1 files changed, 197 insertions, 37 deletions

diff --git a/src/vcl/vppcom.c b/src/vcl/vppcom.c
index a557093e897..4568ac618d5 100644
--- a/src/vcl/vppcom.c
+++ b/src/vcl/vppcom.c
@@ -351,10 +351,16 @@ vcl_session_accepted_handler (vcl_worker_t * wrk, session_accepted_msg_t * mp,
 
   session->vpp_handle = mp->handle;
   session->session_state = VCL_STATE_READY;
-  if (mp->rmt.is_ip4)
+  if (mp->rmt.is_ip4 && mp->original_dst_port)
     {
-      session->original_dst_ip4 = mp->original_dst_ip4;
-      session->original_dst_port = mp->original_dst_port;
+      transport_endpt_attr_t *tep_attr;
+      vec_add2 (session->tep_attrs, tep_attr, 1);
+      /* Expecting to receive this on accepted connections
+       * and the external transport endpoint received is
+       * the local one, prior to something like nat */
+      tep_attr->type = TRANSPORT_ENDPT_ATTR_EXT_ENDPT;
+      tep_attr->ext_endpt.port = mp->original_dst_port;
+      tep_attr->ext_endpt.ip.ip4.as_u32 = mp->original_dst_ip4;
     }
   session->transport.rmt_port = mp->rmt.port;
   session->transport.is_ip4 = mp->rmt.is_ip4;
@@ -989,6 +995,24 @@ vcl_worker_rpc_handler (vcl_worker_t * wrk, void *data)
 }
 
 static void
+vcl_session_transport_attr_handler (vcl_worker_t *wrk, void *data)
+{
+  session_transport_attr_msg_t *mp = (session_transport_attr_msg_t *) data;
+  vcl_session_t *s;
+
+  s = vcl_session_get_w_vpp_handle (wrk, mp->handle);
+  if (!s)
+    {
+      VDBG (0, "session transport attr with wrong handle %llx", mp->handle);
+      return;
+    }
+
+  VDBG (0, "session %u [0x%llx]: transport attr %u", s->session_index,
+	s->vpp_handle, mp->attr.type);
+  vec_add1 (s->tep_attrs, mp->attr);
+}
+
+static void
 vcl_session_transport_attr_reply_handler (vcl_worker_t *wrk, void *data)
 {
   session_transport_attr_reply_msg_t *mp;
@@ -1129,6 +1153,9 @@ vcl_handle_mq_event (vcl_worker_t * wrk, session_event_t * e)
     case SESSION_CTRL_EVT_APP_WRK_RPC:
       vcl_worker_rpc_handler (wrk, e->data);
       break;
+    case SESSION_CTRL_EVT_TRANSPORT_ATTR:
+      vcl_session_transport_attr_handler (wrk, e->data);
+      break;
     case SESSION_CTRL_EVT_TRANSPORT_ATTR_REPLY:
       vcl_session_transport_attr_reply_handler (wrk, e->data);
       break;
@@ -1754,6 +1781,10 @@ vppcom_unformat_proto (uint8_t * proto, char *proto_str)
     *proto = VPPCOM_PROTO_SRTP;
   else if (!strcmp (proto_str, "srtp"))
     *proto = VPPCOM_PROTO_SRTP;
+  else if (!strcmp (proto_str, "HTTP"))
+    *proto = VPPCOM_PROTO_HTTP;
+  else if (!strcmp (proto_str, "http"))
+    *proto = VPPCOM_PROTO_HTTP;
   else
     return 1;
   return 0;
@@ -2345,6 +2376,75 @@ vppcom_session_write_inline (vcl_worker_t *wrk, vcl_session_t *s, void *buf,
 }
 
