session/vcl: improve fifo tx notifications

Change-Id: Ied476c417c8ba5b4d7c615125c36eecd9ca64fbc
Signed-off-by: Florin Coras <fcoras@cisco.com>

5 files changed, 144 insertions, 83 deletions

diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index 9bba85fddf4..613e0cad5eb 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -41,6 +41,13 @@ format_function_t format_ooo_list;
 #define SVM_FIFO_INVALID_SESSION_INDEX 	((u32)~0)
 #define SVM_FIFO_INVALID_INDEX		((u32)~0)
 
+enum
+{
+  SVM_FIFO_NO_TX_NOTIF = 0,
+  SVM_FIFO_WANT_TX_NOTIF = 1,
+  SVM_FIFO_WANT_TX_NOTIF_IF_FULL = 2,
+};
+
 typedef struct
 {
   u32 offset;
@@ -66,7 +73,8 @@ typedef struct _svm_fifo
   u32 ct_session_index;		/**< Local session index for vpp */
     CLIB_CACHE_LINE_ALIGN_MARK (end_shared);
   u32 head;
-  volatile u32 want_tx_evt;	/**< producer wants nudge */
+  volatile u32 want_tx_ntf;	/**< producer wants nudge */
+  volatile u32 has_tx_ntf;
     CLIB_CACHE_LINE_ALIGN_MARK (end_consumer);
 
   /* producer */
@@ -174,18 +182,6 @@ svm_fifo_unset_event (svm_fifo_t * f)
   clib_atomic_release (&f->has_event);
 }
 
-static inline void
-svm_fifo_set_want_tx_evt (svm_fifo_t * f, u8 want_evt)
-{
-  f->want_tx_evt = want_evt;
-}
-
-static inline u8
-svm_fifo_want_tx_evt (svm_fifo_t * f)
-{
-  return f->want_tx_evt;
-}
-
 svm_fifo_t *svm_fifo_create (u32 data_size_in_bytes);
 void svm_fifo_free (svm_fifo_t * f);
 
@@ -200,27 +196,10 @@ int svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes);
 void svm_fifo_dequeue_drop_all (svm_fifo_t * f);
 int svm_fifo_segments (svm_fifo_t * f, svm_fifo_segment_t * fs);
 void svm_fifo_segments_free (svm_fifo_t * f, svm_fifo_segment_t * fs);
-u32 svm_fifo_number_ooo_segments (svm_fifo_t * f);
-ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f);
 void svm_fifo_init_pointers (svm_fifo_t * f, u32 pointer);
 void svm_fifo_overwrite_head (svm_fifo_t * f, u8 * data, u32 len);
-
 format_function_t format_svm_fifo;
 
-always_inline ooo_segment_t *
-svm_fifo_newest_ooo_segment (svm_fifo_t * f)
-{
-  if (f->ooos_newest == OOO_SEGMENT_INVALID_INDEX)
-    return 0;
-  return pool_elt_at_index (f->ooo_segments, f->ooos_newest);
-}
-
-always_inline void
-svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f)
-{
-  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
-}
-
 /**
  * Max contiguous chunk of data that can be read
  */
@@ -265,6 +244,77 @@ svm_fifo_tail (svm_fifo_t * f)
 }
 
