net: skb_shared_info optimization

skb_dma_unmap() is quite expensive for small packets,
because we use two different cache lines from skb_shared_info.

One to access nr_frags, one to access dma_maps[0]

Instead of dma_maps being an array of MAX_SKB_FRAGS + 1 elements,
let dma_head alone in a new dma_head field, close to nr_frags,
to reduce cache lines misses.

Tested on my dev machine (bnx2 & tg3 adapters), nice speedup !

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 3 insertions, 3 deletions

diff --git a/drivers/net/bnx2.c b/drivers/net/bnx2.c
index f53017250e09..391d2d47089c 100644
--- a/drivers/net/bnx2.c
+++ b/drivers/net/bnx2.c
@@ -5487,7 +5487,7 @@ bnx2_run_loopback(struct bnx2 *bp, int loopback_mode)
                 dev_kfree_skb(skb);
                 return -EIO;
         }
-        map = skb_shinfo(skb)->dma_maps[0];
+        map = skb_shinfo(skb)->dma_head;
 
         REG_WR(bp, BNX2_HC_COMMAND,
                bp->hc_cmd | BNX2_HC_COMMAND_COAL_NOW_WO_INT);
@@ -6167,7 +6167,7 @@ bnx2_start_xmit(struct sk_buff *skb, struct net_device *dev)
         }
 
         sp = skb_shinfo(skb);
-        mapping = sp->dma_maps[0];
+        mapping = sp->dma_head;
 
         tx_buf = &txr->tx_buf_ring[ring_prod];
         tx_buf->skb = skb;
@@ -6191,7 +6191,7 @@ bnx2_start_xmit(struct sk_buff *skb, struct net_device *dev)
                 txbd = &txr->tx_desc_ring[ring_prod];
 
                 len = frag->size;
-                mapping = sp->dma_maps[i + 1];
+                mapping = sp->dma_maps[i];
 
                 txbd->tx_bd_haddr_hi = (u64) mapping >> 32;
                 txbd->tx_bd_haddr_lo = (u64) mapping & 0xffffffff;