gro: __napi_gro_receive() optimizations
Eric Dumazet [Fri, 27 Aug 2010 05:03:08 +0000 (22:03 -0700)]
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
net/8021q/vlan_core.c
net/core/dev.c

index 2308fbb..fb6aa60 100644 (file)
@@ -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 */
index 07eeb5b..3438c01 100644 (file)
@@ -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;
        }
 
index 859e30f..63bd20a 100644 (file)
@@ -3169,16 +3169,18 @@ normal:
 }
 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;
        }