ipvs: allow connection reuse for unconfirmed conntrack

[ Upstream commit f0a5e4d7a594e0fe237d3dfafb069bb82f80f42f ]

YangYuxi is reporting that connection reuse
is causing one-second delay when SYN hits
existing connection in TIME_WAIT state.
Such delay was added to give time to expire
both the IPVS connection and the corresponding
conntrack. This was considered a rare case
at that time but it is causing problem for
some environments such as Kubernetes.

As nf_conntrack_tcp_packet() can decide to
release the conntrack in TIME_WAIT state and
to replace it with a fresh NEW conntrack, we
can use this to allow rescheduling just by
tuning our check: if the conntrack is
confirmed we can not schedule it to different
real server and the one-second delay still
applies but if new conntrack was created,
we are free to select new real server without
any delays.

YangYuxi lists some of the problem reports:

- One second connection delay in masquerading mode:
https://marc.info/?t=151683118100004&r=1&w=2

- IPVS low throughput #70747
kubernetes/kubernetes#70747

- Apache Bench can fill up ipvs service proxy in seconds #544
cloudnativelabs/kube-router#544

- Additional 1s latency in `host -> service IP -> pod`
kubernetes/kubernetes#90854

Fixes: f719e37 ("ipvs: drop first packet to redirect conntrack")
Co-developed-by: YangYuxi <[email protected]>
Signed-off-by: YangYuxi <[email protected]>
Signed-off-by: Julian Anastasov <[email protected]>
Reviewed-by: Simon Horman <[email protected]>
Signed-off-by: Pablo Neira Ayuso <[email protected]>
Signed-off-by: Sasha Levin <[email protected]>

include/net/ip_vs.h

@@ -1607,18 +1607,16 @@ static inline void ip_vs_conn_drop_conntrack(struct ip_vs_conn *cp)
 }
 #endif /* CONFIG_IP_VS_NFCT */
 
-/* Really using conntrack? */
-static inline bool ip_vs_conn_uses_conntrack(struct ip_vs_conn *cp,
-					     struct sk_buff *skb)
+/* Using old conntrack that can not be redirected to another real server? */
+static inline bool ip_vs_conn_uses_old_conntrack(struct ip_vs_conn *cp,
+						 struct sk_buff *skb)
 {
 #ifdef CONFIG_IP_VS_NFCT
 	enum ip_conntrack_info ctinfo;
 	struct nf_conn *ct;
 
-	if (!(cp->flags & IP_VS_CONN_F_NFCT))
-		return false;
 	ct = nf_ct_get(skb, &ctinfo);
-	if (ct)
+	if (ct && nf_ct_is_confirmed(ct))
 		return true;
 #endif
 	return false;

net/netfilter/ipvs/ip_vs_core.c

@@ -1916,30 +1916,32 @@ ip_vs_in(struct netns_ipvs *ipvs, unsigned int hooknum, struct sk_buff *skb, int
 
 	conn_reuse_mode = sysctl_conn_reuse_mode(ipvs);
 	if (conn_reuse_mode && !iph.fragoffs && is_new_conn(skb, &iph) && cp) {
-		bool uses_ct = false, resched = false;
+		bool old_ct = false, resched = false;
 
 		if (unlikely(sysctl_expire_nodest_conn(ipvs)) && cp->dest &&
 		    unlikely(!atomic_read(&cp->dest->weight))) {
 			resched = true;
-			uses_ct = ip_vs_conn_uses_conntrack(cp, skb);
+			old_ct = ip_vs_conn_uses_old_conntrack(cp, skb);
 		} else if (is_new_conn_expected(cp, conn_reuse_mode)) {
-			uses_ct = ip_vs_conn_uses_conntrack(cp, skb);
+			old_ct = ip_vs_conn_uses_old_conntrack(cp, skb);
 			if (!atomic_read(&cp->n_control)) {
 				resched = true;
 			} else {
 				/* Do not reschedule controlling connection
 				 * that uses conntrack while it is still
 				 * referenced by controlled connection(s).
 				 */
-				resched = !uses_ct;
+				resched = !old_ct;
 			}
 		}
 
 		if (resched) {
+			if (!old_ct)
+				cp->flags &= ~IP_VS_CONN_F_NFCT;
 			if (!atomic_read(&cp->n_control))
 				ip_vs_conn_expire_now(cp);
 			__ip_vs_conn_put(cp);
-			if (uses_ct)
+			if (old_ct)
 				return NF_DROP;
 			cp = NULL;
 		}