 int
+vppcom_session_write_segments (uint32_t session_handle,
+			       vppcom_data_segment_t *ds, uint32_t n_segments)
+{
+  vcl_worker_t *wrk = vcl_worker_get_current ();
+  int n_write = 0, n_bytes = 0, is_nonblocking;
+  vcl_session_t *s = 0;
+  svm_fifo_t *tx_fifo;
+  svm_msg_q_t *mq;
+  u8 is_ct;
+  u32 i;
+
+  if (PREDICT_FALSE (!ds))
+    return VPPCOM_EFAULT;
+
+  /* Accept zero length writes but just return */
+  if (PREDICT_FALSE (ds[0].len == 0))
+    return VPPCOM_OK;
+
+  s = vcl_session_get_w_handle (wrk, session_handle);
+  if (PREDICT_FALSE (!s || (s->flags & VCL_SESSION_F_IS_VEP)))
+    return VPPCOM_EBADFD;
+
+  if (PREDICT_FALSE (!vcl_session_is_open (s)))
+    return vcl_session_closed_error (s);
+
+  if (PREDICT_FALSE (s->flags & VCL_SESSION_F_WR_SHUTDOWN))
+    return VPPCOM_EPIPE;
+
+  is_nonblocking = vcl_session_has_attr (s, VCL_SESS_ATTR_NONBLOCK);
+  is_ct = vcl_session_is_ct (s);
+  mq = wrk->app_event_queue;
+  tx_fifo = is_ct ? s->ct_tx_fifo : s->tx_fifo;
+
+  for (i = 0; i < n_segments; i++)
+    n_bytes += ds[i].len;
+
+  if (svm_fifo_max_enqueue_prod (tx_fifo) < n_bytes)
+    {
+      if (is_nonblocking)
+	{
+	  return VPPCOM_EWOULDBLOCK;
+	}
+      while (svm_fifo_max_enqueue_prod (tx_fifo) < n_bytes)
+	{
+	  svm_fifo_add_want_deq_ntf (tx_fifo, SVM_FIFO_WANT_DEQ_NOTIF);
+	  if (vcl_session_is_closing (s))
+	    return vcl_session_closing_error (s);
+
+	  svm_msg_q_wait (mq, SVM_MQ_WAIT_EMPTY);
+	  vcl_worker_flush_mq_events (wrk);
+	}
+    }
+
+  n_write = svm_fifo_enqueue_segments (tx_fifo, (svm_fifo_seg_t *) ds,
+				       n_segments, 0 /* allow_partial */);
+
+  /* The underlying fifo segment can run out of memory */
+  if (PREDICT_FALSE (n_write < 0))
+    return VPPCOM_EAGAIN;
+
+  if (svm_fifo_set_event (s->tx_fifo))
+    app_send_io_evt_to_vpp (s->vpp_evt_q,
+			    s->tx_fifo->shr->master_session_index,
+			    SESSION_IO_EVT_TX, SVM_Q_WAIT);
+
+  return n_write;
+}
+
+int
 vppcom_session_write (uint32_t session_handle, void *buf, size_t n)
 {
   vcl_worker_t *wrk = vcl_worker_get_current ();
@@ -2534,6 +2634,9 @@ vcl_select_handle_mq_event (vcl_worker_t * wrk, session_event_t * e,
     case SESSION_CTRL_EVT_APP_WRK_RPC:
       vcl_worker_rpc_handler (wrk, e->data);
       break;
+    case SESSION_CTRL_EVT_TRANSPORT_ATTR:
+      vcl_session_transport_attr_handler (wrk, e->data);
+      break;
     default:
       clib_warning ("unhandled: %u", e->event_type);
       break;
@@ -2619,6 +2722,7 @@ vppcom_select_eventfd (vcl_worker_t * wrk, int n_bits,
   vcl_mq_evt_conn_t *mqc;
   int __clib_unused n_read;
   int n_mq_evts, i;
+  double end = -1;
   u64 buf;
 
   if (PREDICT_FALSE (wrk->api_client_handle == ~0))
@@ -2628,23 +2732,45 @@ vppcom_select_eventfd (vcl_worker_t * wrk, int n_bits,
     }
 
