src/vnet/session/application_local.c

/*
 * Copyright (c) 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 <vnet/session/application_local.h>
#include <vnet/session/session.h>

typedef enum ct_segment_flags_
{
  CT_SEGMENT_F_CLIENT_DETACHED = 1 << 0,
  CT_SEGMENT_F_SERVER_DETACHED = 1 << 1,
} ct_segment_flags_t;

typedef struct ct_segment_
{
  u32 client_n_sessions;
  u32 server_n_sessions;
  u32 seg_ctx_index;
  u32 ct_seg_index;
  u32 segment_index;
  ct_segment_flags_t flags;
} ct_segment_t;

typedef struct ct_segments_
{
  u32 sm_index;
  u32 server_wrk;
  u32 client_wrk;
  u32 fifo_pair_bytes;
  ct_segment_t *segments;
} ct_segments_ctx_t;

typedef struct ct_cleanup_req_
{
  u32 ct_index;
} ct_cleanup_req_t;

typedef struct ct_worker_
{
  ct_connection_t *connections;	      /**< Per-worker connection pools */
  u32 *pending_connects;	      /**< Fifo of pending ho indices */
  ct_cleanup_req_t *pending_cleanups; /**< Fifo of pending indices */
  u8 have_connects;		      /**< Set if connect rpc pending */
  u8 have_cleanups;		      /**< Set if cleanup rpc pending */
  clib_spinlock_t pending_connects_lock; /**< Lock for pending connects */
  u32 *new_connects;			 /**< Burst of connects to be done */
} ct_worker_t;

typedef struct ct_main_
{
  ct_worker_t *wrk;			/**< Per-worker state */
  u32 n_workers;			/**< Number of vpp workers */
  u32 n_sessions;			/**< Cumulative sessions counter */
  u32 *ho_reusable;			/**< Vector of reusable ho indices */
  clib_spinlock_t ho_reuseable_lock;	/**< Lock for reusable ho indices */
  clib_rwlock_t app_segs_lock;		/**< RW lock for seg contexts */
  uword *app_segs_ctxs_table;		/**< App handle to segment pool map */
  ct_segments_ctx_t *app_seg_ctxs;	/**< Pool of ct segment contexts */
  u32 **fwrk_pending_connects;		/**< First wrk pending half-opens */
  u32 fwrk_thread;			/**< First worker thread */
  u8 fwrk_have_flush;			/**< Flag for connect flush rpc */
  u8 is_init;
} ct_main_t;

static ct_main_t ct_main;

static inline ct_worker_t *
ct_worker_get (u32 thread_index)
{
  return &ct_main.wrk[thread_index];
}

static ct_connection_t *
ct_connection_alloc (u32 thread_index)
{
  ct_worker_t *wrk = ct_worker_get (thread_index);
  ct_connection_t *ct;

  pool_get_aligned_safe (wrk->connections, ct, CLIB_CACHE_LINE_BYTES);
  clib_memset (ct, 0, sizeof (*ct));
  ct->c_c_index = ct - wrk->connections;
  ct->c_thread_index = thread_index;
  ct->client_wrk = ~0;
  ct->server_wrk = ~0;
  ct->seg_ctx_index = ~0;
  ct->ct_seg_index = ~0;
  return ct;
}

static ct_connection_t *
ct_connection_get (u32 ct_index, u32 thread_index)
{
  ct_worker_t *wrk = ct_worker_get (thread_index);

  if (pool_is_free_index (wrk->connections, ct_index))
    return 0;
  return pool_elt_at_index (wrk->connections, ct_index);
}

static void
ct_connection_free (ct_connection_t * ct)
{
  ct_worker_t *wrk = ct_worker_get (ct->c_thread_index);

  if (CLIB_DEBUG)
    {
      clib_memset (ct, 0xfc, sizeof (*ct));
      pool_put (wrk->connections, ct);
      return;
    }
  pool_put (wrk->connections, ct);
}

static ct_connection_t *
ct_half_open_alloc (void)
{
  ct_main_t *cm = &ct_main;
  u32 *hip;

  clib_spinlock_lock (&cm->ho_reuseable_lock);
  vec_foreach (hip, cm->ho_reusable)
    pool_put_index (cm->wrk[cm->fwrk_thread].connections, *hip);
  vec_reset_length (cm->ho_reusable);
  clib_spinlock_unlock (&cm->ho_reuseable_lock);

  return ct_connection_alloc (cm->fwrk_thread);
}

static ct_connection_t *
ct_half_open_get (u32 ho_index)
{
  ct_main_t *cm = &ct_main;
  return ct_connection_get (ho_index, cm->fwrk_thread);
}

void
ct_half_open_add_reusable (u32 ho_index)
{
  ct_main_t *cm = &ct_main;

  clib_spinlock_lock (&cm->ho_reuseable_lock);
  vec_add1 (cm->ho_reusable, ho_index);
  clib_spinlock_unlock (&cm->ho_reuseable_lock);
}

session_t *
ct_session_get_peer (session_t * s)
{
  ct_connection_t *ct, *peer_ct;
  ct = ct_connection_get (s->connection_index, s->thread_index);
  peer_ct = ct_connection_get (ct->peer_index, s->thread_index);
  return session_get (peer_ct->c_s_index, s->thread_index);
}

void
ct_session_endpoint (session_t * ll, session_endpoint_t * sep)
{
  ct_connection_t *ct;
  ct = (ct_connection_t *) session_get_transport (ll);
  sep->transport_proto = ct->actual_tp;
  sep->port = ct->c_lcl_port;
  sep->is_ip4 = ct->c_is_ip4;
  ip_copy (&sep->ip, &ct->c_lcl_ip, ct->c_is_ip4);
}

static void
ct_set_invalid_app_wrk (ct_connection_t *ct, u8 is_client)
{
  ct_connection_t *peer_ct;

  peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index);

  if (is_client)
    {
      ct->client_wrk = APP_INVALID_INDEX;
      if (peer_ct)
	ct->client_wrk = APP_INVALID_INDEX;
    }
  else
    {
      ct->server_wrk = APP_INVALID_INDEX;
      if (peer_ct)
	ct->server_wrk = APP_INVALID_INDEX;
    }
}

static inline u64
ct_client_seg_handle (u64 server_sh, u32 client_wrk_index)
{
  return (((u64) client_wrk_index << 56) | server_sh);
}

static void
ct_session_dealloc_fifos (ct_connection_t *ct, svm_fifo_t *rx_fifo,
			  svm_fifo_t *tx_fifo)
{
  ct_segments_ctx_t *seg_ctx;
  ct_main_t *cm = &ct_main;
  segment_manager_t *sm;
  app_worker_t *app_wrk;
  ct_segment_t *ct_seg;
  fifo_segment_t *fs;
  u32 seg_index;
  session_t *s;
  int cnt;

