tap: add gro support

Type: feature

Change-Id: I5868dd267aa26aa97aec5fd70e70c5956ac52277
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>

8 files changed, 87 insertions, 14 deletions

diff --git a/src/vnet/devices/tap/cli.c b/src/vnet/devices/tap/cli.c
index 34897ea7c5d..fa5fa91e7a6 100644
--- a/src/vnet/devices/tap/cli.c
+++ b/src/vnet/devices/tap/cli.c
@@ -201,7 +201,7 @@ tap_offload_command_fn (vlib_main_t * vm, unformat_input_t * input,
   unformat_input_t _line_input, *line_input = &_line_input;
   u32 sw_if_index = ~0;
   vnet_main_t *vnm = vnet_get_main ();
-  int gso_enable = 0, gso_disable = 0;
+  int gso_enable = 0, gso_disable = 0, is_gro_coalesce = 0;
   int csum_offload_enable = 0, csum_offload_disable = 0;
   int rv = 0;
 
@@ -217,7 +217,11 @@ tap_offload_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			 vnm, &sw_if_index))
 	;
       else if (unformat (line_input, "gso-enable"))
-	gso_enable = 1;
+	{
+	  gso_enable = 1;
+	  if (unformat (line_input, "gro-coalesce"))
+	    is_gro_coalesce = 1;
+	}
       else if (unformat (line_input, "gso-disable"))
 	gso_disable = 1;
       else if (unformat (line_input, "csum-offload-enable"))
@@ -235,11 +239,11 @@ tap_offload_command_fn (vlib_main_t * vm, unformat_input_t * input,
 			      "please specify interface name or sw_if_index");
 
   if (gso_enable)
-    rv = tap_gso_enable_disable (vm, sw_if_index, 1);
+    rv = tap_gso_enable_disable (vm, sw_if_index, 1, is_gro_coalesce);
   else if (csum_offload_enable)
     rv = tap_csum_offload_enable_disable (vm, sw_if_index, 1);
   else if (gso_disable)
-    rv = tap_gso_enable_disable (vm, sw_if_index, 0);
+    rv = tap_gso_enable_disable (vm, sw_if_index, 0, 0);
   else if (csum_offload_disable)
     rv = tap_csum_offload_enable_disable (vm, sw_if_index, 0);
 
@@ -256,7 +260,8 @@ VLIB_CLI_COMMAND (tap_offload_command, static) =
 {
   .path = "set tap offload",
   .short_help = "set tap offload {<interface> | sw_if_index <sw_idx>}"
-    " <gso-enable | gso-disable | csum-offload-enable | csum-offload-disable>",
+    " <gso-enable [gro-coalesce]  | gso-disable | csum-offload-enable |"
+    "csum-offload-disable>",
   .function = tap_offload_command_fn,
 };
 /* *INDENT-ON* */
diff --git a/src/vnet/devices/tap/tap.c b/src/vnet/devices/tap/tap.c
index fa5eef19c0d..a76b4e0c16a 100644
--- a/src/vnet/devices/tap/tap.c
+++ b/src/vnet/devices/tap/tap.c
@@ -730,6 +730,12 @@ tap_create_if (vlib_main_t * vm, tap_create_if_args_t * args)
     {
       hw->flags |= VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD;
     }
+  if ((args->tap_flags & TAP_FLAG_GSO)
+      && (args->tap_flags & TAP_FLAG_GRO_COALESCE))
+    {
+      vif->packet_coalesce = 1;
+      virtio_set_packet_coalesce (vif);
+    }
   vnet_hw_interface_set_input_node (vnm, vif->hw_if_index,
 				    virtio_input_node.index);
 
