snat_worker_details_t *rmp;
+  snat_main_t * sm = &snat_main;
+  vlib_worker_thread_t *w =
+    vlib_worker_threads + worker_index + sm->first_worker_index;
+
+  rmp = vl_msg_api_alloc (sizeof (*rmp));
+  memset (rmp, 0, sizeof (*rmp));
+  rmp->_vl_msg_id = ntohs (VL_API_SNAT_WORKER_DETAILS+sm->msg_id_base);
+  rmp->context = context;
+  rmp->worker_index = htonl (worker_index);
+  rmp->lcore_id = htonl (w->lcore_id);
+  strncpy ((char *) rmp->name, (char *) w->name, ARRAY_LEN (rmp->name) - 1);
+
+  vl_msg_api_send_shmem (q, (u8 *) & rmp);
+}
+
+static void
+vl_api_snat_worker_dump_t_handler
+(vl_api_snat_worker_dump_t * mp)
+{
+  unix_shared_memory_queue_t *q;
+  snat_main_t * sm = &snat_main;
+  u32 * worker_index;
+
+  q = vl_api_client_index_to_input_queue (mp->client_index);
+  if (q == 0)
+    return;
+
+  vec_foreach (worker_index, sm->workers)
+    {
+      send_snat_worker_details(*worker_index, q, mp->context);
+    }
+}
+
+static void *vl_api_snat_worker_dump_t_print
+(vl_api_snat_worker_dump_t *mp, void * handle)
+{
+  u8 *s;
+
+  s = format (0, "SCRIPT: snat_worker_dump ");
+
+  FINISH;
+}
+
 /* List of message types that this plugin understands */
 #define foreach_snat_plugin_api_msg                                     \
 _(SNAT_ADD_ADDRESS_RANGE, snat_add_address_range)                       \
@@ -934,7 +1072,9 @@ _(SNAT_CONTROL_PING, snat_control_ping)                                 \
 _(SNAT_STATIC_MAPPING_DUMP, snat_static_mapping_dump)                   \
 _(SNAT_SHOW_CONFIG, snat_show_config)                                   \
 _(SNAT_ADDRESS_DUMP, snat_address_dump)                                 \
-_(SNAT_INTERFACE_DUMP, snat_interface_dump)
+_(SNAT_INTERFACE_DUMP, snat_interface_dump)                             \
+_(SNAT_SET_WORKERS, snat_set_workers)                                   \
+_(SNAT_WORKER_DUMP, snat_worker_dump)
 
 /* Set up the API message handling tables */
 static clib_error_t *
@@ -984,6 +1124,11 @@ static clib_error_t * snat_init (vlib_main_t * vm)
   ip4_main_t * im = &ip4_main;
   ip_lookup_main_t * lm = &im->lookup_main;
   u8 * name;
+  uword *p;
+  vlib_thread_registration_t *tr;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
+  uword *bitmap = 0;
+  u32 i;
 
   name = format (0, "snat_%08x%c", api_version, 0);
 
@@ -996,6 +1141,32 @@ static clib_error_t * snat_init (vlib_main_t * vm)
   sm->ip4_main = im;
   sm->ip4_lookup_main = lm;
   sm->api_main = &api_main;
+  sm->first_worker_index = 0;
+  sm->next_worker = 0;
+  sm->num_workers = 0;
+  sm->workers = 0;
+  sm->fq_in2out_index = ~0;
+  sm->fq_out2in_index = ~0;
+
+  p = hash_get_mem (tm->thread_registrations_by_name, "workers");
+  if (p)
+    {
+      tr = (vlib_thread_registration_t *) p[0];
+      if (tr)
+        {
+          sm->num_workers = tr->count;
+          sm->first_worker_index = tr->first_index;
+        }
+    }
+
+  /* Use all available workers by default */
+  if (sm->num_workers > 1)
+    {
+      for (i=0; i < sm->num_workers; i++)
+        bitmap = clib_bitmap_set (bitmap, i, 1);
+      snat_set_workers(bitmap);
+      clib_bitmap_free (bitmap);
+    }
 
   error = snat_plugin_api_hookup (vm);
 