 always_inline u32
+svm_fifo_nitems (svm_fifo_t * f)
+{
+  return f->nitems;
+}
+
+static inline void
+svm_fifo_add_want_tx_ntf (svm_fifo_t * f, u8 ntf_type)
+{
+  f->want_tx_ntf |= ntf_type;
+}
+
+static inline void
+svm_fifo_del_want_tx_ntf (svm_fifo_t * f, u8 ntf_type)
+{
+  f->want_tx_ntf &= ~ntf_type;
+}
+
+static inline void
+svm_fifo_clear_tx_ntf (svm_fifo_t * f)
+{
+  /* Set the flag if want_tx_notif_if_full was the only ntf requested */
+  f->has_tx_ntf = f->want_tx_ntf == SVM_FIFO_WANT_TX_NOTIF_IF_FULL;
+  svm_fifo_del_want_tx_ntf (f, SVM_FIFO_WANT_TX_NOTIF);
+}
+
+static inline void
+svm_fifo_reset_tx_ntf (svm_fifo_t * f)
+{
+  f->has_tx_ntf = 0;
+}
+
+static inline u8
+svm_fifo_needs_tx_ntf (svm_fifo_t * f, u32 n_last_deq)
+{
+  u8 want_ntf = f->want_tx_ntf;
+
+  if (PREDICT_TRUE (want_ntf == SVM_FIFO_NO_TX_NOTIF))
+    return 0;
+  else if (want_ntf & SVM_FIFO_WANT_TX_NOTIF)
+    return 1;
+  else if (want_ntf & SVM_FIFO_WANT_TX_NOTIF_IF_FULL)
+    {
+      u32 max_deq = svm_fifo_max_dequeue (f);
+      u32 nitems = svm_fifo_nitems (f);
+      if (!f->has_tx_ntf && max_deq < nitems
+	  && max_deq + n_last_deq >= nitems)
+	return 1;
+
+      return 0;
+    }
+  return 0;
+}
+
+u32 svm_fifo_number_ooo_segments (svm_fifo_t * f);
+ooo_segment_t *svm_fifo_first_ooo_segment (svm_fifo_t * f);
+
+always_inline ooo_segment_t *
+svm_fifo_newest_ooo_segment (svm_fifo_t * f)
+{
+  if (f->ooos_newest == OOO_SEGMENT_INVALID_INDEX)
+    return 0;
+  return pool_elt_at_index (f->ooo_segments, f->ooos_newest);
+}
+
+always_inline void
+svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f)
+{
+  f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
+}
+
+always_inline u32
 ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
 {
   /* Ambiguous. Assumption is that ooo segments don't touch tail */
diff --git a/src/vcl/vppcom.c b/src/vcl/vppcom.c
index 54c5f10284d..1797d93c683 100644
--- a/src/vcl/vppcom.c
+++ b/src/vcl/vppcom.c
@@ -1744,9 +1744,9 @@ vppcom_session_read_internal (uint32_t session_handle, void *buf, int n,
   if (svm_fifo_is_empty (rx_fifo))
     svm_fifo_unset_event (rx_fifo);
 
-  if (is_ct && svm_fifo_want_tx_evt (rx_fifo))
+  if (is_ct && svm_fifo_needs_tx_ntf (rx_fifo, n_read))
     {
-      svm_fifo_set_want_tx_evt (s->rx_fifo, 0);
+      svm_fifo_clear_tx_ntf (s->rx_fifo);
       app_send_io_evt_to_vpp (s->vpp_evt_q, s->rx_fifo, SESSION_IO_EVT_CT_RX,
 			      SVM_Q_WAIT);
     }
@@ -1959,7 +1959,7 @@ vppcom_session_write_inline (uint32_t session_handle, void *buf, size_t n,
 	}
       while (svm_fifo_is_full (tx_fifo))
 	{
-	  svm_fifo_set_want_tx_evt (tx_fifo, 1);
+	  svm_fifo_add_want_tx_ntf (tx_fifo, SVM_FIFO_WANT_TX_NOTIF);
 	  svm_msg_q_lock (mq);
 	  if (svm_msg_q_is_empty (mq))
 	    svm_msg_q_wait (mq);
@@ -2351,7 +2351,7 @@ vppcom_select (int n_bits, vcl_si_set * read_map, vcl_si_set * write_map,
         bits_set++;
       }
     else
-      svm_fifo_set_want_tx_evt (session->tx_fifo, 1);
+      svm_fifo_add_want_tx_ntf (session->tx_fifo, SVM_FIFO_WANT_TX_NOTIF);
   }));
 
 check_rd:
@@ -2570,6 +2570,10 @@ vppcom_epoll_ctl (uint32_t vep_handle, int op, uint32_t session_handle,
       session->is_vep_session = 1;
       vep_session->vep.next_sh = session_handle;
 
