Use Quad-loop to improve gtpu-encap performance

Change-Id: I440fae09e77502392af5cde45dd9c532afa5d976
Signed-off-by: Hongjun Ni <hongjun.ni@intel.com>

1 files changed, 194 insertions, 49 deletions

diff --git a/src/plugins/gtpu/gtpu_encap.c b/src/plugins/gtpu/gtpu_encap.c
index adacf4c9a1a..8ad53c537a6 100644
--- a/src/plugins/gtpu/gtpu_encap.c
+++ b/src/plugins/gtpu/gtpu_encap.c
@@ -36,8 +36,15 @@ typedef enum {
     GTPU_ENCAP_N_ERROR,
 } gtpu_encap_error_t;
 
+#define foreach_gtpu_encap_next        \
+_(DROP, "error-drop")                  \
+_(IP4_LOOKUP, "ip4-lookup")             \
+_(IP6_LOOKUP, "ip6-lookup")
+
 typedef enum {
     GTPU_ENCAP_NEXT_DROP,
+    GTPU_ENCAP_NEXT_IP4_LOOKUP,
+    GTPU_ENCAP_NEXT_IP6_LOOKUP,
     GTPU_ENCAP_N_NEXT,
 } gtpu_encap_next_t;
 
@@ -76,13 +83,13 @@ gtpu_encap_inline (vlib_main_t * vm,
   vnet_main_t * vnm = gtm->vnet_main;
   vnet_interface_main_t * im = &vnm->interface_main;
   u32 pkts_encapsulated = 0;
-  u16 old_l0 = 0, old_l1 = 0;
+  u16 old_l0 = 0, old_l1 = 0, old_l2 = 0, old_l3 = 0;
   u32 thread_index = vlib_get_thread_index();
   u32 stats_sw_if_index, stats_n_packets, stats_n_bytes;
-  u32 sw_if_index0 = 0, sw_if_index1 = 0;
-  u32 next0 = 0, next1 = 0;
-  vnet_hw_interface_t * hi0, * hi1;
-  gtpu_tunnel_t * t0 = NULL, * t1 = NULL;
+  u32 sw_if_index0 = 0, sw_if_index1 = 0, sw_if_index2 = 0, sw_if_index3 = 0;
+  u32 next0 = 0, next1 = 0, next2 = 0, next3 = 0;
+  vnet_hw_interface_t * hi0, * hi1, * hi2, * hi3;
+  gtpu_tunnel_t * t0 = NULL, * t1 = NULL, * t2 = NULL, * t3 = NULL;
 
   from = vlib_frame_vector_args (from_frame);
   n_left_from = from_frame->n_vectors;
@@ -98,82 +105,117 @@ gtpu_encap_inline (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)
+      while (n_left_from >= 8 && n_left_to_next >= 4)
 	{
-          u32 bi0, bi1;
-	  vlib_buffer_t * b0, * b1;
-          u32 flow_hash0, flow_hash1;
-	  u32 len0, len1;
-          ip4_header_t * ip4_0, * ip4_1;
-          ip6_header_t * ip6_0, * ip6_1;
-          udp_header_t * udp0, * udp1;
-          gtpu_header_t * gtpu0, * gtpu1;
+          u32 bi0, bi1, bi2, bi3;
+	  vlib_buffer_t * b0, * b1, * b2, * b3;
+          u32 flow_hash0, flow_hash1, flow_hash2, flow_hash3;
+	  u32 len0, len1, len2, len3;
+          ip4_header_t * ip4_0, * ip4_1, * ip4_2, * ip4_3;
+          ip6_header_t * ip6_0, * ip6_1, * ip6_2, * ip6_3;
+          udp_header_t * udp0, * udp1, * udp2, * udp3;
+          gtpu_header_t * gtpu0, * gtpu1, * gtpu2, * gtpu3;
           u64 * copy_src0, * copy_dst0;
           u64 * copy_src1, * copy_dst1;
+          u64 * copy_src2, * copy_dst2;
+          u64 * copy_src3, * copy_dst3;
           u32 * copy_src_last0, * copy_dst_last0;
           u32 * copy_src_last1, * copy_dst_last1;
-          u16 new_l0, new_l1;
-          ip_csum_t sum0, sum1;
+          u32 * copy_src_last2, * copy_dst_last2;
+          u32 * copy_src_last3, * copy_dst_last3;
+          u16 new_l0, new_l1, new_l2, new_l3;
+          ip_csum_t sum0, sum1, sum2, sum3;
 
