rdma: more batching, compile rdma-core in release mode

rx: add batching for WC processing and release
tx: improve batching for WC submission and processing
rdma-core: compile in release mode to remove assert()

Change-Id: I5fb8736db36b50f8b758cd688100477b67e72d80
Signed-off-by: Benoît Ganne <bganne@cisco.com>

4 files changed, 300 insertions, 102 deletions

diff --git a/build/external/packages/rdma-core.mk b/build/external/packages/rdma-core.mk
index d3363e141bc..82a5f0fbdb0 100644
--- a/build/external/packages/rdma-core.mk
+++ b/build/external/packages/rdma-core.mk
@@ -30,7 +30,7 @@ define  rdma-core_config_cmds
 	cd $(rdma-core_build_dir) && \
 	  $(CMAKE) -G Ninja $(rdma-core_src_dir) \
 	    -DENABLE_STATIC=1 -DENABLE_RESOLVE_NEIGH=0 -DNO_PYVERBS=1 -DENABLE_VALGRIND=0 \
-	    -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+	    -DCMAKE_BUILD_TYPE=Release \
 	    -DCMAKE_C_FLAGS=-fPIC > $(rdma-core_config_log)
 endef
 
diff --git a/src/plugins/rdma/device.c b/src/plugins/rdma/device.c
index 4098d3c8bb2..09816644e72 100644
--- a/src/plugins/rdma/device.c
+++ b/src/plugins/rdma/device.c
@@ -268,11 +268,11 @@ rdma_rxq_init (vlib_main_t * vm, rdma_device_t * rd, u16 qid, u32 n_desc)
     return clib_error_return_unix (0, "Create CQ Failed");
 
   memset (&qpia, 0, sizeof (qpia));
-  qpia.qp_type = IBV_QPT_RAW_PACKET;
   qpia.send_cq = rxq->cq;
   qpia.recv_cq = rxq->cq;
   qpia.cap.max_recv_wr = n_desc;
   qpia.cap.max_recv_sge = 1;
+  qpia.qp_type = IBV_QPT_RAW_PACKET;
 
   if ((rxq->qp = ibv_create_qp (rd->pd, &qpia)) == 0)
     return clib_error_return_unix (0, "Queue Pair create failed");
@@ -309,11 +309,12 @@ rdma_txq_init (vlib_main_t * vm, rdma_device_t * rd, u16 qid, u32 n_desc)
     return clib_error_return_unix (0, "Create CQ Failed");
 
   memset (&qpia, 0, sizeof (qpia));
-  qpia.qp_type = IBV_QPT_RAW_PACKET;
   qpia.send_cq = txq->cq;
   qpia.recv_cq = txq->cq;
   qpia.cap.max_send_wr = n_desc;
   qpia.cap.max_send_sge = 1;
+  qpia.qp_type = IBV_QPT_RAW_PACKET;
+  qpia.sq_sig_all = 1;
 
   if ((txq->qp = ibv_create_qp (rd->pd, &qpia)) == 0)
     return clib_error_return_unix (0, "Queue Pair create failed");
@@ -430,7 +431,7 @@ rdma_create_if (vlib_main_t * vm, rdma_create_if_args_t * args)
 
   pool_get_zero (rm->devices, rd);
   rd->dev_instance = rd - rm->devices;
-  rd->per_interface_next_index = ~0;
+  rd->per_interface_next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
   rd->name = vec_dup (args->name);
 
   /* check if device exist and if it is bound to mlx5_core */
diff --git a/src/plugins/rdma/input.c b/src/plugins/rdma/input.c
index 001d1c5d493..2aa98f07e8c 100644
--- a/src/plugins/rdma/input.c
+++ b/src/plugins/rdma/input.c
@@ -45,108 +45,283 @@ rdma_device_input_refill (vlib_main_t * vm, rdma_device_t * rd,
 			  rdma_rxq_t * rxq)
 {
   u32 n_alloc, n;
-  struct ibv_sge sg_entry;
-  struct ibv_recv_wr wr, *bad_wr;
-  u32 buffers[VLIB_FRAME_SIZE];
+  u32 buffers[VLIB_FRAME_SIZE], *bi = buffers;
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b = bufs;
+  struct ibv_recv_wr wr[VLIB_FRAME_SIZE], *w = wr;
+  struct ibv_sge sge[VLIB_FRAME_SIZE], *s = sge;
 