  /*
   * Cleanup fifos
   */

  sm = segment_manager_get (rx_fifo->segment_manager);
  seg_index = rx_fifo->segment_index;

  fs = segment_manager_get_segment_w_lock (sm, seg_index);
  fifo_segment_free_fifo (fs, rx_fifo);
  fifo_segment_free_fifo (fs, tx_fifo);
  segment_manager_segment_reader_unlock (sm);

  /*
   * Atomically update segment context with readers lock
   */

  clib_rwlock_reader_lock (&cm->app_segs_lock);

  seg_ctx = pool_elt_at_index (cm->app_seg_ctxs, ct->seg_ctx_index);
  ct_seg = pool_elt_at_index (seg_ctx->segments, ct->ct_seg_index);

  if (ct->flags & CT_CONN_F_CLIENT)
    {
      cnt =
	__atomic_sub_fetch (&ct_seg->client_n_sessions, 1, __ATOMIC_RELAXED);
    }
  else
    {
      cnt =
	__atomic_sub_fetch (&ct_seg->server_n_sessions, 1, __ATOMIC_RELAXED);
    }

  clib_rwlock_reader_unlock (&cm->app_segs_lock);

  /*
   * No need to do any app updates, return
   */
  ASSERT (cnt >= 0);
  if (cnt)
    return;

  /*
   * Grab exclusive lock and update flags unless some other thread
   * added more sessions
   */
  clib_rwlock_writer_lock (&cm->app_segs_lock);

  seg_ctx = pool_elt_at_index (cm->app_seg_ctxs, ct->seg_ctx_index);
  ct_seg = pool_elt_at_index (seg_ctx->segments, ct->ct_seg_index);
  if (ct->flags & CT_CONN_F_CLIENT)
    {
      cnt = ct_seg->client_n_sessions;
      if (cnt)
	goto done;
      ct_seg->flags |= CT_SEGMENT_F_CLIENT_DETACHED;
      s = session_get (ct->c_s_index, ct->c_thread_index);
      if (s->app_wrk_index == APP_INVALID_INDEX)
	ct_set_invalid_app_wrk (ct, 1 /* is_client */);
    }
  else
    {
      cnt = ct_seg->server_n_sessions;
      if (cnt)
	goto done;
      ct_seg->flags |= CT_SEGMENT_F_SERVER_DETACHED;
      s = session_get (ct->c_s_index, ct->c_thread_index);
      if (s->app_wrk_index == APP_INVALID_INDEX)
	ct_set_invalid_app_wrk (ct, 0 /* is_client */);
    }

  if (!(ct_seg->flags & CT_SEGMENT_F_CLIENT_DETACHED) ||
      !(ct_seg->flags & CT_SEGMENT_F_SERVER_DETACHED))
    goto done;

  /*
   * Remove segment context because both client and server detached
   */

  pool_put_index (seg_ctx->segments, ct->ct_seg_index);

  /*
   * No more segment indices left, remove the segments context
   */
  if (!pool_elts (seg_ctx->segments))
    {
      u64 table_handle = seg_ctx->client_wrk << 16 | seg_ctx->server_wrk;
      table_handle = (u64) seg_ctx->sm_index << 32 | table_handle;
      hash_unset (cm->app_segs_ctxs_table, table_handle);
      pool_free (seg_ctx->segments);
      pool_put_index (cm->app_seg_ctxs, ct->seg_ctx_index);
    }

  /*
   * Segment to be removed so notify both apps
   */

  app_wrk = app_worker_get_if_valid (ct->client_wrk);
  /* Determine if client app still needs notification, i.e., if it is
   * still attached. If client detached and this is the last ct session
   * on this segment, then its connects segment manager should also be
   * detached, so do not send notification */
  if (app_wrk)
    {
      segment_manager_t *csm;
      csm = app_worker_get_connect_segment_manager (app_wrk);
      if (!segment_manager_app_detached (csm))
	app_worker_del_segment_notify (
	  app_wrk, ct_client_seg_handle (ct->segment_handle, ct->client_wrk));
    }

  /* Notify server app and free segment */
  segment_manager_lock_and_del_segment (sm, seg_index);

  /* Cleanup segment manager if needed. If server detaches there's a chance
   * the client's sessions will hold up segment removal */
  if (segment_manager_app_detached (sm) && !segment_manager_has_fifos (sm))
    segment_manager_free_safe (sm);

done:

  clib_rwlock_writer_unlock (&cm->app_segs_lock);
}

static void
ct_session_force_disconnect_server (ct_connection_t *sct)
{
  sct->peer_index = ~0;
  session_transport_closing_notify (&sct->connection);
}

int
ct_session_connect_notify (session_t *ss, session_error_t err)
{
  u32 ss_index, opaque, thread_index;
  ct_connection_t *sct, *cct;
  app_worker_t *client_wrk;
  session_t *cs;

  ss_index = ss->session_index;
  thread_index = ss->thread_index;
  sct = (ct_connection_t *) session_get_transport (ss);
  client_wrk = app_worker_get (sct->client_wrk);
  opaque = sct->client_opaque;

  cct = ct_connection_get (sct->peer_index, thread_index);

  /* Client closed while waiting for reply from server */
  if (PREDICT_FALSE (!cct))
    {
      ct_session_force_disconnect_server (sct);
      return 0;
    }

  session_half_open_delete_notify (&cct->connection);
  cct->flags &= ~CT_CONN_F_HALF_OPEN;

  if (PREDICT_FALSE (err))
    goto connect_error;