@@ -827,6 +833,7 @@ tap_csum_offload_enable_disable (vlib_main_t * vm, u32 sw_if_index,
   vec_foreach_index (i, vif->tap_fds)
     _IOCTL (vif->tap_fds[i], TUNSETOFFLOAD, offload);
   vif->gso_enabled = 0;
+  vif->packet_coalesce = 0;
   vif->csum_offload_enabled = enable_disable ? 1 : 0;
 
   if ((hw->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO) != 0)
@@ -862,7 +869,8 @@ error:
 }
 
 int
-tap_gso_enable_disable (vlib_main_t * vm, u32 sw_if_index, int enable_disable)
+tap_gso_enable_disable (vlib_main_t * vm, u32 sw_if_index, int enable_disable,
+			int is_packet_coalesce)
 {
   vnet_main_t *vnm = vnet_get_main ();
   virtio_main_t *mm = &virtio_main;
@@ -892,6 +900,11 @@ tap_gso_enable_disable (vlib_main_t * vm, u32 sw_if_index, int enable_disable)
 	  hw->flags |= VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO |
 	    VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD;
 	}
+      if (is_packet_coalesce)
+	{
+	  vif->packet_coalesce = 1;
+	  virtio_set_packet_coalesce (vif);
+	}
     }
   else
     {
@@ -900,6 +913,7 @@ tap_gso_enable_disable (vlib_main_t * vm, u32 sw_if_index, int enable_disable)
 	  hw->flags &= ~(VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO |
 			 VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD);
 	}
+      vif->packet_coalesce = 0;
     }
 
 error:
diff --git a/src/vnet/devices/tap/tap.h b/src/vnet/devices/tap/tap.h
index a9cd4766c3c..93b5da51dc5 100644
--- a/src/vnet/devices/tap/tap.h
+++ b/src/vnet/devices/tap/tap.h
@@ -101,7 +101,7 @@ typedef struct
 void tap_create_if (vlib_main_t * vm, tap_create_if_args_t * args);
 int tap_delete_if (vlib_main_t * vm, u32 sw_if_index);
 int tap_gso_enable_disable (vlib_main_t * vm, u32 sw_if_index,
-			    int enable_disable);
+			    int enable_disable, int packet_coalesce);
 int
 tap_csum_offload_enable_disable (vlib_main_t * vm, u32 sw_if_index,
 				 int enable_disable);
diff --git a/src/vnet/devices/virtio/device.c b/src/vnet/devices/virtio/device.c
index 45c390d415d..99157cd6166 100644
--- a/src/vnet/devices/virtio/device.c
+++ b/src/vnet/devices/virtio/device.c
@@ -23,6 +23,7 @@
 #include <vlib/unix/unix.h>
 #include <vnet/vnet.h>
 #include <vnet/ethernet/ethernet.h>
+#include <vnet/gso/gro_func.h>
 #include <vnet/gso/hdr_offset_parser.h>
 #include <vnet/ip/ip4_packet.h>
 #include <vnet/ip/ip6_packet.h>
@@ -502,7 +503,7 @@ static_always_inline uword
 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)
+				int csum_offload, int do_gro)
 {
   u16 n_left = frame->n_vectors;
   virtio_vring_t *vring;
@@ -513,6 +514,7 @@ virtio_interface_tx_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u16 mask = sz - 1;
   u16 retry_count = 2;
   u32 *buffers = vlib_frame_vector_args (frame);
+  u32 to[GRO_TO_VECTOR_SIZE (n_left)];
 
   clib_spinlock_lock_if_init (&vring->lockp);
 
@@ -520,6 +522,12 @@ virtio_interface_tx_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
       (vring->last_kick_avail_idx != vring->avail->idx))
     virtio_kick (vm, vring, vif);
 
+  if (do_gro)
+    {
+      n_left = vnet_gro_inline (vm, vring->flow_table, buffers, n_left, to);
+      buffers = to;
+    }
+
 retry:
   /* free consumed buffers */
   virtio_free_used_device_desc (vm, vring, node->node_index);
@@ -621,7 +629,8 @@ retry:
       if (retry_count--)
 	goto retry;
 
-      virtio_interface_drop_inline (vm, node->node_index, buffers, n_left,
+      virtio_interface_drop_inline (vm, node->node_index,
+				    buffers, n_left,
 				    VIRTIO_TX_ERROR_NO_FREE_SLOTS);
     }
 
