src/vnet/geneve

/*
 * Copyright (c) 2017-2019 Cisco and/or its affiliates.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at:
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include <math.h>
#include <vlib/vlib.h>
#include <vnet/vnet.h>
#include <vppinfra/elog.h>
#include <vnet/session/transport.h>
#include <vnet/session/session.h>
#include <vnet/session/application.h>
#include <vnet/session/application_interface.h>
#include <vnet/session/application_local.h>
#include <vnet/session/session_debug.h>
#include <svm/queue.h>

#define app_check_thread_and_barrier(_fn, _arg)				\
  if (!vlib_thread_is_main_w_barrier ())				\
    {									\
     vlib_rpc_call_main_thread (_fn, (u8 *) _arg, sizeof(*_arg));	\
      return;								\
   }

static void
session_mq_listen_handler (void *data)
{
  session_listen_msg_t *mp = (session_listen_msg_t *) data;
  vnet_listen_args_t _a, *a = &_a;
  app_worker_t *app_wrk;
  application_t *app;
  int rv;

  app_check_thread_and_barrier (session_mq_listen_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  clib_memset (a, 0, sizeof (*a));
  a->sep.is_ip4 = mp->is_ip4;
  clib_memcpy_fast (&a->sep.ip, &mp->ip, sizeof (mp->ip));
  a->sep.port = mp->port;
  a->sep.fib_index = mp->vrf;
  a->sep.sw_if_index = ENDPOINT_INVALID_INDEX;
  a->sep.transport_proto = mp->proto;
  a->sep_ext.ckpair_index = mp->ckpair_index;
  a->app_index = app->app_index;
  a->wrk_map_index = mp->wrk_index;

  if ((rv = vnet_listen (a)))
    clib_warning ("listen returned: %d", rv);

  app_wrk = application_get_worker (app, mp->wrk_index);
  mq_send_session_bound_cb (app_wrk->wrk_index, mp->context, a->handle, rv);
  return;
}

static void
session_mq_listen_uri_handler (void *data)
{
  session_listen_uri_msg_t *mp = (session_listen_uri_msg_t *) data;
  vnet_listen_args_t _a, *a = &_a;
  app_worker_t *app_wrk;
  application_t *app;
  int rv;

  app_check_thread_and_barrier (session_mq_listen_uri_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  clib_memset (a, 0, sizeof (*a));
  a->uri = (char *) mp->uri;
  a->app_index = app->app_index;
  rv = vnet_bind_uri (a);

  app_wrk = application_get_worker (app, 0);
  mq_send_session_bound_cb (app_wrk->wrk_index, mp->context, a->handle, rv);
}

static void
session_mq_connect_handler (void *data)
{
  session_connect_msg_t *mp = (session_connect_msg_t *) data;
  vnet_connect_args_t _a, *a = &_a;
  app_worker_t *app_wrk;
  application_t *app;
  int rv;

  app_check_thread_and_barrier (session_mq_connect_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  clib_memset (a, 0, sizeof (*a));
  a->sep.is_ip4 = mp->is_ip4;
  clib_memcpy_fast (&a->sep.ip, &mp->ip, sizeof (mp->ip));
  a->sep.port = mp->port;
  a->sep.transport_proto = mp->proto;
  a->sep.peer.fib_index = mp->vrf;
  a->sep.peer.sw_if_index = ENDPOINT_INVALID_INDEX;
  a->sep_ext.parent_handle = mp->parent_handle;
  a->sep_ext.ckpair_index = mp->ckpair_index;
  if (mp->hostname_len)
    {
      vec_validate (a->sep_ext.hostname, mp->hostname_len - 1);
      clib_memcpy_fast (a->sep_ext.hostname, mp->hostname, mp->hostname_len);
    }
  a->api_context = mp->context;
  a->app_index = app->app_index;
  a->wrk_map_index = mp->wrk_index;

  if ((rv = vnet_connect (a)))
    {
      clib_warning ("connect returned: %U", format_vnet_api_errno, rv);
      app_wrk = application_get_worker (app, mp->wrk_index);
      mq_send_session_connected_cb (app_wrk->wrk_index, mp->context, 0,
				    /* is_fail */ 1);
    }

  vec_free (a->sep_ext.hostname);
}

static void
session_mq_connect_uri_handler (void *data)
{
  session_connect_uri_msg_t *mp = (session_connect_uri_msg_t *) data;
  vnet_connect_args_t _a, *a = &_a;
  app_worker_t *app_wrk;
  application_t *app;
  int rv;

  app_check_thread_and_barrier (session_mq_connect_uri_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  clib_memset (a, 0, sizeof (*a));
  a->uri = (char *) mp->uri;
  a->api_context = mp->context;
  a->app_index = app->app_index;
  if ((rv = vnet_connect_uri (a)))
    {
      clib_warning ("connect_uri returned: %d", rv);
      app_wrk = application_get_worker (app, 0 /* default wrk only */ );
      mq_send_session_connected_cb (app_wrk->wrk_index, mp->context, 0,
				    /* is_fail */ 1);
    }
}

static void
session_mq_disconnect_handler (void *data)
{
  session_disconnect_msg_t *mp = (session_disconnect_msg_t *) data;
  vnet_disconnect_args_t _a, *a = &_a;
  application_t *app;

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  a->app_index = app->app_index;
  a->handle = mp->handle;
  vnet_disconnect_session (a);
}

static void
app_mq_detach_handler (void *data)
{
  session_app_detach_msg_t *mp = (session_app_detach_msg_t *) data;
  vnet_app_detach_args_t _a, *a = &_a;
  application_t *app;

  app_check_thread_and_barrier (app_mq_detach_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  a->app_index = app->app_index;
  a->api_client_index = mp->client_index;
  vnet_application_detach (a);
}

static void
session_mq_unlisten_handler (void *data)
{
  session_unlisten_msg_t *mp = (session_unlisten_msg_t *) data;
  vnet_unlisten_args_t _a, *a = &_a;
  app_worker_t *app_wrk;
  application_t *app;
  int rv;

  app_check_thread_and_barrier (session_mq_unlisten_handler, mp);

  app = application_lookup (mp->client_index);
  if (!app)
    return;

  clib_memset (a, 0, sizeof (*a));
  a->app_index = app->app_index;
  a->handle = mp->handle;
  a->wrk_map_index = mp->wrk_index;
  if ((rv = vnet_unlisten (a)))
    clib_warning ("unlisten returned: %d", rv);

  app_wrk = application_get_worker (app, a->wrk_map_index);
  if (!app_wrk)
    return;

  mq_send_unlisten_reply (app_wrk, mp->handle, mp->context, rv);
}

static void
session_mq_accepted_reply_handler (void *data)
{
  session_accepted_reply_msg_t *mp = (session_accepted_reply_msg_t *) data;
  vnet_disconnect_args_t _a = { 0 }, *a = &_a;
  session_state_t old_state;
  app_worker_t *app_wrk;
  session_t *s;

  /* Server isn't interested, kill the session */
  if (mp->retval)
    {
      a->app_index = mp->context;
      a->handle = mp->handle;
      vnet_disconnect_session (a);
      return;
    }