-  if (rxq->n_enq >= rxq->size)
+  if (PREDICT_FALSE (rxq->n_enq >= rxq->size))
     return;
 
   n_alloc = clib_min (VLIB_FRAME_SIZE, rxq->size - rxq->n_enq);
-  n_alloc = vlib_buffer_alloc (vm, buffers, n_alloc);
-
-  sg_entry.length = vlib_buffer_get_default_data_size (vm);
-  sg_entry.lkey = rd->mr->lkey;
-  wr.num_sge = 1;
-  wr.sg_list = &sg_entry;
-  wr.next = NULL;
-  for (n = 0; n < n_alloc; n++)
+  n_alloc = n = vlib_buffer_alloc (vm, buffers, n_alloc);
+  vlib_get_buffers (vm, buffers, bufs, n_alloc);
+
+  while (n >= 4)
     {
-      vlib_buffer_t *b = vlib_get_buffer (vm, buffers[n]);
-      sg_entry.addr = vlib_buffer_get_va (b);
-      wr.wr_id = buffers[n];
-      if (ibv_post_recv (rxq->qp, &wr, &bad_wr) != 0)
-	vlib_buffer_free (vm, buffers + n, 1);
-      else
-	rxq->n_enq++;
+      if (PREDICT_TRUE (n >= 8))
+	{
+	  CLIB_PREFETCH (&s[4 + 0], 4 * sizeof (s[0]), STORE);
+	  CLIB_PREFETCH (&w[4 + 0], 4 * sizeof (w[0]), STORE);
+	}
+
+      s[0].addr = vlib_buffer_get_va (b[0]);
+      s[0].length = vlib_buffer_get_default_data_size (vm);
+      s[0].lkey = rd->mr->lkey;
+
+      s[1].addr = vlib_buffer_get_va (b[1]);
+      s[1].length = vlib_buffer_get_default_data_size (vm);
+      s[1].lkey = rd->mr->lkey;
+
+      s[2].addr = vlib_buffer_get_va (b[2]);
+      s[2].length = vlib_buffer_get_default_data_size (vm);
+      s[2].lkey = rd->mr->lkey;
+
+      s[3].addr = vlib_buffer_get_va (b[3]);
+      s[3].length = vlib_buffer_get_default_data_size (vm);
+      s[3].lkey = rd->mr->lkey;
+
+      w[0].wr_id = bi[0];
+      w[0].next = &w[0] + 1;
+      w[0].sg_list = &s[0];
+      w[0].num_sge = 1;
+
+      w[1].wr_id = bi[1];
+      w[1].next = &w[1] + 1;
+      w[1].sg_list = &s[1];
+      w[1].num_sge = 1;
+
+      w[2].wr_id = bi[2];
+      w[2].next = &w[2] + 1;
+      w[2].sg_list = &s[2];
+      w[2].num_sge = 1;
+
+      w[3].wr_id = bi[3];
+      w[3].next = &w[3] + 1;
+      w[3].sg_list = &s[3];
+      w[3].num_sge = 1;
+
+      s += 4;
+      bi += 4;
+      w += 4;
+      b += 4;
+      n -= 4;
     }
+
+  while (n >= 1)
+    {
+      s[0].addr = vlib_buffer_get_va (b[0]);
+      s[0].length = vlib_buffer_get_default_data_size (vm);
+      s[0].lkey = rd->mr->lkey;
+
+      w[0].wr_id = bi[0];
+      w[0].next = &w[0] + 1;
+      w[0].sg_list = &s[0];
+      w[0].num_sge = 1;
+
+      s += 1;
+      bi += 1;
+      w += 1;
+      b += 1;
+      n -= 1;
+    }
+
+  w[-1].next = 0;		/* fix next pointer in WR linked-list last item */
+
+  w = wr;
+  ibv_post_recv (rxq->qp, wr, &w);
+  n = wr == w ? n_alloc : (uintptr_t) (w - wr);
+
+  if (PREDICT_FALSE (n != n_alloc))
+    vlib_buffer_free (vm, buffers + n, n_alloc - n);
+
+  rxq->n_enq += n;
 }
 
