Make adjacencies shareable

Change-Id: I620871ca715b751d2e487f37341b7118797c9176
Signed-off-by: Damjan Marion <damarion@cisco.com>

5 files changed, 235 insertions, 11 deletions

diff --git a/vnet/vnet/ip/ip4_forward.c b/vnet/vnet/ip/ip4_forward.c
index 45bc22e7550..c5b3e9a5a0a 100644
--- a/vnet/vnet/ip/ip4_forward.c
+++ b/vnet/vnet/ip/ip4_forward.c
@@ -380,6 +380,14 @@ void ip4_add_del_route (ip4_main_t * im, ip4_add_del_route_args_t * a)
 
   old_adj_index = fib->old_hash_values[0];
 
+  /* Avoid spurious reference count increments */
+  if (old_adj_index == adj_index)
+    {
+      ip_adjacency_t * adj = ip_get_adjacency (lm, adj_index);
+      if (adj->share_count > 0)
+        adj->share_count --;
+    }
+
   ip4_fib_mtrie_add_del_route (fib, a->dst_address, dst_address_length,
 			       is_del ? old_adj_index : adj_index,
 			       is_del);
diff --git a/vnet/vnet/ip/ip6_forward.c b/vnet/vnet/ip/ip6_forward.c
index fd6874d4def..1d0e21e528c 100644
--- a/vnet/vnet/ip/ip6_forward.c
+++ b/vnet/vnet/ip/ip6_forward.c
@@ -265,6 +265,14 @@ void ip6_add_del_route (ip6_main_t * im, ip6_add_del_route_args_t * a)
       BV(clib_bihash_add_del) (&im->ip6_lookup_table, &kv, 1 /* is_add */);
     }
 
+  /* Avoid spurious reference count increments */
+  if (old_adj_index == adj_index)
+    {
+      ip_adjacency_t * adj = ip_get_adjacency (lm, adj_index);
+      if (adj->share_count > 0)
+        adj->share_count --;
+    }
+
   /* Delete old adjacency index if present and changed. */
   {
     if (! (a->flags & IP6_ROUTE_FLAG_KEEP_OLD_ADJACENCY)
diff --git a/vnet/vnet/ip/lookup.c b/vnet/vnet/ip/lookup.c
index d0ec4947d7a..9e34bfa9064 100644
--- a/vnet/vnet/ip/lookup.c
+++ b/vnet/vnet/ip/lookup.c
@@ -60,6 +60,55 @@ ip_add_adjacency (ip_lookup_main_t * lm,
   ip_adjacency_t * adj;
   u32 ai, i, handle;
 
+  /* See if we know enough to attempt to share an existing adjacency */
+  if (copy_adj && n_adj == 1)
+    {
+      uword signature;
+      uword * p;
+
+      switch (copy_adj->lookup_next_index)
+        {
+        case IP_LOOKUP_NEXT_DROP:
+          if (lm->drop_adj_index)
+            {
+              adj = ip_get_adjacency (lm, lm->drop_adj_index);
+              *adj_index_return = lm->drop_adj_index;
+              return (adj);
+            }
+          break;
+
+        case IP_LOOKUP_NEXT_LOCAL:
+          if (lm->local_adj_index)
+            {
+              adj = ip_get_adjacency (lm, lm->local_adj_index);
+              *adj_index_return = lm->local_adj_index;
+              return (adj);
+            }
+        default:
+          break;
+        }
+
+      signature = vnet_ip_adjacency_signature (copy_adj);
+      p = hash_get (lm->adj_index_by_signature, signature);
+      if (p)
+        {
+          adj = heap_elt_at_index (lm->adjacency_heap, p[0]);
+          while (1)
+            {
+              if (vnet_ip_adjacency_share_compare (adj, copy_adj))
+                {
+                  adj->share_count++;
+                  *adj_index_return = p[0];
+                  return adj;
+                }
+              if (adj->next_adj_with_signature == 0)
+                break;
+              adj = heap_elt_at_index (lm->adjacency_heap,
+                                       adj->next_adj_with_signature);
+            }
+        }
+    }
+
   ai = heap_alloc (lm->adjacency_heap, n_adj, handle);
   adj = heap_elt_at_index (lm->adjacency_heap, ai);
 
