Simplify adjacency rewrite code

Using memcpy instead of complex specific copy logic. This simplify
the implementation and also improve perf slightly.
Also move adjacency data from tail to head of buffer, which improves
cache locality (header and data share the same cacheline)
Finally, fix VxLAN which used to workaround vnet_rewrite logic.

Change-Id: I770ddad9846f7ee505aa99ad417e6a61d5cbbefa
Signed-off-by: Benoît Ganne <bganne@cisco.com>

1 files changed, 23 insertions, 115 deletions

diff --git a/src/vnet/adj/rewrite.h b/src/vnet/adj/rewrite.h
index 32781134ecc..05a722fc764 100644
--- a/src/vnet/adj/rewrite.h
+++ b/src/vnet/adj/rewrite.h
@@ -131,8 +131,8 @@ vnet_rewrite_set_data_internal (vnet_rewrite_header_t * rw,
   ASSERT ((data_bytes >= 0) && (data_bytes < max_size));
 
   rw->data_bytes = data_bytes;
-  clib_memcpy_fast (rw->data + max_size - data_bytes, data, data_bytes);
-  clib_memset (rw->data, 0xfe, max_size - data_bytes);
+  clib_memcpy_fast (rw->data, data, data_bytes);
+  clib_memset (rw->data + data_bytes, 0xfe, max_size - data_bytes);
 }
 
 #define vnet_rewrite_set_data(rw,data,data_bytes)		\
@@ -145,76 +145,28 @@ always_inline void *
 vnet_rewrite_get_data_internal (vnet_rewrite_header_t * rw, int max_size)
 {
   ASSERT (rw->data_bytes <= max_size);
-  return rw->data + max_size - rw->data_bytes;
+  return rw->data;
 }
 
 #define vnet_rewrite_get_data(rw) \
   vnet_rewrite_get_data_internal (&((rw).rewrite_header), sizeof ((rw).rewrite_data))
 
 always_inline void
-vnet_rewrite_copy_one (vnet_rewrite_data_t * p0, vnet_rewrite_data_t * rw0,
-		       int i)
-{
-  p0[-i] = rw0[-i];
-}
-
-void vnet_rewrite_copy_slow_path (vnet_rewrite_data_t * p0,
-				  vnet_rewrite_data_t * rw0,
-				  word n_left, uword most_likely_size);
-
-/* *INDENT-OFF* */
-typedef CLIB_PACKED (struct {
-  u64 a;
-  u32 b;
-  u16 c;
-}) eh_copy_t;
-/* *INDENT-ON* */
-
-always_inline void
 _vnet_rewrite_one_header (vnet_rewrite_header_t * h0,
 			  void *packet0, int max_size, int most_likely_size)
 {
-  vnet_rewrite_data_t *p0 = packet0;
-  vnet_rewrite_data_t *rw0 = (vnet_rewrite_data_t *) (h0->data + max_size);
-  word n_left0;
-
   /* 0xfefe => poisoned adjacency => crash */
   ASSERT (h0->data_bytes != 0xfefe);
-
-  if (PREDICT_TRUE (h0->data_bytes == sizeof (eh_copy_t)))
+  if (PREDICT_TRUE (most_likely_size == h0->data_bytes))
     {
-      eh_copy_t *s, *d;
-      s = (eh_copy_t *) (h0->data + max_size - sizeof (eh_copy_t));
-      d = (eh_copy_t *) (((u8 *) packet0) - sizeof (eh_copy_t));
-      clib_memcpy (d, s, sizeof (eh_copy_t));
-      return;
+      clib_memcpy_fast ((u8 *) packet0 - most_likely_size,
+			h0->data, most_likely_size);
+    }
+  else
+    {
+      clib_memcpy_fast ((u8 *) packet0 - h0->data_bytes,
+			h0->data, h0->data_bytes);
     }
-  /*
-   * Stop now if the data_bytes field is zero, to avoid the cache
-   * miss consequences of spraying [functionally harmless] junk into
-   * un-prefetched rewrite space.
-   */
-  if (PREDICT_FALSE (h0->data_bytes == 0))
-    return;
-
-#define _(i)								\
-  do {									\
-    if (most_likely_size > ((i)-1)*sizeof (vnet_rewrite_data_t))	\
-      vnet_rewrite_copy_one (p0, rw0, (i));				\
-  } while (0)
-
-  _(4);
-  _(3);
-  _(2);
-  _(1);
-
-#undef _
-
-  n_left0 = (int)
-    (((int) h0->data_bytes - most_likely_size) + (sizeof (rw0[0]) - 1))
-    / (int) sizeof (rw0[0]);
-  if (PREDICT_FALSE (n_left0 > 0))
-    vnet_rewrite_copy_slow_path (p0, rw0, n_left0, most_likely_size);
 }
 