@@ -1363,6 +1534,66 @@ VLIB_CLI_COMMAND (add_static_mapping_command, static) = {
 };
 
 static clib_error_t *
+set_workers_command_fn (vlib_main_t * vm,
+                        unformat_input_t * input,
+                        vlib_cli_command_t * cmd)
+{
+  unformat_input_t _line_input, *line_input = &_line_input;
+  uword *bitmap = 0;
+  int rv = 0;
+
+  /* Get a line of input. */
+  if (!unformat_user (input, unformat_line_input, line_input))
+    return 0;
+
+  while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (line_input, "%U", unformat_bitmap_list, &bitmap))
+        ;
+      else
+        return clib_error_return (0, "unknown input '%U'",
+          format_unformat_error, input);
+     }
+  unformat_free (line_input);
+
+  if (bitmap == 0)
+    return clib_error_return (0, "List of workers must be specified.");
+
+  rv = snat_set_workers(bitmap);
+
+  clib_bitmap_free (bitmap);
+
+  switch (rv)
+    {
+    case VNET_API_ERROR_INVALID_WORKER:
+      return clib_error_return (0, "Invalid worker(s).");
+      break;
+    case VNET_API_ERROR_FEATURE_DISABLED:
+      return clib_error_return (0,
+        "Supported only if 2 or more workes available.");
+      break;
+    default:
+      break;
+    }
+
+  return 0;
+}
+
+/*?
+ * @cliexpar
+ * @cliexstart{set snat workers}
+ * Set SNAT workers if 2 or more workers available, use:
+ *  vpp# set snat workers 0-2,5
+ * @cliexend
+?*/
+VLIB_CLI_COMMAND (set_workers_command, static) = {
+  .path = "set snat workers",
+  .function = set_workers_command_fn,
+  .short_help =
+    "set snat workers <workers-list>",
+};
+
+static clib_error_t *
 snat_config (vlib_main_t * vm, unformat_input_t * input)
 {
   snat_main_t * sm = &snat_main;
@@ -1377,6 +1608,7 @@ snat_config (vlib_main_t * vm, unformat_input_t * input)
   u32 static_mapping_memory_size = 64<<20;
   u8 static_mapping_only = 0;
   u8 static_mapping_connection_tracking = 0;
+  vlib_thread_main_t *tm = vlib_get_thread_main ();
 
   while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
     {
@@ -1425,6 +1657,14 @@ snat_config (vlib_main_t * vm, unformat_input_t * input)
   if (!static_mapping_only ||
       (static_mapping_only && static_mapping_connection_tracking))
     {
+      clib_bihash_init_8_8 (&sm->worker_by_in, "worker-by-in", user_buckets,
+                            user_memory_size);
+
+      clib_bihash_init_8_8 (&sm->worker_by_out, "worker-by-out", user_buckets,
+                            user_memory_size);
+
+      vec_validate (sm->per_thread_data, tm->n_vlib_mains - 1);
+
       clib_bihash_init_8_8 (&sm->in2out, "in2out", translation_buckets,
                             translation_memory_size);
 
@@ -1485,7 +1725,7 @@ u8 * format_snat_session (u8 * s, va_list * args)
 
 u8 * format_snat_user (u8 * s, va_list * args)
 {
-  snat_main_t * sm = va_arg (*args, snat_main_t *);
+  snat_main_per_thread_data_t * sm = va_arg (*args, snat_main_per_thread_data_t *);
   snat_user_t * u = va_arg (*args, snat_user_t *);
   int verbose = va_arg (*args, int);
   dlist_elt_t * head, * elt;
@@ -1552,6 +1792,9 @@ show_snat_command_fn (vlib_main_t * vm,
   snat_static_mapping_t *m;
   snat_interface_t *i;
   vnet_main_t *vnm = vnet_get_main();
+  snat_main_per_thread_data_t *tsm;
+  u32 users_num = 0, sessions_num = 0, *worker;
+  uword j = 0;
 
   if (unformat (input, "detail"))
     verbose = 1;
@@ -1581,6 +1824,20 @@ show_snat_command_fn (vlib_main_t * vm,
       }));
     }
 
