dpdk: fix wrong tx ring size calculations

At many places code was using constant ring size od 4096
which was defined in macro DPDK_TX_RING_SIZE.

As we support differnet ring size and default value s now
1024, we need to remove DPDK_TX_RING_SIZE and use
value stored in device structure.

For that reason dpdk_device_t.nb_tx_desc is moved to
first cacheline.

Change-Id: I2c2ac11f0f5e8ae779d34f9a9104eaf2921ec34c
Signed-off-by: Damjan Marion <damarion@cisco.com>

4 files changed, 22 insertions, 27 deletions

diff --git a/vnet/vnet/devices/dpdk/device.c b/vnet/vnet/devices/dpdk/device.c
index 119d6039f74..3649178f388 100644
--- a/vnet/vnet/devices/dpdk/device.c
+++ b/vnet/vnet/devices/dpdk/device.c
@@ -281,7 +281,7 @@ static_always_inline
 
   n_packets = ring->tx_head - ring->tx_tail;
 
-  tx_head = ring->tx_head % DPDK_TX_RING_SIZE;
+  tx_head = ring->tx_head % xd->nb_tx_desc;
 
   /*
    * Ensure rte_eth_tx_burst is not called with 0 packets, which can lead to
@@ -296,7 +296,7 @@ static_always_inline
    * a bit because it decreases the probability of having to issue two tx_burst
    * calls due to a ring wrap.
    */
-  ASSERT (n_packets < DPDK_TX_RING_SIZE);
+  ASSERT (n_packets < xd->nb_tx_desc);
 
   /*
    * If there is no flowcontrol callback, there is only temporary buffering
@@ -317,7 +317,7 @@ static_always_inline
   do
     {
       /* start the burst at the tail */
-      tx_tail = ring->tx_tail % DPDK_TX_RING_SIZE;
+      tx_tail = ring->tx_tail % xd->nb_tx_desc;
 
       /*
        * This device only supports one TX queue,
@@ -354,15 +354,14 @@ static_always_inline
 	      rv = rte_eth_tx_burst (xd->device_index,
 				     (uint16_t) queue_id,
 				     &tx_vector[tx_tail],
-				     (uint16_t) (DPDK_TX_RING_SIZE -
-						 tx_tail));
+				     (uint16_t) (xd->nb_tx_desc - tx_tail));
 
 	      /*
 	       * If we transmitted everything we wanted, then allow 1 retry
 	       * so we can try to transmit the rest. If we didn't transmit
 	       * everything, stop now.
 	       */
-	      n_retry = (rv == DPDK_TX_RING_SIZE - tx_tail) ? 1 : 0;
+	      n_retry = (rv == xd->nb_tx_desc - tx_tail) ? 1 : 0;
 	    }
 	}
 #if DPDK_VHOST_USER
@@ -438,7 +437,7 @@ static_always_inline
 	      int i;
 	      u32 bytes = 0;
 	      struct rte_mbuf **pkts = &tx_vector[tx_tail];
-	      for (i = 0; i < (DPDK_TX_RING_SIZE - tx_tail); i++)
+	      for (i = 0; i < (xd->nb_tx_desc - tx_tail); i++)
 		{
 		  struct rte_mbuf *buff = pkts[i];
 		  bytes += rte_pktmbuf_data_len (buff);
@@ -447,7 +446,7 @@ static_always_inline
 		rte_vhost_enqueue_burst (&xd->vu_vhost_dev,
 					 offset + VIRTIO_RXQ,
 					 &tx_vector[tx_tail],
-					 (uint16_t) (DPDK_TX_RING_SIZE -
+					 (uint16_t) (xd->nb_tx_desc -
 						     tx_tail));
 
 	      if (PREDICT_TRUE (rv > 0))
@@ -476,7 +475,7 @@ static_always_inline
 		    rte_pktmbuf_free (tx_vector[tx_tail + c]);
 		}
 
-	      n_retry = (rv == DPDK_TX_RING_SIZE - tx_tail) ? 1 : 0;
+	      n_retry = (rv == xd->nb_tx_desc - tx_tail) ? 1 : 0;
 	    }
 
 	  if (xd->need_txlock)
@@ -504,15 +503,14 @@ static_always_inline
 	       */
 	      rv = rte_kni_tx_burst (xd->kni,
 				     &tx_vector[tx_tail],
-				     (uint16_t) (DPDK_TX_RING_SIZE -
-						 tx_tail));
+				     (uint16_t) (xd->nb_tx_desc - tx_tail));
 
 	      /*
 	       * If we transmitted everything we wanted, then allow 1 retry
 	       * so we can try to transmit the rest. If we didn't transmit
 	       * everything, stop now.
 	       */
-	      n_retry = (rv == DPDK_TX_RING_SIZE - tx_tail) ? 1 : 0;
+	      n_retry = (rv == xd->nb_tx_desc - tx_tail) ? 1 : 0;
 	    }
 	}
 #endif
@@ -632,7 +630,7 @@ dpdk_interface_tx (vlib_main_t * vm,
 
   ASSERT (n_packets <= VLIB_FRAME_SIZE);
 