  /* Mail this back from the main thread. We're not polling in main
   * thread so we're using other workers for notifications. */
  if (vlib_num_workers () && vlib_get_thread_index () != 0
      && session_thread_from_handle (mp->handle) == 0)
    {
      vlib_rpc_call_main_thread (session_mq_accepted_reply_handler,
				 (u8 *) mp, sizeof (*mp));
      return;
    }

  s = session_get_from_handle_if_valid (mp->handle);
  if (!s)
    return;

  app_wrk = app_worker_get (s->app_wrk_index);
  if (app_wrk->app_index != mp->context)
    {
      clib_warning ("app doesn't own session");
      return;
    }

  if (!session_has_transport (s))
    {
      s->session_state = SESSION_STATE_READY;
      if (ct_session_connect_notify (s))
	return;
    }
  else
    {
      old_state = s->session_state;
      s->session_state = SESSION_STATE_READY;

      if (!svm_fifo_is_empty_prod (s->rx_fifo))
	app_worker_lock_and_send_event (app_wrk, s, SESSION_IO_EVT_RX);

      /* Closed while waiting for app to reply. Resend disconnect */
      if (old_state >= SESSION_STATE_TRANSPORT_CLOSING)
	{
	  app_worker_close_notify (app_wrk, s);
	  s->session_state = old_state;
	  return;
	}
    }
}

static void
session_mq_reset_reply_handler (void *data)
{
  vnet_disconnect_args_t _a = { 0 }, *a = &_a;
  session_reset_reply_msg_t *mp;
  app_worker_t *app_wrk;
  session_t *s;
  application_t *app;
  u32 index, thread_index;

  mp = (session_reset_reply_msg_t *) data;
  app = application_lookup (mp->context);
  if (!app)
    return;

  session_parse_handle (mp->handle, &index, &thread_index);
  s = session_get_if_valid (index, thread_index);

  /* Session was already closed or already cleaned up */
  if (!s || s->session_state != SESSION_STATE_TRANSPORT_CLOSING)
    return;

  app_wrk = app_worker_get (s->app_wrk_index);
  if (!app_wrk || app_wrk->app_index != app->app_index)
    {
      clib_warning ("App %u does not own handle 0x%lx!", app->app_index,
		    mp->handle);
      return;
    }

  /* Client objected to resetting the session, log and continue */
  if (mp->retval)
    {
      clib_warning ("client retval %d", mp->retval);
      return;
    }

  /* This comes as a response to a reset, transport only waiting for
   * confirmation to remove connection state, no need to disconnect */
  a->handle = mp->handle;
  a->app_index = app->app_index;
  vnet_disconnect_session (a);
}

static void
session_mq_disconnected_handler (void *data)
{
  session_disconnected_reply_msg_t *rmp;
  vnet_disconnect_args_t _a, *a = &_a;
  svm_msg_q_msg_t _msg, *msg = &_msg;
  session_disconnected_msg_t *mp;
  app_worker_t *app_wrk;
  session_event_t *evt;
  session_t *s;
  application_t *app;
  int rv = 0;

  mp = (session_disconnected_msg_t *) data;
  if (!(s = session_get_from_handle_if_valid (mp->handle)))
    {
      clib_warning ("could not disconnect handle %llu", mp->handle);
      return;
    }
  app_wrk = app_worker_get (s->app_wrk_index);
  app = application_lookup (mp->client_index);
  if (!(app_wrk && app && app->app_index == app_wrk->app_index))
    {
      clib_warning ("could not disconnect session: %llu app: %u",
		    mp->handle, mp->client_index);
      return;
    }

  a->handle = mp->handle;
  a->app_index = app_wrk->wrk_index;
  rv = vnet_disconnect_session (a);

  svm_msg_q_lock_and_alloc_msg_w_ring (app_wrk->event_queue,
				       SESSION_MQ_CTRL_EVT_RING,
				       SVM_Q_WAIT, msg);
  evt = svm_msg_q_msg_data (app_wrk->event_queue, msg);
  clib_memset (evt, 0, sizeof (*evt));
  evt->event_type = SESSION_CTRL_EVT_DISCONNECTED_REPLY;
  rmp = (session_disconnected_reply_msg_t *) evt->data;
  rmp->handle = mp->handle;
  rmp->context = mp->context;
  rmp->retval = rv;
  svm_msg_q_add_and_unlock (app_wrk->event_queue, msg);
}

static void
session_mq_disconnected_reply_handler (void *data)
{
  session_disconnected_reply_msg_t *mp;
  vnet_disconnect_args_t _a, *a = &_a;
  application_t *app;

  mp = (session_disconnected_reply_msg_t *) data;

  /* Client objected to disconnecting the session, log and continue */
  if (mp->retval)
    {
      clib_warning ("client retval %d", mp->retval);
      return;
    }

  /* Disconnect has been confirmed. Confirm close to transport */
  app = application_lookup (mp->context);
  if (app)
    {
      a->handle = mp->handle;
      a->app_index = app->app_index;
      vnet_disconnect_session (a);
    }
}

static void
session_mq_worker_update_handler (void *data)
{
  session_worker_update_msg_t *mp = (session_worker_update_msg_t *) data;
  session_worker_update_reply_msg_t *rmp;
  svm_msg_q_msg_t _msg, *msg = &_msg;
  app_worker_t *app_wrk;
  u32 owner_app_wrk_map;
  session_event_t *evt;
  session_t *s;
  application_t *app;

  app = application_lookup (mp->client_index);
  if (!app)
    return;
  if (!(s = session_get_from_handle_if_valid (mp->handle)))
    {
      clib_warning ("invalid handle %llu", mp->handle);
      return;
    }
  app_wrk = app_worker_get (s->app_wrk_index);
  if (app_wrk->app_index != app->app_index)
    {
      clib_warning ("app %u does not own session %llu", app->app_index,
		    mp->handle);
      return;
    }
  owner_app_wrk_map = app_wrk->wrk_map_index;
  app_wrk = application_get_worker (app, mp->wrk_index);

  /* This needs to come from the new owner */
  if (mp->req_wrk_index == owner_app_wrk_map)
    {
      session_req_worker_update_msg_t *wump;

      svm_msg_q_lock_and_alloc_msg_w_ring (app_wrk->event_queue,
					   SESSION_MQ_CTRL_EVT_RING,
					   SVM_Q_WAIT, msg);
      evt = svm_msg_q_msg_data (app_wrk->event_queue, msg);
      clib_memset (evt, 0, sizeof (*evt));
      evt->event_type = SESSION_CTRL_EVT_REQ_WORKER_UPDATE;
      wump = (session_req_worker_update_msg_t *) evt->data;
      wump->session_handle = mp->handle;
      svm_msg_q_add_and_unlock (app_wrk->event_queue, msg);
      return;
    }

  app_worker_own_session (app_wrk, s);

  /*
   * Send reply
   */
  svm_msg_q_lock_and_alloc_msg_w_ring (app_wrk->event_queue,
				       SESSION_MQ_CTRL_EVT_RING,
				       SVM_Q_WAIT, msg);
  evt = svm_msg_q_msg_data (app_wrk->event_queue, msg);
  clib_memset (evt, 0, sizeof (*evt));
  evt->event_type = SESSION_CTRL_EVT_WORKER_UPDATE_REPLY;
  rmp = (session_worker_update_reply_msg_t *) evt->data;
  rmp->handle = mp->handle;
  rmp->rx_fifo = pointer_to_uword (s->rx_fifo);
  rmp->tx_fifo = pointer_to_uword (s->tx_fifo);
  rmp->segment_handle = session_segment_handle (s);
  svm_msg_q_add_and_unlock (app_wrk->event_queue, msg);