 always_inline void
@@ -223,68 +175,24 @@ _vnet_rewrite_two_headers (vnet_rewrite_header_t * h0,
 			   void *packet0,
 			   void *packet1, int max_size, int most_likely_size)
 {
-  vnet_rewrite_data_t *p0 = packet0;
-  vnet_rewrite_data_t *p1 = packet1;
-  vnet_rewrite_data_t *rw0 = (vnet_rewrite_data_t *) (h0->data + max_size);
-  vnet_rewrite_data_t *rw1 = (vnet_rewrite_data_t *) (h1->data + max_size);
-  word n_left0, n_left1;
-  int slow_path;
-
   /* 0xfefe => poisoned adjacency => crash */
   ASSERT (h0->data_bytes != 0xfefe);
   ASSERT (h1->data_bytes != 0xfefe);
-
-  /* Arithmetic calculation: bytes0 == bytes1 == 14 */
-  slow_path = h0->data_bytes ^ h1->data_bytes;
-  slow_path += h0->data_bytes ^ sizeof (eh_copy_t);
-
-  if (PREDICT_TRUE (slow_path == 0))
+  if (PREDICT_TRUE
+      (most_likely_size == h0->data_bytes
+       && most_likely_size == h1->data_bytes))
     {
-      eh_copy_t *s0, *d0, *s1, *d1;
-      s0 = (eh_copy_t *) (h0->data + max_size - sizeof (eh_copy_t));
-      d0 = (eh_copy_t *) (((u8 *) packet0) - sizeof (eh_copy_t));
-      clib_memcpy (d0, s0, sizeof (eh_copy_t));
-      s1 = (eh_copy_t *) (h1->data + max_size - sizeof (eh_copy_t));
-      d1 = (eh_copy_t *) (((u8 *) packet1) - sizeof (eh_copy_t));
-      clib_memcpy (d1, s1, sizeof (eh_copy_t));
-      return;
+      clib_memcpy_fast ((u8 *) packet0 - most_likely_size,
+			h0->data, most_likely_size);
+      clib_memcpy_fast ((u8 *) packet1 - most_likely_size,
+			h1->data, most_likely_size);
     }
-
-  /*
-   * Stop now if both rewrite data_bytes fields are zero, to avoid the cache
-   * miss consequences of spraying [functionally harmless] junk into
-   * un-prefetched rewrite space.
-   */
-  if (PREDICT_FALSE (h0->data_bytes + h1->data_bytes == 0))
-    return;
-
-#define _(i)								\
-  do {									\
-    if (most_likely_size > ((i)-1)*sizeof (vnet_rewrite_data_t))	\
-      {									\
-	vnet_rewrite_copy_one (p0, rw0, (i));				\
-	vnet_rewrite_copy_one (p1, rw1, (i));				\
-      }									\
-  } while (0)
-
-  _(4);
-  _(3);
-  _(2);
-  _(1);
-
-#undef _
-
-  n_left0 = (int)
-    (((int) h0->data_bytes - most_likely_size) + (sizeof (rw0[0]) - 1))
-    / (int) sizeof (rw0[0]);
-  n_left1 = (int)
-    (((int) h1->data_bytes - most_likely_size) + (sizeof (rw1[0]) - 1))
-    / (int) sizeof (rw1[0]);
-
-  if (PREDICT_FALSE (n_left0 > 0 || n_left1 > 0))
+  else
     {
-      vnet_rewrite_copy_slow_path (p0, rw0, n_left0, most_likely_size);
-      vnet_rewrite_copy_slow_path (p1, rw1, n_left1, most_likely_size);
+      clib_memcpy_fast ((u8 *) packet0 - h0->data_bytes,
+			h0->data, h0->data_bytes);
+      clib_memcpy_fast ((u8 *) packet1 - h1->data_bytes,
+			h1->data, h1->data_bytes);
     }
 }