@@ -82,11 +131,37 @@ ip_add_adjacency (ip_lookup_main_t * lm,
 
       adj[i].heap_handle = handle;
       adj[i].n_adj = n_adj;
+      adj[i].share_count = 0;
+      adj[i].next_adj_with_signature = 0;
 
       /* Zero possibly stale counters for re-used adjacencies. */
       vlib_zero_combined_counter (&lm->adjacency_counters, ai + i);
     }
 
+  /* Set up to share the adj later */
+  if (copy_adj && n_adj == 1)
+    {
+      uword * p;
+      u32 old_ai;
+      uword signature = vnet_ip_adjacency_signature (adj);
+
+      p = hash_get (lm->adj_index_by_signature, signature);
+      /* Hash collision? */
+      if (p)
+        {
+          /* Save the adj index, p[0] will be toast after the unset! */
+          old_ai = p[0];
+          hash_unset (lm->adj_index_by_signature, signature);
+          hash_set (lm->adj_index_by_signature, signature, ai);
+          adj->next_adj_with_signature = old_ai;
+        }
+      else
+        {
+          adj->next_adj_with_signature = 0;
+          hash_set (lm->adj_index_by_signature, signature, ai);
+        }
+    }
+
   *adj_index_return = ai;
   return adj;
 }
@@ -101,6 +176,69 @@ static void ip_del_adjacency2 (ip_lookup_main_t * lm, u32 adj_index, u32 delete_
   adj = ip_get_adjacency (lm, adj_index);
   handle = adj->heap_handle;
 
+  /* Special-case local, drop adjs */
+  switch (adj->lookup_next_index)
+    {
+    case IP_LOOKUP_NEXT_LOCAL:
+    case IP_LOOKUP_NEXT_DROP:
+      return;
+    default:
+      break;
+    }
+
+
+  if (adj->n_adj == 1)
+    {
+      uword signature;
+      uword * p;
+      u32 this_ai;
+      ip_adjacency_t * this_adj, * prev_adj = 0;
+      if (adj->share_count > 0)
+        {
+          adj->share_count --;
+          return;
+        }
+
+      signature = vnet_ip_adjacency_signature (adj);
+      p = hash_get (lm->adj_index_by_signature, signature);
+      if (p == 0)
+        {
+          clib_warning ("adj 0x%llx signature %llx not in table",
+                        adj, signature);
+          goto bag_it;
+        }
+      this_ai = p[0];
+      /* At the top of the signature chain (likely)? */
+      if (this_ai == adj_index)
+        {
+          if (adj->next_adj_with_signature == 0)
+            {
+              hash_unset (lm->adj_index_by_signature, signature);
+              goto bag_it;
+            }
+          else
+            {
+              this_adj = ip_get_adjacency (lm, adj->next_adj_with_signature);
+              hash_unset (lm->adj_index_by_signature, signature);
+              hash_set (lm->adj_index_by_signature, signature,
+                        this_adj->heap_handle);
+            }
+        }
+      else                      /* walk signature chain */
+        {
+          this_adj = ip_get_adjacency (lm, this_ai);
+          while (this_adj != adj)
+            {
+              prev_adj = this_adj;
+              this_adj = ip_get_adjacency
+                (lm, this_adj->next_adj_with_signature);
+              ASSERT(this_adj->heap_handle != 0);
+            }
+          prev_adj->next_adj_with_signature = this_adj->next_adj_with_signature;
+        }
+    }
+
+ bag_it:
   if (delete_multipath_adjacency)
     ip_multipath_del_adjacency (lm, adj_index);
 
