Dual Loop Load-Balance Nodes

Change-Id: I23bc477aa03f97dd546747b3297a878055d32e1c
Signed-off-by: Neale Ranns <nranns@cisco.com>

3 files changed, 561 insertions, 129 deletions

diff --git a/vnet/vnet/ip/ip4_forward.c b/vnet/vnet/ip/ip4_forward.c
index 85ad10e62b6..fa8ec4e644c 100644
--- a/vnet/vnet/ip/ip4_forward.c
+++ b/vnet/vnet/ip/ip4_forward.c
@@ -517,6 +517,101 @@ ip4_load_balance (vlib_main_t * vm,
 			   to_next, n_left_to_next);
 
 
+      while (n_left_from >= 4 && n_left_to_next >= 2)
+      	{
+ 	  ip_lookup_next_t next0, next1;
+	  const load_balance_t *lb0, *lb1;
+	  vlib_buffer_t * p0, *p1;
+	  u32 pi0, lbi0, hc0, pi1, lbi1, hc1;
+	  const ip4_header_t *ip0, *ip1;
+	  const dpo_id_t *dpo0, *dpo1;
+
+      	  /* Prefetch next iteration. */
+      	  {
+            vlib_buffer_t * p2, * p3;
+
+	    p2 = vlib_get_buffer (vm, from[2]);
+	    p3 = vlib_get_buffer (vm, from[3]);
+
+	    vlib_prefetch_buffer_header (p2, STORE);
+	    vlib_prefetch_buffer_header (p3, STORE);
+
+	    CLIB_PREFETCH (p2->data, sizeof (ip0[0]), STORE);
+	    CLIB_PREFETCH (p3->data, sizeof (ip0[0]), STORE);
+      	  }
+
+	  pi0 = to_next[0] = from[0];
+	  pi1 = to_next[1] = from[1];
+
+	  from += 2;
+	  n_left_from -= 2;
+	  to_next += 2;
+	  n_left_to_next -= 2;
+
+	  p0 = vlib_get_buffer (vm, pi0);
+	  p1 = vlib_get_buffer (vm, pi1);
+
+	  ip0 = vlib_buffer_get_current (p0);
+	  ip1 = vlib_buffer_get_current (p1);
+	  lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+	  lbi1 = vnet_buffer (p1)->ip.adj_index[VLIB_TX];
+
+	  lb0 = load_balance_get(lbi0);
+	  lb1 = load_balance_get(lbi1);
+
+          /*
+           * this node is for via FIBs we can re-use the hash value from the
+           * to node if present.
+           * We don't want to use the same hash value at each level in the recursion
+           * graph as that would lead to polarisation
+           */
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          hc1 = vnet_buffer (p1)->ip.flow_hash = 0;
+
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = ip4_compute_flow_hash(ip0, hc0);
+              }
+          }
+          if (PREDICT_FALSE (lb1->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p1)->ip.flow_hash))
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = vnet_buffer(p1)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = ip4_compute_flow_hash(ip1, hc1);
+              }
+          }
+
+	  dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
+	  dpo1 = load_balance_get_bucket_i(lb1, hc1 & (lb1->lb_n_buckets_minus_1));
+
+	  next0 = dpo0->dpoi_next_node;
+	  next1 = dpo1->dpoi_next_node;
+
+	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+	  vnet_buffer (p1)->ip.adj_index[VLIB_TX] = dpo1->dpoi_index;
+
+	  vlib_increment_combined_counter
+              (cm, cpu_index, lbi0, 1,
+               vlib_buffer_length_in_chain (vm, p0));
+	  vlib_increment_combined_counter
+              (cm, cpu_index, lbi1, 1,
+               vlib_buffer_length_in_chain (vm, p1));
+
+	  vlib_validate_buffer_enqueue_x2 (vm, node, next,
+					   to_next, n_left_to_next,
+					   pi0, pi1, next0, next1);
+       }
+
       while (n_left_from > 0 && n_left_to_next > 0)
 	{
 	  ip_lookup_next_t next0;
@@ -528,6 +623,10 @@ ip4_load_balance (vlib_main_t * vm,
 
 	  pi0 = from[0];
 	  to_next[0] = pi0;
+	  from += 1;
+	  to_next += 1;
+	  n_left_to_next -= 1;
+	  n_left_from -= 1;
 
 	  p0 = vlib_get_buffer (vm, pi0);
 
