virtio: interface type based improvements

Type: improvement

Change-Id: I220ea6ab609e3c1628f5210be441d0d5e825a32c
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>

3 files changed, 142 insertions, 75 deletions

diff --git a/src/vnet/devices/virtio/device.c b/src/vnet/devices/virtio/device.c
index 9a35bb8162f..87fd6fdcf5f 100644
--- a/src/vnet/devices/virtio/device.c
+++ b/src/vnet/devices/virtio/device.c
@@ -255,7 +255,8 @@ set_gso_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr, int is_l2)
 
 static_always_inline u16
 add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
-		    virtio_vring_t * vring, u32 bi, u16 free_desc_count,
+		    virtio_if_type_t type, virtio_vring_t * vring,
+		    u32 bi, u16 free_desc_count,
 		    u16 avail, u16 next, u16 mask, int do_gso,
 		    int csum_offload, uword node_index)
 {
@@ -265,7 +266,7 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
   d = &vring->desc[next];
   vlib_buffer_t *b = vlib_get_buffer (vm, bi);
   struct virtio_net_hdr_v1 *hdr = vlib_buffer_get_current (b) - hdr_sz;
-  int is_l2 = (vif->type & (VIRTIO_IF_TYPE_TAP | VIRTIO_IF_TYPE_PCI));
+  int is_l2 = (type & (VIRTIO_IF_TYPE_TAP | VIRTIO_IF_TYPE_PCI));
 
   clib_memset (hdr, 0, hdr_sz);
 
@@ -296,8 +297,8 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
   if (PREDICT_TRUE ((b->flags & VLIB_BUFFER_NEXT_PRESENT) == 0))
     {
       d->addr =
-	((vif->type == VIRTIO_IF_TYPE_PCI) ? vlib_buffer_get_current_pa (vm,
-									 b) :
+	((type == VIRTIO_IF_TYPE_PCI) ? vlib_buffer_get_current_pa (vm,
+								    b) :
 	 pointer_to_uword (vlib_buffer_get_current (b))) - hdr_sz;
       d->len = b->current_length + hdr_sz;
       d->flags = 0;
@@ -329,7 +330,7 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
       struct vring_desc *id =
 	(struct vring_desc *) vlib_buffer_get_current (indirect_desc);
       u32 count = 1;
-      if (vif->type == VIRTIO_IF_TYPE_PCI)
+      if (type == VIRTIO_IF_TYPE_PCI)
 	{
 	  d->addr = vlib_physmem_get_pa (vm, id);
 	  id->addr = vlib_buffer_get_current_pa (vm, b) - hdr_sz;
@@ -386,7 +387,7 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
       d->len = count * sizeof (struct vring_desc);
       d->flags = VRING_DESC_F_INDIRECT;
     }
-  else if (vif->type == VIRTIO_IF_TYPE_PCI)
+  else if (type == VIRTIO_IF_TYPE_PCI)
     {
       u16 count = next;
       vlib_buffer_t *b_temp = b;
@@ -478,9 +479,10 @@ virtio_find_free_desc (virtio_vring_t * vring, u16 size, u16 mask,
 }
 
 static_always_inline uword
-virtio_interface_tx_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
-			    vlib_frame_t * frame, virtio_if_t * vif,
-			    int do_gso, int csum_offload)
+virtio_interface_tx_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+				vlib_frame_t * frame, virtio_if_t * vif,
+				virtio_if_type_t type, int do_gso,
+				int csum_offload)
 {
   u16 n_left = frame->n_vectors;
   virtio_vring_t *vring;
@@ -525,7 +527,7 @@ retry:
     {
       u16 n_added = 0;
       n_added =
-	add_buffer_to_slot (vm, vif, vring, buffers[0], free_desc_count,
+	add_buffer_to_slot (vm, vif, type, vring, buffers[0], free_desc_count,
 			    avail, next, mask, do_gso, csum_offload,
 			    node->node_index);
 
