avf: add TSO support

Change-Id: Ica663e1d07225acf53fd74b0051a5a2a34174960
Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
Type: improvement

3 files changed, 120 insertions, 21 deletions

diff --git a/src/plugins/avf/avf.h b/src/plugins/avf/avf.h
index 66a791ec9b2..025fa6ea4e9 100644
--- a/src/plugins/avf/avf.h
+++ b/src/plugins/avf/avf.h
@@ -20,7 +20,14 @@
 
 #include <avf/virtchnl.h>
 
+#include <vppinfra/types.h>
+#include <vppinfra/error_bootstrap.h>
+#include <vppinfra/lock.h>
+
 #include <vlib/log.h>
+#include <vlib/pci/pci.h>
+
+#include <vnet/interface.h>
 
 #define AVF_QUEUE_SZ_MAX                4096
 #define AVF_QUEUE_SZ_MIN                64
@@ -66,6 +73,13 @@
 #define AVF_TXD_OFFSET_IPLEN(val)       AVF_TXD_OFFSET( 7, 4, val)
 #define AVF_TXD_OFFSET_L4LEN(val)       AVF_TXD_OFFSET(14, 4, val)
 
+#define AVF_TXD_DTYP_CTX                0x1ULL
+#define AVF_TXD_CTX_CMD_TSO             AVF_TXD_CMD(0)
+#define AVF_TXD_CTX_SEG(val,x)          (((u64)val) << (30 + x))
+#define AVF_TXD_CTX_SEG_TLEN(val)       AVF_TXD_CTX_SEG(val,0)
+#define AVF_TXD_CTX_SEG_MSS(val)        AVF_TXD_CTX_SEG(val,20)
+
+
 extern vlib_log_class_registration_t avf_log;
 
 #define avf_log_err(dev, f, ...)                        \
@@ -158,6 +172,7 @@ typedef struct
   volatile u32 *qtx_tail;
   u16 next;
   u16 size;
+  u32 ctx_desc_placeholder_bi;
   clib_spinlock_t lock;
   avf_tx_desc_t *descs;
   u32 *bufs;
diff --git a/src/plugins/avf/device.c b/src/plugins/avf/device.c
index bb2dcf74e07..c6bf2df4443 100644
--- a/src/plugins/avf/device.c
+++ b/src/plugins/avf/device.c
@@ -298,6 +298,8 @@ avf_txq_init (vlib_main_t * vm, avf_device_t * ad, u16 qid, u16 txq_size)
 {
   clib_error_t *err;
   avf_txq_t *txq;
+  u8 bpi = vlib_buffer_pool_get_default_for_numa (vm,
+						  ad->numa_node);
 
   if (qid >= ad->num_queue_pairs)
     {
@@ -313,6 +315,14 @@ avf_txq_init (vlib_main_t * vm, avf_device_t * ad, u16 qid, u16 txq_size)
   txq = vec_elt_at_index (ad->txqs, qid);
   txq->size = txq_size;
   txq->next = 0;
+
+  /* Prepare a placeholder buffer to maintain a 1-1
+     relationship between bufs and descs when a context
+     descriptor is added in descs */
+  if (!vlib_buffer_alloc_from_pool
+      (vm, &txq->ctx_desc_placeholder_bi, 1, bpi))
+    return clib_error_return (0, "buffer allocation error");
+
   txq->descs = vlib_physmem_alloc_aligned_on_numa (vm, txq->size *
 						   sizeof (avf_tx_desc_t),
 						   2 * CLIB_CACHE_LINE_BYTES,
@@ -1442,6 +1452,8 @@ avf_delete_if (vlib_main_t * vm, avf_device_t * ad, int with_barrier)
 	  vlib_buffer_free_from_ring (vm, txq->bufs, first, txq->size,
 				      txq->n_enqueued);
 	}
+      /* Free the placeholder buffer */
+      vlib_buffer_free_one(vm, txq->ctx_desc_placeholder_bi);
       vec_free (txq->bufs);
       clib_ring_free (txq->rs_slots);
     }
@@ -1638,7 +1650,8 @@ avf_create_if (vlib_main_t * vm, avf_create_if_args_t * args)
   vnet_hw_interface_t *hi = vnet_get_hw_interface (vnm, ad->hw_if_index);
   hi->flags |=
     VNET_HW_INTERFACE_FLAG_SUPPORTS_MAC_FILTER |
-    VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD;
+    VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD |
+    VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO;
   ethernet_set_flags (vnm, ad->hw_if_index,
 		      ETHERNET_INTERFACE_FLAG_DEFAULT_L3);
 
diff --git a/src/plugins/avf/output.c b/src/plugins/avf/output.c
index b1b5bd005ff..90cc56fc7d8 100644
--- a/src/plugins/avf/output.c
+++ b/src/plugins/avf/output.c
@@ -88,6 +88,14 @@ avf_tx_prepare_cksum (vlib_buffer_t * b, u8 is_tso)
   if (is_ip4)
     ip4->checksum = 0;
 