@@ -641,15 +650,18 @@ virtio_interface_tx_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   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 */ );
+					   1 /* checksum offload */ ,
+					   vif->packet_coalesce);
   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 */ );
+					   1 /* checksum offload */ ,
+					   0 /* do_gro */ );
   else
     return virtio_interface_tx_gso_inline (vm, node, frame, vif, type,
 					   0 /* no do_gso */ ,
-					   0 /* no checksum offload */ );
+					   0 /* no checksum offload */ ,
+					   0 /* do_gro */ );
 }
 
 VNET_DEVICE_CLASS_TX_FN (virtio_device_class) (vlib_main_t * vm,
@@ -717,9 +729,16 @@ virtio_interface_rx_mode_change (vnet_main_t * vnm, u32 hw_if_index, u32 qid,
     }
 
   if (mode == VNET_HW_INTERFACE_RX_MODE_POLLING)
-    vring->avail->flags |= VIRTIO_RING_FLAG_MASK_INT;
+    {
+      /* only enable packet coalesce in poll mode */
+      gro_flow_table_set_is_enable (vring->flow_table, 1);
+      vring->avail->flags |= VIRTIO_RING_FLAG_MASK_INT;
+    }
   else
-    vring->avail->flags &= ~VIRTIO_RING_FLAG_MASK_INT;
+    {
+      gro_flow_table_set_is_enable (vring->flow_table, 0);
+      vring->avail->flags &= ~VIRTIO_RING_FLAG_MASK_INT;
+    }
 
   return 0;
 }
diff --git a/src/vnet/devices/virtio/node.c b/src/vnet/devices/virtio/node.c
index 63e40da53b7..a27131a6398 100644
--- a/src/vnet/devices/virtio/node.c
+++ b/src/vnet/devices/virtio/node.c
@@ -28,6 +28,7 @@
 #include <vnet/ethernet/ethernet.h>
 #include <vnet/devices/devices.h>
 #include <vnet/feature/feature.h>
+#include <vnet/gso/gro_func.h>
 #include <vnet/ip/ip4_packet.h>
 #include <vnet/ip/ip6_packet.h>
 #include <vnet/udp/udp_packet.h>
@@ -267,6 +268,8 @@ virtio_device_input_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 thread_index = vm->thread_index;
   uword n_trace = vlib_get_trace_count (vm, node);
   virtio_vring_t *vring = vec_elt_at_index (vif->rxq_vrings, qid);
+  u16 txq_id = thread_index % vif->num_txqs;
+  virtio_vring_t *txq_vring = vec_elt_at_index (vif->txq_vrings, txq_id);
   u32 next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
   const int hdr_sz = vif->virtio_net_hdr_sz;
   u32 *to_next = 0;
@@ -276,6 +279,12 @@ virtio_device_input_gso_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   u16 last = vring->last_used_idx;
   u16 n_left = vring->used->idx - last;
 
+  if (vif->packet_coalesce)
+    {
+      vnet_gro_flow_table_schedule_node_on_dispatcher (vm,
+						       txq_vring->flow_table);
+    }
+
   if ((vring->used->flags & VIRTIO_RING_FLAG_MASK_INT) == 0 &&
       vring->last_kick_avail_idx != vring->avail->idx)
     virtio_kick (vm, vring, vif);
diff --git a/src/vnet/devices/virtio/pci.c b/src/vnet/devices/virtio/pci.c
index cb3a3a73632..d28c7464e5a 100644
--- a/src/vnet/devices/virtio/pci.c
+++ b/src/vnet/devices/virtio/pci.c
@@ -824,6 +824,7 @@ virtio_pci_vring_init (vlib_main_t * vm, virtio_if_t * vif, u16 queue_num)
   vring->used = vr.used;
   vring->queue_id = queue_num;
   vring->avail->flags = VIRTIO_RING_FLAG_MASK_INT;
+  vring->flow_table = 0;
 