 	  /* 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, LOAD);
-	    vlib_prefetch_buffer_header (p3, LOAD);
-
-	    CLIB_PREFETCH (p2->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
-	    CLIB_PREFETCH (p3->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
+	    vlib_buffer_t * p4, * p5, * p6, * p7;
+
+	    p4 = vlib_get_buffer (vm, from[4]);
+	    p5 = vlib_get_buffer (vm, from[5]);
+	    p6 = vlib_get_buffer (vm, from[6]);
+	    p7 = vlib_get_buffer (vm, from[7]);
+
+	    vlib_prefetch_buffer_header (p4, LOAD);
+	    vlib_prefetch_buffer_header (p5, LOAD);
+	    vlib_prefetch_buffer_header (p6, LOAD);
+	    vlib_prefetch_buffer_header (p7, LOAD);
+
+	    CLIB_PREFETCH (p4->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
+	    CLIB_PREFETCH (p5->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
+	    CLIB_PREFETCH (p6->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
+	    CLIB_PREFETCH (p7->data, 2*CLIB_CACHE_LINE_BYTES, LOAD);
 	  }
 
 	  bi0 = from[0];
 	  bi1 = from[1];
+	  bi2 = from[2];
+	  bi3 = from[3];
 	  to_next[0] = bi0;
 	  to_next[1] = bi1;
-	  from += 2;
-	  to_next += 2;
-	  n_left_to_next -= 2;
-	  n_left_from -= 2;
+	  to_next[2] = bi2;
+	  to_next[3] = bi3;
+	  from += 4;
+	  to_next += 4;
+	  n_left_to_next -= 4;
+	  n_left_from -= 4;
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b1 = vlib_get_buffer (vm, bi1);
+	  b2 = vlib_get_buffer (vm, bi2);
+	  b3 = vlib_get_buffer (vm, bi3);
 
           flow_hash0 = vnet_l2_compute_flow_hash (b0);
           flow_hash1 = vnet_l2_compute_flow_hash (b1);
+          flow_hash2 = vnet_l2_compute_flow_hash (b2);
+          flow_hash3 = vnet_l2_compute_flow_hash (b3);
 
 	  /* Get next node index and adj index from tunnel next_dpo */
 	  sw_if_index0 = vnet_buffer(b0)->sw_if_index[VLIB_TX];
+	  sw_if_index1 = vnet_buffer(b1)->sw_if_index[VLIB_TX];
+	  sw_if_index2 = vnet_buffer(b2)->sw_if_index[VLIB_TX];
+	  sw_if_index3 = vnet_buffer(b3)->sw_if_index[VLIB_TX];
 	  hi0 = vnet_get_sup_hw_interface (vnm, sw_if_index0);
+	  hi1 = vnet_get_sup_hw_interface (vnm, sw_if_index1);
+	  hi2 = vnet_get_sup_hw_interface (vnm, sw_if_index2);
+	  hi3 = vnet_get_sup_hw_interface (vnm, sw_if_index3);
 	  t0 = &gtm->tunnels[hi0->dev_instance];
+	  t1 = &gtm->tunnels[hi1->dev_instance];
+	  t2 = &gtm->tunnels[hi2->dev_instance];
+	  t3 = &gtm->tunnels[hi3->dev_instance];
+
 	  /* Note: change to always set next0 if it may be set to drop */
 	  next0 = t0->next_dpo.dpoi_next_node;
           vnet_buffer(b0)->ip.adj_index[VLIB_TX] = t0->next_dpo.dpoi_index;
-
-	  /* Get next node index and adj index from tunnel next_dpo */
-	  sw_if_index1 = vnet_buffer(b1)->sw_if_index[VLIB_TX];
-	  hi1 = vnet_get_sup_hw_interface (vnm, sw_if_index1);
-	  t1 = &gtm->tunnels[hi1->dev_instance];
-	  /* Note: change to always set next1 if it may be set to drop */
 	  next1 = t1->next_dpo.dpoi_next_node;
           vnet_buffer(b1)->ip.adj_index[VLIB_TX] = t1->next_dpo.dpoi_index;
+	  next2 = t2->next_dpo.dpoi_next_node;
+          vnet_buffer(b2)->ip.adj_index[VLIB_TX] = t2->next_dpo.dpoi_index;
+	  next3 = t3->next_dpo.dpoi_next_node;
+          vnet_buffer(b3)->ip.adj_index[VLIB_TX] = t3->next_dpo.dpoi_index;
 
