udp: preallocate ports sparse vec map

Not ideal. The sparse vector used to map ports to next nodes assumes
only a few ports are ever used. When udp transport is enabled this does
not hold and, to make matters worse, ports are consumed in a random
order.

This can lead to a lot of slow updates to internal data structures
which in turn can slow udp connection allocations until all ports are
eventually consumed.

Consequently, reallocate sparse vector, preallocate all ports and have
them point to UDP_NO_NODE_SET. We could consider switching the sparse
vector to a preallocated vector but that would increase memory
consumption for vpp deployments that do not rely on host stack.

For reference, populating one of the v4 or v6 sparse vectors in reverse
order takes about 9.8s on a skylake cpu.

Type: improvement

Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id795e1805d0d3ba54f56a152a9506a7a2a06ecbc

1 files changed, 80 insertions, 2 deletions

diff --git a/src/vnet/udp/udp.c b/src/vnet/udp/udp.c
index e6ee9d8b37d..2400b254fd0 100644
--- a/src/vnet/udp/udp.c
+++ b/src/vnet/udp/udp.c
@@ -493,8 +493,87 @@ format_udp_listener_session (u8 * s, va_list * args)
   return format (s, "%U", format_udp_connection, uc, verbose);
 }
 
-/* *INDENT-OFF* */
+static void
+udp_realloc_ports_sv (u16 **ports_nh_svp)
+{
+  u16 port, port_no, *ports_nh_sv, *mc;
+  u32 *ports = 0, *nh = 0, msum, i;
+  sparse_vec_header_t *h;
+  uword sv_index, *mb;
+
+  ports_nh_sv = *ports_nh_svp;
+
+  for (port = 1; port < 65535; port++)
+    {
+      port_no = clib_host_to_net_u16 (port);
+
+      sv_index = sparse_vec_index (ports_nh_sv, port_no);
+      if (sv_index != SPARSE_VEC_INVALID_INDEX)
+	{
+	  vec_add1 (ports, port_no);
+	  vec_add1 (nh, ports_nh_sv[sv_index]);
+	}
+    }
+
+  sparse_vec_free (ports_nh_sv);
+
+  ports_nh_sv =
+    sparse_vec_new (/* elt bytes */ sizeof (ports_nh_sv[0]),
+		    /* bits in index */ BITS (((udp_header_t *) 0)->dst_port));
+
+  vec_resize (ports_nh_sv, 65535);
+
+  for (port = 1; port < 65535; port++)
+    ports_nh_sv[port] = ~0;
+
+  for (i = 0; i < vec_len (ports); i++)
+    ports_nh_sv[ports[i]] = nh[i];
+
+  h = sparse_vec_header (ports_nh_sv);
+  vec_foreach (mb, h->is_member_bitmap)
+    *mb = (uword) ~0;
+
+  msum = 0;
+  vec_foreach (mc, h->member_counts)
+    {
+      *mc = msum;
+      msum += msum == 0 ? 63 : 64;
+    }
+
+  vec_free (ports);
+  vec_free (nh);
+
+  *ports_nh_svp = ports_nh_sv;
+}
+
+static clib_error_t *
+udp_enable_disable (vlib_main_t *vm, u8 is_en)
+{
+  udp_main_t *um = &udp_main;
+
+  /* Not ideal. The sparse vector used to map ports to next nodes assumes
+   * only a few ports are ever used. When udp transport is enabled this does
+   * not hold and, to make matters worse, ports are consumed in a random
+   * order.
+   *
+   * This can lead to a lot of slow updates to internal data structures
+   * which in turn can slow udp connection allocations until all ports are
+   * eventually consumed.
+   *
+   * Consequently, reallocate sparse vector, preallocate all ports and have
+   * them point to UDP_NO_NODE_SET. We could consider switching the sparse
+   * vector to a preallocated vector but that would increase memory
+   * consumption for vpp deployments that do not rely on host stack.
+   */
+
+  udp_realloc_ports_sv (&um->next_by_dst_port4);
+  udp_realloc_ports_sv (&um->next_by_dst_port6);
+
+  return 0;
+}
+
 static const transport_proto_vft_t udp_proto = {
+  .enable = udp_enable_disable,
   .start_listen = udp_session_bind,
   .connect = udp_open_connection,
   .stop_listen = udp_session_unbind,
@@ -515,7 +594,6 @@ static const transport_proto_vft_t udp_proto = {
     .service_type = TRANSPORT_SERVICE_CL,
   },
 };
-/* *INDENT-ON* */
 
 static clib_error_t *
 udp_init (vlib_main_t * vm)