-  if (PREDICT_FALSE (n_on_ring + n_packets > DPDK_TX_RING_SIZE))
+  if (PREDICT_FALSE (n_on_ring + n_packets > xd->nb_tx_desc))
     {
       /*
        * Overflowing the ring should never happen.
@@ -668,7 +666,7 @@ dpdk_interface_tx (vlib_main_t * vm,
 
   from = vlib_frame_vector_args (f);
   n_left = n_packets;
-  i = ring->tx_head % DPDK_TX_RING_SIZE;
+  i = ring->tx_head % xd->nb_tx_desc;
 
   while (n_left >= 4)
     {
@@ -770,9 +768,9 @@ dpdk_interface_tx (vlib_main_t * vm,
 
       if (PREDICT_TRUE (any_clone == 0))
 	{
-	  tx_vector[i % DPDK_TX_RING_SIZE] = mb0;
+	  tx_vector[i % xd->nb_tx_desc] = mb0;
 	  i++;
-	  tx_vector[i % DPDK_TX_RING_SIZE] = mb1;
+	  tx_vector[i % xd->nb_tx_desc] = mb1;
 	  i++;
 	}
       else
@@ -780,12 +778,12 @@ dpdk_interface_tx (vlib_main_t * vm,
 	  /* cloning was done, need to check for failure */
 	  if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0))
 	    {
-	      tx_vector[i % DPDK_TX_RING_SIZE] = mb0;
+	      tx_vector[i % xd->nb_tx_desc] = mb0;
 	      i++;
 	    }
 	  if (PREDICT_TRUE ((b1->flags & VLIB_BUFFER_REPL_FAIL) == 0))
 	    {
-	      tx_vector[i % DPDK_TX_RING_SIZE] = mb1;
+	      tx_vector[i % xd->nb_tx_desc] = mb1;
 	      i++;
 	    }
 	}
@@ -839,7 +837,7 @@ dpdk_interface_tx (vlib_main_t * vm,
 
       if (PREDICT_TRUE ((b0->flags & VLIB_BUFFER_REPL_FAIL) == 0))
 	{
-	  tx_vector[i % DPDK_TX_RING_SIZE] = mb0;
+	  tx_vector[i % xd->nb_tx_desc] = mb0;
 	  i++;
 	}
       n_left--;
diff --git a/vnet/vnet/devices/dpdk/dpdk.h b/vnet/vnet/devices/dpdk/dpdk.h
index 2cb301ce291..48072560eed 100644
--- a/vnet/vnet/devices/dpdk/dpdk.h
+++ b/vnet/vnet/devices/dpdk/dpdk.h
@@ -214,6 +214,7 @@ typedef struct
 #define DPDK_DEVICE_FLAG_VHOST_USER     (1 << 4)
 #define DPDK_DEVICE_FLAG_HAVE_SUBIF     (1 << 5)
 
+  u16 nb_tx_desc;
     CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
 
   u8 *interface_name_suffix;
@@ -225,7 +226,6 @@ typedef struct
   u16 tx_q_used;
   u16 rx_q_used;
   u16 nb_rx_desc;
-  u16 nb_tx_desc;
   u16 *cpu_socket_id_by_queue;
   struct rte_eth_conf port_conf;
   struct rte_eth_txconf tx_conf;
@@ -265,9 +265,6 @@ typedef struct
   u8 need_txlock;		/* Used by VNET_DPDK_DEV_VHOST_USER */
 } dpdk_device_t;
 
-
-#define DPDK_TX_RING_SIZE (4 * 1024)
-
 #define DPDK_STATS_POLL_INTERVAL      (10.0)
 #define DPDK_MIN_STATS_POLL_INTERVAL  (0.001)	/* 1msec */
 
diff --git a/vnet/vnet/devices/dpdk/init.c b/vnet/vnet/devices/dpdk/init.c
index a2cc8849f41..44b4dc31097 100644
--- a/vnet/vnet/devices/dpdk/init.c
+++ b/vnet/vnet/devices/dpdk/init.c
@@ -666,7 +666,7 @@ dpdk_lib_init (dpdk_main_t * dm)
 			    CLIB_CACHE_LINE_BYTES);
       for (j = 0; j < tm->n_vlib_mains; j++)
 	{
-	  vec_validate_ha (xd->tx_vectors[j], DPDK_TX_RING_SIZE,
+	  vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc,
 			   sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES);
 	  vec_reset_length (xd->tx_vectors[j]);
 	}
@@ -775,7 +775,7 @@ dpdk_lib_init (dpdk_main_t * dm)
 				CLIB_CACHE_LINE_BYTES);
 	  for (j = 0; j < tm->n_vlib_mains; j++)
 	    {
-	      vec_validate_ha (xd->tx_vectors[j], DPDK_TX_RING_SIZE,
+	      vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc,
 			       sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES);
 	      vec_reset_length (xd->tx_vectors[j]);
 	    }
diff --git a/vnet/vnet/devices/dpdk/vhost_user.c b/vnet/vnet/devices/dpdk/vhost_user.c
index f8910ad65d7..946c6e1f1db 100644
--- a/vnet/vnet/devices/dpdk/vhost_user.c
+++ b/vnet/vnet/devices/dpdk/vhost_user.c
@@ -266,7 +266,7 @@ dpdk_create_vhost_user_if_internal (u32 * hw_if_index, u32 if_id, u8 * hwaddr)
       // reset tx vectors
       for (j = 0; j < tm->n_vlib_mains; j++)
 	{
-	  vec_validate_ha (xd->tx_vectors[j], DPDK_TX_RING_SIZE,
+	  vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc,
 			   sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES);
 	  vec_reset_length (xd->tx_vectors[j]);
 	}
@@ -338,7 +338,7 @@ dpdk_create_vhost_user_if_internal (u32 * hw_if_index, u32 if_id, u8 * hwaddr)
 
       for (j = 0; j < tm->n_vlib_mains; j++)
 	{
-	  vec_validate_ha (xd->tx_vectors[j], DPDK_TX_RING_SIZE,
+	  vec_validate_ha (xd->tx_vectors[j], xd->nb_tx_desc,
 			   sizeof (tx_ring_hdr_t), CLIB_CACHE_LINE_BYTES);
 	  vec_reset_length (xd->tx_vectors[j]);
 	}