@@ -570,25 +572,49 @@ retry:
   return frame->n_vectors - n_left;
 }
 
+static_always_inline uword
+virtio_interface_tx_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+			    vlib_frame_t * frame, virtio_if_t * vif,
+			    virtio_if_type_t type)
+{
+  vnet_main_t *vnm = vnet_get_main ();
+  vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, vif->hw_if_index);
+
+  if (hw->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO)
+    return virtio_interface_tx_gso_inline (vm, node, frame, vif, type,
+					   1 /* do_gso */ ,
+					   1 /* checksum offload */ );
+  else if (hw->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD)
+    return virtio_interface_tx_gso_inline (vm, node, frame, vif, type,
+					   0 /* no do_gso */ ,
+					   1 /* checksum offload */ );
+  else
+    return virtio_interface_tx_gso_inline (vm, node, frame, vif, type,
+					   0 /* no do_gso */ ,
+					   0 /* no checksum offload */ );
+}
+
 VNET_DEVICE_CLASS_TX_FN (virtio_device_class) (vlib_main_t * vm,
 					       vlib_node_runtime_t * node,
 					       vlib_frame_t * frame)
 {
-  vnet_main_t *vnm = vnet_get_main ();
   virtio_main_t *nm = &virtio_main;
   vnet_interface_output_runtime_t *rund = (void *) node->runtime_data;
   virtio_if_t *vif = pool_elt_at_index (nm->interfaces, rund->dev_instance);
-  vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, vif->hw_if_index);
 
-  if (hw->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO)
-    return virtio_interface_tx_inline (vm, node, frame, vif, 1 /* do_gso */ ,
-				       1);
-  else if (hw->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD)
+  if (vif->type == VIRTIO_IF_TYPE_TAP)
     return virtio_interface_tx_inline (vm, node, frame, vif,
-				       0 /* no do_gso */ , 1);
-  else
+				       VIRTIO_IF_TYPE_TAP);
+  else if (vif->type == VIRTIO_IF_TYPE_PCI)
+    return virtio_interface_tx_inline (vm, node, frame, vif,
+				       VIRTIO_IF_TYPE_PCI);
+  else if (vif->type == VIRTIO_IF_TYPE_TUN)
     return virtio_interface_tx_inline (vm, node, frame, vif,
-				       0 /* no do_gso */ , 0);
+				       VIRTIO_IF_TYPE_TUN);
+  else
+    ASSERT (0);
+
+  return 0;
 }
 
 static void
diff --git a/src/vnet/devices/virtio/node.c b/src/vnet/devices/virtio/node.c
index 8d7c9360b00..06641852316 100644
--- a/src/vnet/devices/virtio/node.c
+++ b/src/vnet/devices/virtio/node.c
@@ -80,7 +80,8 @@ format_virtio_input_trace (u8 * s, va_list * args)
 
 static_always_inline void
 virtio_refill_vring (vlib_main_t * vm, virtio_if_t * vif,
-		     virtio_vring_t * vring, const int hdr_sz)
+		     virtio_if_type_t type, virtio_vring_t * vring,
+		     const int hdr_sz)
 {
   u16 used, next, avail, n_slots;
   u16 sz = vring->size;
@@ -114,12 +115,11 @@ more:
        * previous offset. Here we want to make sure, it should be 0
        * initialized.
        */
-      b->current_data = 0;
-      b->current_data -= hdr_sz;
+      b->current_data = -hdr_sz;
       memset (vlib_buffer_get_current (b), 0, hdr_sz);
       d->addr =
-	((vif->type == VIRTIO_IF_TYPE_PCI) ? vlib_buffer_get_current_pa (vm,
-									 b) :
+	((type == VIRTIO_IF_TYPE_PCI) ? vlib_buffer_get_current_pa (vm,
+								    b) :
 	 pointer_to_uword (vlib_buffer_get_current (b)));
       d->len = vlib_buffer_get_default_data_size (vm) + hdr_sz;
       d->flags = VRING_DESC_F_WRITE;
@@ -146,10 +146,22 @@ virtio_needs_csum (vlib_buffer_t * b0, struct virtio_net_hdr_v1 *hdr,
 		   u8 * l4_proto, u8 * l4_hdr_sz, virtio_if_type_t type)
 {
   if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
-
     {
       u16 ethertype = 0, l2hdr_sz = 0;
-      if (type != VIRTIO_IF_TYPE_TUN)
+
+      if (type == VIRTIO_IF_TYPE_TUN)
+	{
+	  switch (b0->data[0] & 0xf0)
+	    {
+	    case 0x40:
+	      ethertype = ETHERNET_TYPE_IP4;
+	      break;
+	    case 0x60:
+	      ethertype = ETHERNET_TYPE_IP6;
+	      break;
+	    }
+	}
+      else
 	{
 	  ethernet_header_t *eh =
 	    (ethernet_header_t *) vlib_buffer_get_current (b0);
@@ -171,21 +183,10 @@ virtio_needs_csum (vlib_buffer_t * b0, struct virtio_net_hdr_v1 *hdr,
 		}
 	    }
 	}
-      else
-	{
-	  switch (b0->data[0] & 0xf0)
-	    {
-	    case 0x40:
-	      ethertype = ETHERNET_TYPE_IP4;
-	      break;
-	    case 0x60:
-	      ethertype = ETHERNET_TYPE_IP6;
-	      break;
-	    }
-	}
 
