add no_route_to_host_fix switch to turn on/off the fix in bpf mode

Test case
	1) conn_reuse_mode = 1 && no_route_to_host_fix==1
	result: wrk performance good
	2) conn_reuse_mode = 1 && no_route_to_host_fix==0
	result: wrk performance bad, packet loss
	3) conn_reuse_mode = 0 && no_route_to_host_fix==0
	result: wrk performance good. icmp no route to host error
	4) conn_reuse_mode = 0 && no_route_to_host_fix==1
	result: wrk performance good. no icmp no route to host error

Signed-off-by: jianmingfan <jianmingfan@tencent.com>

include/net/ip_vs.h

@@ -1704,6 +1704,7 @@ struct bpf_sym_addrs {
 extern struct bpf_sym_addrs resolve_addrs;
 extern struct bpf_map *conntrack_map;
 extern bool bpf_mode_on;
+extern bool no_route_to_host_fix;
 extern __be32 major_nic_ip;
 
 int ip_vs_bpf_put(void);

net/netfilter/ipvs/ip_vs_bpf_proc.c

@@ -54,7 +54,7 @@ __be32 major_nic_ip;
 static int ip_local_port_begin;
 static int ip_local_port_end;
 struct bpf_map  *conntrack_map;
-
+bool no_route_to_host_fix;
 struct ip_vs_iter_state {
 	struct seq_net_private	p;
 	struct hlist_head	*l;
@@ -667,6 +667,58 @@ static ssize_t ip_vs_port_write(struct file *file,
 	return count;
 }
 
+static ssize_t ip_vs_bpf_fix_write(struct file *file,
+				   const char __user *ubuf,
+				   size_t count,
+				   loff_t *ppos)
+{
+	char buf[10];
+	int fix;
+
+	memset(buf, 0, sizeof(buf));
+	if (*ppos > 0) {
+		pr_err("%s %d\n", __func__, __LINE__);
+		return -EFAULT;
+	}
+	if (copy_from_user(buf, ubuf, count)) {
+		pr_err("%s %d\n", __FILE__, __LINE__);
+		return -EFAULT;
+	}
+	if (kstrtoint(buf, 0, &fix) != 0) {
+		pr_err("%s %d\n", __FILE__, __LINE__);
+		return -EINVAL;
+	}
+	if (fix != 1 && fix != 0) {
+		pr_err("%s %d\n", __FILE__, __LINE__);
+		return -EINVAL;
+	}
+
+	no_route_to_host_fix = fix;
+	return count;
+}
+
+static ssize_t ip_vs_bpf_fix_read(struct file *file,
+				  char __user *ubuf,
+				  size_t count,
+				  loff_t *ppos)
+{
+	int len = 0;
+	char buf[10];
+
+	memset(buf, 0, sizeof(buf));
+	if (*ppos > 0) {
+		/* return 0 to sigal eof to user */
+		return 0;
+	}
+	/* snprintf need tail to save null bytes */
+	len = snprintf(buf, sizeof(buf), "0x%d\n", no_route_to_host_fix);
+	if (copy_to_user(ubuf, buf, len))
+		return -EFAULT;
+	/* so next time ppos will be bigger than 0 and return 0 */
+	*ppos = len;
+	return len;
+}
+
 static const struct file_operations ip_vs_svc_fops = {
 	.owner	 = THIS_MODULE,
 	.open    = ip_vs_svc_open,
@@ -694,6 +746,12 @@ static const struct file_operations ip_vs_port_fops = {
 	.write =  ip_vs_port_write,
 };
 
+static const struct file_operations ip_vs_bpf_fix_fops = {
+	.owner	 = THIS_MODULE,
+	.read    = ip_vs_bpf_fix_read,
+	.write   = ip_vs_bpf_fix_write,
+};
+
 struct ip_vs_err_map {
 	int errno;
 	const char *name;
@@ -774,9 +832,14 @@ int ip_vs_svc_proc_init(struct netns_ipvs *ipvs)
 	if (!proc_create("ip_vs_bpf_stat", S_IRUGO, ipvs->net->proc_net,
 			 &ip_vs_bpf_stat_fops))
 		goto out_bpf_stat;
+	if (!proc_create("ip_vs_bpf_fix", 0600,
+			 ipvs->net->proc_net, &ip_vs_bpf_fix_fops))
+		goto out_bpf_fix;
 
 	return 0;
 
+out_bpf_fix:
+	remove_proc_entry("ip_vs_bpf_stat", ipvs->net->proc_net);
 out_bpf_stat:
 	remove_proc_entry("ip_vs_port_range", ipvs->net->proc_net);
 out_port_range:
@@ -796,5 +859,6 @@ int ip_vs_svc_proc_cleanup(struct netns_ipvs *ipvs)
 	remove_proc_entry("ip_vs_devip", ipvs->net->proc_net);
 	remove_proc_entry("ip_vs_port_range", ipvs->net->proc_net);
 	remove_proc_entry("ip_vs_bpf_stat", ipvs->net->proc_net);
+	remove_proc_entry("ip_vs_bpf_fix", ipvs->net->proc_net);
 	return 0;
 }

net/netfilter/ipvs/ip_vs_conn.c

@@ -270,7 +270,7 @@ static void ip_vs_unlink_bpf(struct ip_vs_conn *cp)
 			return;
 		}
 
-		if (atomic_dec_return(&v->ref) != 0)
+		if (no_route_to_host_fix && atomic_dec_return(&v->ref) != 0)
 			return;
 		reply.sip = cp->dest->addr.ip;
 		reply.sport = cp->dest->port;
@@ -1125,10 +1125,12 @@ static bool ip_vs_conn_new_bpf(struct ip_vs_dest *dest,
 	 * otherwise, it doesn't reschedule. in such case, the following code
 	 * will not run.
 	 */
-	v = map->ops->map_lookup_elem(map, &key);
-	if (v) {
-		atomic_inc(&v->ref);
-		return true;
+	if (no_route_to_host_fix) {
+		v = map->ops->map_lookup_elem(map, &key);
+		if (v) {
+			atomic_inc(&v->ref);
+			return true;
+		}
 	}
 
 	lip = alloc_localip();

net/netfilter/ipvs/ip_vs_core.c

@@ -1994,8 +1994,9 @@ ip_vs_in(struct netns_ipvs *ipvs, unsigned int hooknum, struct sk_buff *skb, int
 	 * as zero rs may be killed already.
 	 * Only enable in bpf mode currently. Shall promote to IPVS mode later.
 	 */
-	if (bpf_mode_on && cp && unlikely(!atomic_read(&cp->dest->weight)) &&
-	   is_new_conn(skb, &iph) && !iph.fragoffs) {
+	if (no_route_to_host_fix && bpf_mode_on && cp &&
+	    unlikely(!atomic_read(&cp->dest->weight)) &&
+	    is_new_conn(skb, &iph) && !iph.fragoffs) {
 		ip_vs_conn_expire_now(cp);
 		__ip_vs_conn_put(cp);
 		cp = NULL;