gro: __napi_gro_receive() optimizations

compare_ether_header() can have a special implementation on 64 bit
arches if CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is defined.

__napi_gro_receive() and vlan_gro_common() can avoid a conditional
branch to perform device match.

On x86_64, __napi_gro_receive() has now 38 instructions instead of 53

As gcc-4.4.3 still choose to not inline it, add inline keyword to this
performance critical function.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/etherdevice.h

@@ -237,13 +237,29 @@ static inline bool is_etherdev_addr(const struct net_device *dev,
  * entry points.
  */
 
-static inline int compare_ether_header(const void *a, const void *b)
+static inline unsigned long compare_ether_header(const void *a, const void *b)
 {
+#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
+	unsigned long fold;
+
+	/*
+	 * We want to compare 14 bytes:
+	 *  [a0 ... a13] ^ [b0 ... b13]
+	 * Use two long XOR, ORed together, with an overlap of two bytes.
+	 *  [a0  a1  a2  a3  a4  a5  a6  a7 ] ^ [b0  b1  b2  b3  b4  b5  b6  b7 ] |
+	 *  [a6  a7  a8  a9  a10 a11 a12 a13] ^ [b6  b7  b8  b9  b10 b11 b12 b13]
+	 * This means the [a6 a7] ^ [b6 b7] part is done two times.
+	*/
+	fold = *(unsigned long *)a ^ *(unsigned long *)b;
+	fold |= *(unsigned long *)(a + 6) ^ *(unsigned long *)(b + 6);
+	return fold;
+#else
 	u32 *a32 = (u32 *)((u8 *)a + 2);
 	u32 *b32 = (u32 *)((u8 *)b + 2);
 
 	return (*(u16 *)a ^ *(u16 *)b) | (a32[0] ^ b32[0]) |
 	       (a32[1] ^ b32[1]) | (a32[2] ^ b32[2]);
+#endif
 }
 
 #endif	/* _LINUX_ETHERDEVICE_H */

net/8021q/vlan_core.c

@@ -105,9 +105,12 @@ vlan_gro_common(struct napi_struct *napi, struct vlan_group *grp,
 		goto drop;
 
 	for (p = napi->gro_list; p; p = p->next) {
-		NAPI_GRO_CB(p)->same_flow =
-			p->dev == skb->dev && !compare_ether_header(
-				skb_mac_header(p), skb_gro_mac_header(skb));
+		unsigned long diffs;
+
+		diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
+		diffs |= compare_ether_header(skb_mac_header(p),
+					      skb_gro_mac_header(skb));
+		NAPI_GRO_CB(p)->same_flow = !diffs;
 		NAPI_GRO_CB(p)->flush = 0;
 	}
 

net/core/dev.c

@@ -3169,16 +3169,18 @@ enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 }
 EXPORT_SYMBOL(dev_gro_receive);
 
-static gro_result_t
+static inline gro_result_t
 __napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
 {
 	struct sk_buff *p;
 
 	for (p = napi->gro_list; p; p = p->next) {
-		NAPI_GRO_CB(p)->same_flow =
-			(p->dev == skb->dev) &&
-			!compare_ether_header(skb_mac_header(p),
+		unsigned long diffs;
+
+		diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
+		diffs |= compare_ether_header(skb_mac_header(p),
 					      skb_gro_mac_header(skb));
+		NAPI_GRO_CB(p)->same_flow = !diffs;
 		NAPI_GRO_CB(p)->flush = 0;
 	}