+  if (sm->num_workers > 1)
+    {
+      vlib_cli_output (vm, "%d workers", vec_len (sm->workers));
+      if (verbose > 0)
+        {
+          vec_foreach (worker, sm->workers)
+            {
+              vlib_worker_thread_t *w =
+                vlib_worker_threads + *worker + sm->first_worker_index;
+              vlib_cli_output (vm, "  %s", w->name);
+            }
+        }
+    }
+
   if (sm->static_mapping_only && !(sm->static_mapping_connection_tracking))
     {
       vlib_cli_output (vm, "%d static mappings",
@@ -1596,11 +1853,17 @@ show_snat_command_fn (vlib_main_t * vm,
     }
   else
     {
+      vec_foreach (tsm, sm->per_thread_data)
+        {
+          users_num += pool_elts (tsm->users);
+          sessions_num += pool_elts (tsm->sessions);
+        }
+
       vlib_cli_output (vm, "%d users, %d outside addresses, %d active sessions,"
                        " %d static mappings",
-                       pool_elts (sm->users),
+                       users_num,
                        vec_len (sm->addresses),
-                       pool_elts (sm->sessions),
+                       sessions_num,
                        pool_elts (sm->static_mappings));
 
       if (verbose > 0)
@@ -1609,13 +1872,29 @@ show_snat_command_fn (vlib_main_t * vm,
                            verbose - 1);
           vlib_cli_output (vm, "%U", format_bihash_8_8, &sm->out2in,
                            verbose - 1);
-          vlib_cli_output (vm, "%d list pool elements",
-                           pool_elts (sm->list_pool));
+          vlib_cli_output (vm, "%U", format_bihash_8_8, &sm->worker_by_in,
+                           verbose - 1);
+          vlib_cli_output (vm, "%U", format_bihash_8_8, &sm->worker_by_out,
+                           verbose - 1);
+          vec_foreach_index (j, sm->per_thread_data)
+            {
+              tsm = vec_elt_at_index (sm->per_thread_data, j);
 
-          pool_foreach (u, sm->users,
-          ({
-            vlib_cli_output (vm, "%U", format_snat_user, sm, u, verbose - 1);
-          }));
+              if (pool_elts (tsm->users) == 0)
+                continue;
+
+              vlib_worker_thread_t *w = vlib_worker_threads + j;
+              vlib_cli_output (vm, "Thread %d (%s at lcore %u):", j, w->name,
+                               w->lcore_id);
+              vlib_cli_output (vm, "  %d list pool elements",
+                               pool_elts (tsm->list_pool));
+
+              pool_foreach (u, tsm->users,
+              ({
+                vlib_cli_output (vm, "  %U", format_snat_user, tsm, u,
+                                 verbose - 1);
+              }));
+            }
 
           if (pool_elts (sm->static_mappings))
             {
diff --git a/plugins/snat-plugin/snat/snat.h b/plugins/snat-plugin/snat/snat.h
index ca10b76adce..cb31dc51423 100644
--- a/plugins/snat-plugin/snat/snat.h
+++ b/plugins/snat-plugin/snat/snat.h
@@ -132,6 +132,17 @@ typedef struct {
 } snat_interface_t;
 
 typedef struct {
+  /* User pool */
+  snat_user_t * users;
+
+  /* Session pool */
+  snat_session_t * sessions;
+
+  /* Pool of doubly-linked list elements */
+  dlist_elt_t * list_pool;
+} snat_main_per_thread_data_t;
+
+typedef struct {
   /* Main lookup tables */
   clib_bihash_8_8_t out2in;
   clib_bihash_8_8_t in2out;
@@ -139,18 +150,26 @@ typedef struct {
   /* Find-a-user => src address lookup */
   clib_bihash_8_8_t user_hash;
 