-static_always_inline uword
-rdma_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
-			  vlib_frame_t * frame, rdma_device_t * rd, u16 qid)
+static_always_inline void
+rdma_device_input_trace (vlib_main_t * vm, vlib_node_runtime_t * node,
+			 const rdma_device_t * rd, u32 n_left, const u32 * bi)
 {
-  vnet_main_t *vnm = vnet_get_main ();
-  rdma_rxq_t *rxq = vec_elt_at_index (rd->rxqs, qid);
-  u32 n_trace;
-  struct ibv_wc wc[VLIB_FRAME_SIZE];
-  u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
-  u32 *bi, *to_next, n_left_to_next;
-  int i;
-  u32 n_rx_packets = 0, n_rx_bytes = 0;
+  u32 n_trace, i;
 
-  n_rx_packets = ibv_poll_cq (rxq->cq, VLIB_FRAME_SIZE, wc);
+  if (PREDICT_TRUE (0 == (n_trace = vlib_get_trace_count (vm, node))))
+    return;
+
+  i = 0;
+  while (n_trace && n_left)
+    {
+      vlib_buffer_t *b;
+      rdma_input_trace_t *tr;
+      b = vlib_get_buffer (vm, bi[0]);
+      vlib_trace_buffer (vm, node, rd->per_interface_next_index, b,
+			 /* follow_chain */ 0);
+      tr = vlib_add_trace (vm, node, b, sizeof (*tr));
+      tr->next_index = rd->per_interface_next_index;
+      tr->hw_if_index = rd->hw_if_index;
+
+      /* next */
+      n_trace--;
+      n_left--;
+      bi++;
+      i++;
+    }
+  vlib_set_trace_count (vm, node, n_trace);
+}
+
+static_always_inline void
+rdma_device_input_ethernet (vlib_main_t * vm, vlib_node_runtime_t * node,
+			    const rdma_device_t * rd)
+{
+  vlib_next_frame_t *nf;
+  vlib_frame_t *f;
+  ethernet_input_frame_t *ef;
 
-  if (n_rx_packets <= 0)
-    rdma_device_input_refill (vm, rd, rxq);
+  if (PREDICT_FALSE
+      (VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT != rd->per_interface_next_index))
+    return;
 
-  if (PREDICT_FALSE (rd->per_interface_next_index != ~0))
-    next_index = rd->per_interface_next_index;
+  nf =
+    vlib_node_runtime_get_next_frame (vm, node, rd->per_interface_next_index);
+  f = vlib_get_frame (vm, nf->frame_index);
+  f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX;
+  /* FIXME: f->flags |= ETH_INPUT_FRAME_F_IP4_CKSUM_OK; */
 
-  vlib_get_new_next_frame (vm, node, next_index, to_next, n_left_to_next);
+  ef = vlib_frame_scalar_args (f);
+  ef->sw_if_index = rd->sw_if_index;
+  ef->hw_if_index = rd->hw_if_index;
+}
 