  /*
   * Alloc client session, server session assumed to be established
   */

  ASSERT (ss->session_state >= SESSION_STATE_READY);

  cs = session_alloc (thread_index);
  ss = session_get (ss_index, thread_index);
  cs->session_type = ss->session_type;
  cs->listener_handle = SESSION_INVALID_HANDLE;
  session_set_state (cs, SESSION_STATE_CONNECTING);
  cs->app_wrk_index = client_wrk->wrk_index;
  cs->connection_index = cct->c_c_index;
  cs->opaque = opaque;
  cct->c_s_index = cs->session_index;

  /* This will allocate fifos for the session. They won't be used for
   * exchanging data but they will be used to close the connection if
   * the segment manager/worker is freed */
  if ((err = app_worker_init_connected (client_wrk, cs)))
    {
      session_free (cs);
      ct_session_force_disconnect_server (sct);
      err = SESSION_E_ALLOC;
      goto connect_error;
    }

  session_set_state (cs, SESSION_STATE_CONNECTING);

  if (app_worker_connect_notify (client_wrk, cs, 0, opaque))
    {
      segment_manager_dealloc_fifos (cs->rx_fifo, cs->tx_fifo);
      session_free (cs);
      ct_session_force_disconnect_server (sct);
      goto cleanup_client;
    }

  cs = session_get (cct->c_s_index, cct->c_thread_index);
  session_set_state (cs, SESSION_STATE_READY);

  return 0;

connect_error:

  app_worker_connect_notify (client_wrk, 0, err, cct->client_opaque);

cleanup_client:

  if (cct->client_rx_fifo)
    ct_session_dealloc_fifos (cct, cct->client_rx_fifo, cct->client_tx_fifo);
  ct_connection_free (cct);
  return -1;
}

static inline ct_segment_t *
ct_lookup_free_segment (ct_main_t *cm, segment_manager_t *sm,
			u32 seg_ctx_index)
{
  uword free_bytes, max_free_bytes;
  ct_segment_t *ct_seg, *res = 0;
  ct_segments_ctx_t *seg_ctx;
  fifo_segment_t *fs;
  u32 max_fifos;

  seg_ctx = pool_elt_at_index (cm->app_seg_ctxs, seg_ctx_index);
  max_free_bytes = seg_ctx->fifo_pair_bytes;

  pool_foreach (ct_seg, seg_ctx->segments)
    {
      /* Client or server has detached so segment cannot be used */
      fs = segment_manager_get_segment (sm, ct_seg->segment_index);
      free_bytes = fifo_segment_available_bytes (fs);
      max_fifos = fifo_segment_size (fs) / seg_ctx->fifo_pair_bytes;
      if (free_bytes > max_free_bytes &&
	  fifo_segment_num_fifos (fs) / 2 < max_fifos)
	{
	  max_free_bytes = free_bytes;
	  res = ct_seg;
	}
    }

  return res;
}

static ct_segment_t *
ct_alloc_segment (ct_main_t *cm, app_worker_t *server_wrk, u64 table_handle,
		  segment_manager_t *sm, u32 client_wrk_index)
{
  u32 seg_ctx_index = ~0, sm_index, pair_bytes;
  u64 seg_size, seg_handle, client_seg_handle;
  segment_manager_props_t *props;
  const u32 margin = 16 << 10;
  ct_segments_ctx_t *seg_ctx;
  app_worker_t *client_wrk;
  application_t *server;
  ct_segment_t *ct_seg;
  uword *spp;
  int fs_index;

  server = application_get (server_wrk->app_index);
  props = application_segment_manager_properties (server);
  sm_index = segment_manager_index (sm);
  pair_bytes = props->rx_fifo_size + props->tx_fifo_size + margin;

  /*
   * Make sure another thread did not alloc a segment while acquiring the lock
   */

  spp = hash_get (cm->app_segs_ctxs_table, table_handle);
  if (spp)
    {
      seg_ctx_index = *spp;
      ct_seg = ct_lookup_free_segment (cm, sm, seg_ctx_index);
      if (ct_seg)
	return ct_seg;
    }

  /*
   * No segment, try to alloc one and notify the server and the client.
   * Make sure the segment is not used for other fifos
   */
  seg_size = clib_max (props->segment_size, 128 << 20);
  fs_index =
    segment_manager_add_segment2 (sm, seg_size, FIFO_SEGMENT_F_CUSTOM_USE);
  if (fs_index < 0)
    return 0;

  if (seg_ctx_index == ~0)
    {
      pool_get_zero (cm->app_seg_ctxs, seg_ctx);
      seg_ctx_index = seg_ctx - cm->app_seg_ctxs;
      hash_set (cm->app_segs_ctxs_table, table_handle, seg_ctx_index);
      seg_ctx->server_wrk = server_wrk->wrk_index;
      seg_ctx->client_wrk = client_wrk_index;
      seg_ctx->sm_index = sm_index;
      seg_ctx->fifo_pair_bytes = pair_bytes;
    }
  else
    {
      seg_ctx = pool_elt_at_index (cm->app_seg_ctxs, seg_ctx_index);
    }

  pool_get_zero (seg_ctx->segments, ct_seg);
  ct_seg->segment_index = fs_index;
  ct_seg->server_n_sessions = 0;
  ct_seg->client_n_sessions = 0;
  ct_seg->ct_seg_index = ct_seg - seg_ctx->segments;
  ct_seg->seg_ctx_index = seg_ctx_index;