+  /* Non-translated packets worker lookup => src address + VRF */
+  clib_bihash_8_8_t worker_by_in;
+
+  /* Translated packets worker lookup => IP address + port number */
+  clib_bihash_8_8_t worker_by_out;
+
+  u32 num_workers;
+  u32 first_worker_index;
+  u32 next_worker;
+  u32 * workers;
+
+  /* Per thread data */
+  snat_main_per_thread_data_t * per_thread_data;
+
   /* Find a static mapping by local */
   clib_bihash_8_8_t static_mapping_by_local;
 
   /* Find a static mapping by external */
   clib_bihash_8_8_t static_mapping_by_external;
 
-  /* User pool */
-  snat_user_t * users;
-
-  /* Session pool */
-  snat_session_t * sessions;
-
   /* Static mapping pool */
   snat_static_mapping_t * static_mappings;
 
@@ -160,12 +179,13 @@ typedef struct {
   /* Vector of outside addresses */
   snat_address_t * addresses;
 
-  /* Pool of doubly-linked list elements */
-  dlist_elt_t * list_pool;
-
   /* Randomize port allocation order */
   u32 random_seed;
 
+  /* Worker handoff index */
+  u32 fq_in2out_index;
+  u32 fq_out2in_index;
+
   /* Config parameters */
   u8 static_mapping_only;
   u8 static_mapping_connection_tracking;
@@ -196,6 +216,8 @@ extern vlib_node_registration_t snat_in2out_node;
 extern vlib_node_registration_t snat_out2in_node;
 extern vlib_node_registration_t snat_in2out_fast_node;
 extern vlib_node_registration_t snat_out2in_fast_node;
+extern vlib_node_registration_t snat_in2out_worker_handoff_node;
+extern vlib_node_registration_t snat_out2in_worker_handoff_node;
 
 void snat_free_outside_address_and_port (snat_main_t * sm, 
                                          snat_session_key_t * k, 
diff --git a/plugins/snat-plugin/snat/snat_test.c b/plugins/snat-plugin/snat/snat_test.c
index e03c9eeccdd..2a003ba60c6 100644
--- a/plugins/snat-plugin/snat/snat_test.c
+++ b/plugins/snat-plugin/snat/snat_test.c
@@ -60,7 +60,8 @@ snat_test_main_t snat_test_main;
 #define foreach_standard_reply_retval_handler   \
 _(snat_add_address_range_reply)                 \
 _(snat_interface_add_del_feature_reply)         \
-_(snat_add_static_mapping_reply)
+_(snat_add_static_mapping_reply)                \
+_(snat_set_workers_reply)
 
 #define _(n)                                            \
     static void vl_api_##n##_t_handler                  \
@@ -91,7 +92,9 @@ _(SNAT_CONTROL_PING_REPLY, snat_control_ping_reply)             \
 _(SNAT_STATIC_MAPPING_DETAILS, snat_static_mapping_details)     \
 _(SNAT_SHOW_CONFIG_REPLY, snat_show_config_reply)               \
 _(SNAT_ADDRESS_DETAILS, snat_address_details)                   \
-_(SNAT_INTERFACE_DETAILS, snat_interface_details)
+_(SNAT_INTERFACE_DETAILS, snat_interface_details)               \
+_(SNAT_SET_WORKERS_REPLY, snat_set_workers_reply)               \
+_(SNAT_WORKER_DETAILS, snat_worker_details)
 
 /* M: construct, but don't yet send a message */
 #define M(T,t)                                                  \
@@ -473,6 +476,69 @@ static int api_snat_interface_dump(vat_main_t * vam)
   return 0;
 }
 