  /*
   * Retransmit messages that may have been lost
   */
  if (s->tx_fifo && !svm_fifo_is_empty (s->tx_fifo))
    session_send_io_evt_to_thread (s->tx_fifo, SESSION_IO_EVT_TX);

  if (s->rx_fifo && !svm_fifo_is_empty (s->rx_fifo))
    app_worker_lock_and_send_event (app_wrk, s, SESSION_IO_EVT_RX);

  if (s->session_state >= SESSION_STATE_TRANSPORT_CLOSING)
    app_worker_close_notify (app_wrk, s);
}

vlib_node_registration_t session_queue_node;

typedef struct
{
  u32 session_index;
  u32 server_thread_index;
} session_queue_trace_t;

/* packet trace format function */
static u8 *
format_session_queue_trace (u8 * s, va_list * args)
{
  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
  CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
  session_queue_trace_t *t = va_arg (*args, session_queue_trace_t *);

  s = format (s, "SESSION_QUEUE: session index %d, server thread index %d",
	      t->session_index, t->server_thread_index);
  return s;
}

#define foreach_session_queue_error		\
_(TX, "Packets transmitted")                  	\
_(TIMER, "Timer events")			\
_(NO_BUFFER, "Out of buffers")

typedef enum
{
#define _(sym,str) SESSION_QUEUE_ERROR_##sym,
  foreach_session_queue_error
#undef _
    SESSION_QUEUE_N_ERROR,
} session_queue_error_t;

static char *session_queue_error_strings[] = {
#define _(sym,string) string,
  foreach_session_queue_error
#undef _
};

enum
{
  SESSION_TX_NO_BUFFERS = -2,
  SESSION_TX_NO_DATA,
  SESSION_TX_OK
};

static void
session_tx_trace_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
			u32 next_index, u32 * to_next, u16 n_segs,
			session_t * s, u32 n_trace)
{
  session_queue_trace_t *t;
  vlib_buffer_t *b;
  int i;

  for (i = 0; i < clib_min (n_trace, n_segs); i++)
    {
      b = vlib_get_buffer (vm, to_next[i - n_segs]);
      vlib_trace_buffer (vm, node, next_index, b, 1 /* follow_chain */ );
      t = vlib_add_trace (vm, node, b, sizeof (*t));
      t->session_index = s->session_index;
      t->server_thread_index = s->thread_index;
    }
  vlib_set_trace_count (vm, node, n_trace - i);
}

always_inline void
session_tx_fifo_chain_tail (vlib_main_t * vm, session_tx_context_t * ctx,
			    vlib_buffer_t * b, u16 * n_bufs, u8 peek_data)
{
  vlib_buffer_t *chain_b, *prev_b;
  u32 chain_bi0, to_deq, left_from_seg;
  session_worker_t *wrk;
  u16 len_to_deq, n_bytes_read;
  u8 *data, j;

  wrk = session_main_get_worker (ctx->s->thread_index);
  b->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
  b->total_length_not_including_first_buffer = 0;

  chain_b = b;
  left_from_seg = clib_min (ctx->snd_mss - b->current_length,
			    ctx->left_to_snd);
  to_deq = left_from_seg;
  for (j = 1; j < ctx->n_bufs_per_seg; j++)
    {
      prev_b = chain_b;
      len_to_deq = clib_min (to_deq, ctx->deq_per_buf);

      *n_bufs -= 1;
      chain_bi0 = wrk->tx_buffers[*n_bufs];
      chain_b = vlib_get_buffer (vm, chain_bi0);
      chain_b->current_data = 0;
      data = vlib_buffer_get_current (chain_b);
      if (peek_data)
	{
	  n_bytes_read = svm_fifo_peek (ctx->s->tx_fifo,
					ctx->tx_offset, len_to_deq, data);
	  ctx->tx_offset += n_bytes_read;
	}
      else
	{
	  if (ctx->transport_vft->transport_options.tx_type ==
	      TRANSPORT_TX_DGRAM)
	    {
	      svm_fifo_t *f = ctx->s->tx_fifo;
	      session_dgram_hdr_t *hdr = &ctx->hdr;
	      u16 deq_now;
	      deq_now = clib_min (hdr->data_length - hdr->data_offset,
				  len_to_deq);
	      n_bytes_read = svm_fifo_peek (f, hdr->data_offset, deq_now,
					    data);
	      ASSERT (n_bytes_read > 0);

	      hdr->data_offset += n_bytes_read;
	      if (hdr->data_offset == hdr->data_length)
		{
		  u32 offset = hdr->data_length + SESSION_CONN_HDR_LEN;
		  svm_fifo_dequeue_drop (f, offset);
		}
	    }
	  else
	    n_bytes_read = svm_fifo_dequeue (ctx->s->tx_fifo,
					     len_to_deq, data);
	}
      ASSERT (n_bytes_read == len_to_deq);
      chain_b->current_length = n_bytes_read;
      b->total_length_not_including_first_buffer += chain_b->current_length;

      /* update previous buffer */
      prev_b->next_buffer = chain_bi0;
      prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;

      /* update current buffer */
      chain_b->next_buffer = 0;

      to_deq -= n_bytes_read;
      if (to_deq == 0)
	break;
    }
  ASSERT (to_deq == 0
	  && b->total_length_not_including_first_buffer == left_from_seg);
  ctx->left_to_snd -= left_from_seg;
}

always_inline void
session_tx_fill_buffer (vlib_main_t * vm, session_tx_context_t * ctx,
			vlib_buffer_t * b, u16 * n_bufs, u8 peek_data)
{
  u32 len_to_deq;
  u8 *data0;
  int n_bytes_read;

  /*
   * Start with the first buffer in chain
   */
  b->error = 0;
  b->flags = VNET_BUFFER_F_LOCALLY_ORIGINATED;
  b->current_data = 0;

  data0 = vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
  len_to_deq = clib_min (ctx->left_to_snd, ctx->deq_per_first_buf);

  if (peek_data)
    {
      n_bytes_read = svm_fifo_peek (ctx->s->tx_fifo, ctx->tx_offset,
				    len_to_deq, data0);
      ASSERT (n_bytes_read > 0);
      /* Keep track of progress locally, transport is also supposed to
       * increment it independently when pushing the header */
      ctx->tx_offset += n_bytes_read;
    }
  else
    {
      if (ctx->transport_vft->transport_options.tx_type == TRANSPORT_TX_DGRAM)
	{
	  session_dgram_hdr_t *hdr = &ctx->hdr;
	  svm_fifo_t *f = ctx->s->tx_fifo;
	  u16 deq_now;
	  u32 offset;