@@ -535,12 +634,21 @@ ip4_load_balance (vlib_main_t * vm,
 	  lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
 
 	  lb0 = load_balance_get(lbi0);
-	  hc0 = lb0->lb_hash_config;
-	  vnet_buffer(p0)->ip.flow_hash = ip4_compute_flow_hash(ip0, hc0);
 
-	  dpo0 = load_balance_get_bucket_i(lb0,
-					   vnet_buffer(p0)->ip.flow_hash &
-					   (lb0->lb_n_buckets_minus_1));
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = ip4_compute_flow_hash(ip0, hc0);
+              }
+          }
+
+	  dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
 
 	  next0 = dpo0->dpoi_next_node;
 	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
@@ -549,22 +657,9 @@ ip4_load_balance (vlib_main_t * vm,
               (cm, cpu_index, lbi0, 1,
                vlib_buffer_length_in_chain (vm, p0));
 
-	  from += 1;
-	  to_next += 1;
-	  n_left_to_next -= 1;
-	  n_left_from -= 1;
-
-	  if (PREDICT_FALSE (next0 != next))
-	    {
-	      n_left_to_next += 1;
-	      vlib_put_next_frame (vm, node, next, n_left_to_next);
-	      next = next0;
-	      vlib_get_next_frame (vm, node, next,
-				   to_next, n_left_to_next);
-	      to_next[0] = pi0;
-	      to_next += 1;
-	      n_left_to_next -= 1;
-	    }
+	  vlib_validate_buffer_enqueue_x1 (vm, node, next,
+					   to_next, n_left_to_next,
+					   pi0, next0);
 	}
 
       vlib_put_next_frame (vm, node, next, n_left_to_next);
@@ -573,15 +668,13 @@ ip4_load_balance (vlib_main_t * vm,
   return frame->n_vectors;
 }
 
-static u8 * format_ip4_forward_next_trace (u8 * s, va_list * args);
-
 VLIB_REGISTER_NODE (ip4_load_balance_node) = {
   .function = ip4_load_balance,
   .name = "ip4-load-balance",
   .vector_size = sizeof (u32),
   .sibling_of = "ip4-lookup",
 
-  .format_trace = format_ip4_forward_next_trace,
+  .format_trace = format_ip4_lookup_trace,
 };
 
 VLIB_NODE_FUNCTION_MULTIARCH (ip4_load_balance_node, ip4_load_balance)
diff --git a/vnet/vnet/ip/ip6_forward.c b/vnet/vnet/ip/ip6_forward.c
index 325ef9b7b67..bc346786283 100644
--- a/vnet/vnet/ip/ip6_forward.c
+++ b/vnet/vnet/ip/ip6_forward.c
@@ -703,8 +703,8 @@ VLIB_NODE_FUNCTION_MULTIARCH (ip6_lookup_node, ip6_lookup)
 
 always_inline uword
 ip6_load_balance (vlib_main_t * vm,
-		  vlib_node_runtime_t * node,
-		  vlib_frame_t * frame)
+                  vlib_node_runtime_t * node,
+                  vlib_frame_t * frame)
 {
   vlib_combined_counter_main_t * cm = &load_balance_main.lbm_via_counters;
   u32 n_left_from, n_left_to_next, * from, * to_next;
@@ -722,63 +722,172 @@ ip6_load_balance (vlib_main_t * vm,
   while (n_left_from > 0)
     {
       vlib_get_next_frame (vm, node, next,
-			   to_next, n_left_to_next);
+                           to_next, n_left_to_next);
 
