avf: use write combining store for queues tail update

Performance improvement: on supported platforms, currently only Intel Tremont,
use a write combining store to update the tail pointers.
Also, Tremont node variant is added for all.

Type: improvement

Signed-off-by: Radu Nicolau <radu.nicolau@intel.com>
Change-Id: Ie9606e403b7d9655184f778e3ffee3027c8c9edd

3 files changed, 13 insertions, 2 deletions

diff --git a/src/plugins/avf/avf.h b/src/plugins/avf/avf.h
index b3fcc259206..23cc36c4882 100644
--- a/src/plugins/avf/avf.h
+++ b/src/plugins/avf/avf.h
@@ -373,6 +373,17 @@ avf_reg_flush (avf_device_t * ad)
   asm volatile ("":::"memory");
 }
 
+static inline void
+avf_tail_write (volatile u32 *addr, u32 val)
+{
+#ifdef __MOVDIRI__
+  _mm_sfence ();
+  _directstoreu_u32 ((void *) addr, val);
+#else
+  clib_atomic_store_rel_n (addr, val);
+#endif
+}
+
 static_always_inline int
 avf_rxd_is_not_eop (avf_rx_desc_t * d)
 {
diff --git a/src/plugins/avf/input.c b/src/plugins/avf/input.c
index 5041f6ef4e7..221b54b19a9 100644
--- a/src/plugins/avf/input.c
+++ b/src/plugins/avf/input.c
@@ -125,7 +125,7 @@ avf_rxq_refill (vlib_main_t * vm, vlib_node_runtime_t * node, avf_rxq_t * rxq,
       n_alloc -= 8;
     }
 
-  clib_atomic_store_rel_n (rxq->qrx_tail, slot);
+  avf_tail_write (rxq->qrx_tail, slot);
 }
 
 
diff --git a/src/plugins/avf/output.c b/src/plugins/avf/output.c
index 952d151117c..c53fc42a850 100644
--- a/src/plugins/avf/output.c
+++ b/src/plugins/avf/output.c
@@ -441,7 +441,7 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
     }
 
   txq->next = next & mask;
-  clib_atomic_store_rel_n (txq->qtx_tail, txq->next);
+  avf_tail_write (txq->qtx_tail, txq->next);
   txq->n_enqueued += n_desc;
   return n_packets - n_packets_left;
 }