-  for (i = 0; i < n_rx_packets; i++)
+static_always_inline u32
+rdma_device_input_load_wc (u32 n_left_from, struct ibv_wc * wc, u32 * to_next,
+			   u32 * bufsz)
+{
+  u32 n_rx_bytes[4] = { 0 };
+
+  while (n_left_from >= 4)
     {
-      u32 bi = wc[i].wr_id;
-      vlib_buffer_t *b = vlib_get_buffer (vm, bi);
-      b->current_length = wc[i].byte_len;
-      vnet_buffer (b)->sw_if_index[VLIB_RX] = rd->sw_if_index;
-      vnet_buffer (b)->sw_if_index[VLIB_TX] = ~0;
-      to_next[i] = bi;
-      n_rx_bytes += wc[i].byte_len;
+      if (PREDICT_TRUE (n_left_from >= 8))
+	{
+	  CLIB_PREFETCH (&wc[4 + 0], CLIB_CACHE_LINE_BYTES, LOAD);
+	  CLIB_PREFETCH (&wc[4 + 1], CLIB_CACHE_LINE_BYTES, LOAD);
+	  CLIB_PREFETCH (&wc[4 + 2], CLIB_CACHE_LINE_BYTES, LOAD);
+	  CLIB_PREFETCH (&wc[4 + 3], CLIB_CACHE_LINE_BYTES, LOAD);
+	  CLIB_PREFETCH (&bufsz[4 + 0], 4 * sizeof (bufsz[0]), STORE);
+	  CLIB_PREFETCH (&to_next[4 + 0], 4 * sizeof (to_next[0]), STORE);
+	}
+
+      to_next[0] = wc[0].wr_id;
+      to_next[1] = wc[1].wr_id;
+      to_next[2] = wc[2].wr_id;
+      to_next[3] = wc[3].wr_id;
+
+      bufsz[0] = wc[0].byte_len;
+      bufsz[1] = wc[1].byte_len;
+      bufsz[2] = wc[2].byte_len;
+      bufsz[3] = wc[3].byte_len;
+
+      n_rx_bytes[0] += wc[0].byte_len;
+      n_rx_bytes[1] += wc[1].byte_len;
+      n_rx_bytes[2] += wc[2].byte_len;
+      n_rx_bytes[3] += wc[3].byte_len;
+
+      wc += 4;
+      to_next += 4;
+      bufsz += 4;
+      n_left_from -= 4;
     }
 
-  if (PREDICT_FALSE ((n_trace = vlib_get_trace_count (vm, node))))
+  while (n_left_from >= 1)
     {
-      u32 n_left = n_rx_packets, i = 0;
-      bi = to_next;
+      to_next[0] = wc[0].wr_id;
+      bufsz[0] = wc[0].byte_len;
+      n_rx_bytes[0] += wc[0].byte_len;
 
-      while (n_trace && n_left)
+      wc += 1;
+      to_next += 1;
+      bufsz += 1;
+      n_left_from -= 1;
+    }
+
+  return n_rx_bytes[0] + n_rx_bytes[1] + n_rx_bytes[2] + n_rx_bytes[3];
+}
+
+static_always_inline void
+rdma_device_input_bufs_init (u32 n_left_from, vlib_buffer_t ** bufs,
+			     u32 * bufsz, u32 sw_if_index)
+{
+  while (n_left_from >= 4)
+    {
+      if (PREDICT_TRUE (n_left_from >= 8))
 	{
-	  vlib_buffer_t *b;
-	  rdma_input_trace_t *tr;
-	  b = vlib_get_buffer (vm, bi[0]);
-	  vlib_trace_buffer (vm, node, next_index, b, /* follow_chain */ 0);
-	  tr = vlib_add_trace (vm, node, b, sizeof (*tr));
-	  tr->next_index = next_index;
-	  tr->hw_if_index = rd->hw_if_index;
-
-	  /* next */
-	  n_trace--;
-	  n_left--;
-	  bi++;
-	  i++;
+	  vlib_prefetch_buffer_header (bufs[4 + 0], STORE);
+	  vlib_prefetch_buffer_header (bufs[4 + 1], STORE);
+	  vlib_prefetch_buffer_header (bufs[4 + 2], STORE);
+	  vlib_prefetch_buffer_header (bufs[4 + 3], STORE);
+	  CLIB_PREFETCH (&bufsz[4 + 0], 4 * sizeof (bufsz[0]), LOAD);
 	}
-      vlib_set_trace_count (vm, node, n_trace);
+
+      bufs[0]->current_length = bufsz[0];
+      bufs[1]->current_length = bufsz[1];
+      bufs[2]->current_length = bufsz[2];
+      bufs[3]->current_length = bufsz[3];
+
+      vnet_buffer (bufs[0])->sw_if_index[VLIB_RX] = sw_if_index;
+      vnet_buffer (bufs[1])->sw_if_index[VLIB_RX] = sw_if_index;
+      vnet_buffer (bufs[2])->sw_if_index[VLIB_RX] = sw_if_index;
+      vnet_buffer (bufs[3])->sw_if_index[VLIB_RX] = sw_if_index;
+
+      vnet_buffer (bufs[0])->sw_if_index[VLIB_TX] = ~0;
+      vnet_buffer (bufs[1])->sw_if_index[VLIB_TX] = ~0;
+      vnet_buffer (bufs[2])->sw_if_index[VLIB_TX] = ~0;
+      vnet_buffer (bufs[3])->sw_if_index[VLIB_TX] = ~0;
+
+      bufs += 4;
+      bufsz += 4;
+      n_left_from -= 4;
     }
 