	  ASSERT (hdr->data_length > hdr->data_offset);
	  deq_now = clib_min (hdr->data_length - hdr->data_offset,
			      len_to_deq);
	  offset = hdr->data_offset + SESSION_CONN_HDR_LEN;
	  n_bytes_read = svm_fifo_peek (f, offset, deq_now, data0);
	  ASSERT (n_bytes_read > 0);

	  if (ctx->s->session_state == SESSION_STATE_LISTENING)
	    {
	      ip_copy (&ctx->tc->rmt_ip, &hdr->rmt_ip, ctx->tc->is_ip4);
	      ctx->tc->rmt_port = hdr->rmt_port;
	    }
	  hdr->data_offset += n_bytes_read;
	  if (hdr->data_offset == hdr->data_length)
	    {
	      offset = hdr->data_length + SESSION_CONN_HDR_LEN;
	      svm_fifo_dequeue_drop (f, offset);
	    }
	}
      else
	{
	  n_bytes_read = svm_fifo_dequeue (ctx->s->tx_fifo,
					   len_to_deq, data0);
	  ASSERT (n_bytes_read > 0);
	}
    }
  b->current_length = n_bytes_read;
  ctx->left_to_snd -= n_bytes_read;

  /*
   * Fill in the remaining buffers in the chain, if any
   */
  if (PREDICT_FALSE (ctx->n_bufs_per_seg > 1 && ctx->left_to_snd))
    session_tx_fifo_chain_tail (vm, ctx, b, n_bufs, peek_data);
}

always_inline u8
session_tx_not_ready (session_t * s, u8 peek_data)
{
  if (peek_data)
    {
      if (PREDICT_TRUE (s->session_state == SESSION_STATE_READY))
	return 0;
      /* Can retransmit for closed sessions but can't send new data if
       * session is not ready or closed */
      else if (s->session_state < SESSION_STATE_READY)
	return 1;
      else if (s->session_state >= SESSION_STATE_TRANSPORT_CLOSED)
	{
	  /* Allow closed transports to still send custom packets.
	   * For instance, tcp may want to send acks in time-wait. */
	  if (s->session_state != SESSION_STATE_TRANSPORT_DELETED
	      && (s->flags & SESSION_F_CUSTOM_TX))
	    return 0;
	  return 2;
	}
    }
  return 0;
}

always_inline transport_connection_t *
session_tx_get_transport (session_tx_context_t * ctx, u8 peek_data)
{
  if (peek_data)
    {
      return ctx->transport_vft->get_connection (ctx->s->connection_index,
						 ctx->s->thread_index);
    }
  else
    {
      if (ctx->s->session_state == SESSION_STATE_LISTENING)
	return ctx->transport_vft->get_listener (ctx->s->connection_index);
      else
	{
	  return ctx->transport_vft->get_connection (ctx->s->connection_index,
						     ctx->s->thread_index);
	}
    }
}

always_inline void
session_tx_set_dequeue_params (vlib_main_t * vm, session_tx_context_t * ctx,
			       u32 max_segs, u8 peek_data)
{
  u32 n_bytes_per_buf, n_bytes_per_seg;
  ctx->max_dequeue = svm_fifo_max_dequeue_cons (ctx->s->tx_fifo);
  if (peek_data)
    {
      /* Offset in rx fifo from where to peek data */
      ctx->tx_offset = ctx->transport_vft->tx_fifo_offset (ctx->tc);
      if (PREDICT_FALSE (ctx->tx_offset >= ctx->max_dequeue))
	{
	  ctx->max_len_to_snd = 0;
	  return;
	}
      ctx->max_dequeue -= ctx->tx_offset;
    }
  else
    {
      if (ctx->transport_vft->transport_options.tx_type == TRANSPORT_TX_DGRAM)
	{
	  if (ctx->max_dequeue <= sizeof (ctx->hdr))
	    {
	      ctx->max_len_to_snd = 0;
	      return;
	    }
	  svm_fifo_peek (ctx->s->tx_fifo, 0, sizeof (ctx->hdr),
			 (u8 *) & ctx->hdr);
	  ASSERT (ctx->hdr.data_length > ctx->hdr.data_offset);
	  ctx->max_dequeue = ctx->hdr.data_length - ctx->hdr.data_offset;
	}
    }
  ASSERT (ctx->max_dequeue > 0);

  /* Ensure we're not writing more than transport window allows */
  if (ctx->max_dequeue < ctx->snd_space)
    {
      /* Constrained by tx queue. Try to send only fully formed segments */
      ctx->max_len_to_snd =
	(ctx->max_dequeue > ctx->snd_mss) ?
	ctx->max_dequeue - ctx->max_dequeue % ctx->snd_mss : ctx->max_dequeue;
      /* TODO Nagle ? */
    }
  else
    {
      /* Expectation is that snd_space0 is already a multiple of snd_mss */
      ctx->max_len_to_snd = ctx->snd_space;
    }

  /* Check if we're tx constrained by the node */
  ctx->n_segs_per_evt = ceil ((f64) ctx->max_len_to_snd / ctx->snd_mss);
  if (ctx->n_segs_per_evt > max_segs)
    {
      ctx->n_segs_per_evt = max_segs;
      ctx->max_len_to_snd = max_segs * ctx->snd_mss;
    }

  n_bytes_per_buf = vlib_buffer_get_default_data_size (vm);
  ASSERT (n_bytes_per_buf > TRANSPORT_MAX_HDRS_LEN);
  n_bytes_per_seg = TRANSPORT_MAX_HDRS_LEN + ctx->snd_mss;
  ctx->n_bufs_per_seg = ceil ((f64) n_bytes_per_seg / n_bytes_per_buf);
  ctx->deq_per_buf = clib_min (ctx->snd_mss, n_bytes_per_buf);
  ctx->deq_per_first_buf = clib_min (ctx->snd_mss,
				     n_bytes_per_buf -
				     TRANSPORT_MAX_HDRS_LEN);
}

always_inline int
session_tx_fifo_read_and_snd_i (session_worker_t * wrk,
				vlib_node_runtime_t * node,
				session_evt_elt_t * elt,
				int *n_tx_packets, u8 peek_data)
{
  u32 next_index, next0, next1, *to_next, n_left_to_next, max_burst;
  u32 n_trace, n_bufs_needed = 0, n_left, pbi;
  session_tx_context_t *ctx = &wrk->ctx;
  session_main_t *smm = &session_main;
  session_event_t *e = &elt->evt;
  vlib_main_t *vm = wrk->vm;
  transport_proto_t tp;
  vlib_buffer_t *pb;
  u16 n_bufs, rv;

  if (PREDICT_FALSE ((rv = session_tx_not_ready (ctx->s, peek_data))))
    {
      if (rv < 2)
	session_evt_add_old (wrk, elt);
      return SESSION_TX_NO_DATA;
    }

  next_index = smm->session_type_to_next[ctx->s->session_type];
  next0 = next1 = next_index;
  max_burst = VLIB_FRAME_SIZE - *n_tx_packets;

  tp = session_get_transport_proto (ctx->s);
  ctx->transport_vft = transport_protocol_get_vft (tp);
  ctx->tc = session_tx_get_transport (ctx, peek_data);

  if (PREDICT_FALSE (e->event_type == SESSION_IO_EVT_TX_FLUSH))
    {
      if (ctx->transport_vft->flush_data)
	ctx->transport_vft->flush_data (ctx->tc);
    }