           /* Apply the rewrite string. $$$$ vnet_rewrite? */
           vlib_buffer_advance (b0, -(word)_vec_len(t0->rewrite));
           vlib_buffer_advance (b1, -(word)_vec_len(t1->rewrite));
+          vlib_buffer_advance (b2, -(word)_vec_len(t2->rewrite));
+          vlib_buffer_advance (b3, -(word)_vec_len(t3->rewrite));
 
 	  if (is_ip4)
 	    {
 	      ip4_0 = vlib_buffer_get_current(b0);
 	      ip4_1 = vlib_buffer_get_current(b1);
+	      ip4_2 = vlib_buffer_get_current(b2);
+	      ip4_3 = vlib_buffer_get_current(b3);
 
 	      /* Copy the fixed header */
 	      copy_dst0 = (u64 *) ip4_0;
 	      copy_src0 = (u64 *) t0->rewrite;
 	      copy_dst1 = (u64 *) ip4_1;
 	      copy_src1 = (u64 *) t1->rewrite;
+	      copy_dst2 = (u64 *) ip4_2;
+	      copy_src2 = (u64 *) t2->rewrite;
+	      copy_dst3 = (u64 *) ip4_3;
+	      copy_src3 = (u64 *) t3->rewrite;
+
 	      /* Copy first 32 octets 8-bytes at a time */
 #define _(offs) copy_dst0[offs] = copy_src0[offs];
 	      foreach_fixed_header4_offset;
@@ -181,6 +223,12 @@ gtpu_encap_inline (vlib_main_t * vm,
 #define _(offs) copy_dst1[offs] = copy_src1[offs];
 	      foreach_fixed_header4_offset;
 #undef _
+#define _(offs) copy_dst2[offs] = copy_src2[offs];
+	      foreach_fixed_header4_offset;
+#undef _
+#define _(offs) copy_dst3[offs] = copy_src3[offs];
+	      foreach_fixed_header4_offset;
+#undef _
 	      /* Last 4 octets. Hopefully gcc will be our friend */
               copy_dst_last0 = (u32 *)(&copy_dst0[4]);
               copy_src_last0 = (u32 *)(&copy_src0[4]);
@@ -188,6 +236,12 @@ gtpu_encap_inline (vlib_main_t * vm,
               copy_dst_last1 = (u32 *)(&copy_dst1[4]);
               copy_src_last1 = (u32 *)(&copy_src1[4]);
               copy_dst_last1[0] = copy_src_last1[0];
+              copy_dst_last2 = (u32 *)(&copy_dst2[4]);
+              copy_src_last2 = (u32 *)(&copy_src2[4]);
+              copy_dst_last2[0] = copy_src_last2[0];
+              copy_dst_last3 = (u32 *)(&copy_dst3[4]);
+              copy_src_last3 = (u32 *)(&copy_src3[4]);
+              copy_dst_last3[0] = copy_src_last3[0];
 
 	      /* Fix the IP4 checksum and length */
 	      sum0 = ip4_0->checksum;
@@ -204,6 +258,20 @@ gtpu_encap_inline (vlib_main_t * vm,
 				     length /* changed member */);
 	      ip4_1->checksum = ip_csum_fold (sum1);
 	      ip4_1->length = new_l1;
+	      sum2 = ip4_2->checksum;
+	      new_l2 = /* old_l0 always 0, see the rewrite setup */
+                clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b2));
+              sum2 = ip_csum_update (sum2, old_l2, new_l2, ip4_header_t,
+				     length /* changed member */);
+	      ip4_2->checksum = ip_csum_fold (sum2);
+	      ip4_2->length = new_l2;
+	      sum3 = ip4_3->checksum;
+	      new_l3 = /* old_l1 always 0, see the rewrite setup */
+                clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b3));
+              sum3 = ip_csum_update (sum3, old_l3, new_l3, ip4_header_t,
+				     length /* changed member */);
+	      ip4_3->checksum = ip_csum_fold (sum3);
+	      ip4_3->length = new_l3;
 
 	      /* Fix UDP length and set source port */
 	      udp0 = (udp_header_t *)(ip4_0+1);
@@ -216,6 +284,16 @@ gtpu_encap_inline (vlib_main_t * vm,
 					     - sizeof (*ip4_1));
 	      udp1->length = new_l1;
 	      udp1->src_port = flow_hash1;
+	      udp2 = (udp_header_t *)(ip4_2+1);
+	      new_l2 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b2)
+					     - sizeof (*ip4_2));
+	      udp2->length = new_l2;
+	      udp2->src_port = flow_hash2;
+	      udp3 = (udp_header_t *)(ip4_3+1);
+	      new_l3 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b3)
+					     - sizeof (*ip4_3));
+	      udp3->length = new_l3;
+	      udp3->src_port = flow_hash3;
 
 	      /* Fix GTPU length */
 	      gtpu0 = (gtpu_header_t *)(udp0+1);
@@ -226,6 +304,14 @@ gtpu_encap_inline (vlib_main_t * vm,
 	      new_l1 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b1)
 					     - sizeof (*ip4_1) - sizeof(*udp1));
 	      gtpu1->length = new_l1;
+	      gtpu2 = (gtpu_header_t *)(udp2+1);
+	      new_l2 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b2)
+					     - sizeof (*ip4_2) - sizeof(*udp2));
+	      gtpu2->length = new_l2;
+	      gtpu3 = (gtpu_header_t *)(udp1+3);
+	      new_l3 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b3)
+					     - sizeof (*ip4_3) - sizeof(*udp3));
+	      gtpu3->length = new_l3;
 	    }
 	  else /* ipv6 */
 	    {
@@ -233,12 +319,18 @@ gtpu_encap_inline (vlib_main_t * vm,
 
 	      ip6_0 = vlib_buffer_get_current(b0);
 	      ip6_1 = vlib_buffer_get_current(b1);
+	      ip6_2 = vlib_buffer_get_current(b2);
+	      ip6_3 = vlib_buffer_get_current(b3);
 
 	      /* Copy the fixed header */
 	      copy_dst0 = (u64 *) ip6_0;
 	      copy_src0 = (u64 *) t0->rewrite;
 	      copy_dst1 = (u64 *) ip6_1;
 	      copy_src1 = (u64 *) t1->rewrite;
+	      copy_dst2 = (u64 *) ip6_2;
+	      copy_src2 = (u64 *) t2->rewrite;
+	      copy_dst3 = (u64 *) ip6_3;
+	      copy_src3 = (u64 *) t3->rewrite;
 	      /* Copy first 56 (ip6) octets 8-bytes at a time */
 #define _(offs) copy_dst0[offs] = copy_src0[offs];
 	      foreach_fixed_header6_offset;
@@ -246,6 +338,12 @@ gtpu_encap_inline (vlib_main_t * vm,
 #define _(offs) copy_dst1[offs] = copy_src1[offs];
 	      foreach_fixed_header6_offset;
 #undef _
+#define _(offs) copy_dst2[offs] = copy_src2[offs];
+	      foreach_fixed_header6_offset;
+#undef _
+#define _(offs) copy_dst3[offs] = copy_src3[offs];
+	      foreach_fixed_header6_offset;
+#undef _
 	      /* Fix IP6 payload length */
 	      new_l0 =
                 clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b0)
@@ -255,6 +353,14 @@ gtpu_encap_inline (vlib_main_t * vm,
                 clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b1)
 				      - sizeof(*ip6_1));
 	      ip6_1->payload_length = new_l1;
+	      new_l2 =
+                clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b2)
+				      - sizeof(*ip6_2));
+	      ip6_2->payload_length = new_l2;
+	      new_l3 =
+                clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, b3)
+				      - sizeof(*ip6_3));
+	      ip6_3->payload_length = new_l3;
 
 	      /* Fix UDP length  and set source port */
 	      udp0 = (udp_header_t *)(ip6_0+1);
@@ -263,6 +369,12 @@ gtpu_encap_inline (vlib_main_t * vm,
 	      udp1 = (udp_header_t *)(ip6_1+1);
 	      udp1->length = new_l1;
 	      udp1->src_port = flow_hash1;
+	      udp2 = (udp_header_t *)(ip6_2+1);
+	      udp2->length = new_l2;
+	      udp2->src_port = flow_hash2;
+	      udp3 = (udp_header_t *)(ip6_3+1);
+	      udp3->length = new_l3;
+	      udp3->src_port = flow_hash3;
 
 	      /* IPv6 UDP checksum is mandatory */
 	      udp0->checksum = ip6_tcp_udp_icmp_compute_checksum(vm, b0,
@@ -273,6 +385,14 @@ gtpu_encap_inline (vlib_main_t * vm,
                                                         ip6_1, &bogus);
 	      if (udp1->checksum == 0)
 		udp1->checksum = 0xffff;
+	      udp2->checksum = ip6_tcp_udp_icmp_compute_checksum(vm, b2,
+								 ip6_2, &bogus);
+	      if (udp2->checksum == 0)
+		udp2->checksum = 0xffff;
+	      udp3->checksum = ip6_tcp_udp_icmp_compute_checksum(vm, b3,
+                                                        ip6_3, &bogus);
+	      if (udp3->checksum == 0)
+		udp3->checksum = 0xffff;
 
 	      /* Fix GTPU length */
 	      gtpu0 = (gtpu_header_t *)(udp0+1);
@@ -283,24 +403,38 @@ gtpu_encap_inline (vlib_main_t * vm,
 	      new_l1 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b1)
 					     - sizeof (*ip4_1) - sizeof(*udp1));
 	      gtpu1->length = new_l1;
+	      gtpu2 = (gtpu_header_t *)(udp2+1);
+	      new_l2 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b2)
+					     - sizeof (*ip4_2) - sizeof(*udp2));
+	      gtpu2->length = new_l2;
+	      gtpu3 = (gtpu_header_t *)(udp3+1);
+	      new_l3 = clib_host_to_net_u16 (vlib_buffer_length_in_chain(vm, b3)
+					     - sizeof (*ip4_3) - sizeof(*udp3));
+	      gtpu3->length = new_l3;
 	    }
 
-          pkts_encapsulated += 2;
+          pkts_encapsulated += 4;
  	  len0 = vlib_buffer_length_in_chain (vm, b0);
  	  len1 = vlib_buffer_length_in_chain (vm, b1);
-	  stats_n_packets += 2;
-	  stats_n_bytes += len0 + len1;
+ 	  len2 = vlib_buffer_length_in_chain (vm, b2);
+ 	  len3 = vlib_buffer_length_in_chain (vm, b3);
+	  stats_n_packets += 4;
+	  stats_n_bytes += len0 + len1 + len2 + len3;
 