   vec_validate (wrk->mq_events, pool_elts (wrk->mq_evt_conns));
-  n_mq_evts = epoll_wait (wrk->mqs_epfd, wrk->mq_events,
-			  vec_len (wrk->mq_events), time_to_wait);
-  for (i = 0; i < n_mq_evts; i++)
+  if (time_to_wait > 0)
+    end = clib_time_now (&wrk->clib_time) + (time_to_wait / 1e3);
+
+  do
     {
-      if (PREDICT_FALSE (wrk->mq_events[i].data.u32 == ~0))
+      n_mq_evts = epoll_wait (wrk->mqs_epfd, wrk->mq_events,
+			      vec_len (wrk->mq_events), time_to_wait);
+      if (n_mq_evts < 0)
 	{
-	  vcl_api_handle_disconnect (wrk);
-	  continue;
+	  if (errno == EINTR)
+	    continue;
+
+	  VDBG (0, "epoll_wait error %u", errno);
+	  return 0;
+	}
+
+      if (n_mq_evts == 0)
+	return 0;
+
+      for (i = 0; i < n_mq_evts; i++)
+	{
+	  if (PREDICT_FALSE (wrk->mq_events[i].data.u32 == ~0))
+	    {
+	      vcl_api_handle_disconnect (wrk);
+	      continue;
+	    }
+
+	  mqc = vcl_mq_evt_conn_get (wrk, wrk->mq_events[i].data.u32);
+	  n_read = read (mqc->mq_fd, &buf, sizeof (buf));
+	  vcl_select_handle_mq (wrk, mqc->mq, n_bits, read_map, write_map,
+				except_map, 0, bits_set);
 	}
 
-      mqc = vcl_mq_evt_conn_get (wrk, wrk->mq_events[i].data.u32);
-      n_read = read (mqc->mq_fd, &buf, sizeof (buf));
-      vcl_select_handle_mq (wrk, mqc->mq, n_bits, read_map, write_map,
-			    except_map, 0, bits_set);
+      if (*bits_set || !time_to_wait)
+	return (int) *bits_set;
     }
+  while (end == -1 || clib_time_now (&wrk->clib_time) < end);
 
-  return (n_mq_evts > 0 ? (int) *bits_set : 0);
+  return 0;
 }
 
 int
@@ -3286,6 +3412,9 @@ vcl_epoll_wait_handle_mq_event (vcl_worker_t * wrk, session_event_t * e,
     case SESSION_CTRL_EVT_APP_WRK_RPC:
       vcl_worker_rpc_handler (wrk, e->data);
       break;
+    case SESSION_CTRL_EVT_TRANSPORT_ATTR:
+      vcl_session_transport_attr_handler (wrk, e->data);
+      break;
     default:
       VDBG (0, "unhandled: %u", e->event_type);
       break;
@@ -3407,6 +3536,9 @@ vppcom_epoll_wait_eventfd (vcl_worker_t *wrk, struct epoll_event *events,
 			      vec_len (wrk->mq_events), timeout_ms);
       if (n_mq_evts < 0)
 	{
+	  if (errno == EINTR)
+	    continue;
+
 	  VDBG (0, "epoll_wait error %u", errno);
 	  return n_evts;
 	}
@@ -3573,7 +3705,7 @@ vppcom_session_attr (uint32_t session_handle, uint32_t op,
   vcl_worker_t *wrk = vcl_worker_get_current ();
   u32 *flags = buffer;
   vppcom_endpt_t *ep = buffer;
-  transport_endpt_attr_t tea;
+  transport_endpt_attr_t tea, *tepap;
   vcl_session_t *session;
   int rv = VPPCOM_OK;
 