  if (ctx->s->flags & SESSION_F_CUSTOM_TX)
    {
      u32 n_custom_tx;
      ctx->s->flags &= ~SESSION_F_CUSTOM_TX;
      n_custom_tx = ctx->transport_vft->custom_tx (ctx->tc, max_burst);
      *n_tx_packets += n_custom_tx;
      if (PREDICT_FALSE
	  (ctx->s->session_state >= SESSION_STATE_TRANSPORT_CLOSED))
	return SESSION_TX_OK;
      max_burst -= n_custom_tx;
      if (!max_burst)
	{
	  session_evt_add_old (wrk, elt);
	  return SESSION_TX_OK;
	}
    }

  ctx->snd_mss = ctx->transport_vft->send_mss (ctx->tc);
  ctx->snd_space = transport_connection_snd_space (ctx->tc,
						   vm->clib_time.
						   last_cpu_time,
						   ctx->snd_mss);

  if (ctx->snd_space == 0 || ctx->snd_mss == 0)
    {
      session_evt_add_old (wrk, elt);
      return SESSION_TX_NO_DATA;
    }

  /* Allow enqueuing of a new event */
  svm_fifo_unset_event (ctx->s->tx_fifo);

  /* Check how much we can pull. */
  session_tx_set_dequeue_params (vm, ctx, max_burst, peek_data);

  if (PREDICT_FALSE (!ctx->max_len_to_snd))
    return SESSION_TX_NO_DATA;

  n_bufs_needed = ctx->n_segs_per_evt * ctx->n_bufs_per_seg;
  vec_validate_aligned (wrk->tx_buffers, n_bufs_needed - 1,
			CLIB_CACHE_LINE_BYTES);
  n_bufs = vlib_buffer_alloc (vm, wrk->tx_buffers, n_bufs_needed);
  if (PREDICT_FALSE (n_bufs < n_bufs_needed))
    {
      if (n_bufs)
	vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
      session_evt_add_old (wrk, elt);
      vlib_node_increment_counter (wrk->vm, node->node_index,
				   SESSION_QUEUE_ERROR_NO_BUFFER, 1);
      return SESSION_TX_NO_BUFFERS;
    }

  /*
   * Write until we fill up a frame
   */
  vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
  if (PREDICT_FALSE (ctx->n_segs_per_evt > n_left_to_next))
    {
      ctx->n_segs_per_evt = n_left_to_next;
      ctx->max_len_to_snd = ctx->snd_mss * n_left_to_next;
    }
  ctx->left_to_snd = ctx->max_len_to_snd;
  n_left = ctx->n_segs_per_evt;

  while (n_left >= 4)
    {
      vlib_buffer_t *b0, *b1;
      u32 bi0, bi1;

      pbi = wrk->tx_buffers[n_bufs - 3];
      pb = vlib_get_buffer (vm, pbi);
      vlib_prefetch_buffer_header (pb, STORE);
      pbi = wrk->tx_buffers[n_bufs - 4];
      pb = vlib_get_buffer (vm, pbi);
      vlib_prefetch_buffer_header (pb, STORE);

      to_next[0] = bi0 = wrk->tx_buffers[--n_bufs];
      to_next[1] = bi1 = wrk->tx_buffers[--n_bufs];

      b0 = vlib_get_buffer (vm, bi0);
      b1 = vlib_get_buffer (vm, bi1);

      session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data);
      session_tx_fill_buffer (vm, ctx, b1, &n_bufs, peek_data);

      ctx->transport_vft->push_header (ctx->tc, b0);
      ctx->transport_vft->push_header (ctx->tc, b1);

      to_next += 2;
      n_left_to_next -= 2;
      n_left -= 2;

      VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);
      VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b1);

      vlib_validate_buffer_enqueue_x2 (vm, node, next_index, to_next,
				       n_left_to_next, bi0, bi1, next0,
				       next1);
    }
  while (n_left)
    {
      vlib_buffer_t *b0;
      u32 bi0;

      if (n_left > 1)
	{
	  pbi = wrk->tx_buffers[n_bufs - 2];
	  pb = vlib_get_buffer (vm, pbi);
	  vlib_prefetch_buffer_header (pb, STORE);
	}

      to_next[0] = bi0 = wrk->tx_buffers[--n_bufs];
      b0 = vlib_get_buffer (vm, bi0);
      session_tx_fill_buffer (vm, ctx, b0, &n_bufs, peek_data);

      /* Ask transport to push header after current_length and
       * total_length_not_including_first_buffer are updated */
      ctx->transport_vft->push_header (ctx->tc, b0);

      to_next += 1;
      n_left_to_next -= 1;
      n_left -= 1;

      VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);

      vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
				       n_left_to_next, bi0, next0);
    }

  if (PREDICT_FALSE ((n_trace = vlib_get_trace_count (vm, node)) > 0))
    session_tx_trace_frame (vm, node, next_index, to_next,
			    ctx->n_segs_per_evt, ctx->s, n_trace);

  if (PREDICT_FALSE (n_bufs))
    vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);

  *n_tx_packets += ctx->n_segs_per_evt;
  transport_connection_update_tx_bytes (ctx->tc, ctx->max_len_to_snd);
  vlib_put_next_frame (vm, node, next_index, n_left_to_next);

  SESSION_EVT (SESSION_EVT_DEQ, ctx->s, ctx->max_len_to_snd, ctx->max_dequeue,
	       ctx->s->tx_fifo->has_event, wrk->last_vlib_time);

  /* If we couldn't dequeue all bytes mark as partially read */
  ASSERT (ctx->left_to_snd == 0);
  if (ctx->max_len_to_snd < ctx->max_dequeue)
    if (svm_fifo_set_event (ctx->s->tx_fifo))
      session_evt_add_old (wrk, elt);

  if (!peek_data
      && ctx->transport_vft->transport_options.tx_type == TRANSPORT_TX_DGRAM)
    {
      /* Fix dgram pre header */
      if (ctx->max_len_to_snd < ctx->max_dequeue)
	svm_fifo_overwrite_head (ctx->s->tx_fifo, (u8 *) & ctx->hdr,
				 sizeof (session_dgram_pre_hdr_t));
      /* More data needs to be read */
      else if (svm_fifo_max_dequeue_cons (ctx->s->tx_fifo) > 0)
	if (svm_fifo_set_event (ctx->s->tx_fifo))
	  session_evt_add_old (wrk, elt);

      if (svm_fifo_needs_deq_ntf (ctx->s->tx_fifo, ctx->max_len_to_snd))
	session_dequeue_notify (ctx->s);
    }
  return SESSION_TX_OK;
}

int
session_tx_fifo_peek_and_snd (session_worker_t * wrk,
			      vlib_node_runtime_t * node,
			      session_evt_elt_t * e, int *n_tx_packets)
{
  return session_tx_fifo_read_and_snd_i (wrk, node, e, n_tx_packets, 1);
}

int
session_tx_fifo_dequeue_and_snd (session_worker_t * wrk,
				 vlib_node_runtime_t * node,
				 session_evt_elt_t * e, int *n_tx_packets)
{
  return session_tx_fifo_read_and_snd_i (wrk, node, e, n_tx_packets, 0);
}