+  if (is_tso)
+    {
+      if (is_ip4)
+	ip4->length = 0;
+      else
+	ip6->payload_length = 0;
+    }
+
   if (is_tcp || is_udp)
     {
       if (is_ip4)
@@ -120,13 +128,56 @@ avf_tx_prepare_cksum (vlib_buffer_t * b, u8 is_tso)
   return flags;
 }
 
+static_always_inline int
+avf_tx_fill_ctx_desc (vlib_main_t * vm, avf_txq_t * txq, avf_tx_desc_t * d,
+		      vlib_buffer_t * b)
+{
+  vlib_buffer_t *ctx_ph = vlib_get_buffer (vm, txq->ctx_desc_placeholder_bi);
+  if (PREDICT_FALSE (ctx_ph->ref_count == 255))
+    {
+      /* We need a new placeholder buffer */
+      u32 new_bi;
+      u8 bpi = vlib_buffer_pool_get_default_for_numa (vm, vm->numa_node);
+      if (PREDICT_TRUE
+	  (vlib_buffer_alloc_from_pool (vm, &new_bi, 1, bpi) == 1))
+	{
+	  /* Remove our own reference on the current placeholder buffer */
+	  ctx_ph->ref_count--;
+	  /* Replace with the new placeholder buffer */
+	  txq->ctx_desc_placeholder_bi = new_bi;
+	  ctx_ph = vlib_get_buffer (vm, new_bi);
+	}
+      else
+	/* Impossible to enqueue a ctx descriptor, fail */
+	return 1;
+    }
+
+  /* Acquire a reference on the placeholder buffer */
+  ctx_ph->ref_count++;
+
+  u16 l234hdr_sz =
+    vnet_buffer (b)->l4_hdr_offset -
+    vnet_buffer (b)->l2_hdr_offset + vnet_buffer2 (b)->gso_l4_hdr_sz;
+  u16 tlen = vlib_buffer_length_in_chain (vm, b) - l234hdr_sz;
+  d[0].qword[0] = 0;
+  d[0].qword[1] = AVF_TXD_DTYP_CTX | AVF_TXD_CTX_CMD_TSO
+    | AVF_TXD_CTX_SEG_MSS (vnet_buffer2 (b)->gso_size) |
+    AVF_TXD_CTX_SEG_TLEN (tlen);
+  return 0;
+}
+
+
 static_always_inline u16
 avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 		u32 * buffers, u32 n_packets, int use_va_dma)
 {
   u16 next = txq->next;
   u64 bits = AVF_TXD_CMD_EOP | AVF_TXD_CMD_RSV;
+  const u32 offload_mask = VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
+    VNET_BUFFER_F_OFFLOAD_TCP_CKSUM | VNET_BUFFER_F_OFFLOAD_UDP_CKSUM |
+    VNET_BUFFER_F_GSO;
   u64 one_by_one_offload_flags = 0;
+  int is_tso;
   u16 n_desc = 0;
   u16 *slot, n_desc_left, n_packets_left = n_packets;
   u16 mask = txq->size - 1;
@@ -160,9 +211,7 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 
       or_flags = b[0]->flags | b[1]->flags | b[2]->flags | b[3]->flags;
 
-      if (or_flags &
-	  (VLIB_BUFFER_NEXT_PRESENT | VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
-	   VNET_BUFFER_F_OFFLOAD_TCP_CKSUM | VNET_BUFFER_F_OFFLOAD_UDP_CKSUM))
+      if (or_flags & (VLIB_BUFFER_NEXT_PRESENT | offload_mask))
 	goto one_by_one;
 
       vlib_buffer_copy_indices (txq->bufs + next, buffers, 4);
@@ -199,17 +248,14 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
       one_by_one_offload_flags = 0;
       txq->bufs[next] = buffers[0];
       b[0] = vlib_get_buffer (vm, buffers[0]);
-
-      if (PREDICT_FALSE (b[0]->flags & (VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
-					VNET_BUFFER_F_OFFLOAD_TCP_CKSUM |
-					VNET_BUFFER_F_OFFLOAD_UDP_CKSUM)))
-	one_by_one_offload_flags |=
-	  avf_tx_prepare_cksum (b[0], 0 /* no TSO */ );
+      is_tso = ! !(b[0]->flags & VNET_BUFFER_F_GSO);
+      if (PREDICT_FALSE (is_tso || b[0]->flags & offload_mask))
+	one_by_one_offload_flags |= avf_tx_prepare_cksum (b[0], is_tso);
 