  /* New segment, notify the server and client */
  seg_handle = segment_manager_make_segment_handle (sm_index, fs_index);
  if (app_worker_add_segment_notify (server_wrk, seg_handle))
    goto error;

  client_wrk = app_worker_get (client_wrk_index);
  /* Make sure client workers do not have overlapping segment handles.
   * Ideally, we should attach fs to client worker segment manager and
   * create a new handle but that's not currently possible. */
  client_seg_handle = ct_client_seg_handle (seg_handle, client_wrk_index);
  if (app_worker_add_segment_notify (client_wrk, client_seg_handle))
    {
      app_worker_del_segment_notify (server_wrk, seg_handle);
      goto error;
    }

  return ct_seg;

error:

  segment_manager_lock_and_del_segment (sm, fs_index);
  pool_put_index (seg_ctx->segments, ct_seg->seg_ctx_index);
  return 0;
}

static int
ct_init_accepted_session (app_worker_t *server_wrk, ct_connection_t *ct,
			  session_t *ls, session_t *ll)
{
  segment_manager_props_t *props;
  u64 seg_handle, table_handle;
  u32 sm_index, fs_index = ~0;
  ct_segments_ctx_t *seg_ctx;
  ct_main_t *cm = &ct_main;
  application_t *server;
  segment_manager_t *sm;
  ct_segment_t *ct_seg;
  fifo_segment_t *fs;
  uword *spp;
  int rv;

  sm = app_worker_get_listen_segment_manager (server_wrk, ll);
  sm_index = segment_manager_index (sm);
  server = application_get (server_wrk->app_index);
  props = application_segment_manager_properties (server);

  table_handle = ct->client_wrk << 16 | server_wrk->wrk_index;
  table_handle = (u64) sm_index << 32 | table_handle;

  /*
   * Check if we already have a segment that can hold the fifos
   */

  clib_rwlock_reader_lock (&cm->app_segs_lock);

  spp = hash_get (cm->app_segs_ctxs_table, table_handle);
  if (spp)
    {
      ct_seg = ct_lookup_free_segment (cm, sm, *spp);
      if (ct_seg)
	{
	  ct->seg_ctx_index = ct_seg->seg_ctx_index;
	  ct->ct_seg_index = ct_seg->ct_seg_index;
	  fs_index = ct_seg->segment_index;
	  ct_seg->flags &=
	    ~(CT_SEGMENT_F_SERVER_DETACHED | CT_SEGMENT_F_CLIENT_DETACHED);
	  __atomic_add_fetch (&ct_seg->server_n_sessions, 1, __ATOMIC_RELAXED);
	  __atomic_add_fetch (&ct_seg->client_n_sessions, 1, __ATOMIC_RELAXED);
	}
    }

  clib_rwlock_reader_unlock (&cm->app_segs_lock);

  /*
   * If not, grab exclusive lock and allocate segment
   */
  if (fs_index == ~0)
    {
      clib_rwlock_writer_lock (&cm->app_segs_lock);

      ct_seg =
	ct_alloc_segment (cm, server_wrk, table_handle, sm, ct->client_wrk);
      if (!ct_seg)
	{
	  clib_rwlock_writer_unlock (&cm->app_segs_lock);
	  return -1;
	}

      ct->seg_ctx_index = ct_seg->seg_ctx_index;
      ct->ct_seg_index = ct_seg->ct_seg_index;
      ct_seg->server_n_sessions += 1;
      ct_seg->client_n_sessions += 1;
      fs_index = ct_seg->segment_index;

      clib_rwlock_writer_unlock (&cm->app_segs_lock);
    }

  /*
   * Allocate and initialize the fifos
   */
  fs = segment_manager_get_segment_w_lock (sm, fs_index);
  rv = segment_manager_try_alloc_fifos (
    fs, ls->thread_index, props->rx_fifo_size, props->tx_fifo_size,
    &ls->rx_fifo, &ls->tx_fifo);
  if (rv)
    {
      segment_manager_segment_reader_unlock (sm);

      clib_rwlock_reader_lock (&cm->app_segs_lock);

      seg_ctx = pool_elt_at_index (cm->app_seg_ctxs, ct->seg_ctx_index);
      ct_seg = pool_elt_at_index (seg_ctx->segments, ct->ct_seg_index);
      __atomic_sub_fetch (&ct_seg->server_n_sessions, 1, __ATOMIC_RELAXED);
      __atomic_sub_fetch (&ct_seg->client_n_sessions, 1, __ATOMIC_RELAXED);

      clib_rwlock_reader_unlock (&cm->app_segs_lock);

      return rv;
    }

  ls->rx_fifo->shr->master_session_index = ls->session_index;
  ls->tx_fifo->shr->master_session_index = ls->session_index;
  ls->rx_fifo->master_thread_index = ls->thread_index;
  ls->tx_fifo->master_thread_index = ls->thread_index;

  seg_handle = segment_manager_segment_handle (sm, fs);
  segment_manager_segment_reader_unlock (sm);

  ct->segment_handle = seg_handle;

  return 0;
}

static void
ct_accept_one (u32 thread_index, u32 ho_index)
{
  ct_connection_t *sct, *cct, *ho;
  transport_connection_t *ll_ct;
  app_worker_t *server_wrk;
  u32 cct_index, ll_index;
  session_t *ss, *ll;

  /*
   * Alloc client ct and initialize from ho
   */
  cct = ct_connection_alloc (thread_index);
  cct_index = cct->c_c_index;

  ho = ct_half_open_get (ho_index);

  /* Unlikely but half-open session and transport could have been freed */
  if (PREDICT_FALSE (!ho))
    {
      ct_connection_free (cct);
      return;
    }

  clib_memcpy (cct, ho, sizeof (*ho));
  cct->c_c_index = cct_index;
  cct->c_thread_index = thread_index;
  cct->flags |= CT_CONN_F_HALF_OPEN;

  /* Notify session layer that half-open is on a different thread
   * and mark ho connection index reusable. Avoids another rpc
   */
  session_half_open_migrate_notify (&cct->connection);
  session_half_open_migrated_notify (&cct->connection);
  ct_half_open_add_reusable (ho_index);

  /*
   * Alloc and init server transport
   */

  ll_index = cct->peer_index;
  ll = listen_session_get (ll_index);
  sct = ct_connection_alloc (thread_index);
  /* Transport not necessarily ct but it might, so grab after sct alloc */
  ll_ct = listen_session_get_transport (ll);

  /* Make sure cct is valid after sct alloc */
  cct = ct_connection_get (cct_index, thread_index);

  sct->c_rmt_port = 0;
  sct->c_lcl_port = ll_ct->lcl_port;
  sct->c_is_ip4 = cct->c_is_ip4;
  clib_memcpy (&sct->c_lcl_ip, &cct->c_rmt_ip, sizeof (cct->c_rmt_ip));
  sct->client_wrk = cct->client_wrk;
  sct->c_proto = TRANSPORT_PROTO_NONE;
  sct->client_opaque = cct->client_opaque;
  sct->actual_tp = cct->actual_tp;

  sct->peer_index = cct->c_c_index;
  cct->peer_index = sct->c_c_index;