       vnet_buffer (b0)->l2_hdr_offset = 0;
       vnet_buffer (b0)->l3_hdr_offset = l2hdr_sz;
+
       if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP4))
 	{
 	  ip4_header_t *ip4 =
@@ -252,9 +253,10 @@ fill_gso_buffer_flags (vlib_buffer_t * b0, struct virtio_net_hdr_v1 *hdr,
 }
 
 static_always_inline uword
-virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
-			    vlib_frame_t * frame, virtio_if_t * vif, u16 qid,
-			    int gso_enabled, int checksum_offload_enabled)
+virtio_device_input_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+				vlib_frame_t * frame, virtio_if_t * vif,
+				u16 qid, virtio_if_type_t type,
+				int gso_enabled, int checksum_offload_enabled)
 {
   vnet_main_t *vnm = vnet_get_main ();
   u32 thread_index = vm->thread_index;
@@ -276,10 +278,14 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   if (n_left == 0)
     goto refill;
 
+  if (type == VIRTIO_IF_TYPE_TUN)
+    next_index = VNET_DEVICE_INPUT_NEXT_IP4_INPUT;
+
   while (n_left)
     {
       u32 n_left_to_next;
       u32 next0 = next_index;
+
       vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
 
       while (n_left && n_left_to_next)
@@ -296,13 +302,11 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  if (hdr_sz == sizeof (struct virtio_net_hdr_v1))
 	    num_buffers = hdr->num_buffers;
 
-	  b0->current_data += hdr_sz;
+	  b0->current_data = 0;
 	  b0->current_length = len;
-	  b0->total_length_not_including_first_buffer = 0;
-	  b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID;
 
 	  if (checksum_offload_enabled)
-	    virtio_needs_csum (b0, hdr, &l4_proto, &l4_hdr_sz, vif->type);
+	    virtio_needs_csum (b0, hdr, &l4_proto, &l4_hdr_sz, type);
 
 	  if (gso_enabled)
 	    fill_gso_buffer_flags (b0, hdr, l4_proto, l4_hdr_sz);
@@ -315,6 +319,8 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	    {
 	      vlib_buffer_t *pb, *cb;
 	      pb = b0;
+	      b0->total_length_not_including_first_buffer = 0;
+	      b0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID;
 	      while (num_buffers > 1)
 		{
 		  last++;
@@ -337,8 +343,10 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		  num_buffers--;
 		  n_left--;
 		}
+	      len += b0->total_length_not_including_first_buffer;
 	    }
-	  if (PREDICT_FALSE (vif->type == VIRTIO_IF_TYPE_TUN))
+
+	  if (type == VIRTIO_IF_TYPE_TUN)
 	    {
 	      switch (b0->data[0] & 0xf0)
 		{
@@ -354,12 +362,15 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 		}
 	    }
 
-
 	  if (PREDICT_FALSE (vif->per_interface_next_index != ~0))
 	    next0 = vif->per_interface_next_index;
 
-	  /* redirect if feature path enabled */
-	  vnet_feature_start_device_input_x1 (vif->sw_if_index, &next0, b0);
+	  if (type != VIRTIO_IF_TYPE_TUN)
+	    {
+	      /* only for l2, redirect if feature path enabled */
+	      vnet_feature_start_device_input_x1 (vif->sw_if_index, &next0,
+						  b0);
+	    }
 
 	  /* trace */
 	  VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b0);
@@ -373,7 +384,7 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	      tr = vlib_add_trace (vm, node, b0, sizeof (*tr));
 	      tr->next_index = next0;
 	      tr->hw_if_index = vif->hw_if_index;
-	      tr->len = len + b0->total_length_not_including_first_buffer;
+	      tr->len = len;
 	      clib_memcpy_fast (&tr->hdr, hdr, hdr_sz);
 	    }
 
@@ -391,7 +402,7 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  /* next packet */
 	  n_rx_packets++;
-	  n_rx_bytes += (len + b0->total_length_not_including_first_buffer);
+	  n_rx_bytes += len;
 	}
       vlib_put_next_frame (vm, node, next_index, n_left_to_next);
     }
@@ -403,11 +414,29 @@ virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 				   n_rx_bytes);
 
 refill:
-  virtio_refill_vring (vm, vif, vring, hdr_sz);
+  virtio_refill_vring (vm, vif, type, vring, hdr_sz);
 