+static int api_snat_set_workers (vat_main_t * vam)
+{
+  snat_test_main_t * sm = &snat_test_main;
+  unformat_input_t * i = vam->input;
+  f64 timeout;
+  vl_api_snat_set_workers_t * mp;
+  uword *bitmap;
+
+  while (unformat_check_input (i) != UNFORMAT_END_OF_INPUT)
+    {
+      if (unformat (i, "%U", unformat_bitmap_list, &bitmap))
+        ;
+      else
+        {
+          clib_warning("unknown input '%U'", format_unformat_error, i);
+          return -99;
+        }
+    }
+
+  M(SNAT_SET_WORKERS, snat_set_workers);
+  mp->worker_mask = clib_host_to_net_u64 (bitmap[0]);
+
+  S; W;
+
+  /* NOTREACHED */
+  return 0;
+}
+
+static void vl_api_snat_worker_details_t_handler
+  (vl_api_snat_worker_details_t *mp)
+{
+  snat_test_main_t * sm = &snat_test_main;
+  vat_main_t *vam = sm->vat_main;
+
+  fformat (vam->ofp, "worker_index %d (%s at lcore %u)\n",
+           ntohl (mp->worker_index), mp->name, ntohl (mp->lcore_id));
+}
+
+static int api_snat_worker_dump(vat_main_t * vam)
+{
+  snat_test_main_t * sm = &snat_test_main;
+  f64 timeout;
+  vl_api_snat_worker_dump_t * mp;
+
+  if (vam->json_output)
+    {
+      clib_warning ("JSON output not supported for snat_address_dump");
+      return -99;
+    }
+
+  M(SNAT_WORKER_DUMP, snat_worker_dump);
+  S;
+  /* Use a control ping for synchronization */
+  {
+    vl_api_snat_control_ping_t *mp;
+    M (SNAT_CONTROL_PING, snat_control_ping);
+    S;
+  }
+  W;
+  /* NOTREACHED */
+  return 0;
+}
+
 /* 
  * List of messages that the api test plugin sends,
  * and that the data plane plugin processes
@@ -483,10 +549,12 @@ _(snat_interface_add_del_feature,                                \
   "<intfc> | sw_if_index <id> [in] [out] [del]")                 \
 _(snat_add_static_mapping, "local_addr <ip> external_addr <ip> " \
   "[local_port <n>] [external_port <n>] [vrf <table-id>] [del]") \
+_(snat_set_workers, "<wokrers_bitmap>")                          \
 _(snat_static_mapping_dump, "")                                  \
 _(snat_show_config, "")                                          \
 _(snat_address_dump, "")                                         \
-_(snat_interface_dump, "")
+_(snat_interface_dump, "")                                       \
+_(snat_worker_dump, "")
 
 void vat_api_hookup (vat_main_t *vam)
 {
diff --git a/vnet/etc/scripts/snat b/vnet/etc/scripts/snat
new file mode 100644
index 00000000000..87fd699ee51
--- /dev/null
+++ b/vnet/etc/scripts/snat
@@ -0,0 +1,34 @@
+packet-generator new {
+  name f1
+  limit 1000000
+  node ip4-input
+  size 64-64
+  no-recycle
+  worker 0
+  data {
+    UDP: 10.0.0.3 -> 172.16.1.2
+    UDP: 3000 -> 3001
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+
+packet-generator new {
+  name f2
+  limit 1000000
+  node ip4-input
+  size 64-64
+  no-recycle
+  worker 1
+  data {
+    UDP: 10.0.0.3 -> 172.16.1.2
+    UDP: 3005 -> 3006
+    length 128 checksum 0 incrementing 1
+  }
+}
+
+snat add address 172.16.1.3
+ip route 172.16.1.2/32 via drop
+set int ip address pg0 10.0.0.1/24
+set int snat in pg0
+trace add pg-input 10