   ASSERT (vring->buffers == 0);
   vec_validate_aligned (vring->buffers, queue_size, CLIB_CACHE_LINE_BYTES);
diff --git a/src/vnet/devices/virtio/virtio.c b/src/vnet/devices/virtio/virtio.c
index 17fbd7920da..f8b6a39ea47 100644
--- a/src/vnet/devices/virtio/virtio.c
+++ b/src/vnet/devices/virtio/virtio.c
@@ -217,11 +217,26 @@ virtio_vring_free_tx (vlib_main_t * vm, virtio_if_t * vif, u32 idx)
   if (vring->avail)
     clib_mem_free (vring->avail);
   vec_free (vring->buffers);
+  gro_flow_table_free (vring->flow_table);
   clib_spinlock_free (&vring->lockp);
   return 0;
 }
 
 void
+virtio_set_packet_coalesce (virtio_if_t * vif)
+{
+  vnet_main_t *vnm = vnet_get_main ();
+  vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, vif->hw_if_index);
+  virtio_vring_t *vring;
+  vec_foreach (vring, vif->txq_vrings)
+  {
+    gro_flow_table_init (&vring->flow_table,
+			 vif->type & (VIRTIO_IF_TYPE_TAP |
+				      VIRTIO_IF_TYPE_PCI), hw->tx_node_index);
+  }
+}
+
+void
 virtio_vring_set_numa_node (vlib_main_t * vm, virtio_if_t * vif, u32 idx)
 {
   vnet_main_t *vnm = vnet_get_main ();
@@ -320,6 +335,7 @@ virtio_show (vlib_main_t * vm, u32 * hw_if_indices, u8 show_descr, u32 type)
 	}
       vlib_cli_output (vm, "  gso-enabled %d", vif->gso_enabled);
       vlib_cli_output (vm, "  csum-enabled %d", vif->csum_offload_enabled);
+      vlib_cli_output (vm, "  packet-coalesce %d", vif->packet_coalesce);
       if (type & (VIRTIO_IF_TYPE_TAP | VIRTIO_IF_TYPE_PCI))
 	vlib_cli_output (vm, "  Mac Address: %U", format_ethernet_address,
 			 vif->mac_addr);
@@ -412,6 +428,11 @@ virtio_show (vlib_main_t * vm, u32 * hw_if_indices, u8 show_descr, u32 type)
 	    vlib_cli_output (vm, "    kickfd %d, callfd %d", vring->kick_fd,
 			     vring->call_fd);
 	  }
+	if (vring->flow_table)
+	  {
+	    vlib_cli_output (vm, "    %U", gro_flow_table_format,
+			     vring->flow_table);
+	  }
 	if (show_descr)
 	  {
 	    vlib_cli_output (vm, "\n  descriptor table:\n");
diff --git a/src/vnet/devices/virtio/virtio.h b/src/vnet/devices/virtio/virtio.h
index 04896ed6e2a..b57c6ebd77d 100644
--- a/src/vnet/devices/virtio/virtio.h
+++ b/src/vnet/devices/virtio/virtio.h
@@ -22,6 +22,7 @@
 #include <linux/virtio_net.h>
 #include <linux/virtio_pci.h>
 #include <linux/virtio_ring.h>
+#include <vnet/gso/gro.h>
 
 #define foreach_virtio_net_features      \
   _ (VIRTIO_NET_F_CSUM, 0)	/* Host handles pkts w/ partial csum */ \
@@ -123,6 +124,7 @@ typedef struct
   u16 last_used_idx;
   u16 last_kick_avail_idx;
   u32 call_file_index;
+  gro_flow_table_t *flow_table;
 } virtio_vring_t;
 
 typedef union
@@ -165,6 +167,7 @@ typedef struct
   u32 sw_if_index;
 
     CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
+  int packet_coalesce;
   union
   {
     u32 id;
@@ -221,6 +224,7 @@ extern void virtio_free_rx_buffers (vlib_main_t * vm, virtio_vring_t * vring);
 extern void virtio_set_net_hdr_size (virtio_if_t * vif);
 extern void virtio_show (vlib_main_t * vm, u32 * hw_if_indices, u8 show_descr,
 			 u32 type);
+extern void virtio_set_packet_coalesce (virtio_if_t * vif);
 extern void virtio_pci_legacy_notify_queue (vlib_main_t * vm,
 					    virtio_if_t * vif, u16 queue_id);
 format_function_t format_virtio_device_name;