   return n_rx_packets;
 }
 
+static_always_inline uword
+virtio_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
+			    vlib_frame_t * frame, virtio_if_t * vif, u16 qid,
+			    virtio_if_type_t type)
+{
+
+  if (vif->gso_enabled)
+    return virtio_device_input_gso_inline (vm, node, frame, vif,
+					   qid, type, 1, 1);
+  else if (vif->csum_offload_enabled)
+    return virtio_device_input_gso_inline (vm, node, frame, vif,
+					   qid, type, 0, 1);
+  else
+    return virtio_device_input_gso_inline (vm, node, frame, vif,
+					   qid, type, 0, 0);
+  return 0;
+}
+
 VLIB_NODE_FN (virtio_input_node) (vlib_main_t * vm,
 				  vlib_node_runtime_t * node,
 				  vlib_frame_t * frame)
@@ -423,15 +452,18 @@ VLIB_NODE_FN (virtio_input_node) (vlib_main_t * vm,
     vif = vec_elt_at_index (nm->interfaces, dq->dev_instance);
     if (vif->flags & VIRTIO_IF_FLAG_ADMIN_UP)
       {
-	if (vif->gso_enabled)
+	if (vif->type == VIRTIO_IF_TYPE_TAP)
 	  n_rx += virtio_device_input_inline (vm, node, frame, vif,
-					      dq->queue_id, 1, 1);
-	else if (vif->csum_offload_enabled)
+					      dq->queue_id,
+					      VIRTIO_IF_TYPE_TAP);
+	else if (vif->type == VIRTIO_IF_TYPE_PCI)
 	  n_rx += virtio_device_input_inline (vm, node, frame, vif,
-					      dq->queue_id, 0, 1);
-	else
+					      dq->queue_id,
+					      VIRTIO_IF_TYPE_PCI);
+	else if (vif->type == VIRTIO_IF_TYPE_TUN)
 	  n_rx += virtio_device_input_inline (vm, node, frame, vif,
-					      dq->queue_id, 0, 0);
+					      dq->queue_id,
+					      VIRTIO_IF_TYPE_TUN);
       }
   }
 
diff --git a/src/vnet/devices/virtio/virtio.h b/src/vnet/devices/virtio/virtio.h
index f5dc1c86930..68462d8075d 100644
--- a/src/vnet/devices/virtio/virtio.h
+++ b/src/vnet/devices/virtio/virtio.h
@@ -106,10 +106,10 @@ typedef struct
 typedef struct
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
+  clib_spinlock_t lockp;
   struct vring_desc *desc;
   struct vring_used *used;
   struct vring_avail *avail;
-  clib_spinlock_t lockp;
   u16 desc_in_use;
   u16 desc_next;
   int kick_fd;
@@ -119,10 +119,10 @@ typedef struct
   u16 queue_id;
 #define VRING_TX_OUT_OF_ORDER 1
   u16 flags;
-  u32 call_file_index;
   u32 *buffers;
   u16 last_used_idx;
   u16 last_kick_avail_idx;
+  u32 call_file_index;
 } virtio_vring_t;
 
 typedef union
@@ -140,34 +140,45 @@ typedef union
 typedef struct
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
+  u64 features;
   u32 flags;
+  u32 per_interface_next_index;
+  u16 num_rxqs;
+  u16 num_txqs;
+  virtio_vring_t *rxq_vrings;
+  virtio_vring_t *txq_vrings;
+  int gso_enabled;
+  int csum_offload_enabled;
+  union
+  {
+    int *tap_fds;
+    struct
+    {
+      u32 pci_dev_handle;
+      u32 msix_enabled;
+    };
+  };
+  u16 virtio_net_hdr_sz;
+  virtio_if_type_t type;
 
-  u32 dev_instance;
   u32 hw_if_index;
   u32 sw_if_index;
-  u32 numa_node;
-  u16 virtio_net_hdr_sz;
-  virtio_if_type_t type;
+
+    CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
   union
   {
     u32 id;
     pci_addr_t pci_addr;
   };
-  u32 per_interface_next_index;
   int *vhost_fds;
-  int *tap_fds;
-  u32 msix_enabled;
-  u32 pci_dev_handle;
-  virtio_vring_t *rxq_vrings;
-  virtio_vring_t *txq_vrings;
-  u64 features, remote_features;
+  u32 dev_instance;
+  u32 numa_node;
+  u64 remote_features;
 
   /* error */
   clib_error_t *error;
   u8 support_int_mode;		/* support interrupt mode */
   u16 max_queue_pairs;
-  u16 num_rxqs;
-  u16 num_txqs;
   u8 status;
   u8 mac_addr[6];
   u8 *host_if_name;
@@ -179,8 +190,6 @@ typedef struct
   ip6_address_t host_ip6_addr;
   u8 host_ip6_prefix_len;
   u32 host_mtu_size;
-  int gso_enabled;
-  int csum_offload_enabled;
   int ifindex;
   virtio_vring_t *cxq_vring;
 } virtio_if_t;