ipv6_gro_receive() compares 34 bytes using slow memcmp(), while handcoding with a couple of ipv6_addr_equal() is much faster.
Before this patch, "perf top -e cycles:pp -C <cpu>" would see memcmp() using ~10% of cpu cycles on a 40Gbit NIC receiving IPv6 TCP traffic. Signed-off-by: Eric Dumazet <[email protected]> --- net/ipv6/ip6_offload.c | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/net/ipv6/ip6_offload.c b/net/ipv6/ip6_offload.c index c7e495f1201105f1ac1724a7b8fd82399efcce32..70f525c33cb6c1f375919b94a7afc45cc6bdcd5f 100644 --- a/net/ipv6/ip6_offload.c +++ b/net/ipv6/ip6_offload.c @@ -229,14 +229,21 @@ static struct sk_buff *ipv6_gro_receive(struct list_head *head, * XXX skbs on the gro_list have all been parsed and pulled * already so we don't need to compare nlen * (nlen != (sizeof(*iph2) + ipv6_exthdrs_len(iph2, &ops))) - * memcmp() alone below is suffcient, right? + * memcmp() alone below is sufficient, right? */ if ((first_word & htonl(0xF00FFFFF)) || - memcmp(&iph->nexthdr, &iph2->nexthdr, - nlen - offsetof(struct ipv6hdr, nexthdr))) { + !ipv6_addr_equal(&iph->saddr, &iph2->saddr) || + !ipv6_addr_equal(&iph->daddr, &iph2->daddr) || + *(u16 *)&iph->nexthdr != *(u16 *)&iph2->nexthdr) { +not_same_flow: NAPI_GRO_CB(p)->same_flow = 0; continue; } + if (unlikely(nlen > sizeof(struct ipv6hdr))) { + if (memcmp(iph + 1, iph2 + 1, + nlen - sizeof(struct ipv6hdr))) + goto not_same_flow; + } /* flush if Traffic Class fields are different */ NAPI_GRO_CB(p)->flush |= !!(first_word & htonl(0x0FF00000)); NAPI_GRO_CB(p)->flush |= flush; -- 2.19.1.930.g4563a0d9d0-goog