+      if (session->tx_fifo)
+	svm_fifo_add_want_tx_ntf (session->tx_fifo,
+				  SVM_FIFO_WANT_TX_NOTIF_IF_FULL);
+
       VDBG (1, "EPOLL_CTL_ADD: vep_sh %u, sh %u, events 0x%x, data 0x%llx!",
 	    vep_handle, session_handle, event->events, event->data.u64);
       vcl_evt (VCL_EVT_EPOLL_CTLADD, session, event->events, event->data.u64);
@@ -2655,6 +2659,10 @@ vppcom_epoll_ctl (uint32_t vep_handle, int op, uint32_t session_handle,
       session->vep.prev_sh = ~0;
       session->vep.vep_sh = ~0;
       session->is_vep_session = 0;
+
+      if (session->tx_fifo)
+	svm_fifo_del_want_tx_ntf (session->tx_fifo, SVM_FIFO_NO_TX_NOTIF);
+
       VDBG (1, "EPOLL_CTL_DEL: vep_idx %u, sid %u!", vep_handle,
 	    session_handle);
       vcl_evt (VCL_EVT_EPOLL_CTLDEL, session, vep_sh);
@@ -2708,6 +2716,7 @@ vcl_epoll_wait_handle_mq_event (vcl_worker_t * wrk, session_event_t * e,
       add_event = 1;
       events[*num_ev].events |= EPOLLOUT;
       session_evt_data = session->vep.ev.data.u64;
+      svm_fifo_reset_tx_ntf (session->tx_fifo);
       break;
     case SESSION_IO_EVT_CT_TX:
       vcl_fifo_rx_evt_valid_or_break (e->fifo);
@@ -2734,6 +2743,7 @@ vcl_epoll_wait_handle_mq_event (vcl_worker_t * wrk, session_event_t * e,
       add_event = 1;
       events[*num_ev].events |= EPOLLOUT;
       session_evt_data = session->vep.ev.data.u64;
+      svm_fifo_reset_tx_ntf (session->tx_fifo);
       break;
     case SESSION_CTRL_EVT_ACCEPTED:
       session = vcl_session_accepted (wrk,
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 4cf0f9e7e49..c56712bbf87 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -557,7 +557,11 @@ session_dequeue_notify (stream_session_t * s)
   if (PREDICT_FALSE (!app))
     return -1;
 
-  return app_worker_lock_and_send_event (app, s, FIFO_EVENT_APP_TX);
+  if (app_worker_lock_and_send_event (app, s, FIFO_EVENT_APP_TX))
+    return -1;
+
+  svm_fifo_clear_tx_ntf (s->server_tx_fifo);
+  return 0;
 }
 
 /**
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index cf1b3e99f4d..2bbc380282c 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -176,20 +176,6 @@ typedef struct session_tx_context_
   session_dgram_hdr_t hdr;
 } session_tx_context_t;
 
-/* Forward definition */
-typedef struct _session_manager_main session_manager_main_t;
-
-typedef int
-  (session_fifo_rx_fn) (vlib_main_t * vm, vlib_node_runtime_t * node,
-			session_event_t * e0, stream_session_t * s0,
-			int *n_tx_pkts);
-
-extern session_fifo_rx_fn session_tx_fifo_peek_and_snd;
-extern session_fifo_rx_fn session_tx_fifo_dequeue_and_snd;
-extern session_fifo_rx_fn session_tx_fifo_dequeue_internal;
-
-u8 session_node_lookup_fifo_event (svm_fifo_t * f, session_event_t * e);
-
 typedef struct session_manager_worker_
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
@@ -237,7 +223,18 @@ typedef struct session_manager_worker_
 
 } session_manager_worker_t;
 
-struct _session_manager_main
+typedef int (session_fifo_rx_fn) (vlib_main_t * vm,
+				  vlib_node_runtime_t * node,
+				  session_manager_worker_t * wrk,
+				  session_event_t * e, int *n_tx_pkts);
+
+extern session_fifo_rx_fn session_tx_fifo_peek_and_snd;
+extern session_fifo_rx_fn session_tx_fifo_dequeue_and_snd;
+extern session_fifo_rx_fn session_tx_fifo_dequeue_internal;
+
+u8 session_node_lookup_fifo_event (svm_fifo_t * f, session_event_t * e);
+
+typedef struct session_manager_main_
 {
   /** Worker contexts */
   session_manager_worker_t *wrk;
@@ -297,7 +294,7 @@ struct _session_manager_main
   f64 *last_event_poll_by_thread;
 #endif
 
-};
+} session_manager_main_t;
 
 extern session_manager_main_t session_manager_main;
 extern vlib_node_registration_t session_queue_node;
diff --git a/src/vnet/session/session_node.c b/src/vnet/session/session_node.c
index 880f16388b8..f5e5efeaf0f 100644
--- a/src/vnet/session/session_node.c
+++ b/src/vnet/session/session_node.c
@@ -632,32 +632,30 @@ session_tx_set_dequeue_params (vlib_main_t * vm, session_tx_context_t * ctx,
 
 always_inline int
 session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
-				session_event_t * e,
-				stream_session_t * s, int *n_tx_packets,
+				session_manager_worker_t * wrk,
+				session_event_t * e, int *n_tx_packets,
 				u8 peek_data)
 {
   u32 next_index, next0, next1, *to_next, n_left_to_next;
   u32 n_trace = vlib_get_trace_count (vm, node), n_bufs_needed = 0;
-  u32 thread_index = s->thread_index, n_left, pbi;
+  u32 thread_index = vm->thread_index, n_left, pbi;
   session_manager_main_t *smm = &session_manager_main;
-  session_manager_worker_t *wrk = &smm->wrk[thread_index];
   session_tx_context_t *ctx = &wrk->ctx;
   transport_proto_t tp;
   vlib_buffer_t *pb;
   u16 n_bufs, rv;
 
-  if (PREDICT_FALSE ((rv = session_tx_not_ready (s, peek_data))))
+  if (PREDICT_FALSE ((rv = session_tx_not_ready (ctx->s, peek_data))))
     {
       if (rv < 2)
 	vec_add1 (wrk->pending_event_vector, *e);
       return SESSION_TX_NO_DATA;
     }
 
-  next_index = smm->session_type_to_next[s->session_type];
+  next_index = smm->session_type_to_next[ctx->s->session_type];
   next0 = next1 = next_index;
 
-  tp = session_get_transport_proto (s);
-  ctx->s = s;
+  tp = session_get_transport_proto (ctx->s);
   ctx->transport_vft = transport_protocol_get_vft (tp);
   ctx->tc = session_tx_get_transport (ctx, peek_data);
   ctx->snd_mss = ctx->transport_vft->send_mss (ctx->tc);
@@ -679,7 +677,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
     }
 
   /* Allow enqueuing of a new event */
-  svm_fifo_unset_event (s->server_tx_fifo);
+  svm_fifo_unset_event (ctx->s->server_tx_fifo);
 
   /* Check how much we can pull. */
   session_tx_set_dequeue_params (vm, ctx, VLIB_FRAME_SIZE - *n_tx_packets,
@@ -784,7 +782,7 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 
   if (PREDICT_FALSE (n_trace > 0))
     session_tx_trace_frame (vm, node, next_index, to_next,
-			    ctx->n_segs_per_evt, s, n_trace);
+			    ctx->n_segs_per_evt, ctx->s, n_trace);
 
   _vec_len (wrk->tx_buffers) = n_bufs;
   *n_tx_packets += ctx->n_segs_per_evt;
@@ -794,18 +792,18 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
   /* 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 (s->server_tx_fifo))
+    if (svm_fifo_set_event (ctx->s->server_tx_fifo))
       vec_add1 (wrk->pending_event_vector, *e);
 
   if (!peek_data && ctx->transport_vft->tx_type == TRANSPORT_TX_DGRAM)
     {
       /* Fix dgram pre header */
       if (ctx->max_len_to_snd < ctx->max_dequeue)
-	svm_fifo_overwrite_head (s->server_tx_fifo, (u8 *) & ctx->hdr,
+	svm_fifo_overwrite_head (ctx->s->server_tx_fifo, (u8 *) & ctx->hdr,
 				 sizeof (session_dgram_pre_hdr_t));
       /* More data needs to be read */
-      else if (svm_fifo_max_dequeue (s->server_tx_fifo) > 0)
-	if (svm_fifo_set_event (s->server_tx_fifo))
+      else if (svm_fifo_max_dequeue (ctx->s->server_tx_fifo) > 0)
+	if (svm_fifo_set_event (ctx->s->server_tx_fifo))
 	  vec_add1 (wrk->pending_event_vector, *e);
     }
   return SESSION_TX_OK;
@@ -813,27 +811,29 @@ session_tx_fifo_read_and_snd_i (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 int
 session_tx_fifo_peek_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
-			      session_event_t * e,
-			      stream_session_t * s, int *n_tx_pkts)
+			      session_manager_worker_t * wrk,
+			      session_event_t * e, int *n_tx_pkts)
 {
-  return session_tx_fifo_read_and_snd_i (vm, node, e, s, n_tx_pkts, 1);
+  return session_tx_fifo_read_and_snd_i (vm, node, wrk, e, n_tx_pkts, 1);
 }
 
 int
 session_tx_fifo_dequeue_and_snd (vlib_main_t * vm, vlib_node_runtime_t * node,
-				 session_event_t * e,
-				 stream_session_t * s, int *n_tx_pkts)
+				 session_manager_worker_t * wrk,
+				 session_event_t * e, int *n_tx_pkts)
 {
-  return session_tx_fifo_read_and_snd_i (vm, node, e, s, n_tx_pkts, 0);
+  return session_tx_fifo_read_and_snd_i (vm, node, wrk, e, n_tx_pkts, 0);
 }
 
 int
 session_tx_fifo_dequeue_internal (vlib_main_t * vm,
 				  vlib_node_runtime_t * node,
-				  session_event_t * e,
-				  stream_session_t * s, int *n_tx_pkts)
+				  session_manager_worker_t * wrk,
+				  session_event_t * e, int *n_tx_pkts)
 {
+  stream_session_t *s = wrk->ctx.s;
   application_t *app;
+
   if (PREDICT_FALSE (s->session_state == SESSION_STATE_CLOSED))
     return 0;
   app = application_get (s->t_app_index);
@@ -923,7 +923,7 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
     {
       stream_session_t *s;	/* $$$ prefetch 1 ahead maybe */
       session_event_t *e;
-      u8 want_tx_evt;
+      u8 need_tx_ntf;
 
       e = &fifo_events[i];
       switch (e->event_type)
@@ -943,19 +943,17 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      clib_warning ("session was freed!");
 	      continue;
 	    }
-
-	  want_tx_evt = svm_fifo_want_tx_evt (s->server_tx_fifo);
+	  wrk->ctx.s = s;
 	  /* Spray packets in per session type frames, since they go to
 	   * different nodes */
-	  rv = (smm->session_tx_fns[s->session_type]) (vm, node, e, s,
+	  rv = (smm->session_tx_fns[s->session_type]) (vm, node, wrk, e,
 						       &n_tx_packets);
 	  if (PREDICT_TRUE (rv == SESSION_TX_OK))
 	    {
-	      if (PREDICT_FALSE (want_tx_evt))
-		{
-		  svm_fifo_set_want_tx_evt (s->server_tx_fifo, 0);
-		  session_dequeue_notify (s);
-		}
+	      need_tx_ntf = svm_fifo_needs_tx_ntf (s->server_tx_fifo,
+						   wrk->ctx.max_len_to_snd);
+	      if (PREDICT_FALSE (need_tx_ntf))
+		session_dequeue_notify (s);
 	    }
 	  else if (PREDICT_FALSE (rv == SESSION_TX_NO_BUFFERS))
 	    {
@@ -995,8 +993,10 @@ session_queue_node_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  break;
 	case FIFO_EVENT_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 (vm, node, e, s, &n_tx_packets);
+	    session_tx_fifo_dequeue_internal (vm, node, wrk, e,
+					      &n_tx_packets);
 	  break;
 	case FIFO_EVENT_RPC:
 	  fp = e->rpc_args.fp;