 
-      while (n_left_from > 0 && n_left_to_next > 0)
-	{
-	  ip_lookup_next_t next0;
-	  const load_balance_t *lb0;
-	  vlib_buffer_t * p0;
-	  u32 pi0, lbi0, hc0;
-	  const ip6_header_t *ip0;
-	  const dpo_id_t *dpo0;
-
-	  pi0 = from[0];
-	  to_next[0] = pi0;
+      while (n_left_from >= 4 && n_left_to_next >= 2)
+        {
+          ip_lookup_next_t next0, next1;
+          const load_balance_t *lb0, *lb1;
+          vlib_buffer_t * p0, *p1;
+          u32 pi0, lbi0, hc0, pi1, lbi1, hc1;
+          const ip6_header_t *ip0, *ip1;
+          const dpo_id_t *dpo0, *dpo1;
+
+          /* Prefetch next iteration. */
+          {
+            vlib_buffer_t * p2, * p3;
+
+            p2 = vlib_get_buffer (vm, from[2]);
+            p3 = vlib_get_buffer (vm, from[3]);
+
+            vlib_prefetch_buffer_header (p2, STORE);
+            vlib_prefetch_buffer_header (p3, STORE);
+
+            CLIB_PREFETCH (p2->data, sizeof (ip0[0]), STORE);
+            CLIB_PREFETCH (p3->data, sizeof (ip0[0]), STORE);
+          }
+
+          pi0 = to_next[0] = from[0];
+          pi1 = to_next[1] = from[1];
+
+          from += 2;
+          n_left_from -= 2;
+          to_next += 2;
+          n_left_to_next -= 2;
+
+          p0 = vlib_get_buffer (vm, pi0);
+          p1 = vlib_get_buffer (vm, pi1);
+
+          ip0 = vlib_buffer_get_current (p0);
+          ip1 = vlib_buffer_get_current (p1);
+          lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+          lbi1 = vnet_buffer (p1)->ip.adj_index[VLIB_TX];
+
+          lb0 = load_balance_get(lbi0);
+          lb1 = load_balance_get(lbi1);
+
+          /*
+           * this node is for via FIBs we can re-use the hash value from the
+           * to node if present.
+           * We don't want to use the same hash value at each level in the recursion
+           * graph as that would lead to polarisation
+           */
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          hc1 = vnet_buffer (p1)->ip.flow_hash = 0;
+
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = ip6_compute_flow_hash(ip0, hc0);
+              }
+          }
+          if (PREDICT_FALSE (lb1->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p1)->ip.flow_hash))
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = vnet_buffer(p1)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = ip6_compute_flow_hash(ip1, hc1);
+              }
+          }
+
+          dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
+          dpo1 = load_balance_get_bucket_i(lb1, hc1 & (lb1->lb_n_buckets_minus_1));
+
+          next0 = dpo0->dpoi_next_node;
+          next1 = dpo1->dpoi_next_node;
+
+          /* Only process the HBH Option Header if explicitly configured to do so */
+          if (PREDICT_FALSE(ip0->protocol == IP_PROTOCOL_IP6_HOP_BY_HOP_OPTIONS))
+            {
+              next0 = (dpo_is_adj(dpo0) && im->hbh_enabled) ?
+                (ip_lookup_next_t) IP6_LOOKUP_NEXT_HOP_BY_HOP : next0;
+            }
+          /* Only process the HBH Option Header if explicitly configured to do so */
+          if (PREDICT_FALSE(ip1->protocol == IP_PROTOCOL_IP6_HOP_BY_HOP_OPTIONS))
+            {
+              next1 = (dpo_is_adj(dpo1) && im->hbh_enabled) ?
+                (ip_lookup_next_t) IP6_LOOKUP_NEXT_HOP_BY_HOP : next1;
+            }
 
-	  p0 = vlib_get_buffer (vm, pi0);
+          vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+          vnet_buffer (p1)->ip.adj_index[VLIB_TX] = dpo1->dpoi_index;
 
-	  ip0 = vlib_buffer_get_current (p0);
-	  lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi0, 1,
+               vlib_buffer_length_in_chain (vm, p0));
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi1, 1,
+               vlib_buffer_length_in_chain (vm, p1));
 
-	  lb0 = load_balance_get(lbi0);
-	  hc0 = lb0->lb_hash_config;
-	  vnet_buffer(p0)->ip.flow_hash = ip6_compute_flow_hash(ip0, hc0);
+          vlib_validate_buffer_enqueue_x2 (vm, node, next,
+                                           to_next, n_left_to_next,
+                                           pi0, pi1, next0, next1);
+       }
 
