dpdk: add multi-txq support

Type: improvement

Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: I9f061a05d947bc2867e1b962bf0522ad344bcc1a

src/plugins/dpdk/device/common.c

@@ -58,7 +58,6 @@ dpdk_device_setup (dpdk_device_t * xd)
 {
   vlib_main_t *vm = vlib_get_main ();
   vnet_main_t *vnm = vnet_get_main ();
-  vlib_thread_main_t *tm = vlib_get_thread_main ();
   vnet_sw_interface_t *sw = vnet_get_sw_interface (vnm, xd->sw_if_index);
   vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, xd->hw_if_index);
   u16 buf_sz = vlib_buffer_get_default_data_size (vm);
@@ -222,8 +221,7 @@ retry:
       if (rv < 0)
 	dpdk_device_error (xd, "rte_eth_tx_queue_setup", rv);
 
-      if (xd->conf.n_tx_queues < tm->n_vlib_mains)
-	clib_spinlock_init (&vec_elt (xd->tx_queues, j).lock);
+      clib_spinlock_init (&vec_elt (xd->tx_queues, j).lock);
     }
 
   vec_validate_aligned (xd->rx_queues, xd->conf.n_rx_queues - 1,

src/plugins/dpdk/device/device.c

@@ -152,28 +152,28 @@ dpdk_validate_rte_mbuf (vlib_main_t * vm, vlib_buffer_t * b,
  * support multiple queues. It returns the number of packets untransmitted
  * If all packets are transmitted (the normal case), the function returns 0.
  */
-static_always_inline
-  u32 tx_burst_vector_internal (vlib_main_t * vm,
-				dpdk_device_t * xd,
-				struct rte_mbuf **mb, u32 n_left)
+static_always_inline u32
+tx_burst_vector_internal (vlib_main_t *vm, dpdk_device_t *xd,
+			  struct rte_mbuf **mb, u32 n_left, int queue_id,
+			  u8 is_shared)
 {
   dpdk_tx_queue_t *txq;
   u32 n_retry;
   int n_sent = 0;
-  int queue_id;
 
   n_retry = 16;
-  queue_id = vm->thread_index % xd->conf.n_tx_queues;
   txq = vec_elt_at_index (xd->tx_queues, queue_id);
 
   do
     {
-      clib_spinlock_lock_if_init (&txq->lock);
+      if (is_shared)
+	clib_spinlock_lock (&txq->lock);
 
       /* no wrap, transmit in one burst */
       n_sent = rte_eth_tx_burst (xd->port_id, queue_id, mb, n_left);
 
-      clib_spinlock_unlock_if_init (&txq->lock);
+      if (is_shared)
+	clib_spinlock_unlock (&txq->lock);
 
       n_retry--;
       n_left -= n_sent;
@@ -276,10 +276,12 @@ VNET_DEVICE_CLASS_TX_FN (dpdk_device_class) (vlib_main_t * vm,
   dpdk_main_t *dm = &dpdk_main;
   vnet_interface_output_runtime_t *rd = (void *) node->runtime_data;
   dpdk_device_t *xd = vec_elt_at_index (dm->devices, rd->dev_instance);
+  vnet_hw_if_tx_frame_t *tf = vlib_frame_scalar_args (f);
   u32 n_packets = f->n_vectors;
   u32 n_left;
   u32 thread_index = vm->thread_index;
-  int queue_id = thread_index;
+  int queue_id = tf->queue_id;
+  u8 is_shared = tf->shared_queue;
   u32 tx_pkts = 0;
   dpdk_per_thread_data_t *ptd = vec_elt_at_index (dm->per_thread_data,
 						  thread_index);
@@ -418,7 +420,8 @@ VNET_DEVICE_CLASS_TX_FN (dpdk_device_class) (vlib_main_t * vm,
 
   /* transmit as many packets as possible */
   tx_pkts = n_packets = mb - ptd->mbufs;
-  n_left = tx_burst_vector_internal (vm, xd, ptd->mbufs, n_packets);
+  n_left = tx_burst_vector_internal (vm, xd, ptd->mbufs, n_packets, queue_id,
+				     is_shared);
 
   {
     /* If there is no callback then drop any non-transmitted packets */

src/plugins/dpdk/device/dpdk.h

@@ -94,6 +94,7 @@ typedef struct
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
   clib_spinlock_t lock;
+  u32 queue_index;
 } dpdk_tx_queue_t;
 
 typedef struct

src/plugins/dpdk/device/init.c

@@ -24,6 +24,7 @@
 #include <vnet/vnet.h>
 #include <vnet/ethernet/ethernet.h>
 #include <vnet/interface/rx_queue_funcs.h>
+#include <vnet/interface/tx_queue_funcs.h>
 #include <dpdk/buffer.h>
 #include <dpdk/device/dpdk.h>
 #include <dpdk/cryptodev/cryptodev.h>
@@ -429,6 +430,8 @@ dpdk_lib_init (dpdk_main_t * dm)
 
       vec_validate_aligned (xd->rx_queues, xd->conf.n_rx_queues - 1,
 			    CLIB_CACHE_LINE_BYTES);
+      vec_validate_aligned (xd->tx_queues, xd->conf.n_tx_queues - 1,
+			    CLIB_CACHE_LINE_BYTES);
 
       rte_eth_macaddr_get (port_id, (void *) addr);
 
@@ -469,6 +472,19 @@ dpdk_lib_init (dpdk_main_t * dm)
 	      vnm, xd->hw_if_index, q, VNET_HW_IF_RXQ_THREAD_ANY);
 	  }
 
+      for (q = 0; q < xd->conf.n_tx_queues; q++)
+	{
+	  dpdk_tx_queue_t *txq = vec_elt_at_index (xd->tx_queues, q);
+	  txq->queue_index =
+	    vnet_hw_if_register_tx_queue (vnm, xd->hw_if_index, q);
+	}
+
+      for (q = 0; q < tm->n_vlib_mains; q++)
+	{
+	  u32 qi = xd->tx_queues[q % xd->conf.n_tx_queues].queue_index;
+	  vnet_hw_if_tx_queue_assign_thread (vnm, qi, q);
+	}
+
       if (devconf->tso == DPDK_DEVICE_TSO_ON)
 	{
 	  /*tcp_udp checksum must be enabled*/