int
session_tx_fifo_dequeue_internal (session_worker_t * wrk,
				  vlib_node_runtime_t * node,
				  session_evt_elt_t * e, int *n_tx_packets)
{
  session_t *s = wrk->ctx.s;

  if (PREDICT_FALSE (s->session_state >= SESSION_STATE_TRANSPORT_CLOSED))
    return 0;
  svm_fifo_unset_event (s->tx_fifo);
  return transport_custom_tx (session_get_transport_proto (s), s,
			      VLIB_FRAME_SIZE - *n_tx_packets);
}

always_inline session_t *
session_event_get_session (session_event_t * e, u8 thread_index)
{
  return session_get_if_valid (e->session_index, thread_index);
}

always_inline void
session_event_dispatch_ctrl (session_worker_t * wrk, session_evt_elt_t * elt)
{
  clib_llist_index_t ei;
  void (*fp) (void *);
  session_event_t *e;
  session_t *s;

  ei = clib_llist_entry_index (wrk->event_elts, elt);
  e = &elt->evt;

  switch (e->event_type)
    {
    case SESSION_CTRL_EVT_RPC:
      fp = e->rpc_args.fp;
      (*fp) (e->rpc_args.arg);
      break;
    case SESSION_CTRL_EVT_CLOSE:
      s = session_get_from_handle_if_valid (e->session_handle);
      if (PREDICT_FALSE (!s))
	break;
      session_transport_close (s);
      break;
    case SESSION_CTRL_EVT_RESET:
      s = session_get_from_handle_if_valid (e->session_handle);
      if (PREDICT_FALSE (!s))
	break;
      session_transport_reset (s);
      break;
    case SESSION_CTRL_EVT_LISTEN:
      session_mq_listen_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_LISTEN_URI:
      session_mq_listen_uri_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_UNLISTEN:
      session_mq_unlisten_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_CONNECT:
      session_mq_connect_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_CONNECT_URI:
      session_mq_connect_uri_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_DISCONNECT:
      session_mq_disconnect_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_DISCONNECTED:
      session_mq_disconnected_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_ACCEPTED_REPLY:
      session_mq_accepted_reply_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_DISCONNECTED_REPLY:
      session_mq_disconnected_reply_handler (session_evt_ctrl_data (wrk,
								    elt));
      break;
    case SESSION_CTRL_EVT_RESET_REPLY:
      session_mq_reset_reply_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_WORKER_UPDATE:
      session_mq_worker_update_handler (session_evt_ctrl_data (wrk, elt));
      break;
    case SESSION_CTRL_EVT_APP_DETACH:
      app_mq_detach_handler (session_evt_ctrl_data (wrk, elt));
      break;
    default:
      clib_warning ("unhandled event type %d", e->event_type);
    }

  /* Regrab elements in case pool moved */
  elt = pool_elt_at_index (wrk->event_elts, ei);
  if (!clib_llist_elt_is_linked (elt, evt_list))
    {
      e = &elt->evt;
      if (e->event_type >= SESSION_CTRL_EVT_BOUND)
	session_evt_ctrl_data_free (wrk, elt);
      session_evt_elt_free (wrk, elt);
    }
}

always_inline void
session_event_dispatch_io (session_worker_t * wrk, vlib_node_runtime_t * node,
			   session_evt_elt_t * elt, u32 thread_index,
			   int *n_tx_packets)
{
  session_main_t *smm = &session_main;
  app_worker_t *app_wrk;
  clib_llist_index_t ei;
  session_event_t *e;
  session_t *s;

  ei = clib_llist_entry_index (wrk->event_elts, elt);
  e = &elt->evt;

  switch (e->event_type)
    {
    case SESSION_IO_EVT_TX_FLUSH:
    case SESSION_IO_EVT_TX:
      s = session_event_get_session (e, thread_index);
      if (PREDICT_FALSE (!s))
	{
	  clib_warning ("session %u was freed!", e->session_index);
	  break;
	}
      CLIB_PREFETCH (s->tx_fifo, 2 * CLIB_CACHE_LINE_BYTES, LOAD);
      wrk->ctx.s = s;
      /* Spray packets in per session type frames, since they go to
       * different nodes */
      (smm->session_tx_fns[s->session_type]) (wrk, node, elt, n_tx_packets);
      break;
    case SESSION_IO_EVT_RX:
      s = session_event_get_session (e, thread_index);
      if (!s)
	break;
      transport_app_rx_evt (session_get_transport_proto (s),
			    s->connection_index, s->thread_index);
      break;
    case SESSION_IO_EVT_BUILTIN_RX:
      s = session_event_get_session (e, thread_index);
      if (PREDICT_FALSE (!s || s->session_state >= SESSION_STATE_CLOSING))
	break;
      svm_fifo_unset_event (s->rx_fifo);
      app_wrk = app_worker_get (s->app_wrk_index);
      app_worker_builtin_rx (app_wrk, s);
      break;
    case SESSION_IO_EVT_BUILTIN_TX:
      s = session_get_from_handle_if_valid (e->session_handle);
      wrk->ctx.s = s;
      if (PREDICT_TRUE (s != 0))
	session_tx_fifo_dequeue_internal (wrk, node, elt, n_tx_packets);
      break;
    default:
      clib_warning ("unhandled event type %d", e->event_type);
    }

  /* Regrab elements in case pool moved */
  elt = pool_elt_at_index (wrk->event_elts, ei);
  if (!clib_llist_elt_is_linked (elt, evt_list))
    session_evt_elt_free (wrk, elt);
}

/* *INDENT-OFF* */
static const u32 session_evt_msg_sizes[] = {
#define _(symc, sym) 							\
  [SESSION_CTRL_EVT_ ## symc] = sizeof (session_ ## sym ##_msg_t),
  foreach_session_ctrl_evt
#undef _
};
/* *INDENT-ON* */

always_inline void
session_evt_add_to_list (session_worker_t * wrk, session_event_t * evt)
{
  session_evt_elt_t *elt;

  if (evt->event_type >= SESSION_CTRL_EVT_RPC)
    {
      elt = session_evt_alloc_ctrl (wrk);
      if (evt->event_type >= SESSION_CTRL_EVT_BOUND)
	{
	  elt->evt.ctrl_data_index = session_evt_ctrl_data_alloc (wrk);
	  elt->evt.event_type = evt->event_type;
	  clib_memcpy_fast (session_evt_ctrl_data (wrk, elt), evt->data,
			    session_evt_msg_sizes[evt->event_type]);
	}
      else
	{
	  /* Internal control events fit into io events footprint */
	  clib_memcpy_fast (&elt->evt, evt, sizeof (elt->evt));
	}
    }
  else
    {
      elt = session_evt_alloc_new (wrk);
      clib_memcpy_fast (&elt->evt, evt, sizeof (elt->evt));
    }
}

static uword
session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
		       vlib_frame_t * frame)
{
  session_main_t *smm = vnet_get_session_main ();
  u32 thread_index = vm->thread_index, n_to_dequeue;
  session_worker_t *wrk = &smm->wrk[thread_index];
  session_evt_elt_t *elt, *ctrl_he, *new_he, *old_he;
  svm_msg_q_msg_t _msg, *msg = &_msg;
  clib_llist_index_t old_ti;
  int i, n_tx_packets = 0;
  session_event_t *evt;
  svm_msg_q_t *mq;

