rdma: tx: fix stats and add batching

Tx stats are no longer counted twice.
Submit tx packets as a single batch per vector instead of per-packet

Change-Id: I26820b21f23842b3a67ace0b939095f3550d3856
Signed-off-by: Benoît Ganne <bganne@cisco.com>

1 files changed, 88 insertions, 61 deletions

diff --git a/src/plugins/rdma/output.c b/src/plugins/rdma/output.c
index 410784308f3..41b9af1fbe5 100644
--- a/src/plugins/rdma/output.c
+++ b/src/plugins/rdma/output.c
@@ -24,39 +24,6 @@
 
 #include <rdma/rdma.h>
 
-static_always_inline u16
-rdma_device_output_tx (vlib_main_t * vm, rdma_device_t * rd, rdma_txq_t * txq,
-		       u32 * buffers, u16 n_left, u32 * n_tx_packets,
-		       u32 * n_tx_bytes)
-{
-  struct ibv_sge sg_entry;
-  struct ibv_send_wr wr, *bad_wr;
-  u16 i;
-
-  for (i = 0; i < n_left; i++)
-    {
-      vlib_buffer_t *b = vlib_get_buffer (vm, buffers[i]);
-      sg_entry.addr = vlib_buffer_get_current_va (b);
-      sg_entry.length = b->current_length;
-      sg_entry.lkey = rd->mr->lkey;
-
-      memset (&wr, 0, sizeof (wr));
-      wr.num_sge = 1;
-      wr.sg_list = &sg_entry;
-      wr.opcode = IBV_WR_SEND;
-      wr.send_flags = IBV_SEND_SIGNALED;
-      wr.wr_id = buffers[i];
-
-      if (ibv_post_send (txq->qp, &wr, &bad_wr) != 0)
-	break;
-
-      *n_tx_bytes += b->current_length;
-    }
-
-  *n_tx_packets += i;
-  return i;
-}
-
 static_always_inline void
 rdma_device_output_free (vlib_main_t * vm, rdma_txq_t * txq)
 {
@@ -79,49 +46,109 @@ VNET_DEVICE_CLASS_TX_FN (rdma_device_class) (vlib_main_t * vm,
 					     vlib_node_runtime_t * node,
 					     vlib_frame_t * frame)
 {
-  vnet_main_t *vnm = vnet_get_main ();
   rdma_main_t *rm = &rdma_main;
   vnet_interface_output_runtime_t *ord = (void *) node->runtime_data;
   rdma_device_t *rd = pool_elt_at_index (rm->devices, ord->dev_instance);
   u32 thread_index = vm->thread_index;
-  u8 qid = thread_index;
-  rdma_txq_t *txq = vec_elt_at_index (rd->txqs, qid % vec_len (rd->txqs));
-  u32 *buffers = vlib_frame_vector_args (frame);
-  u16 n_left;
-  u16 n_retry = 5;
-  u32 n_tx_packets = 0, n_tx_bytes = 0;
+  rdma_txq_t *txq =
+    vec_elt_at_index (rd->txqs, thread_index % vec_len (rd->txqs));
+  u32 *from, *f, n_left_from;
+  u32 n_tx_packets, n_tx_failed;
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b = bufs;
+  struct ibv_send_wr wr[VLIB_FRAME_SIZE], *w = wr;
+  struct ibv_sge sge[VLIB_FRAME_SIZE], *s = sge;
+  int i;
 
-  clib_spinlock_lock_if_init (&txq->lock);
+  f = from = vlib_frame_vector_args (frame);
+  n_left_from = frame->n_vectors;
+  vlib_get_buffers (vm, from, bufs, n_left_from);
 
-  n_left = frame->n_vectors;
+  memset (w, 0, n_left_from * sizeof (w[0]));
 
-  while (n_left)
+  while (n_left_from >= 2)
     {
-      u16 n;
-      rdma_device_output_free (vm, txq);
-      n =
-	rdma_device_output_tx (vm, rd, txq, buffers, n_left, &n_tx_packets,
-			       &n_tx_bytes);
-      n_left -= n;
-      buffers += n;
-
-      if (n_left && n_retry--)
+      if (PREDICT_TRUE (n_left_from >= 4))
 	{
-	  vlib_buffer_free (vm, buffers, n_left);
-	  vlib_error_count (vm, node->node_index,
-			    RDMA_TX_ERROR_NO_FREE_SLOTS, n_left);
-	  break;
+	  vlib_prefetch_buffer_header (b[2 + 0], LOAD);
+	  vlib_prefetch_buffer_header (b[2 + 1], LOAD);
+	  CLIB_PREFETCH (&s[2 + 0], sizeof (s[0]), STORE);
+	  CLIB_PREFETCH (&s[2 + 1], sizeof (s[0]), STORE);
+	  CLIB_PREFETCH (&w[2 + 0], sizeof (w[0]), STORE);
+	  CLIB_PREFETCH (&w[2 + 1], sizeof (w[0]), STORE);
 	}
+
+      s[0].addr = vlib_buffer_get_current_va (b[0]);
+      s[0].length = b[0]->current_length;
+      s[0].lkey = rd->mr->lkey;
+
+      s[1].addr = vlib_buffer_get_current_va (b[1]);
+      s[1].length = b[1]->current_length;
+      s[1].lkey = rd->mr->lkey;
+
+      w[0].wr_id = f[0];
+      w[0].next = &w[1 + 0];
+      w[0].sg_list = &s[0];
+      w[0].num_sge = 1;
+      w[0].opcode = IBV_WR_SEND;
+      w[0].send_flags = IBV_SEND_SIGNALED;
+
+      w[1].wr_id = f[1];
+      w[1].next = &w[1 + 1];
+      w[1].sg_list = &s[1];
+      w[1].num_sge = 1;
+      w[1].opcode = IBV_WR_SEND;
+      w[1].send_flags = IBV_SEND_SIGNALED;
+
+      s += 2;
+      f += 2;
+      w += 2;
+      b += 2;
+      n_left_from -= 2;
+    }
+
+  while (n_left_from >= 1)
+    {
+      s[0].addr = vlib_buffer_get_current_va (b[0]);
+      s[0].length = b[0]->current_length;
+      s[0].lkey = rd->mr->lkey;
+
+      w[0].wr_id = f[0];
+      w[0].next = &w[1 + 0];
+      w[0].sg_list = &s[0];
+      w[0].num_sge = 1;
+      w[0].opcode = IBV_WR_SEND;
+      w[0].send_flags = IBV_SEND_SIGNALED;
+
+      s += 1;
+      f += 1;
+      w += 1;
+      b += 1;
+      n_left_from -= 1;
     }
 
+  w[-1].next = 0;		/* fix next pointer in WR linked-list last item */
+
+  w = wr;
+  clib_spinlock_lock_if_init (&txq->lock);
+  for (i = 0; i < 5; i++)
+    {
+      rdma_device_output_free (vm, txq);
+      if (0 == ibv_post_send (txq->qp, w, &w))
+	break;
+    }
   clib_spinlock_unlock_if_init (&txq->lock);
 
-  vlib_increment_combined_counter
-    (vnm->interface_main.combined_sw_if_counters +
-     VNET_INTERFACE_COUNTER_TX, thread_index,
-     rd->hw_if_index, n_tx_packets, n_tx_bytes);
+  n_tx_packets = w == wr ? frame->n_vectors : w - wr;
+  n_tx_failed = frame->n_vectors - n_tx_packets;
+
+  if (PREDICT_FALSE (n_tx_failed))
+    {
+      vlib_buffer_free (vm, &from[n_tx_packets], n_tx_failed);
+      vlib_error_count (vm, node->node_index,
+			RDMA_TX_ERROR_NO_FREE_SLOTS, n_tx_failed);
+    }
 
-  return frame->n_vectors - n_left;
+  return n_tx_packets;
 }
 
 /*