net/mlx4_en: Improve stack xmit function
authorTariq Toukan <tariqt@mellanox.com>
Thu, 15 Jun 2017 11:35:35 +0000 (14:35 +0300)
committerDavid S. Miller <davem@davemloft.net>
Fri, 16 Jun 2017 02:53:23 +0000 (22:53 -0400)
Several small code and performance improvements in stack TX datapath,
including:
- Compiler branch predictor hints.
- Minimize variables scope.
- Move tx_info non-inline flow handling to a separate function.
- Calculate data_offset in compile time rather than in runtime
  (for !lso_header_size branch).
- Avoid trinary-operator ("?") when value can be preset in a matching
  branch.

Performance tests:
Tested on ConnectX3Pro, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz

Gain is too small to be measurable, no degradation sensed.
Results are similar for IPv4 and IPv6.

Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Cc: kernel-team@fb.com
Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx4/en_tx.c

index 2d5e0da1de2fda5e1f73ace30c7bc648b176081a..58f4b322587bd87df907beaf3cd7b73e426e9cfe 100644 (file)
@@ -774,37 +774,101 @@ static void mlx4_en_tx_write_desc(struct mlx4_en_tx_ring *ring,
        }
 }
 
+static bool mlx4_en_build_dma_wqe(struct mlx4_en_priv *priv,
+                                 struct skb_shared_info *shinfo,
+                                 struct mlx4_wqe_data_seg *data,
+                                 struct sk_buff *skb,
+                                 int lso_header_size,
+                                 __be32 mr_key,
+                                 struct mlx4_en_tx_info *tx_info)
+{
+       struct device *ddev = priv->ddev;
+       dma_addr_t dma = 0;
+       u32 byte_count = 0;
+       int i_frag;
+
+       /* Map fragments if any */
+       for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
+               const struct skb_frag_struct *frag;
+
+               frag = &shinfo->frags[i_frag];
+               byte_count = skb_frag_size(frag);
+               dma = skb_frag_dma_map(ddev, frag,
+                                      0, byte_count,
+                                      DMA_TO_DEVICE);
+               if (dma_mapping_error(ddev, dma))
+                       goto tx_drop_unmap;
+
+               data->addr = cpu_to_be64(dma);
+               data->lkey = mr_key;
+               dma_wmb();
+               data->byte_count = cpu_to_be32(byte_count);
+               --data;
+       }
+
+       /* Map linear part if needed */
+       if (tx_info->linear) {
+               byte_count = skb_headlen(skb) - lso_header_size;
+
+               dma = dma_map_single(ddev, skb->data +
+                                    lso_header_size, byte_count,
+                                    PCI_DMA_TODEVICE);
+               if (dma_mapping_error(ddev, dma))
+                       goto tx_drop_unmap;
+
+               data->addr = cpu_to_be64(dma);
+               data->lkey = mr_key;
+               dma_wmb();
+               data->byte_count = cpu_to_be32(byte_count);
+       }
+       /* tx completion can avoid cache line miss for common cases */
+       tx_info->map0_dma = dma;
+       tx_info->map0_byte_count = byte_count;
+
+       return true;
+
+tx_drop_unmap:
+       en_err(priv, "DMA mapping error\n");
+
+       while (++i_frag < shinfo->nr_frags) {
+               ++data;
+               dma_unmap_page(ddev, (dma_addr_t)be64_to_cpu(data->addr),
+                              be32_to_cpu(data->byte_count),
+                              PCI_DMA_TODEVICE);
+       }
+
+       return false;
+}
+
 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 {
        struct skb_shared_info *shinfo = skb_shinfo(skb);
        struct mlx4_en_priv *priv = netdev_priv(dev);
        union mlx4_wqe_qpn_vlan qpn_vlan = {};
-       struct device *ddev = priv->ddev;
        struct mlx4_en_tx_ring *ring;
        struct mlx4_en_tx_desc *tx_desc;
        struct mlx4_wqe_data_seg *data;
        struct mlx4_en_tx_info *tx_info;
-       int tx_ind = 0;
+       int tx_ind;
        int nr_txbb;
        int desc_size;
        int real_size;
        u32 index, bf_index;
        __be32 op_own;
-       u16 vlan_proto = 0;
-       int i_frag;
        int lso_header_size;
        void *fragptr = NULL;
        bool bounce = false;
        bool send_doorbell;
        bool stop_queue;
        bool inline_ok;
+       u8 data_offset;
        u32 ring_cons;
        bool bf_ok;
 
        tx_ind = skb_get_queue_mapping(skb);
        ring = priv->tx_ring[TX][tx_ind];
 