  SESSION_EVT (SESSION_EVT_DISPATCH_START, wrk);

  wrk->last_vlib_time = vlib_time_now (vm);

  /*
   *  Update transport time
   */
  transport_update_time (wrk->last_vlib_time, thread_index);

  /*
   *  Dequeue and handle new events
   */

  /* Try to dequeue what is available. Don't wait for lock.
   * XXX: we may need priorities here */
  mq = wrk->vpp_event_queue;
  n_to_dequeue = svm_msg_q_size (mq);
  if (n_to_dequeue && svm_msg_q_try_lock (mq) == 0)
    {
      for (i = 0; i < n_to_dequeue; i++)
	{
	  svm_msg_q_sub_w_lock (mq, msg);
	  evt = svm_msg_q_msg_data (mq, msg);
	  session_evt_add_to_list (wrk, evt);
	  svm_msg_q_free_msg (mq, msg);
	}
      svm_msg_q_unlock (mq);
    }

  /*
   * Handle control events
   */

  ctrl_he = pool_elt_at_index (wrk->event_elts, wrk->ctrl_head);

  /* *INDENT-OFF* */
  clib_llist_foreach_safe (wrk->event_elts, evt_list, ctrl_he, elt, ({
    clib_llist_remove (wrk->event_elts, evt_list, elt);
    session_event_dispatch_ctrl (wrk, elt);
  }));
  /* *INDENT-ON* */

  /*
   * Handle the new io events.
   */

  new_he = pool_elt_at_index (wrk->event_elts, wrk->new_head);

  /* *INDENT-OFF* */
  clib_llist_foreach_safe (wrk->event_elts, evt_list, new_he, elt, ({
    session_evt_type_t et;

    et = elt->evt.event_type;
    clib_llist_remove (wrk->event_elts, evt_list, elt);

    /* Postpone tx events if we can't handle them this dispatch cycle */
    if (n_tx_packets >= VLIB_FRAME_SIZE
	&& (et == SESSION_IO_EVT_TX || et == SESSION_IO_EVT_TX_FLUSH))
      {
	clib_llist_add (wrk->event_elts, evt_list, elt, new_he);
	continue;
      }

    session_event_dispatch_io (wrk, node, elt, thread_index, &n_tx_packets);
  }));
  /* *INDENT-ON* */

  /*
   * Handle the old io events
   */

  old_he = pool_elt_at_index (wrk->event_elts, wrk->old_head);
  old_ti = clib_llist_prev_index (old_he, evt_list);

  while (n_tx_packets < VLIB_FRAME_SIZE
	 && !clib_llist_is_empty (wrk->event_elts, evt_list, old_he))
    {
      clib_llist_index_t ei;

      clib_llist_pop_first (wrk->event_elts, evt_list, elt, old_he);
      ei = clib_llist_entry_index (wrk->event_elts, elt);
      session_event_dispatch_io (wrk, node, elt, thread_index, &n_tx_packets);

      old_he = pool_elt_at_index (wrk->event_elts, wrk->old_head);
      if (ei == old_ti)
	break;
    };

  vlib_node_increment_counter (vm, session_queue_node.index,
			       SESSION_QUEUE_ERROR_TX, n_tx_packets);

  SESSION_EVT (SESSION_EVT_DISPATCH_END, wrk, n_tx_packets);

  return n_tx_packets;
}

/* *INDENT-OFF* */
VLIB_REGISTER_NODE (session_queue_node) =
{
  .function = session_queue_node_fn,
  .flags = VLIB_NODE_FLAG_TRACE_SUPPORTED,
  .name = "session-queue",
  .format_trace = format_session_queue_trace,
  .type = VLIB_NODE_TYPE_INPUT,
  .n_errors = ARRAY_LEN (session_queue_error_strings),
  .error_strings = session_queue_error_strings,
  .state = VLIB_NODE_STATE_DISABLED,
};
/* *INDENT-ON* */

void
dump_thread_0_event_queue (void)
{
  session_main_t *smm = vnet_get_session_main ();
  vlib_main_t *vm = &vlib_global_main;
  u32 my_thread_index = vm->thread_index;
  session_event_t _e, *e = &_e;
  svm_msg_q_ring_t *ring;
  session_t *s0;
  svm_msg_q_msg_t *msg;
  svm_msg_q_t *mq;
  int i, index;

  mq = smm->wrk[my_thread_index].vpp_event_queue;
  index = mq->q->head;

  for (i = 0; i < mq->q->cursize; i++)
    {
      msg = (svm_msg_q_msg_t *) (&mq->q->data[0] + mq->q->elsize * index);
      ring = svm_msg_q_ring (mq, msg->ring_index);
      clib_memcpy_fast (e, svm_msg_q_msg_data (mq, msg), ring->elsize);

      switch (e->event_type)
	{
	case SESSION_IO_EVT_TX:
	  s0 = session_event_get_session (e, my_thread_index);
	  fformat (stdout, "[%04d] TX session %d\n", i, s0->session_index);
	  break;

	case SESSION_CTRL_EVT_CLOSE:
	  s0 = session_get_from_handle (e->session_handle);
	  fformat (stdout, "[%04d] disconnect session %d\n", i,
		   s0->session_index);
	  break;

	case SESSION_IO_EVT_BUILTIN_RX:
	  s0 = session_event_get_session (e, my_thread_index);
	  fformat (stdout, "[%04d] builtin_rx %d\n", i, s0->session_index);
	  break;

	case SESSION_CTRL_EVT_RPC:
	  fformat (stdout, "[%04d] RPC call %llx with %llx\n",
		   i, (u64) (uword) (e->rpc_args.fp),
		   (u64) (uword) (e->rpc_args.arg));
	  break;

	default:
	  fformat (stdout, "[%04d] unhandled event type %d\n",
		   i, e->event_type);
	  break;
	}

      index++;

      if (index == mq->q->maxsize)
	index = 0;
    }
}

static u8
session_node_cmp_event (session_event_t * e, svm_fifo_t * f)
{
  session_t *s;
  switch (e->event_type)
    {
    case SESSION_IO_EVT_RX:
    case SESSION_IO_EVT_TX:
    case SESSION_IO_EVT_BUILTIN_RX:
      if (e->session_index == f->master_session_index)
	return 1;
      break;
    case SESSION_CTRL_EVT_CLOSE:
      break;
    case SESSION_CTRL_EVT_RPC:
      s = session_get_from_handle (e->session_handle);
      if (!s)
	{
	  clib_warning ("session has event but doesn't exist!");
	  break;
	}
      if (s->rx_fifo == f || s->tx_fifo == f)
	return 1;
      break;
    default:
      break;
    }
  return 0;
}

u8
session_node_lookup_fifo_event (svm_fifo_t * f, session_event_t * e)
{
  session_evt_elt_t *elt;
  session_worker_t *wrk;
  int i, index, found = 0;
  svm_msg_q_msg_t *msg;
  svm_msg_q_ring_t *ring;
  svm_msg_q_t *mq;
  u8 thread_index;

  ASSERT (e);
  thread_index = f->master_thread_index;
  wrk = session_main_get_worker (thread_index);