  /*
   * Accept server session. Client session is created only after
   * server confirms accept.
   */
  ss = session_alloc (thread_index);
  ll = listen_session_get (ll_index);
  ss->session_type = session_type_from_proto_and_ip (TRANSPORT_PROTO_NONE,
						     sct->c_is_ip4);
  ss->connection_index = sct->c_c_index;
  ss->listener_handle = listen_session_get_handle (ll);
  session_set_state (ss, SESSION_STATE_CREATED);

  server_wrk = application_listener_select_worker (ll);
  ss->app_wrk_index = server_wrk->wrk_index;

  sct->c_s_index = ss->session_index;
  sct->server_wrk = ss->app_wrk_index;

  if (ct_init_accepted_session (server_wrk, sct, ss, ll))
    {
      ct_session_connect_notify (ss, SESSION_E_ALLOC);
      ct_connection_free (sct);
      session_free (ss);
      return;
    }

  cct->server_wrk = sct->server_wrk;
  cct->seg_ctx_index = sct->seg_ctx_index;
  cct->ct_seg_index = sct->ct_seg_index;
  cct->client_rx_fifo = ss->tx_fifo;
  cct->client_tx_fifo = ss->rx_fifo;
  cct->client_rx_fifo->refcnt++;
  cct->client_tx_fifo->refcnt++;
  cct->segment_handle =
    ct_client_seg_handle (sct->segment_handle, cct->client_wrk);

  session_set_state (ss, SESSION_STATE_ACCEPTING);
  if (app_worker_accept_notify (server_wrk, ss))
    {
      ct_session_connect_notify (ss, SESSION_E_REFUSED);
      ct_session_dealloc_fifos (sct, ss->rx_fifo, ss->tx_fifo);
      ct_connection_free (sct);
      session_free (ss);
    }
}

static void
ct_accept_rpc_wrk_handler (void *rpc_args)
{
  u32 thread_index, n_connects, i, n_pending;
  const u32 max_connects = 32;
  ct_worker_t *wrk;
  u8 need_rpc = 0;

  thread_index = pointer_to_uword (rpc_args);
  wrk = ct_worker_get (thread_index);

  /* Connects could be handled without worker barrier so grab lock */
  clib_spinlock_lock (&wrk->pending_connects_lock);

  n_pending = clib_fifo_elts (wrk->pending_connects);
  n_connects = clib_min (n_pending, max_connects);
  vec_validate (wrk->new_connects, n_connects);

  for (i = 0; i < n_connects; i++)
    clib_fifo_sub1 (wrk->pending_connects, wrk->new_connects[i]);

  if (n_pending == n_connects)
    wrk->have_connects = 0;
  else
    need_rpc = 1;

  clib_spinlock_unlock (&wrk->pending_connects_lock);

  for (i = 0; i < n_connects; i++)
    ct_accept_one (thread_index, wrk->new_connects[i]);

  if (need_rpc)
    session_send_rpc_evt_to_thread_force (
      thread_index, ct_accept_rpc_wrk_handler,
      uword_to_pointer (thread_index, void *));
}

static void
ct_fwrk_flush_connects (void *rpc_args)
{
  u32 thread_index, fwrk_index, n_workers;
  ct_main_t *cm = &ct_main;
  ct_worker_t *wrk;
  u8 need_rpc;

  fwrk_index = cm->fwrk_thread;
  n_workers = vec_len (cm->fwrk_pending_connects);

  for (thread_index = fwrk_index; thread_index < n_workers; thread_index++)
    {
      if (!vec_len (cm->fwrk_pending_connects[thread_index]))
	continue;

      wrk = ct_worker_get (thread_index);

      /* Connects can be done without worker barrier, grab dst worker lock */
      if (thread_index != fwrk_index)
	clib_spinlock_lock (&wrk->pending_connects_lock);

      clib_fifo_add (wrk->pending_connects,
		     cm->fwrk_pending_connects[thread_index],
		     vec_len (cm->fwrk_pending_connects[thread_index]));
      if (!wrk->have_connects)
	{
	  wrk->have_connects = 1;
	  need_rpc = 1;
	}

      if (thread_index != fwrk_index)
	clib_spinlock_unlock (&wrk->pending_connects_lock);

      vec_reset_length (cm->fwrk_pending_connects[thread_index]);

      if (need_rpc)
	session_send_rpc_evt_to_thread_force (
	  thread_index, ct_accept_rpc_wrk_handler,
	  uword_to_pointer (thread_index, void *));
    }

  cm->fwrk_have_flush = 0;
}

static void
ct_program_connect_to_wrk (u32 ho_index)
{
  ct_main_t *cm = &ct_main;
  u32 thread_index;

  /* Simple round-robin policy for spreading sessions over workers. We skip
   * thread index 0, i.e., offset the index by 1, when we have workers as it
   * is the one dedicated to main thread. Note that n_workers does not include
   * main thread */
  cm->n_sessions += 1;
  thread_index = cm->n_workers ? (cm->n_sessions % cm->n_workers) + 1 : 0;

  /* Pospone flushing of connect request to dst worker until after session
   * layer fully initializes the half-open session. */
  vec_add1 (cm->fwrk_pending_connects[thread_index], ho_index);
  if (!cm->fwrk_have_flush)
    {
      session_send_rpc_evt_to_thread_force (
	cm->fwrk_thread, ct_fwrk_flush_connects,
	uword_to_pointer (thread_index, void *));
      cm->fwrk_have_flush = 1;
    }
}

static int
ct_connect (app_worker_t *client_wrk, session_t *ll,
	    session_endpoint_cfg_t *sep)
{
  ct_connection_t *ho;
  u32 ho_index;

