VPP-130: LoadBal: Add lookup bypass and fix adjacency format function

Change-Id: Ib83c6ba04c41dbb80603c6a08c324d71eb44a102
Signed-off-by: Pierre Pfister <ppfister@cisco.com>

4 files changed, 125 insertions, 5 deletions

diff --git a/plugins/lb-plugin/lb/cli.c b/plugins/lb-plugin/lb/cli.c
index b59c6426241..398572ce396 100644
--- a/plugins/lb-plugin/lb/cli.c
+++ b/plugins/lb-plugin/lb/cli.c
@@ -17,6 +17,47 @@
 #include <lb/util.h>
 
 static clib_error_t *
+lb_bypass_command_fn (vlib_main_t * vm,
+              unformat_input_t * input, vlib_cli_command_t * cmd)
+{
+  unformat_input_t _line_input, *line_input = &_line_input;
+  ip46_address_t vip_prefix, as_addr;
+  u8 vip_plen;
+  u32 vip_index;
+  u8 disable = 0;
+  int ret;
+
+  if (!unformat_user (input, unformat_line_input, line_input))
+    return 0;
+
+  if (!unformat(line_input, "%U", unformat_ip46_prefix, &vip_prefix, &vip_plen, IP46_TYPE_ANY))
+    return clib_error_return (0, "invalid vip prefix: '%U'",
+                              format_unformat_error, line_input);
+
+  if ((ret = lb_vip_find_index(&vip_prefix, vip_plen, &vip_index)))
+    return clib_error_return (0, "lb_vip_find_index error %d", ret);
+
+  if (!unformat(line_input, "%U", unformat_ip46_address, &as_addr, IP46_TYPE_ANY))
+    return clib_error_return (0, "invalid as address: '%U'",
+                                  format_unformat_error, line_input);
+
+  if (unformat(line_input, "disable"))
+    disable = 1;
+
+  if ((ret = lb_as_lookup_bypass(vip_index, &as_addr, disable)))
+    return clib_error_return (0, "lb_as_lookup_bypass error %d", ret);
+
+  return 0;
+}
+
+VLIB_CLI_COMMAND (lb_bypass_command, static) =
+{
+  .path = "lb bypass",
+  .short_help = "lb bypass <prefix> <address> [disable]",
+  .function = lb_bypass_command_fn,
+};
+
+static clib_error_t *
 lb_vip_command_fn (vlib_main_t * vm,
               unformat_input_t * input, vlib_cli_command_t * cmd)
 {
diff --git a/plugins/lb-plugin/lb/lb.c b/plugins/lb-plugin/lb/lb.c
index 5eeeac5cfb7..e91bdf0bc10 100644
--- a/plugins/lb-plugin/lb/lb.c
+++ b/plugins/lb-plugin/lb/lb.c
@@ -143,10 +143,12 @@ u8 *format_lb_vip_detailed (u8 * s, va_list * args)
   u32 *as_index;
   pool_foreach(as_index, vip->as_indexes, {
       as = &lbm->ass[*as_index];
-      s = format(s, "%U    %U %d buckets   %d flows  %s\n", format_white_space, indent,
+      s = format(s, "%U    %U %d buckets   %d flows  adj:%u %s\n",
+                   format_white_space, indent,
                    format_ip46_address, &as->address, IP46_TYPE_ANY,
                    count[as - lbm->ass],
                    vlib_refcount_get(&lbm->as_refcount, as - lbm->ass),
+                   as->adj_index,
                    (as->flags & LB_AS_FLAGS_USED)?"used":" removed");
   });
 
@@ -447,6 +449,7 @@ next:
   //Update reused ASs
   vec_foreach(ip, to_be_updated) {
     lbm->ass[*ip].flags = LB_AS_FLAGS_USED;
+    lbm->ass[*ip].adj_index = ~0;
   }
   vec_free(to_be_updated);
 