-	  dpo0 = load_balance_get_bucket_i(lb0,
-					   vnet_buffer(p0)->ip.flow_hash &
-					   (lb0->lb_n_buckets - 1));
-	  next0 = dpo0->dpoi_next_node;
- 	  /* Only process the HBH Option Header if explicitly configured to do so */
-	  if (PREDICT_FALSE(ip0->protocol == IP_PROTOCOL_IP6_HOP_BY_HOP_OPTIONS))
-	    {
-	      next0 = (dpo_is_adj(dpo0) && im->hbh_enabled) ?
-		(ip_lookup_next_t) IP6_LOOKUP_NEXT_HOP_BY_HOP : next0;
-	    }
-	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+      while (n_left_from > 0 && n_left_to_next > 0)
+        {
+          ip_lookup_next_t next0;
+          const load_balance_t *lb0;
+          vlib_buffer_t * p0;
+          u32 pi0, lbi0, hc0;
+          const ip6_header_t *ip0;
+          const dpo_id_t *dpo0;
+
+          pi0 = from[0];
+          to_next[0] = pi0;
+          from += 1;
+          to_next += 1;
+          n_left_to_next -= 1;
+          n_left_from -= 1;
+
+          p0 = vlib_get_buffer (vm, pi0);
+
+          ip0 = vlib_buffer_get_current (p0);
+          lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+
+          lb0 = load_balance_get(lbi0);
+
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = ip6_compute_flow_hash(ip0, hc0);
+              }
+          }
+          dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
+
+          next0 = dpo0->dpoi_next_node;
+          vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+
+          /* Only process the HBH Option Header if explicitly configured to do so */
+          if (PREDICT_FALSE(ip0->protocol == IP_PROTOCOL_IP6_HOP_BY_HOP_OPTIONS))
+            {
+              next0 = (dpo_is_adj(dpo0) && im->hbh_enabled) ?
+                (ip_lookup_next_t) IP6_LOOKUP_NEXT_HOP_BY_HOP : next0;
+            }
 
-	  vlib_increment_combined_counter
+          vlib_increment_combined_counter
               (cm, cpu_index, lbi0, 1,
                vlib_buffer_length_in_chain (vm, p0));
 
-	  from += 1;
-	  to_next += 1;
-	  n_left_to_next -= 1;
-	  n_left_from -= 1;
-
-	  if (PREDICT_FALSE (next0 != next))
-	    {
-	      n_left_to_next += 1;
-	      vlib_put_next_frame (vm, node, next, n_left_to_next);
-	      next = next0;
-	      vlib_get_next_frame (vm, node, next,
-				   to_next, n_left_to_next);
-	      to_next[0] = pi0;
-	      to_next += 1;
-	      n_left_to_next -= 1;
-	    }
-	}
+          vlib_validate_buffer_enqueue_x1 (vm, node, next,
+                                           to_next, n_left_to_next,
+                                           pi0, next0);
+        }
 
       vlib_put_next_frame (vm, node, next, n_left_to_next);
     }
@@ -791,8 +900,8 @@ VLIB_REGISTER_NODE (ip6_load_balance_node) = {
   .name = "ip6-load-balance",
   .vector_size = sizeof (u32),
   .sibling_of = "ip6-lookup",
+
   .format_trace = format_ip6_lookup_trace,
-  .n_next_nodes = 0,
 };
 
 VLIB_NODE_FUNCTION_MULTIARCH (ip6_load_balance_node, ip6_load_balance)
diff --git a/vnet/vnet/mpls/mpls_lookup.c b/vnet/vnet/mpls/mpls_lookup.c
index 9d29cecff90..34ba79e40fd 100644
--- a/vnet/vnet/mpls/mpls_lookup.c
+++ b/vnet/vnet/mpls/mpls_lookup.c
@@ -28,6 +28,7 @@ typedef struct {
   u32 lb_index;
   u32 lfib_index;
   u32 label_net_byte_order;
+  u32 hash;
 } mpls_lookup_trace_t;
 
 static u8 *
@@ -37,9 +38,9 @@ format_mpls_lookup_trace (u8 * s, va_list * args)
   CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
   mpls_lookup_trace_t * t = va_arg (*args, mpls_lookup_trace_t *);
 
-  s = format (s, "MPLS: next [%d], lookup fib index %d, LB index %d "
+  s = format (s, "MPLS: next [%d], lookup fib index %d, LB index %d hash %d"
               "label %d eos %d", 
-              t->next_index, t->lfib_index, t->lb_index,
+              t->next_index, t->lfib_index, t->lb_index, t->hash,
               vnet_mpls_uc_get_label(
                   clib_net_to_host_u32(t->label_net_byte_order)),
               vnet_mpls_uc_get_s(t->label_net_byte_order));
@@ -79,6 +80,128 @@ mpls_lookup (vlib_main_t * vm,
       vlib_get_next_frame (vm, node, next_index,
                            to_next, n_left_to_next);
 
+      while (n_left_from >= 4 && n_left_to_next >= 2)
+        {
+          u32 lbi0, next0, lfib_index0, bi0, hash_c0;
+          const mpls_unicast_header_t * h0;
+          const load_balance_t *lb0;
+          const dpo_id_t *dpo0;
+          vlib_buffer_t * b0;
+          u32 lbi1, next1, lfib_index1, bi1, hash_c1;
+          const mpls_unicast_header_t * h1;
+          const load_balance_t *lb1;
+          const dpo_id_t *dpo1;
+          vlib_buffer_t * b1;
+
+           /* Prefetch next iteration. */
+          {
+            vlib_buffer_t * p2, * p3;
+
+            p2 = vlib_get_buffer (vm, from[2]);
+            p3 = vlib_get_buffer (vm, from[3]);
+
+            vlib_prefetch_buffer_header (p2, STORE);
+            vlib_prefetch_buffer_header (p3, STORE);
+
+            CLIB_PREFETCH (p2->data, sizeof (h0[0]), STORE);
+            CLIB_PREFETCH (p3->data, sizeof (h0[0]), STORE);
+          }
+
+          bi0 = to_next[0] = from[0];
+          bi1 = to_next[1] = from[1];
+
+          from += 2;
+          n_left_from -= 2;
+          to_next += 2;
+          n_left_to_next -= 2;
+
+          b0 = vlib_get_buffer (vm, bi0);
+          b1 = vlib_get_buffer (vm, bi1);
+          h0 = vlib_buffer_get_current (b0);
+          h1 = vlib_buffer_get_current (b1);
+
+          lfib_index0 = vec_elt(mm->fib_index_by_sw_if_index,
+                                vnet_buffer(b0)->sw_if_index[VLIB_RX]);
+          lfib_index1 = vec_elt(mm->fib_index_by_sw_if_index,
+                                vnet_buffer(b1)->sw_if_index[VLIB_RX]);
+
+          lbi0 = mpls_fib_table_forwarding_lookup (lfib_index0, h0);
+          lbi1 = mpls_fib_table_forwarding_lookup (lfib_index1, h1);
+          lb0 = load_balance_get(lbi0);
+          lb1 = load_balance_get(lbi1);
+
+          hash_c0 = vnet_buffer(b0)->ip.flow_hash = 0;
+          hash_c1 = vnet_buffer(b1)->ip.flow_hash = 0;
+
+          if (PREDICT_FALSE(lb0->lb_n_buckets > 1))
+          {
+              hash_c0 = vnet_buffer (b0)->ip.flow_hash =
+                  mpls_compute_flow_hash(h0, lb0->lb_hash_config);
+          }
+          if (PREDICT_FALSE(lb1->lb_n_buckets > 1))
+          {
+              hash_c1 = vnet_buffer (b1)->ip.flow_hash =
+                  mpls_compute_flow_hash(h1, lb1->lb_hash_config);
+          }
+
+          ASSERT (lb0->lb_n_buckets > 0);
+          ASSERT (is_pow2 (lb0->lb_n_buckets));
+          ASSERT (lb1->lb_n_buckets > 0);
+          ASSERT (is_pow2 (lb1->lb_n_buckets));
+
+          dpo0 = load_balance_get_bucket_i(lb0,
+                                           (hash_c0 &
+                                            (lb0->lb_n_buckets_minus_1)));
+          dpo1 = load_balance_get_bucket_i(lb1,
+                                           (hash_c1 &
+                                            (lb1->lb_n_buckets_minus_1)));
+
+          next0 = dpo0->dpoi_next_node;
+          next1 = dpo1->dpoi_next_node;
+
+          vnet_buffer (b0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+          vnet_buffer (b1)->ip.adj_index[VLIB_TX] = dpo1->dpoi_index;
+
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi0, 1,
+               vlib_buffer_length_in_chain (vm, b0));
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi1, 1,
+               vlib_buffer_length_in_chain (vm, b1));
+
+          /*
+           * pop the label that was just used in the lookup
+           */
+          vlib_buffer_advance(b0, sizeof(*h0));
+          vlib_buffer_advance(b1, sizeof(*h1));
+
+          if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED))
+          {
+              mpls_lookup_trace_t *tr = vlib_add_trace (vm, node,
+                                                        b0, sizeof (*tr));
+              tr->next_index = next0;
+              tr->lb_index = lbi0;
+              tr->lfib_index = lfib_index0;
+              tr->hash = hash_c0;
+              tr->label_net_byte_order = h0->label_exp_s_ttl;
+          }
+
+          if (PREDICT_FALSE(b1->flags & VLIB_BUFFER_IS_TRACED))
+          {
+              mpls_lookup_trace_t *tr = vlib_add_trace (vm, node,
+                                                        b1, sizeof (*tr));
+              tr->next_index = next1;
+              tr->lb_index = lbi1;
+              tr->lfib_index = lfib_index1;
+              tr->hash = hash_c1;
+              tr->label_net_byte_order = h1->label_exp_s_ttl;
+          }
+
+          vlib_validate_buffer_enqueue_x2 (vm, node, next_index,
+                                           to_next, n_left_to_next,
+                                           bi0, bi1, next0, next1);
+        }
+
       while (n_left_from > 0 && n_left_to_next > 0)
       {
           u32 lbi0, next0, lfib_index0, bi0, hash_c0;
@@ -101,26 +224,26 @@ mpls_lookup (vlib_main_t * vm,
                                 vnet_buffer(b0)->sw_if_index[VLIB_RX]);
 
           lbi0 = mpls_fib_table_forwarding_lookup (lfib_index0, h0);
-	  lb0 = load_balance_get(lbi0);
+          lb0 = load_balance_get(lbi0);
 
           hash_c0 = vnet_buffer(b0)->ip.flow_hash = 0;
           if (PREDICT_FALSE(lb0->lb_n_buckets > 1))
           {
-              hash_c0 = vnet_buffer (b0)->ip.flow_hash = 
+              hash_c0 = vnet_buffer (b0)->ip.flow_hash =
                   mpls_compute_flow_hash(h0, lb0->lb_hash_config);
           }
 
-	  ASSERT (lb0->lb_n_buckets > 0);
-	  ASSERT (is_pow2 (lb0->lb_n_buckets));
+          ASSERT (lb0->lb_n_buckets > 0);
+          ASSERT (is_pow2 (lb0->lb_n_buckets));
 
-	  dpo0 = load_balance_get_bucket_i(lb0,
+          dpo0 = load_balance_get_bucket_i(lb0,
                                            (hash_c0 &
                                             (lb0->lb_n_buckets_minus_1)));
 
-	  next0 = dpo0->dpoi_next_node;
-	  vnet_buffer (b0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+          next0 = dpo0->dpoi_next_node;
+          vnet_buffer (b0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
 
-	  vlib_increment_combined_counter 
+          vlib_increment_combined_counter
               (cm, cpu_index, lbi0, 1,
                vlib_buffer_length_in_chain (vm, b0));
 
@@ -132,10 +255,11 @@ mpls_lookup (vlib_main_t * vm,
           if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED))
           {
               mpls_lookup_trace_t *tr = vlib_add_trace (vm, node,
-							b0, sizeof (*tr));
+                                                        b0, sizeof (*tr));
               tr->next_index = next0;
               tr->lb_index = lbi0;
               tr->lfib_index = lfib_index0;
+              tr->hash = hash_c0;
               tr->label_net_byte_order = h0->label_exp_s_ttl;
           }
 
@@ -177,6 +301,7 @@ VLIB_NODE_FUNCTION_MULTIARCH (mpls_lookup_node, mpls_lookup)
 typedef struct {
   u32 next_index;
   u32 lb_index;
+  u32 hash;
 } mpls_load_balance_trace_t;
 
 static u8 *
@@ -186,20 +311,20 @@ format_mpls_load_balance_trace (u8 * s, va_list * args)
   CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
   mpls_load_balance_trace_t * t = va_arg (*args, mpls_load_balance_trace_t *);
 
-  s = format (s, "MPLS: next [%d], LB index %d ", 
-              t->next_index, t->lb_index);
+  s = format (s, "MPLS: next [%d], LB index %d hash %d",
+              t->next_index, t->lb_index, t->hash);
   return s;
 }
 
 always_inline uword
 mpls_load_balance (vlib_main_t * vm,
-		  vlib_node_runtime_t * node,
-		  vlib_frame_t * frame)
+                  vlib_node_runtime_t * node,
+                  vlib_frame_t * frame)
 {
   vlib_combined_counter_main_t * cm = &load_balance_main.lbm_via_counters;
   u32 n_left_from, n_left_to_next, * from, * to_next;
-  ip_lookup_next_t next;
   u32 cpu_index = os_get_cpu_number();
+  u32 next;
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -208,57 +333,162 @@ mpls_load_balance (vlib_main_t * vm,
   while (n_left_from > 0)
     {
       vlib_get_next_frame (vm, node, next,
-			   to_next, n_left_to_next);
+                           to_next, n_left_to_next);
+
+
+      while (n_left_from >= 4 && n_left_to_next >= 2)
+        {
+          mpls_lookup_next_t next0, next1;
+          const load_balance_t *lb0, *lb1;
+          vlib_buffer_t * p0, *p1;
+          u32 pi0, lbi0, hc0, pi1, lbi1, hc1;
+          const mpls_unicast_header_t *mpls0, *mpls1;
+          const dpo_id_t *dpo0, *dpo1;
+
+          /* Prefetch next iteration. */
+          {
+            vlib_buffer_t * p2, * p3;
+
+            p2 = vlib_get_buffer (vm, from[2]);
+            p3 = vlib_get_buffer (vm, from[3]);
+
+            vlib_prefetch_buffer_header (p2, STORE);
+            vlib_prefetch_buffer_header (p3, STORE);
+
+            CLIB_PREFETCH (p2->data, sizeof (mpls0[0]), STORE);
+            CLIB_PREFETCH (p3->data, sizeof (mpls0[0]), STORE);
+          }
+
+          pi0 = to_next[0] = from[0];
+          pi1 = to_next[1] = from[1];
+
+          from += 2;
+          n_left_from -= 2;
+          to_next += 2;
+          n_left_to_next -= 2;
+
+          p0 = vlib_get_buffer (vm, pi0);
+          p1 = vlib_get_buffer (vm, pi1);
+
+          mpls0 = vlib_buffer_get_current (p0);
+          mpls1 = vlib_buffer_get_current (p1);
+          lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+          lbi1 = vnet_buffer (p1)->ip.adj_index[VLIB_TX];
+
+          lb0 = load_balance_get(lbi0);
+          lb1 = load_balance_get(lbi1);
+
+          /*
+           * this node is for via FIBs we can re-use the hash value from the
+           * to node if present.
+           * We don't want to use the same hash value at each level in the recursion
+           * graph as that would lead to polarisation
+           */
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          hc1 = vnet_buffer (p1)->ip.flow_hash = 0;
+
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = mpls_compute_flow_hash(mpls0, hc0);
+              }
+          }
+          if (PREDICT_FALSE (lb1->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p1)->ip.flow_hash))
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = vnet_buffer(p1)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc1 = vnet_buffer(p1)->ip.flow_hash = mpls_compute_flow_hash(mpls1, hc1);
+              }
+          }
+
+          dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
+          dpo1 = load_balance_get_bucket_i(lb1, hc1 & (lb1->lb_n_buckets_minus_1));
+
+          next0 = dpo0->dpoi_next_node;
+          next1 = dpo1->dpoi_next_node;
+
+          vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+          vnet_buffer (p1)->ip.adj_index[VLIB_TX] = dpo1->dpoi_index;
+
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi0, 1,
+               vlib_buffer_length_in_chain (vm, p0));
+          vlib_increment_combined_counter
+              (cm, cpu_index, lbi1, 1,
+               vlib_buffer_length_in_chain (vm, p1));
+
+          if (PREDICT_FALSE(p0->flags & VLIB_BUFFER_IS_TRACED))
+          {
+              mpls_load_balance_trace_t *tr = vlib_add_trace (vm, node,
+                                                              p0, sizeof (*tr));
+              tr->next_index = next0;
+              tr->lb_index = lbi0;
+              tr->hash = hc0;
+          }
+
+          vlib_validate_buffer_enqueue_x2 (vm, node, next,
+                                           to_next, n_left_to_next,
+                                           pi0, pi1, next0, next1);
+       }
 
-    
       while (n_left_from > 0 && n_left_to_next > 0)
-	{
-	  const mpls_unicast_header_t *hdr0;
-	  const load_balance_t *lb0;
-	  u32 pi0, lbi0, hc0, next0;
-	  const dpo_id_t *dpo0;
-	  vlib_buffer_t * p0;
+        {
+          mpls_lookup_next_t next0;
+          const load_balance_t *lb0;
+          vlib_buffer_t * p0;
+          u32 pi0, lbi0, hc0;
+          const mpls_unicast_header_t *mpls0;
+          const dpo_id_t *dpo0;
 
-	  pi0 = from[0];
-	  to_next[0] = pi0;
+          pi0 = from[0];
+          to_next[0] = pi0;
+          from += 1;
+          to_next += 1;
+          n_left_to_next -= 1;
+          n_left_from -= 1;
+
+          p0 = vlib_get_buffer (vm, pi0);
 
-	  p0 = vlib_get_buffer (vm, pi0);
+          mpls0 = vlib_buffer_get_current (p0);
+          lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
 
-	  hdr0 = vlib_buffer_get_current (p0);
-	  lbi0 = vnet_buffer (p0)->ip.adj_index[VLIB_TX];
+          lb0 = load_balance_get(lbi0);
 
-	  lb0 = load_balance_get(lbi0);
-	  hc0 = lb0->lb_hash_config;
-	  vnet_buffer(p0)->ip.flow_hash = mpls_compute_flow_hash(hdr0, hc0);
+          hc0 = vnet_buffer (p0)->ip.flow_hash = 0;
+          if (PREDICT_FALSE (lb0->lb_n_buckets > 1))
+          {
+              if (PREDICT_TRUE (vnet_buffer(p0)->ip.flow_hash))
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = vnet_buffer(p0)->ip.flow_hash >> 1;
+              }
+              else
+              {
+                  hc0 = vnet_buffer(p0)->ip.flow_hash = mpls_compute_flow_hash(mpls0, hc0);
+              }
+          }
 
-	  dpo0 = load_balance_get_bucket_i(lb0, 
-					   vnet_buffer(p0)->ip.flow_hash &
-					   (lb0->lb_n_buckets_minus_1));
+          dpo0 = load_balance_get_bucket_i(lb0, hc0 & (lb0->lb_n_buckets_minus_1));
 
-	  next0 = dpo0->dpoi_next_node;
-	  vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
+          next0 = dpo0->dpoi_next_node;
+          vnet_buffer (p0)->ip.adj_index[VLIB_TX] = dpo0->dpoi_index;
 
-	  vlib_increment_combined_counter 
+          vlib_increment_combined_counter
               (cm, cpu_index, lbi0, 1,
                vlib_buffer_length_in_chain (vm, p0));
 
-	  from += 1;
-	  to_next += 1;
-	  n_left_to_next -= 1;
-	  n_left_from -= 1;
-
-	  if (PREDICT_FALSE (next0 != next))
-	    {
-	      n_left_to_next += 1;
-	      vlib_put_next_frame (vm, node, next, n_left_to_next);
-	      next = next0;
-	      vlib_get_next_frame (vm, node, next,
-				   to_next, n_left_to_next);
-	      to_next[0] = pi0;
-	      to_next += 1;
-	      n_left_to_next -= 1;
-	    }
-	}
+          vlib_validate_buffer_enqueue_x1 (vm, node, next,
+                                           to_next, n_left_to_next,
+                                           pi0, next0);
+        }
 
       vlib_put_next_frame (vm, node, next, n_left_to_next);
     }