 	  /* Batch stats increment on the same gtpu tunnel so counter is not
 	     incremented per packet. Note stats are still incremented for deleted
 	     and admin-down tunnel where packets are dropped. It is not worthwhile
 	     to check for this rare case and affect normal path performance. */
 	  if (PREDICT_FALSE ((sw_if_index0 != stats_sw_if_index) ||
-			     (sw_if_index1 != stats_sw_if_index)))
+			     (sw_if_index1 != stats_sw_if_index) ||
+			     (sw_if_index2 != stats_sw_if_index) ||
+			     (sw_if_index3 != stats_sw_if_index) ))
 	    {
-	      stats_n_packets -= 2;
-	      stats_n_bytes -= len0 + len1;
-	      if (sw_if_index0 == sw_if_index1)
+	      stats_n_packets -= 4;
+	      stats_n_bytes -= len0 + len1 + len2 + len3;
+	      if ( (sw_if_index0 == sw_if_index1 ) &&
+		   (sw_if_index1 == sw_if_index2 ) &&
+		   (sw_if_index2 == sw_if_index3 ) )
 	        {
 		  if (stats_n_packets)
 		    vlib_increment_combined_counter
@@ -308,8 +442,8 @@ gtpu_encap_inline (vlib_main_t * vm,
 		       thread_index, stats_sw_if_index,
 		       stats_n_packets, stats_n_bytes);
 		  stats_sw_if_index = sw_if_index0;
-		  stats_n_packets = 2;
-		  stats_n_bytes = len0 + len1;
+		  stats_n_packets = 4;
+		  stats_n_bytes = len0 + len1 + len2 + len3;
 	        }
 	      else
 	        {
@@ -319,6 +453,12 @@ gtpu_encap_inline (vlib_main_t * vm,
 		  vlib_increment_combined_counter
 		      (im->combined_sw_if_counters + VNET_INTERFACE_COUNTER_TX,
 		       thread_index, sw_if_index1, 1, len1);
+		  vlib_increment_combined_counter
+		      (im->combined_sw_if_counters + VNET_INTERFACE_COUNTER_TX,
+		       thread_index, sw_if_index2, 1, len2);
+		  vlib_increment_combined_counter
+		      (im->combined_sw_if_counters + VNET_INTERFACE_COUNTER_TX,
+		       thread_index, sw_if_index3, 1, len3);
 		}
 	    }
 
@@ -338,9 +478,10 @@ gtpu_encap_inline (vlib_main_t * vm,
               tr->teid = t1->teid;
             }
 
-	  vlib_validate_buffer_enqueue_x2 (vm, node, next_index,
+	  vlib_validate_buffer_enqueue_x4 (vm, node, next_index,
 					   to_next, n_left_to_next,
-					   bi0, bi1, next0, next1);
+					   bi0, bi1, bi2, bi3,
+					   next0, next1, next2, next3);
 	}
 
       while (n_left_from > 0 && n_left_to_next > 0)
@@ -536,7 +677,9 @@ VLIB_REGISTER_NODE (gtpu4_encap_node) = {
   .error_strings = gtpu_encap_error_strings,
   .n_next_nodes = GTPU_ENCAP_N_NEXT,
   .next_nodes = {
-        [GTPU_ENCAP_NEXT_DROP] = "error-drop",
+#define _(s,n) [GTPU_ENCAP_NEXT_##s] = n,
+    foreach_gtpu_encap_next
+#undef _
   },
 };
 
@@ -552,7 +695,9 @@ VLIB_REGISTER_NODE (gtpu6_encap_node) = {
   .error_strings = gtpu_encap_error_strings,
   .n_next_nodes = GTPU_ENCAP_N_NEXT,
   .next_nodes = {
-        [GTPU_ENCAP_NEXT_DROP] = "error-drop",
+#define _(s,n) [GTPU_ENCAP_NEXT_##s] = n,
+    foreach_gtpu_encap_next
+#undef _
   },
 };