@@ -829,6 +967,14 @@ void unserialize_ip_lookup_main (serialize_main_t * m, va_list * va)
 void ip_lookup_init (ip_lookup_main_t * lm, u32 is_ip6)
 {
   ip_adjacency_t * adj;
+  ip_adjacency_t template_adj;
+
+  /* ensure that adjacency is cacheline aligned and sized */
+  ASSERT(STRUCT_OFFSET_OF(ip_adjacency_t, cacheline0) == 0);
+  ASSERT(STRUCT_OFFSET_OF(ip_adjacency_t, cacheline1) == CLIB_CACHE_LINE_BYTES);
+
+  lm->adj_index_by_signature = hash_create (0, sizeof (uword));
+  memset (&template_adj, 0, sizeof (template_adj));
 
   /* Hand-craft special miss adjacency to use when nothing matches in the
      routing table.  Same for drop adjacency. */
@@ -836,12 +982,14 @@ void ip_lookup_init (ip_lookup_main_t * lm, u32 is_ip6)
   adj->lookup_next_index = IP_LOOKUP_NEXT_MISS;
   ASSERT (lm->miss_adj_index == IP_LOOKUP_MISS_ADJ_INDEX);
 
-  adj = ip_add_adjacency (lm, /* template */ 0, /* n-adj */ 1, &lm->drop_adj_index);
-  adj->lookup_next_index = IP_LOOKUP_NEXT_DROP;
+  /* Make the "drop" adj sharable */
+  template_adj.lookup_next_index = IP_LOOKUP_NEXT_DROP;
+  adj = ip_add_adjacency (lm, &template_adj, /* n-adj */ 1, &lm->drop_adj_index);
 
-  adj = ip_add_adjacency (lm, /* template */ 0, /* n-adj */ 1, &lm->local_adj_index);
-  adj->lookup_next_index = IP_LOOKUP_NEXT_LOCAL;
-  adj->if_address_index = ~0;
+  /* Make the "local" adj sharable */
+  template_adj.lookup_next_index = IP_LOOKUP_NEXT_LOCAL;
+  template_adj.if_address_index = ~0;
+  adj = ip_add_adjacency (lm, &template_adj, /* n-adj */ 1, &lm->local_adj_index);
 
   if (! lm->fib_result_n_bytes)
     lm->fib_result_n_bytes = sizeof (uword);
@@ -983,6 +1131,10 @@ u8 * format_ip_adjacency (u8 * s, va_list * args)
     }
   if (adj->explicit_fib_index != ~0 && adj->explicit_fib_index != 0)
     s = format (s, " lookup fib index %d", adj->explicit_fib_index);
+  if (adj->share_count > 0)
+    s = format (s, " shared %d", adj->share_count + 1);
+  if (adj->next_adj_with_signature)
+    s = format (s, " next_adj_with_signature %d", adj->next_adj_with_signature);
 
   return s;
 }
@@ -1083,11 +1235,17 @@ static uword unformat_ip_adjacency (unformat_input_t * input, va_list * args)
       if (next == IP_LOOKUP_NEXT_LOCAL)
         (void) unformat (input, "%d", &adj->if_address_index);
       else if (next == IP_LOOKUP_NEXT_CLASSIFY)
-        if (!unformat (input, "%d", &adj->classify_table_index))
-          {
-            clib_warning ("classify adj must specify table index");
-            return 0;
-          }
+        {
+          if (!unformat (input, "%d", &adj->classify_table_index))
+            {
+              clib_warning ("classify adj must specify table index");
+              return 0;
+            }
+        }
+      else if (next == IP_LOOKUP_NEXT_DROP)
+        {
+          adj->rewrite_header.node_index = 0;
+        }
     }
 
   else if (unformat_user (input,
diff --git a/vnet/vnet/ip/lookup.h b/vnet/vnet/ip/lookup.h
index ccce88935eb..02ab20d11d9 100644
--- a/vnet/vnet/ip/lookup.h
+++ b/vnet/vnet/ip/lookup.h
@@ -106,9 +106,12 @@ _(reverse, IP_FLOW_HASH_REVERSE_SRC_DST)
 