       /* Deal with chain buffer if present */
-      if (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
+      if (is_tso || b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
 	{
-	  n_desc_needed = 1;
+	  n_desc_needed = 1 + is_tso;
 	  b0 = b[0];
 
 	  /* Wish there were a buffer count for chain buffer */
@@ -220,7 +266,7 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 	    }
 
 	  /* spec says data descriptor is limited to 8 segments */
-	  if (PREDICT_FALSE (n_desc_needed > 8))
+	  if (PREDICT_FALSE (!is_tso && n_desc_needed > 8))
 	    {
 	      vlib_buffer_free_one (vm, buffers[0]);
 	      vlib_error_count (vm, node->node_index,
@@ -237,6 +283,19 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 	     */
 	    break;
 
+	  /* Enqueue a context descriptor if needed */
+	  if (PREDICT_FALSE (is_tso))
+	    {
+	      if (avf_tx_fill_ctx_desc (vm, txq, d, b[0]))
+		/* Failure to acquire ref on ctx placeholder */
+		break;
+	      txq->bufs[next + 1] = txq->bufs[next];
+	      txq->bufs[next] = txq->ctx_desc_placeholder_bi;
+	      next += 1;
+	      n_desc += 1;
+	      n_desc_left -= 1;
+	      d += 1;
+	    }
 	  while (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
 	    {
 	      if (use_va_dma)
@@ -292,16 +351,14 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 	  b[0] = vlib_get_buffer (vm, buffers[0]);
 
 	  one_by_one_offload_flags = 0;
-	  if (PREDICT_FALSE (b[0]->flags & (VNET_BUFFER_F_OFFLOAD_IP_CKSUM |
-					    VNET_BUFFER_F_OFFLOAD_TCP_CKSUM |
-					    VNET_BUFFER_F_OFFLOAD_UDP_CKSUM)))
-	    one_by_one_offload_flags |=
-	      avf_tx_prepare_cksum (b[0], 0 /* no TSO */ );
+	  is_tso = ! !(b[0]->flags & VNET_BUFFER_F_GSO);
+	  if (PREDICT_FALSE (is_tso || b[0]->flags & offload_mask))
+	    one_by_one_offload_flags |= avf_tx_prepare_cksum (b[0], is_tso);
 
 	  /* Deal with chain buffer if present */
-	  if (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
+	  if (is_tso || b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
 	    {
-	      n_desc_needed = 1;
+	      n_desc_needed = 1 + is_tso;
 	      b0 = b[0];
 
 	      while (b0->flags & VLIB_BUFFER_NEXT_PRESENT)
@@ -311,7 +368,7 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 		}
 
 	      /* Spec says data descriptor is limited to 8 segments */
-	      if (PREDICT_FALSE (n_desc_needed > 8))
+	      if (PREDICT_FALSE (!is_tso && n_desc_needed > 8))
 		{
 		  vlib_buffer_free_one (vm, buffers[0]);
 		  vlib_error_count (vm, node->node_index,
@@ -324,6 +381,20 @@ avf_tx_enqueue (vlib_main_t * vm, vlib_node_runtime_t * node, avf_txq_t * txq,
 	      if (PREDICT_FALSE (n_desc_left < n_desc_needed))
 		break;
 
+	      /* Enqueue a context descriptor if needed */
+	      if (PREDICT_FALSE (is_tso))
+		{
+		  if (avf_tx_fill_ctx_desc (vm, txq, d, b[0]))
+		    /* Failure to acquire ref on ctx placeholder */
+		    break;
+
+		  txq->bufs[(next + 1) & mask] = txq->bufs[next & mask];
+		  txq->bufs[next & mask] = txq->ctx_desc_placeholder_bi;
+		  next += 1;
+		  n_desc += 1;
+		  n_desc_left -= 1;
+		  d = txq->descs + (next & mask);
+		}
 	      while (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
 		{
 		  if (use_va_dma)