  /*
   * Alloc and init client half-open transport
   */

  ho = ct_half_open_alloc ();
  ho_index = ho->c_c_index;
  ho->c_rmt_port = sep->port;
  ho->c_lcl_port = 0;
  ho->c_is_ip4 = sep->is_ip4;
  ho->client_opaque = sep->opaque;
  ho->client_wrk = client_wrk->wrk_index;
  ho->peer_index = ll->session_index;
  ho->c_proto = TRANSPORT_PROTO_NONE;
  ho->c_flags |= TRANSPORT_CONNECTION_F_NO_LOOKUP;
  clib_memcpy (&ho->c_rmt_ip, &sep->ip, sizeof (sep->ip));
  ho->flags |= CT_CONN_F_CLIENT;
  ho->c_s_index = ~0;
  ho->actual_tp = sep->original_tp;

  /*
   * Program connect on a worker, connected reply comes
   * after server accepts the connection.
   */
  ct_program_connect_to_wrk (ho_index);

  return ho_index;
}

static u32
ct_start_listen (u32 app_listener_index, transport_endpoint_cfg_t *tep)
{
  session_endpoint_cfg_t *sep;
  ct_connection_t *ct;

  sep = (session_endpoint_cfg_t *) tep;
  ct = ct_connection_alloc (0);
  ct->server_wrk = sep->app_wrk_index;
  ct->c_is_ip4 = sep->is_ip4;
  clib_memcpy (&ct->c_lcl_ip, &sep->ip, sizeof (sep->ip));
  ct->c_lcl_port = sep->port;
  ct->c_s_index = app_listener_index;
  ct->actual_tp = sep->transport_proto;
  return ct->c_c_index;
}

static u32
ct_stop_listen (u32 ct_index)
{
  ct_connection_t *ct;
  ct = ct_connection_get (ct_index, 0);
  ct_connection_free (ct);
  return 0;
}

static transport_connection_t *
ct_listener_get (u32 ct_index)
{
  return (transport_connection_t *) ct_connection_get (ct_index, 0);
}

static transport_connection_t *
ct_session_half_open_get (u32 ct_index)
{
  return (transport_connection_t *) ct_half_open_get (ct_index);
}

static void
ct_session_cleanup (u32 conn_index, u32 thread_index)
{
  ct_connection_t *ct, *peer_ct;

  ct = ct_connection_get (conn_index, thread_index);
  if (!ct)
    return;

  peer_ct = ct_connection_get (ct->peer_index, thread_index);
  if (peer_ct)
    peer_ct->peer_index = ~0;

  ct_connection_free (ct);
}

static void
ct_cleanup_ho (u32 ho_index)
{
  ct_connection_t *ho;

  ho = ct_half_open_get (ho_index);
  ct_connection_free (ho);
}

static int
ct_session_connect (transport_endpoint_cfg_t * tep)
{
  session_endpoint_cfg_t *sep_ext;
  session_endpoint_t _sep, *sep = &_sep;
  app_worker_t *app_wrk;
  session_handle_t lh;
  application_t *app;
  app_listener_t *al;
  u32 table_index;
  session_t *ll;
  u8 fib_proto;

  sep_ext = (session_endpoint_cfg_t *) tep;
  _sep = *(session_endpoint_t *) tep;
  app_wrk = app_worker_get (sep_ext->app_wrk_index);
  app = application_get (app_wrk->app_index);

  sep->transport_proto = sep_ext->original_tp;
  table_index = application_local_session_table (app);
  lh = session_lookup_local_endpoint (table_index, sep);
  if (lh == SESSION_DROP_HANDLE)
    return SESSION_E_FILTERED;

  if (lh == SESSION_INVALID_HANDLE)
    goto global_scope;

  ll = listen_session_get_from_handle (lh);
  al = app_listener_get (ll->al_index);

  /*
   * Break loop if rule in local table points to connecting app. This
   * can happen if client is a generic proxy. Route connect through
   * global table instead.
   */
  if (al->app_index == app->app_index)
    goto global_scope;

  return ct_connect (app_wrk, ll, sep_ext);

  /*
   * If nothing found, check the global scope for locally attached
   * destinations. Make sure first that we're allowed to.
   */

global_scope:
  if (session_endpoint_is_local (sep))
    return SESSION_E_NOROUTE;

  if (!application_has_global_scope (app))
    return SESSION_E_SCOPE;

  fib_proto = session_endpoint_fib_proto (sep);
  table_index = session_lookup_get_index_for_fib (fib_proto, sep->fib_index);
  ll = session_lookup_listener_wildcard (table_index, sep);

  /* Avoid connecting app to own listener */
  if (ll)
    {
      al = app_listener_get (ll->al_index);
      if (al->app_index != app->app_index)
	return ct_connect (app_wrk, ll, sep_ext);
    }

  /* Failed to connect but no error */
  return SESSION_E_LOCAL_CONNECT;
}

static inline int
ct_close_is_reset (ct_connection_t *ct, session_t *s)
{
  if (ct->flags & CT_CONN_F_RESET)
    return 1;
  if (ct->flags & CT_CONN_F_CLIENT)
    return (svm_fifo_max_dequeue (ct->client_rx_fifo) > 0);
  else
    return (svm_fifo_max_dequeue (s->rx_fifo) > 0);
}

static void
ct_session_cleanup_server_session (session_t *s)
{
  ct_connection_t *ct;

  ct = (ct_connection_t *) session_get_transport (s);
  ct_session_dealloc_fifos (ct, s->rx_fifo, s->tx_fifo);
  session_free (s);
  ct_connection_free (ct);
}

static void
ct_session_postponed_cleanup (ct_connection_t *ct)
{
  ct_connection_t *peer_ct;
  app_worker_t *app_wrk;
  session_t *s;

  s = session_get (ct->c_s_index, ct->c_thread_index);
  app_wrk = app_worker_get_if_valid (s->app_wrk_index);

  peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index);
  if (peer_ct)
    {
      if (ct_close_is_reset (ct, s))
	session_transport_reset_notify (&peer_ct->connection);
      else
	session_transport_closing_notify (&peer_ct->connection);
    }
  session_transport_closed_notify (&ct->connection);

  /* It would be cleaner to call session_transport_delete_notify
   * but then we can't control session cleanup lower */
  session_set_state (s, SESSION_STATE_TRANSPORT_DELETED);
  if (app_wrk)
    app_worker_cleanup_notify (app_wrk, s, SESSION_CLEANUP_TRANSPORT);

  if (ct->flags & CT_CONN_F_CLIENT)
    {
      /* Normal free for client session as the fifos are allocated through
       * the connects segment manager in a segment that's not shared with
       * the server */
      ct_session_dealloc_fifos (ct, ct->client_rx_fifo, ct->client_tx_fifo);
      session_program_cleanup (s);
      ct_connection_free (ct);
    }
  else
    {
      /* Manual session and fifo segment cleanup to avoid implicit
       * segment manager cleanups and notifications */
      if (app_wrk)
	{
	  /* Remove custom cleanup notify infra when/if switching to normal
	   * session cleanup. Note that ct is freed in the cb function */
	  app_worker_cleanup_notify_custom (app_wrk, s,
					    SESSION_CLEANUP_SESSION,
					    ct_session_cleanup_server_session);
	}
      else
	{
	  ct_connection_free (ct);
	}
    }
}

static void
ct_handle_cleanups (void *args)
{
  uword thread_index = pointer_to_uword (args);
  const u32 max_cleanups = 100;
  ct_cleanup_req_t *req;
  ct_connection_t *ct;
  u32 n_to_handle = 0;
  ct_worker_t *wrk;
  session_t *s;

  wrk = ct_worker_get (thread_index);
  wrk->have_cleanups = 0;
  n_to_handle = clib_fifo_elts (wrk->pending_cleanups);
  n_to_handle = clib_min (n_to_handle, max_cleanups);

  while (n_to_handle)
    {
      clib_fifo_sub2 (wrk->pending_cleanups, req);
      ct = ct_connection_get (req->ct_index, thread_index);
      s = session_get (ct->c_s_index, ct->c_thread_index);
      if (svm_fifo_has_event (s->tx_fifo) || (s->flags & SESSION_F_RX_EVT))
	clib_fifo_add1 (wrk->pending_cleanups, *req);
      else
	ct_session_postponed_cleanup (ct);
      n_to_handle -= 1;
    }

  if (clib_fifo_elts (wrk->pending_cleanups))
    {
      wrk->have_cleanups = 1;
      session_send_rpc_evt_to_thread_force (
	thread_index, ct_handle_cleanups,
	uword_to_pointer (thread_index, void *));
    }
}

static void
ct_program_cleanup (ct_connection_t *ct)
{
  ct_cleanup_req_t *req;
  uword thread_index;
  ct_worker_t *wrk;

  thread_index = ct->c_thread_index;
  wrk = ct_worker_get (ct->c_thread_index);

  clib_fifo_add2 (wrk->pending_cleanups, req);
  req->ct_index = ct->c_c_index;

  if (wrk->have_cleanups)
    return;

  wrk->have_cleanups = 1;
  session_send_rpc_evt_to_thread_force (
    thread_index, ct_handle_cleanups, uword_to_pointer (thread_index, void *));
}

static void
ct_session_close (u32 ct_index, u32 thread_index)
{
  ct_connection_t *ct, *peer_ct;
  session_t *s;

  ct = ct_connection_get (ct_index, thread_index);
  s = session_get (ct->c_s_index, ct->c_thread_index);
  peer_ct = ct_connection_get (ct->peer_index, thread_index);
  if (peer_ct)
    {
      peer_ct->peer_index = ~0;
      /* Make sure session was allocated */
      if (peer_ct->flags & CT_CONN_F_HALF_OPEN)
	{
	  ct_session_connect_notify (s, SESSION_E_REFUSED);
	  ct->peer_index = ~0;
	}
      else if (peer_ct->c_s_index == ~0)
	{
	  /* should not happen */
	  clib_warning ("ct peer without session");
	  ct_connection_free (peer_ct);
	}
    }

  /* Do not send closed notify to make sure pending tx events are
   * still delivered and program cleanup */
  ct_program_cleanup (ct);
}

static void
ct_session_reset (u32 ct_index, u32 thread_index)
{
  ct_connection_t *ct;
  ct = ct_connection_get (ct_index, thread_index);
  ct->flags |= CT_CONN_F_RESET;
  ct_session_close (ct_index, thread_index);
}

static transport_connection_t *
ct_session_get (u32 ct_index, u32 thread_index)
{
  return (transport_connection_t *) ct_connection_get (ct_index,
						       thread_index);
}

static u8 *
format_ct_connection_id (u8 * s, va_list * args)
{
  ct_connection_t *ct = va_arg (*args, ct_connection_t *);
  if (!ct)
    return s;
  if (ct->c_is_ip4)
    {
      s = format (s, "[%d:%d][CT:%U] %U:%d->%U:%d", ct->c_thread_index,
		  ct->c_s_index, format_transport_proto_short, ct->actual_tp,
		  format_ip4_address, &ct->c_lcl_ip4,
		  clib_net_to_host_u16 (ct->c_lcl_port), format_ip4_address,
		  &ct->c_rmt_ip4, clib_net_to_host_u16 (ct->c_rmt_port));
    }
  else
    {
      s = format (s, "[%d:%d][CT:%U] %U:%d->%U:%d", ct->c_thread_index,
		  ct->c_s_index, format_transport_proto_short, ct->actual_tp,
		  format_ip6_address, &ct->c_lcl_ip6,
		  clib_net_to_host_u16 (ct->c_lcl_port), format_ip6_address,
		  &ct->c_rmt_ip6, clib_net_to_host_u16 (ct->c_rmt_port));
    }

  return s;
}

static int
ct_custom_tx (void *session, transport_send_params_t * sp)
{
  session_t *s = (session_t *) session;
  if (session_has_transport (s))
    return 0;
  /* If event enqueued towards peer, remove from scheduler and remove
   * session tx flag, i.e., accept new tx events. Unset fifo flag now to
   * avoid missing events if peer did not clear fifo flag yet, which is
   * interpreted as successful notification and session is descheduled. */
  svm_fifo_unset_event (s->tx_fifo);
  if (!ct_session_tx (s))
    sp->flags = TRANSPORT_SND_F_DESCHED;

  /* The scheduler uses packet count as a means of upper bounding the amount
   * of work done per dispatch. So make it look like we have sent something */
  return 1;
}

static int
ct_app_rx_evt (transport_connection_t * tc)
{
  ct_connection_t *ct = (ct_connection_t *) tc, *peer_ct;
  session_t *ps, *s;

  s = session_get (ct->c_s_index, ct->c_thread_index);
  if (session_has_transport (s) || s->session_state < SESSION_STATE_READY)
    return -1;
  peer_ct = ct_connection_get (ct->peer_index, tc->thread_index);
  if (!peer_ct || (peer_ct->flags & CT_CONN_F_HALF_OPEN))
    return -1;
  ps = session_get (peer_ct->c_s_index, peer_ct->c_thread_index);
  if (ps->session_state >= SESSION_STATE_TRANSPORT_CLOSING)
    return -1;
  return session_dequeue_notify (ps);
}

static u8 *
format_ct_listener (u8 * s, va_list * args)
{
  u32 tc_index = va_arg (*args, u32);
  u32 __clib_unused thread_index = va_arg (*args, u32);
  u32 __clib_unused verbose = va_arg (*args, u32);
  ct_connection_t *ct = ct_connection_get (tc_index, 0);
  s = format (s, "%-" SESSION_CLI_ID_LEN "U", format_ct_connection_id, ct);
  if (verbose)
    s = format (s, "%-" SESSION_CLI_STATE_LEN "s", "LISTEN");
  return s;
}

static u8 *
format_ct_half_open (u8 *s, va_list *args)
{
  u32 ho_index = va_arg (*args, u32);
  u32 verbose = va_arg (*args, u32);
  ct_connection_t *ct = ct_half_open_get (ho_index);
  s = format (s, "%-" SESSION_CLI_ID_LEN "U", format_ct_connection_id, ct);
  if (verbose)
    s = format (s, "%-" SESSION_CLI_STATE_LEN "s", "HALF-OPEN");
  return s;
}

static u8 *
format_ct_connection (u8 * s, va_list * args)
{
  ct_connection_t *ct = va_arg (*args, ct_connection_t *);
  u32 verbose = va_arg (*args, u32);

  if (!ct)
    return s;
  s = format (s, "%-" SESSION_CLI_ID_LEN "U", format_ct_connection_id, ct);
  if (verbose)
    {
      s = format (s, "%-" SESSION_CLI_STATE_LEN "s", "ESTABLISHED");
      if (verbose > 1)
	{
	  s = format (s, "\n");
	}
    }
  return s;
}

static u8 *
format_ct_session (u8 * s, va_list * args)
{
  u32 ct_index = va_arg (*args, u32);
  u32 thread_index = va_arg (*args, u32);
  u32 verbose = va_arg (*args, u32);
  ct_connection_t *ct;

  ct = ct_connection_get (ct_index, thread_index);
  if (!ct)
    {
      s = format (s, "empty\n");
      return s;
    }

  s = format (s, "%U", format_ct_connection, ct, verbose);
  return s;
}

clib_error_t *
ct_enable_disable (vlib_main_t * vm, u8 is_en)
{
  vlib_thread_main_t *vtm = &vlib_thread_main;
  ct_main_t *cm = &ct_main;
  ct_worker_t *wrk;

  if (is_en == 0)
    return 0;

  if (cm->is_init)
    return 0;

  cm->n_workers = vlib_num_workers ();
  cm->fwrk_thread = transport_cl_thread ();
  vec_validate (cm->wrk, vtm->n_vlib_mains);
  vec_validate (cm->fwrk_pending_connects, cm->n_workers);
  vec_foreach (wrk, cm->wrk)
    clib_spinlock_init (&wrk->pending_connects_lock);
  clib_spinlock_init (&cm->ho_reuseable_lock);
  clib_rwlock_init (&cm->app_segs_lock);
  cm->is_init = 1;

  return 0;
}

static const transport_proto_vft_t cut_thru_proto = {
  .enable = ct_enable_disable,
  .start_listen = ct_start_listen,
  .stop_listen = ct_stop_listen,
  .get_connection = ct_session_get,
  .get_listener = ct_listener_get,
  .get_half_open = ct_session_half_open_get,
  .cleanup = ct_session_cleanup,
  .cleanup_ho = ct_cleanup_ho,
  .connect = ct_session_connect,
  .close = ct_session_close,
  .reset = ct_session_reset,
  .custom_tx = ct_custom_tx,
  .app_rx_evt = ct_app_rx_evt,
  .format_listener = format_ct_listener,
  .format_half_open = format_ct_half_open,
  .format_connection = format_ct_session,
  .transport_options = {
    .name = "ct",
    .short_name = "C",
    .tx_type = TRANSPORT_TX_INTERNAL,
    .service_type = TRANSPORT_SERVICE_VC,
  },
};

static inline int
ct_session_can_tx (session_t *s)
{
  return (s->session_state == SESSION_STATE_READY ||
	  s->session_state == SESSION_STATE_CLOSING ||
	  s->session_state == SESSION_STATE_APP_CLOSED);
}

int
ct_session_tx (session_t * s)
{
  ct_connection_t *ct, *peer_ct;
  session_t *peer_s;

  if (!ct_session_can_tx (s))
    return 0;
  ct = (ct_connection_t *) session_get_transport (s);
  peer_ct = ct_connection_get (ct->peer_index, ct->c_thread_index);
  if (!peer_ct)
    return 0;
  peer_s = session_get (peer_ct->c_s_index, peer_ct->c_thread_index);
  if (peer_s->session_state >= SESSION_STATE_TRANSPORT_CLOSING)
    return 0;
  peer_s->flags |= SESSION_F_RX_EVT;
  return session_enqueue_notify (peer_s);
}

static clib_error_t *
ct_transport_init (vlib_main_t * vm)
{
  transport_register_protocol (TRANSPORT_PROTO_NONE, &cut_thru_proto,
			       FIB_PROTOCOL_IP4, ~0);
  transport_register_protocol (TRANSPORT_PROTO_NONE, &cut_thru_proto,
			       FIB_PROTOCOL_IP6, ~0);
  return 0;
}

VLIB_INIT_FUNCTION (ct_transport_init);

/*
 * fd.io coding-style-patch-verification: ON
 *
 * Local Variables:
 * eval: (c-set-style "gnu")
 * End:
 */