-  if (PREDICT_TRUE (next_index == VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT))
+  while (n_left_from >= 1)
     {
-      vlib_next_frame_t *nf;
-      vlib_frame_t *f;
-      ethernet_input_frame_t *ef;
-      nf = vlib_node_runtime_get_next_frame (vm, node, next_index);
-      f = vlib_get_frame (vm, nf->frame_index);
-      f->flags = ETH_INPUT_FRAME_F_SINGLE_SW_IF_IDX;
-
-      ef = vlib_frame_scalar_args (f);
-      ef->sw_if_index = rd->sw_if_index;
-      ef->hw_if_index = rd->hw_if_index;
-      //f->flags |= ETH_INPUT_FRAME_F_IP4_CKSUM_OK;
+      bufs[0]->current_length = bufsz[0];
+      vnet_buffer (bufs[0])->sw_if_index[VLIB_RX] = sw_if_index;
+      vnet_buffer (bufs[0])->sw_if_index[VLIB_TX] = ~0;
+
+      bufs += 1;
+      bufsz += 1;
+      n_left_from -= 1;
     }
+}
 
-  n_left_to_next -= n_rx_packets;
-  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+static_always_inline uword
+rdma_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+			  vlib_frame_t * frame, rdma_device_t * rd, u16 qid)
+{
+  vnet_main_t *vnm = vnet_get_main ();
+  rdma_rxq_t *rxq = vec_elt_at_index (rd->rxqs, qid);
+  struct ibv_wc wc[VLIB_FRAME_SIZE];
+  u32 bufsz[VLIB_FRAME_SIZE];
+  vlib_buffer_t *bufs[VLIB_FRAME_SIZE];
+  u32 *to_next, n_left_to_next;
+  u32 n_rx_packets, n_rx_bytes;
+
+  n_rx_packets = ibv_poll_cq (rxq->cq, VLIB_FRAME_SIZE, wc);
+
+  if (PREDICT_FALSE (n_rx_packets <= 0))
+    {
+      rdma_device_input_refill (vm, rd, rxq);
+      return 0;
+    }
+
+  vlib_get_new_next_frame (vm, node, rd->per_interface_next_index, to_next,
+			   n_left_to_next);
+  n_rx_bytes = rdma_device_input_load_wc (n_rx_packets, wc, to_next, bufsz);
+  vlib_get_buffers (vm, to_next, bufs, n_rx_packets);
+  rdma_device_input_bufs_init (n_rx_packets, bufs, bufsz, rd->sw_if_index);
+  rdma_device_input_trace (vm, node, rd, n_rx_packets, to_next);
+  rdma_device_input_ethernet (vm, node, rd);
+
+  vlib_put_next_frame (vm, node, rd->per_interface_next_index,
+		       n_left_to_next - n_rx_packets);
 
   vlib_increment_combined_counter
     (vnm->interface_main.combined_sw_if_counters +
@@ -154,6 +329,7 @@ rdma_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
      rd->hw_if_index, n_rx_packets, n_rx_bytes);
 
   rxq->n_enq -= n_rx_packets;
+
   rdma_device_input_refill (vm, rd, rxq);
 