 /* IP unicast adjacency. */
 typedef struct {
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline0);
   /* Handle for this adjacency in adjacency heap. */
   u32 heap_handle;
 
+  STRUCT_MARK(signature_start);
+
   /* Interface address index for this local/arp adjacency. */
   u32 if_address_index;
 
@@ -131,9 +134,51 @@ typedef struct {
   /* Highest possible perf subgraph arc interposition, e.g. for ip6 ioam */
   u16 saved_lookup_next_index;
 
-  vnet_declare_rewrite (VLIB_BUFFER_PRE_DATA_SIZE - 5*sizeof(u32));
+  STRUCT_MARK(signature_end);
+
+  /* Number of FIB entries sharing this adjacency */
+  u32 share_count;
+  /* Use this adjacency instead */
+  u32 next_adj_with_signature;
+
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline1);
+
+  /* Rewrite in second/third cache lines */
+  vnet_declare_rewrite (VLIB_BUFFER_PRE_DATA_SIZE);
 } ip_adjacency_t;
 
+static inline uword
+vnet_ip_adjacency_signature (ip_adjacency_t * adj)
+{
+  uword signature = 0xfeedfaceULL;
+
+  /* Skip heap handle, sum everything up to but not including share_count */
+  signature = hash_memory64
+      (STRUCT_MARK_PTR(adj, signature_start),
+       STRUCT_OFFSET_OF(ip_adjacency_t, signature_end)
+       - STRUCT_OFFSET_OF(ip_adjacency_t, signature_start),
+       signature);
+
+  /* and the rewrite */
+  signature = hash_memory64 (&adj->rewrite_header, VLIB_BUFFER_PRE_DATA_SIZE,
+                             signature);
+  return signature;
+}
+
+static inline int
+vnet_ip_adjacency_share_compare (ip_adjacency_t * a1, ip_adjacency_t *a2)
+{
+  if (memcmp (STRUCT_MARK_PTR(a1, signature_start),
+              STRUCT_MARK_PTR(a2, signature_start),
+              STRUCT_OFFSET_OF(ip_adjacency_t, signature_end)
+              - STRUCT_OFFSET_OF(ip_adjacency_t, signature_start)))
+    return 0;
+  if (memcmp (&a1->rewrite_header, &a2->rewrite_header,
+              VLIB_BUFFER_PRE_DATA_SIZE))
+    return 0;
+  return 1;
+}
+
 /* Index into adjacency table. */
 typedef u32 ip_adjacency_index_t;
 
@@ -261,6 +306,9 @@ typedef struct ip_lookup_main_t {
   /* Indexed by heap_handle from ip_adjacency_t. */
   ip_multipath_adjacency_t * multipath_adjacencies;
 
+  /* Adjacency by signature hash */
+  uword * adj_index_by_signature;
+
   /* Temporary vectors for looking up next hops in hash. */
   ip_multipath_next_hop_t * next_hop_hash_lookup_key;
   ip_multipath_next_hop_t * next_hop_hash_lookup_key_normalized;
diff --git a/vppinfra/vppinfra/clib.h b/vppinfra/vppinfra/clib.h
index e043025a272..9e9a97789db 100644
--- a/vppinfra/vppinfra/clib.h
+++ b/vppinfra/vppinfra/clib.h
@@ -64,6 +64,8 @@
 #define STRUCT_SIZE_OF(t,f)   (sizeof (_STRUCT_FIELD (t, f)))
 #define STRUCT_BITS_OF(t,f)   (BITS (_STRUCT_FIELD (t, f)))
 #define STRUCT_ARRAY_LEN(t,f) ARRAY_LEN (_STRUCT_FIELD (t, f))
+#define STRUCT_MARK(mark)     u8 mark[0]
+#define STRUCT_MARK_PTR(v, f) &(v)->f
 
 /* Stride in bytes between struct array elements. */
 #define STRUCT_STRIDE_OF(t,f)			\