@@ -458,6 +461,7 @@ next:
     as->address = addresses[*ip];
     as->flags = LB_AS_FLAGS_USED;
     as->vip_index = vip_index;
+    as->adj_index = ~0;
     pool_get(vip->as_indexes, as_index);
     *as_index = as - lbm->ass;
   }
@@ -531,6 +535,59 @@ int lb_vip_del_ass(u32 vip_index, ip46_address_t *addresses, u32 n)
   return ret;
 }
 
+int lb_as_lookup_bypass(u32 vip_index, ip46_address_t *address, u8 is_disable)
+{
+  lb_get_writer_lock();
+  lb_main_t *lbm = &lb_main;
+  u32 as_index;
+  lb_as_t *as;
+  lb_vip_t *vip;
+
+  if (!(vip = lb_vip_get_by_index(vip_index)) ||
+      lb_as_find_index_vip(vip, address, &as_index)) {
+    lb_put_writer_lock();
+    return VNET_API_ERROR_NO_SUCH_ENTRY;
+  }
+
+  as = &lbm->ass[as_index];
+
+  if (is_disable) {
+    as->adj_index = ~0;
+  } else if (lb_vip_is_gre4(vip)) {
+    uword *p = ip4_get_route (&ip4_main, 0, 0, as->address.ip4.as_u8, 32);
+    if (p == 0) {
+      lb_put_writer_lock();
+      return VNET_API_ERROR_NO_SUCH_ENTRY;
+    }
+    u32 ai = (u32)p[0];
+    ip_lookup_main_t *lm4 = &ip4_main.lookup_main;
+    ip_adjacency_t *adj4 = ip_get_adjacency (lm4, ai);
+    if (adj4->lookup_next_index != IP_LOOKUP_NEXT_REWRITE) {
+      lb_put_writer_lock();
+      return VNET_API_ERROR_INCORRECT_ADJACENCY_TYPE;
+    }
+
+    as->adj_index = ai;
+  } else {
+    u32 ai = ip6_get_route (&ip6_main, 0, 0, &as->address.ip6, 128);
+    if (ai == 0) {
+      lb_put_writer_lock();
+      return VNET_API_ERROR_NO_SUCH_ENTRY;
+    }
+
+    ip_lookup_main_t *lm6 = &ip6_main.lookup_main;
+    ip_adjacency_t *adj6 = ip_get_adjacency (lm6, ai);
+    if (adj6->lookup_next_index != IP_LOOKUP_NEXT_REWRITE) {
+      lb_put_writer_lock();
+      return VNET_API_ERROR_INCORRECT_ADJACENCY_TYPE;
+    }
+
+    as->adj_index = ai;
+  }
+  lb_put_writer_lock();
+  return 0;
+}
+
 
 /**
  * Add the VIP adjacency to the ip4 or ip6 fib
@@ -733,6 +790,7 @@ lb_init (vlib_main_t * vm)
   lbm->ass = 0;
   pool_get(lbm->ass, default_as);
   default_as->flags = 0;
+  default_as->adj_index = ~0;
   default_as->vip_index = ~0;
   default_as->address.ip6.as_u64[0] = 0xffffffffffffffffL;
   default_as->address.ip6.as_u64[1] = 0xffffffffffffffffL;
diff --git a/plugins/lb-plugin/lb/lb.h b/plugins/lb-plugin/lb/lb.h
index 3d250319422..14a4d8a39e7 100644
--- a/plugins/lb-plugin/lb/lb.h
+++ b/plugins/lb-plugin/lb/lb.h
@@ -55,6 +55,13 @@ typedef struct {
   ip46_address_t address;
 
   /**
+   * Second ip lookup can be avoided by sending directly the packet
+   * to ip-rewrite with a configured adjacency.
+   * When set to ~0, the packets are sent to ip6-lookup.
+   */
+  u32 adj_index;
+
+  /**
    * ASs are indexed by address and VIP Index.
    * Which means there will be duplicated if the same server
    * address is used for multiple VIPs.
@@ -295,6 +302,12 @@ int lb_vip_find_index(ip46_address_t *prefix, u8 plen, u32 *vip_index);
 int lb_vip_add_ass(u32 vip_index, ip46_address_t *addresses, u32 n);
 int lb_vip_del_ass(u32 vip_index, ip46_address_t *addresses, u32 n);
 