  /*
   * Search evt queue
   */
  mq = wrk->vpp_event_queue;
  index = mq->q->head;
  for (i = 0; i < mq->q->cursize; i++)
    {
      msg = (svm_msg_q_msg_t *) (&mq->q->data[0] + mq->q->elsize * index);
      ring = svm_msg_q_ring (mq, msg->ring_index);
      clib_memcpy_fast (e, svm_msg_q_msg_data (mq, msg), ring->elsize);
      found = session_node_cmp_event (e, f);
      if (found)
	return 1;
      if (++index == mq->q->maxsize)
	index = 0;
    }
  /*
   * Search pending events vector
   */

  /* *INDENT-OFF* */
  clib_llist_foreach (wrk->event_elts, evt_list,
                      pool_elt_at_index (wrk->event_elts, wrk->old_head),
                      elt, ({
    found = session_node_cmp_event (&elt->evt, f);
    if (found)
      {
	clib_memcpy_fast (e, &elt->evt, sizeof (*e));
	break;
      }

  }));
  /* *INDENT-ON* */

  return found;
}

static clib_error_t *
session_queue_exit (vlib_main_t * vm)
{
  if (vec_len (vlib_mains) < 2)
    return 0;

  /*
   * Shut off (especially) worker-thread session nodes.
   * Otherwise, vpp can crash as the main thread unmaps the
   * API segment.
   */
  vlib_worker_thread_barrier_sync (vm);
  session_node_enable_disable (0 /* is_enable */ );
  vlib_worker_thread_barrier_release (vm);
  return 0;
}

VLIB_MAIN_LOOP_EXIT_FUNCTION (session_queue_exit);

static uword
session_queue_process (vlib_main_t * vm, vlib_node_runtime_t * rt,
		       vlib_frame_t * f)
{
  f64 now, timeout = 1.0;
  uword *event_data = 0;
  uword event_type;

  while (1)
    {
      vlib_process_wait_for_event_or_clock (vm, timeout);
      now = vlib_time_now (vm);
      event_type = vlib_process_get_events (vm, (uword **) & event_data);

      switch (event_type)
	{
	case SESSION_Q_PROCESS_FLUSH_FRAMES:
	  /* Flush the frames by updating all transports times */
	  transport_update_time (now, 0);
	  break;
	case SESSION_Q_PROCESS_STOP:
	  timeout = 100000.0;
	  break;
	case ~0:
	  /* Timed out. Update time for all transports to trigger all
	   * outstanding retransmits. */
	  transport_update_time (now, 0);
	  break;
	}
      vec_reset_length (event_data);
    }
  return 0;
}

/* *INDENT-OFF* */
VLIB_REGISTER_NODE (session_queue_process_node) =
{
  .function = session_queue_process,
  .type = VLIB_NODE_TYPE_PROCESS,
  .name = "session-queue-process",
  .state = VLIB_NODE_STATE_DISABLED,
};
/* *INDENT-ON* */

static_always_inline uword
session_queue_pre_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
				vlib_frame_t * frame)
{
  session_main_t *sm = &session_main;
  if (!sm->wrk[0].vpp_event_queue)
    return 0;
  return session_queue_node_fn (vm, node, frame);
}

/* *INDENT-OFF* */
VLIB_REGISTER_NODE (session_queue_pre_input_node) =
{
  .function = session_queue_pre_input_inline,
  .type = VLIB_NODE_TYPE_PRE_INPUT,
  .name = "session-queue-main",
  .state = VLIB_NODE_STATE_DISABLED,
};
/* *INDENT-ON* */

/*
 * fd.io coding-style-patch-verification: ON
 *
 * Local Variables:
 * eval: (c-set-style "gnu")
 * End:
 */
Age	Commit message (Expand)	Author	Files	Lines
2020-02-18	misc: fix coverity warnings	Dave Barach	1	-1/+1
2020-02-17	misc: fix coverity warnings	Dave Barach	2	-3/+6
2019-12-20	vxlan: reuse inner packet flow hash for tunnel outer header load balance	Shawn Ji	1	-0/+8
2019-12-17	vxlan geneve gtpu: fix short help	Paul Vinciguerra	1	-1/+1
2019-12-10	api: multiple connections per process	Dave Barach	1	-1/+1
2019-10-10	geneve: API cleanup	Jakub Grajciar	2	-32/+31
2019-07-19	fib: FIB Entry tracking	Neale Ranns	1	-8/+7
2019-07-10	misc: fix buffer issue in geneve_input	Zhiyong Yang	1	-10/+8
2019-06-18	fib: fib api updates	Neale Ranns	1	-7/+5
2019-06-04	features will register udp ports once configured	Jakub Grajciar	1	-5/+11
2019-05-27	geneve: encap performance enhancement	Zhiyong Yang	1	-58/+53
2019-05-21	geneve_input: leverage vlib_get_buffers	Zhiyong Yang	1	-12/+11
2019-04-08	fixing typos	Jim Thompson	1	-1/+1
2019-03-28	Typos. A bunch of typos I've been collecting.	Paul Vinciguerra	1	-4/+4
2019-03-06	geneve: migrate old MULTIARCH macros to VLIB_NODE_FN	Filip Tehlar	4	-57/+44
2018-11-14	Remove c-11 memcpy checks from perf-critical code	Dave Barach	1	-3/+3
2018-10-23	c11 safe string handling support	Dave Barach	3	-6/+6
2018-09-25	L2 BD: introduce a BD interface on which to send UU packets	Neale Ranns	1	-2/+2
2018-09-07	geneve: fix variable initial value	Zhiyong Yang	1	-1/+1
2018-07-19	Remove unused argument to vlib_feature_next	Damjan Marion	1	-6/+3
2018-07-11	avoid using thread local storage for thread index	Damjan Marion	2	-2/+2
2018-05-04	Harmonize vec/pool_get_aligned object sizes and alignment requests	Dave Barach	1	-0/+3
2018-02-01	FIB: Consolidate several copies of fib_ip_proto() into one.	Jon Loeliger	1	-6/+0
2018-01-24	Improve tunnel interface creation performance	John Lo	1	-14/+5
2018-01-23	VPPAPIGEN: vppapigen replacement in Python PLY.	Ole Troan	1	-1/+1
2018-01-13	Improve "show xxx tunnel" and "show int addr" output	John Lo	1	-7/+7
2018-01-11	api: remove transport specific code from handlers	Florin Coras	1	-10/+8
2018-01-09	api: refactor vlibmemory	Florin Coras	1	-2/+2
2018-01-05	Unify and cleanup usage of hash_set/unset_mem by various tunnels	John Lo	1	-26/+7
2017-11-27	Address COVERITY warnings	Marco Varlese	1	-0/+6
2017-11-26	FIB: store the node type not the function pointer.	Neale Ranns	1	-2/+0
2017-11-24	GENEVE: shift/mask for header	Marco Varlese	3	-51/+133
2017-10-24	Add extern to *_main global variable declarations in header files.	Dave Wallace	1	-1/+1
2017-10-17	Add API versioning to GENEVE tunnel implementation.	Marco Varlese	1	-1/+3
2017-10-06	Initial GENEVE TUNNEL implementation and tests.	Marco Varlese	8	-0/+3794