From 86b1871ba212064ceb985be4a6b655ebfe2e32f9 Mon Sep 17 00:00:00 2001
From: "Lijian.Zhang" <Lijian.Zhang@arm.com>
Date: Mon, 8 Jul 2019 10:33:34 +0800
Subject: ip: apply dual loop unrolling in ip4_input

Too many prefetches within loop unrollings induce bottleneck and
performance degradation on some CPUs which have less cache line fill
buffers, e.g, Arm Cortex-A72.
Apply dual loop unrolling and tune prefetches manually to resolve
hot-spot with prefetch instructions.
It saves about 11.5% cycles with ip4_input node on Cortex-A72 CPUs.

Type: feature

Change-Id: I1ac9eb21061a804af2a414b420217fbcda3689c9
Signed-off-by: Lijian Zhang <Lijian.Zhang@arm.com>
---
 src/vnet/ip/ip4_input.c | 68 +++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)

(limited to 'src/vnet/ip')

diff --git a/src/vnet/ip/ip4_input.c b/src/vnet/ip/ip4_input.c
index 91766b4a6c1..94c4aac4271 100644
--- a/src/vnet/ip/ip4_input.c
+++ b/src/vnet/ip/ip4_input.c
@@ -151,6 +151,7 @@ ip4_input_inline (vlib_main_t * vm,
   vlib_get_buffers (vm, from, bufs, n_left_from);
   b = bufs;
   next = nexts;
+#if (CLIB_N_PREFETCHES >= 8)
   while (n_left_from >= 4)
     {
       u32 x = 0;
@@ -233,6 +234,73 @@ ip4_input_inline (vlib_main_t * vm,
       next += 4;
       n_left_from -= 4;
     }
+#elif (CLIB_N_PREFETCHES >= 4)
+  while (n_left_from >= 2)
+    {
+      u32 x = 0;
+      u32 next0, next1;
+
+      /* Prefetch next iteration. */
+      if (n_left_from >= 6)
+	{
+	  vlib_prefetch_buffer_header (b[4], LOAD);
+	  vlib_prefetch_buffer_header (b[5], LOAD);
+
+	  vlib_prefetch_buffer_data (b[2], LOAD);
+	  vlib_prefetch_buffer_data (b[3], LOAD);
+	}
+
+      vnet_buffer (b[0])->ip.adj_index[VLIB_RX] = ~0;
+      vnet_buffer (b[1])->ip.adj_index[VLIB_RX] = ~0;
+
+      sw_if_index[0] = vnet_buffer (b[0])->sw_if_index[VLIB_RX];
+      sw_if_index[1] = vnet_buffer (b[1])->sw_if_index[VLIB_RX];
+
+      x |= sw_if_index[0] ^ last_sw_if_index;
+      x |= sw_if_index[1] ^ last_sw_if_index;
+
+      if (PREDICT_TRUE (x == 0))
+	{
+	  /* we deal with 2 more packets sharing the same sw_if_index
+	     with the previous one, so we can optimize */
+	  cnt += 2;
+	  if (arc_enabled)
+	    {
+	      next0 = ip4_input_set_next (sw_if_index[0], b[0], 1);
+	      next1 = ip4_input_set_next (sw_if_index[1], b[1], 1);
+	    }
+	  else
+	    {
+	      next0 = ip4_input_set_next (sw_if_index[0], b[0], 0);
+	      next1 = ip4_input_set_next (sw_if_index[1], b[1], 0);
+	    }
+	}
+      else
+	{
+	  ip4_input_check_sw_if_index (vm, cm, sw_if_index[0],
+				       &last_sw_if_index, &cnt, &arc_enabled);
+	  ip4_input_check_sw_if_index (vm, cm, sw_if_index[1],
+				       &last_sw_if_index, &cnt, &arc_enabled);
+
+	  next0 = ip4_input_set_next (sw_if_index[0], b[0], 1);
+	  next1 = ip4_input_set_next (sw_if_index[1], b[1], 1);
+	}
+
+      ip[0] = vlib_buffer_get_current (b[0]);
+      ip[1] = vlib_buffer_get_current (b[1]);
+
+      ip4_input_check_x2 (vm, error_node, b[0], b[1], ip[0], ip[1],
+			  &next0, &next1, verify_checksum);
+      next[0] = (u16) next0;
+      next[1] = (u16) next1;
+
+      /* next */
+      b += 2;
+      next += 2;
+      n_left_from -= 2;
+    }
+#endif
+
   while (n_left_from)
     {
       u32 next0;
-- 
cgit 1.2.3-korg