+/**
+ * Updates the adjacency index stored in the AS such that the second
+ * IP lookup (after encap) can be bypassed.
+ */
+int lb_as_lookup_bypass(u32 vip_index, ip46_address_t *address, u8 is_disable);
+
 u32 lb_hash_time_now(vlib_main_t * vm);
 
 void lb_garbage_collection();
diff --git a/plugins/lb-plugin/lb/node.c b/plugins/lb-plugin/lb/node.c
index 4d598104944..8df462c0853 100644
--- a/plugins/lb-plugin/lb/node.c
+++ b/plugins/lb-plugin/lb/node.c
@@ -38,6 +38,7 @@ static char *lb_error_strings[] = {
 
 typedef enum {
   LB_NEXT_LOOKUP,
+  LB_NEXT_REWRITE,
   LB_NEXT_DROP,
   LB_N_NEXT,
 } lb_next_t;
@@ -47,14 +48,14 @@ typedef struct {
   u32 as_index;
 } lb_trace_t;
 
-u8 *lb_format_adjacency(u8 * s,
-                        struct ip_lookup_main_t * lm,
-                        ip_adjacency_t *adj)
+u8 *lb_format_adjacency(u8 * s, va_list * va)
 {
   lb_main_t *lbm = &lb_main;
+  __attribute((unused)) ip_lookup_main_t *lm = va_arg (*va, ip_lookup_main_t *);
+  ip_adjacency_t *adj = va_arg (*va, ip_adjacency_t *);
   lb_adj_data_t *ad = (lb_adj_data_t *) &adj->opaque;
   __attribute__((unused)) lb_vip_t *vip = pool_elt_at_index (lbm->vips, ad->vip_index);
-  return format(s, "idx:%d", ad->vip_index);
+  return format(s, "vip idx:%d", ad->vip_index);
 }
 
 u8 *
@@ -251,6 +252,9 @@ lb_node_fn (vlib_main_t * vm,
           clib_host_to_net_u16(0x0800):
           clib_host_to_net_u16(0x86DD);
 
+      vnet_buffer(p0)->ip.adj_index[VLIB_TX] = as0->adj_index;
+      next0 = (as0->adj_index != ~0)?LB_NEXT_REWRITE:next0;
+
       if (PREDICT_FALSE (p0->flags & VLIB_BUFFER_IS_TRACED))
       {
         lb_trace_t *tr = vlib_add_trace (vm, node, p0, sizeof (*tr));
@@ -310,6 +314,7 @@ VLIB_REGISTER_NODE (lb6_gre6_node) =
   .next_nodes =
   {
       [LB_NEXT_LOOKUP] = "ip6-lookup",
+      [LB_NEXT_REWRITE] = "ip6-rewrite",
       [LB_NEXT_DROP] = "error-drop"
   },
 };
@@ -334,6 +339,7 @@ VLIB_REGISTER_NODE (lb6_gre4_node) =
   .next_nodes =
   {
       [LB_NEXT_LOOKUP] = "ip4-lookup",
+      [LB_NEXT_REWRITE]= "ip4-rewrite-transit",
       [LB_NEXT_DROP] = "error-drop"
   },
 };
@@ -358,6 +364,7 @@ VLIB_REGISTER_NODE (lb4_gre6_node) =
   .next_nodes =
   {
       [LB_NEXT_LOOKUP] = "ip6-lookup",
+      [LB_NEXT_REWRITE] = "ip6-rewrite",
       [LB_NEXT_DROP] = "error-drop"
   },
 };
@@ -382,6 +389,7 @@ VLIB_REGISTER_NODE (lb4_gre4_node) =
   .next_nodes =
   {
       [LB_NEXT_LOOKUP] = "ip4-lookup",
+      [LB_NEXT_REWRITE]= "ip4-rewrite-transit",
       [LB_NEXT_DROP] = "error-drop"
   },
 };