Hello, On Wed, 30 Sep 2020, longguang.yue wrote: > It's ipvs's duty to do traffic statistic if packets get hit, > no matter what mode it is. > > Signed-off-by: longguang.yue <bigclouds@xxxxxxx> > --- > net/netfilter/ipvs/ip_vs_conn.c | 14 ++++++++++++-- > net/netfilter/ipvs/ip_vs_core.c | 5 ++++- > 2 files changed, 16 insertions(+), 3 deletions(-) > > diff --git a/net/netfilter/ipvs/ip_vs_conn.c b/net/netfilter/ipvs/ip_vs_conn.c > index a90b8eac16ac..c4d164ce8ca7 100644 > --- a/net/netfilter/ipvs/ip_vs_conn.c > +++ b/net/netfilter/ipvs/ip_vs_conn.c > @@ -401,6 +401,8 @@ struct ip_vs_conn *ip_vs_ct_in_get(const struct ip_vs_conn_param *p) > struct ip_vs_conn *ip_vs_conn_out_get(const struct ip_vs_conn_param *p) > { > unsigned int hash; > + __be16 cport; > + const union nf_inet_addr *caddr; > struct ip_vs_conn *cp, *ret=NULL; May be we can do it in few rounds, here is a list of some initial notes... caddr/cport are misleading, can be saddr/sport (source) or laddr/lport (local). > /* > @@ -411,10 +413,18 @@ struct ip_vs_conn *ip_vs_conn_out_get(const struct ip_vs_conn_param *p) > rcu_read_lock(); > > hlist_for_each_entry_rcu(cp, &ip_vs_conn_tab[hash], c_list) { Lets first check here for cp->cport before touching more cache lines while traversing the list, it eliminates the cost of next checks: if (p->vport != cp->cport) continue; then if (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) { ... > - if (p->vport == cp->cport && p->cport == cp->dport && > + cport = cp->dport; > + caddr = &cp->daddr; > + > + if (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) { > + cport = cp->vport; > + caddr = &cp->vaddr; > + } Considering the issues solved by commit 3c5ab3f395d6, such check more correctly matches the replies from DR/TUN real server to local clients but also to remote clients if director is used as router. > + > + if (p->vport == cp->cport && p->cport == cport && if (p->cport == sport && ... > cp->af == p->af && > ip_vs_addr_equal(p->af, p->vaddr, &cp->caddr) && > - ip_vs_addr_equal(p->af, p->caddr, &cp->daddr) && > + ip_vs_addr_equal(p->af, p->caddr, caddr) && > p->protocol == cp->protocol && > cp->ipvs == p->ipvs) { > if (!__ip_vs_conn_get(cp)) > diff --git a/net/netfilter/ipvs/ip_vs_core.c b/net/netfilter/ipvs/ip_vs_core.c > index e3668a6e54e4..7ba88dab297a 100644 > --- a/net/netfilter/ipvs/ip_vs_core.c > +++ b/net/netfilter/ipvs/ip_vs_core.c > @@ -1413,8 +1413,11 @@ ip_vs_out(struct netns_ipvs *ipvs, unsigned int hooknum, struct sk_buff *skb, in > ipvs, af, skb, &iph); > > if (likely(cp)) { > - if (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) > + if (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) { > + ip_vs_out_stats(cp, skb); > + skb->ipvs_property = 1; We will also need: if (!(cp->flags & IP_VS_CONN_F_NFCT)) ip_vs_notrack(skb); Similar code is needed in handle_response_icmp(), so that we account ICMP packets, where a jump to new label before ip_vs_out_stats() can work. But such jump is preferred even for handle_response() because the (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) check should be moved from ip_vs_out() into handle_response(). For this to work the ip_vs_set_state() call in handle_response() should be moved before the new label and ip_vs_out_stats() call. > goto ignore_cp; > + } > return handle_response(af, skb, pd, cp, &iph, hooknum); > } Regards -- Julian Anastasov <ja@xxxxxx>