   return n_rx_packets;
@@ -172,9 +348,8 @@ VLIB_NODE_FN (rdma_input_node) (vlib_main_t * vm,
   {
     rdma_device_t *rd;
     rd = vec_elt_at_index (rm->devices, dq->dev_instance);
-    if ((rd->flags & RDMA_DEVICE_F_ADMIN_UP) == 0)
-      continue;
-    n_rx += rdma_device_input_inline (vm, node, frame, rd, dq->queue_id);
+    if (PREDICT_TRUE (rd->flags & RDMA_DEVICE_F_ADMIN_UP))
+      n_rx += rdma_device_input_inline (vm, node, frame, rd, dq->queue_id);
   }
   return n_rx;
 }
diff --git a/src/plugins/rdma/output.c b/src/plugins/rdma/output.c
index 41b9af1fbe5..bb25abe1a98 100644
--- a/src/plugins/rdma/output.c
+++ b/src/plugins/rdma/output.c
@@ -65,16 +65,21 @@ VNET_DEVICE_CLASS_TX_FN (rdma_device_class) (vlib_main_t * vm,
 
   memset (w, 0, n_left_from * sizeof (w[0]));
 
-  while (n_left_from >= 2)
+  while (n_left_from >= 4)
     {
-      if (PREDICT_TRUE (n_left_from >= 4))
+      if (PREDICT_TRUE (n_left_from >= 8))
 	{
-	  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);
+	  vlib_prefetch_buffer_header (b[4 + 0], LOAD);
+	  vlib_prefetch_buffer_header (b[4 + 1], LOAD);
+	  vlib_prefetch_buffer_header (b[4 + 2], LOAD);
+	  vlib_prefetch_buffer_header (b[4 + 3], LOAD);
+
+	  CLIB_PREFETCH (&s[4 + 0], 4 * sizeof (s[0]), STORE);
+
+	  CLIB_PREFETCH (&w[4 + 0], CLIB_CACHE_LINE_BYTES, STORE);
+	  CLIB_PREFETCH (&w[4 + 1], CLIB_CACHE_LINE_BYTES, STORE);
+	  CLIB_PREFETCH (&w[4 + 2], CLIB_CACHE_LINE_BYTES, STORE);
+	  CLIB_PREFETCH (&w[4 + 3], CLIB_CACHE_LINE_BYTES, STORE);
 	}
 
       s[0].addr = vlib_buffer_get_current_va (b[0]);
@@ -85,25 +90,43 @@ VNET_DEVICE_CLASS_TX_FN (rdma_device_class) (vlib_main_t * vm,
       s[1].length = b[1]->current_length;
       s[1].lkey = rd->mr->lkey;
 
+      s[2].addr = vlib_buffer_get_current_va (b[2]);
+      s[2].length = b[2]->current_length;
+      s[2].lkey = rd->mr->lkey;
+
+      s[3].addr = vlib_buffer_get_current_va (b[3]);
+      s[3].length = b[3]->current_length;
+      s[3].lkey = rd->mr->lkey;
+
       w[0].wr_id = f[0];
-      w[0].next = &w[1 + 0];
+      w[0].next = &w[0] + 1;
       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].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;
+      w[2].wr_id = f[2];
+      w[2].next = &w[2] + 1;
+      w[2].sg_list = &s[2];
+      w[2].num_sge = 1;
+      w[2].opcode = IBV_WR_SEND;
+
+      w[3].wr_id = f[3];
+      w[3].next = &w[3] + 1;
+      w[3].sg_list = &s[3];
+      w[3].num_sge = 1;
+      w[3].opcode = IBV_WR_SEND;
+
+      s += 4;
+      f += 4;
+      w += 4;
+      b += 4;
+      n_left_from -= 4;
     }
 
   while (n_left_from >= 1)
@@ -113,11 +136,10 @@ VNET_DEVICE_CLASS_TX_FN (rdma_device_class) (vlib_main_t * vm,
       s[0].lkey = rd->mr->lkey;
 
       w[0].wr_id = f[0];
-      w[0].next = &w[1 + 0];
+      w[0].next = &w[0] + 1;
       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;