3 files changed, 18 insertions, 7 deletions

diff --git a/src/plugins/avf/avf.h b/src/plugins/avf/avf.h
index ea931dc07c4..c03a7c252ae 100644
--- a/src/plugins/avf/avf.h
+++ b/src/plugins/avf/avf.h
@@ -191,6 +191,7 @@ typedef struct
 
   avf_tx_desc_t *tmp_descs;
   u32 *tmp_bufs;
+  u32 queue_index;
 } avf_txq_t;
 
 typedef struct
diff --git a/src/plugins/avf/device.c b/src/plugins/avf/device.c
index 70ea446432c..05946a4087d 100644
--- a/src/plugins/avf/device.c
+++ b/src/plugins/avf/device.c
@@ -21,6 +21,7 @@
 #include <vlib/pci/pci.h>
 #include <vnet/ethernet/ethernet.h>
 #include <vnet/interface/rx_queue_funcs.h>
+#include <vnet/interface/tx_queue_funcs.h>
 
 #include <avf/avf.h>
 
@@ -303,8 +304,7 @@ avf_txq_init (vlib_main_t * vm, avf_device_t * ad, u16 qid, u16 txq_size)
     {
       qid = qid % ad->num_queue_pairs;
       txq = vec_elt_at_index (ad->txqs, qid);
-      if (txq->lock == 0)
-	clib_spinlock_init (&txq->lock);
+      clib_spinlock_init (&txq->lock);
       ad->flags |= AVF_DEVICE_F_SHARED_TXQ_LOCK;
       return 0;
     }
@@ -1748,6 +1748,14 @@ avf_create_if (vlib_main_t * vm, avf_create_if_args_t * args)
 	}
       ad->rxqs[i].queue_index = qi;
     }
+
+  for (i = 0; i < ad->n_tx_queues; i++)
+    {
+      u32 qi = vnet_hw_if_register_tx_queue (vnm, ad->hw_if_index, i);
+      vnet_hw_if_tx_queue_assign_thread (vnm, qi, i);
+      ad->txqs[i].queue_index = qi;
+    }
+
   vnet_hw_if_update_runtime_data (vnm, ad->hw_if_index);
 
   if (pool_elts (am->devices) == 1)
diff --git a/src/plugins/avf/output.c b/src/plugins/avf/output.c
index 78043358e77..4cc9d5a49c1 100644
--- a/src/plugins/avf/output.c
+++ b/src/plugins/avf/output.c
@@ -375,16 +375,17 @@ VNET_DEVICE_CLASS_TX_FN (avf_device_class) (vlib_main_t * vm,
 {
   vnet_interface_output_runtime_t *rd = (void *) node->runtime_data;
   avf_device_t *ad = avf_get_device (rd->dev_instance);
-  u32 thread_index = vm->thread_index;
-  u8 qid = thread_index;
-  avf_txq_t *txq = vec_elt_at_index (ad->txqs, qid % ad->num_queue_pairs);
+  vnet_hw_if_tx_frame_t *tf = vlib_frame_scalar_args (frame);
+  u8 qid = tf->queue_id;
+  avf_txq_t *txq = vec_elt_at_index (ad->txqs, qid);
   u16 next;
   u16 mask = txq->size - 1;
   u32 *buffers = vlib_frame_vector_args (frame);
   u16 n_enq, n_left, n_desc, *slot;
   u16 n_retry = 2;
 
-  clib_spinlock_lock_if_init (&txq->lock);
+  if (tf->shared_queue)
+    clib_spinlock_lock (&txq->lock);
 
   n_left = frame->n_vectors;
 
@@ -474,7 +475,8 @@ retry:
 			AVF_TX_ERROR_NO_FREE_SLOTS, n_left);
     }
 
-  clib_spinlock_unlock_if_init (&txq->lock);
+  if (tf->shared_queue)
+    clib_spinlock_unlock (&txq->lock);
 
   return frame->n_vectors - n_left;
 }