-       if (!priv->port_up)
+       if (unlikely(!priv->port_up))
                goto tx_drop;
 
        /* fetch ring->cons far ahead before needing it to avoid stall */
@@ -826,6 +890,8 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 
        bf_ok = ring->bf_enabled;
        if (skb_vlan_tag_present(skb)) {
+               u16 vlan_proto;
+
                qpn_vlan.vlan_tag = cpu_to_be16(skb_vlan_tag_get(skb));
                vlan_proto = be16_to_cpu(skb->vlan_proto);
                if (vlan_proto == ETH_P_8021AD)
@@ -862,64 +928,31 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
        tx_info->skb = skb;
        tx_info->nr_txbb = nr_txbb;
 
-       data = &tx_desc->data;
-       if (lso_header_size)
-               data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
-                                                     DS_SIZE));
+       if (!lso_header_size) {
+               data = &tx_desc->data;
+               data_offset = offsetof(struct mlx4_en_tx_desc, data);
+       } else {
+               int lso_align = ALIGN(lso_header_size + 4, DS_SIZE);
+
+               data = (void *)&tx_desc->lso + lso_align;
+               data_offset = offsetof(struct mlx4_en_tx_desc, lso) + lso_align;
+       }
 
        /* valid only for none inline segments */
-       tx_info->data_offset = (void *)data - (void *)tx_desc;
+       tx_info->data_offset = data_offset;
 
        tx_info->inl = inline_ok;
 
-       tx_info->linear = (lso_header_size < skb_headlen(skb) &&
-                          !inline_ok) ? 1 : 0;
+       tx_info->linear = lso_header_size < skb_headlen(skb) && !inline_ok;
 
        tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
        data += tx_info->nr_maps - 1;
 
-       if (!tx_info->inl) {
-               dma_addr_t dma = 0;
-               u32 byte_count = 0;
-
-               /* Map fragments if any */
-               for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
-                       const struct skb_frag_struct *frag;
-
-                       frag = &shinfo->frags[i_frag];
-                       byte_count = skb_frag_size(frag);
-                       dma = skb_frag_dma_map(ddev, frag,
-                                              0, byte_count,
-                                              DMA_TO_DEVICE);
-                       if (dma_mapping_error(ddev, dma))
-                               goto tx_drop_unmap;
-
-                       data->addr = cpu_to_be64(dma);
-                       data->lkey = ring->mr_key;
-                       dma_wmb();
-                       data->byte_count = cpu_to_be32(byte_count);
-                       --data;
-               }
-
-               /* Map linear part if needed */
-               if (tx_info->linear) {
-                       byte_count = skb_headlen(skb) - lso_header_size;
-
-                       dma = dma_map_single(ddev, skb->data +
-                                            lso_header_size, byte_count,
-                                            PCI_DMA_TODEVICE);
-                       if (dma_mapping_error(ddev, dma))
-                               goto tx_drop_unmap;
-
-                       data->addr = cpu_to_be64(dma);
-                       data->lkey = ring->mr_key;
-                       dma_wmb();
-                       data->byte_count = cpu_to_be32(byte_count);
-               }
-               /* tx completion can avoid cache line miss for common cases */
-               tx_info->map0_dma = dma;
-               tx_info->map0_byte_count = byte_count;
-       }
+       if (!tx_info->inl)
+               if (!mlx4_en_build_dma_wqe(priv, shinfo, data, skb,
+                                          lso_header_size, ring->mr_key,
+                                          tx_info))
+                       goto tx_drop_count;
 
        /*
         * For timestamping add flag to skb_shinfo and
@@ -1055,16 +1088,6 @@ netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
        }
        return NETDEV_TX_OK;
 
-tx_drop_unmap:
-       en_err(priv, "DMA mapping error\n");
-
-       while (++i_frag < shinfo->nr_frags) {
-               ++data;
-               dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
-                              be32_to_cpu(data->byte_count),
-                              PCI_DMA_TODEVICE);
-       }
-
 tx_drop_count:
        ring->tx_dropped++;
 tx_drop: