Handle dynamic enable/disable VHOST_USER interface

    If number of Qs enabled by the guest is less
    than number of Qs exported to guest, then all
    threads route packet to Q0 with the assumption
    that guest is not interested in performance.
    If all Qs are enabled, each thread queues packet
    in their own TX-Q of the VHOST_USER interface,
    boosting performance.

Change-Id: Ic24bb8c0505b11c7513aeecd21c9ec5da5f90138
Signed-off-by: Shesha Sreenivasamurthy <shesha@cisco.com>

4 files changed, 29 insertions, 11 deletions

diff --git a/vnet/vnet/devices/dpdk/device.c b/vnet/vnet/devices/dpdk/device.c
index 2f9f135a260..327f9dec5a3 100644
--- a/vnet/vnet/devices/dpdk/device.c
+++ b/vnet/vnet/devices/dpdk/device.c
@@ -221,7 +221,8 @@ u32 tx_burst_vector_internal (vlib_main_t * vm,
        * This device only supports one TX queue,
        * and we're running multi-threaded...
        */
-      if (PREDICT_FALSE(xd->lockp != 0))
+      if (PREDICT_FALSE(xd->dev_type != VNET_DPDK_DEV_VHOST_USER &&
+        xd->lockp != 0))
         {
           queue_id = queue_id % xd->tx_q_used;
           while (__sync_lock_test_and_set (xd->lockp[queue_id], 1))
@@ -264,8 +265,12 @@ u32 tx_burst_vector_internal (vlib_main_t * vm,
       else if (xd->dev_type == VNET_DPDK_DEV_VHOST_USER)
         {
           u32 offset = 0;
+          if (xd->need_txlock) {
+            queue_id = 0;
+            while (__sync_lock_test_and_set (xd->lockp[queue_id], 1));
+          }
 #if RTE_VERSION >= RTE_VERSION_NUM(2, 2, 0, 0)
-          if (PREDICT_TRUE(xd->lockp == NULL)) {
+          else {
               dpdk_device_and_queue_t * dq;
               vec_foreach (dq, dm->devices_by_cpu[vm->cpu_index])
               {
@@ -274,8 +279,6 @@ u32 tx_burst_vector_internal (vlib_main_t * vm,
               }
               assert (dq);
               offset = dq->queue_id * VIRTIO_QNUM;
-          } else {
-              offset = queue_id * VIRTIO_QNUM;
           }
 #endif
           if (PREDICT_TRUE(tx_head > tx_tail)) 
@@ -331,6 +334,9 @@ u32 tx_burst_vector_internal (vlib_main_t * vm,
 
               n_retry = (rv == DPDK_TX_RING_SIZE - tx_tail) ? 1 : 0;
             }
+
+          if (xd->need_txlock)
+            *xd->lockp[queue_id] = 0;
         }
 #if RTE_LIBRTE_KNI
       else if (xd->dev_type == VNET_DPDK_DEV_KNI)
@@ -370,7 +376,8 @@ u32 tx_burst_vector_internal (vlib_main_t * vm,
           rv = 0;
         }
 
-      if (PREDICT_FALSE(xd->lockp != 0))
+      if (PREDICT_FALSE(xd->dev_type != VNET_DPDK_DEV_VHOST_USER &&
+            xd->lockp != 0))
           *xd->lockp[queue_id] = 0;
 
       if (PREDICT_FALSE(rv < 0))
diff --git a/vnet/vnet/devices/dpdk/dpdk.h b/vnet/vnet/devices/dpdk/dpdk.h
index 48c17729f50..da4e381b3c1 100644
--- a/vnet/vnet/devices/dpdk/dpdk.h
+++ b/vnet/vnet/devices/dpdk/dpdk.h
@@ -246,6 +246,7 @@ typedef struct {
   dpdk_port_type_t port_type;
 
   dpdk_efd_agent_t efd_agent;
+  u8 need_txlock; /* Used by VNET_DPDK_DEV_VHOST_USER */
 } dpdk_device_t;
 
 #define MAX_NELTS 32
diff --git a/vnet/vnet/devices/dpdk/init.c b/vnet/vnet/devices/dpdk/init.c
index a67ee45804c..6589d2fe14b 100644
--- a/vnet/vnet/devices/dpdk/init.c
+++ b/vnet/vnet/devices/dpdk/init.c
@@ -215,6 +215,7 @@ dpdk_device_lock_init(dpdk_device_t * xd)
                                              CLIB_CACHE_LINE_BYTES);
       memset ((void *) xd->lockp[q], 0, CLIB_CACHE_LINE_BYTES);
   }
+  xd->need_txlock = 1;
 }
 
 void
@@ -226,6 +227,7 @@ dpdk_device_lock_free(dpdk_device_t * xd)
     clib_mem_free((void *) xd->lockp[q]);
   vec_free(xd->lockp);
   xd->lockp = 0;
+  xd->need_txlock = 0;
 }
 
 static clib_error_t *
diff --git a/vnet/vnet/devices/dpdk/vhost_user.c b/vnet/vnet/devices/dpdk/vhost_user.c
index 0efbbef3259..b6b31bafb85 100644
--- a/vnet/vnet/devices/dpdk/vhost_user.c
+++ b/vnet/vnet/devices/dpdk/vhost_user.c
@@ -198,7 +198,7 @@ dpdk_create_vhost_user_if_internal (u32 * hw_if_index, u32 if_id, u8 *hwaddr)
   int num_qpairs = 1;
 
 #if RTE_VERSION >= RTE_VERSION_NUM(2, 2, 0, 0)
-  num_qpairs = dm->use_rss < 1 ? 1 : dm->use_rss;
+  num_qpairs = dm->use_rss < 1 ? 1 : tm->n_vlib_mains;
 #endif
 
   dpdk_device_t * xd = NULL;
@@ -242,9 +242,7 @@ dpdk_create_vhost_user_if_internal (u32 * hw_if_index, u32 if_id, u8 *hwaddr)
 
       // reset lockp
       dpdk_device_lock_free(xd);
-
-      if (xd->tx_q_used < tm->n_vlib_mains)
-        dpdk_device_lock_init(xd);
+      dpdk_device_lock_init(xd);
 
       // reset tx vectors
       for (j = 0; j < tm->n_vlib_mains; j++)
@@ -301,8 +299,7 @@ dpdk_create_vhost_user_if_internal (u32 * hw_if_index, u32 if_id, u8 *hwaddr)
           xd->vu_vhost_dev.virtqueue[j]->backend = -1; 
       }
 
-      if (xd->tx_q_used < dm->input_cpu_count)
-        dpdk_device_lock_init(xd);
+      dpdk_device_lock_init(xd);
 
       DBG_SOCK("tm->n_vlib_mains: %d. TX %d, RX: %d, num_qpairs: %d, Lock: %p",
         tm->n_vlib_mains, xd->tx_q_used, xd->rx_q_used, num_qpairs, xd->lockp);
@@ -706,6 +703,17 @@ int dpdk_vhost_user_set_vring_enable(u32 hw_if_index, u8 idx, int enable)
    */
   vui->vrings[idx].enabled = enable; /* Save local copy */
 
+  int numqs = xd->vu_vhost_dev.virt_qp_nb * VIRTIO_QNUM;
+  while (numqs--) {
+    if (! vui->vrings[numqs].enabled)
+        break;
+  }
+
+  if (numqs == -1) /* All Qs are enabled */
+    xd->need_txlock = 0;
+  else
+    xd->need_txlock = 1;
+
   vq = xd->vu_vhost_dev.virtqueue[idx];
   if (vq->desc && vq->avail && vq->used)
     xd->vu_vhost_dev.virtqueue[idx]->enabled = enable;