@@ -3696,24 +3828,49 @@ vppcom_session_attr (uint32_t session_handle, uint32_t op,
 	  rv = VPPCOM_EAFNOSUPPORT;
 	  break;
 	}
-      if (PREDICT_TRUE (buffer && buflen && (*buflen >= sizeof (*ep)) &&
-			ep->ip))
+      if (PREDICT_FALSE (!buffer || !buflen || (*buflen < sizeof (*ep)) ||
+			 !ep->ip))
 	{
-	  ep->is_ip4 = session->transport.is_ip4;
-	  ep->port = session->original_dst_port;
-	  clib_memcpy_fast (ep->ip, &session->original_dst_ip4,
-			    sizeof (ip4_address_t));
-	  *buflen = sizeof (*ep);
-	  VDBG (1,
-		"VPPCOM_ATTR_GET_ORIGINAL_DST: sh %u, is_ip4 = %u, addr = %U"
-		" port %d",
-		session_handle, ep->is_ip4, vcl_format_ip4_address,
-		(ip4_address_t *) (&session->original_dst_ip4),
-		ep->is_ip4 ? IP46_TYPE_IP4 : IP46_TYPE_IP6,
-		clib_net_to_host_u16 (ep->port));
+	  rv = VPPCOM_EINVAL;
+	  break;
 	}
-      else
-	rv = VPPCOM_EINVAL;
+
+      tepap =
+	vcl_session_tep_attr_get (session, TRANSPORT_ENDPT_ATTR_EXT_ENDPT);
+      if (!tepap)
+	{
+	  rv = VPPCOM_EINVAL;
+	  break;
+	}
+      vcl_ip_copy_to_ep (&tepap->ext_endpt.ip, ep, tepap->ext_endpt.is_ip4);
+      ep->port = tepap->ext_endpt.port;
+      *buflen = sizeof (*ep);
+
+      VDBG (1,
+	    "VPPCOM_ATTR_GET_ORIGINAL_DST: sh %u, is_ip4 = %u, "
+	    "addr = %U port %d",
+	    session_handle, ep->is_ip4, vcl_format_ip4_address,
+	    (ip4_address_t *) ep->ip,
+	    ep->is_ip4 ? IP46_TYPE_IP4 : IP46_TYPE_IP6,
+	    clib_net_to_host_u16 (ep->port));
+      break;
+
+    case VPPCOM_ATTR_GET_EXT_ENDPT:
+      if (PREDICT_FALSE (!buffer || !buflen || (*buflen < sizeof (*ep)) ||
+			 !ep->ip))
+	{
+	  rv = VPPCOM_EINVAL;
+	  break;
+	}
+      tepap =
+	vcl_session_tep_attr_get (session, TRANSPORT_ENDPT_ATTR_EXT_ENDPT);
+      if (!tepap)
+	{
+	  rv = VPPCOM_EINVAL;
+	  break;
+	}
+      vcl_ip_copy_to_ep (&tepap->ext_endpt.ip, ep, tepap->ext_endpt.is_ip4);
+      ep->port = tepap->ext_endpt.port;
       break;
 
     case VPPCOM_ATTR_SET_LCL_ADDR:
@@ -3798,12 +3955,11 @@ vppcom_session_attr (uint32_t session_handle, uint32_t op,
     case VPPCOM_ATTR_GET_TX_FIFO_LEN:
       if (buffer && buflen && (*buflen >= sizeof (u32)))
 	{
-
 	  /* VPP-TBD */
-	  *(size_t *) buffer = (session->sndbuf_size ? session->sndbuf_size :
-				session->tx_fifo ?
-				svm_fifo_size (session->tx_fifo) :
-				vcm->cfg.tx_fifo_size);
+	  *(u32 *) buffer =
+	    (session->sndbuf_size ? session->sndbuf_size :
+	     session->tx_fifo	  ? svm_fifo_size (session->tx_fifo) :
+				    vcm->cfg.tx_fifo_size);
 	  *buflen = sizeof (u32);
 
 	  VDBG (2, "VPPCOM_ATTR_GET_TX_FIFO_LEN: %u (0x%x), buflen %d,"
@@ -4238,7 +4394,8 @@ vppcom_session_attr (uint32_t session_handle, uint32_t op,
 	  break;
 	}
       vcl_session_alloc_ext_cfg (session, TRANSPORT_ENDPT_EXT_CFG_NONE,
-				 *buflen + sizeof (u32));
+				 *buflen +
+				   TRANSPORT_ENDPT_EXT_CFG_HEADER_SIZE);
       clib_memcpy (session->ext_config->data, buffer, *buflen);
       session->ext_config->len = *buflen;
       break;
@@ -4613,6 +4770,9 @@ vppcom_proto_str (vppcom_proto_t proto)
     case VPPCOM_PROTO_SRTP:
       proto_str = "SRTP";
       break;
+    case VPPCOM_PROTO_HTTP:
+      proto_str = "HTTP";
+      break;
     default:
       proto_str = "UNKNOWN";
       break;