net: Add sysctl to toggle early demux for tcp and udp

Certain system process significant unconnected UDP workload.
It would be preferrable to disable UDP early demux for those systems
and enable it for TCP only.

By disabling UDP demux, we see these slight gains on an ARM64 system-
782 -> 788Mbps unconnected single stream UDPv4
633 -> 654Mbps unconnected UDPv4 different sources

The performance impact can change based on CPU architecure and cache
sizes. There will not much difference seen if entire UDP hash table
is in cache.

Both sysctls are enabled by default to preserve existing behavior.

v1->v2: Change function pointer instead of adding conditional as
suggested by Stephen.

v2->v3: Read once in callers to avoid issues due to compiler
optimizations. Also update commit message with the tests.

v3->v4: Store and use read once result instead of querying pointer
again incorrectly.

v4->v5: Refactor to avoid errors due to compilation with IPV6={m,n}

Signed-off-by: Subash Abhinov Kasiviswanathan <[email protected]>
Suggested-by: Eric Dumazet <[email protected]>
Cc: Stephen Hemminger <[email protected]>
Cc: Tom Herbert <[email protected]>
Cc: David Miller <[email protected]>
Signed-off-by: David S. Miller <[email protected]>

Author: [email protected]

Documentation/networking/ip-sysctl.txt

@@ -856,12 +856,21 @@ ip_dynaddr - BOOLEAN
 ip_early_demux - BOOLEAN
 	Optimize input packet processing down to one demux for
 	certain kinds of local sockets.  Currently we only do this
-	for established TCP sockets.
+	for established TCP and connected UDP sockets.
 
 	It may add an additional cost for pure routing workloads that
 	reduces overall throughput, in such case you should disable it.
 	Default: 1
 
+tcp_early_demux - BOOLEAN
+	Enable early demux for established TCP sockets.
+	Default: 1
+
+udp_early_demux - BOOLEAN
+	Enable early demux for connected UDP sockets. Disable this if
+	your system could experience more unconnected load.
+	Default: 1
+
 icmp_echo_ignore_all - BOOLEAN
 	If set non-zero, then the kernel will ignore all ICMP ECHO
 	requests sent to it.

include/net/netns/ipv4.h

@@ -95,6 +95,8 @@ struct netns_ipv4 {
 	/* Shall we try to damage output packets if routing dev changes? */
 	int sysctl_ip_dynaddr;
 	int sysctl_ip_early_demux;
+	int sysctl_tcp_early_demux;
+	int sysctl_udp_early_demux;
 
 	int sysctl_fwmark_reflect;
 	int sysctl_tcp_fwmark_accept;

include/net/protocol.h

@@ -40,6 +40,7 @@
 /* This is used to register protocols. */
 struct net_protocol {
 	void			(*early_demux)(struct sk_buff *skb);
+	void                    (*early_demux_handler)(struct sk_buff *skb);
 	int			(*handler)(struct sk_buff *skb);
 	void			(*err_handler)(struct sk_buff *skb, u32 info);
 	unsigned int		no_policy:1,
@@ -54,7 +55,7 @@ struct net_protocol {
 #if IS_ENABLED(CONFIG_IPV6)
 struct inet6_protocol {
 	void	(*early_demux)(struct sk_buff *skb);
-
+	void    (*early_demux_handler)(struct sk_buff *skb);
 	int	(*handler)(struct sk_buff *skb);
 
 	void	(*err_handler)(struct sk_buff *skb,
@@ -92,12 +93,12 @@ struct inet_protosw {
 #define INET_PROTOSW_PERMANENT 0x02  /* Permanent protocols are unremovable. */
 #define INET_PROTOSW_ICSK      0x04  /* Is this an inet_connection_sock? */
 
-extern const struct net_protocol __rcu *inet_protos[MAX_INET_PROTOS];
+extern struct net_protocol __rcu *inet_protos[MAX_INET_PROTOS];
 extern const struct net_offload __rcu *inet_offloads[MAX_INET_PROTOS];
 extern const struct net_offload __rcu *inet6_offloads[MAX_INET_PROTOS];
 
 #if IS_ENABLED(CONFIG_IPV6)
-extern const struct inet6_protocol __rcu *inet6_protos[MAX_INET_PROTOS];
+extern struct inet6_protocol __rcu *inet6_protos[MAX_INET_PROTOS];
 #endif
 
 int inet_add_protocol(const struct net_protocol *prot, unsigned char num);

include/net/udp.h

@@ -372,4 +372,5 @@ void udp_encap_enable(void);
 #if IS_ENABLED(CONFIG_IPV6)
 void udpv6_encap_enable(void);
 #endif
+
 #endif	/* _UDP_H */

net/ipv4/af_inet.c

@@ -1599,17 +1599,19 @@ static const struct net_protocol igmp_protocol = {
 };
 #endif
 
-static const struct net_protocol tcp_protocol = {
+static struct net_protocol tcp_protocol = {
 	.early_demux	=	tcp_v4_early_demux,
+	.early_demux_handler =  tcp_v4_early_demux,
 	.handler	=	tcp_v4_rcv,
 	.err_handler	=	tcp_v4_err,
 	.no_policy	=	1,
 	.netns_ok	=	1,
 	.icmp_strict_tag_validation = 1,
 };
 
-static const struct net_protocol udp_protocol = {
+static struct net_protocol udp_protocol = {
 	.early_demux =	udp_v4_early_demux,
+	.early_demux_handler =	udp_v4_early_demux,
 	.handler =	udp_rcv,
 	.err_handler =	udp_err,
 	.no_policy =	1,
@@ -1720,6 +1722,8 @@ static __net_init int inet_init_net(struct net *net)
 	net->ipv4.sysctl_ip_default_ttl = IPDEFTTL;
 	net->ipv4.sysctl_ip_dynaddr = 0;
 	net->ipv4.sysctl_ip_early_demux = 1;
+	net->ipv4.sysctl_udp_early_demux = 1;
+	net->ipv4.sysctl_tcp_early_demux = 1;
 #ifdef CONFIG_SYSCTL
 	net->ipv4.sysctl_ip_prot_sock = PROT_SOCK;
 #endif

net/ipv4/ip_input.c

@@ -313,6 +313,7 @@ static int ip_rcv_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
 	const struct iphdr *iph = ip_hdr(skb);
 	struct rtable *rt;
 	struct net_device *dev = skb->dev;
+	void (*edemux)(struct sk_buff *skb);
 
 	/* if ingress device is enslaved to an L3 master device pass the
 	 * skb to its handler for processing
@@ -329,8 +330,8 @@ static int ip_rcv_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
 		int protocol = iph->protocol;
 
 		ipprot = rcu_dereference(inet_protos[protocol]);
-		if (ipprot && ipprot->early_demux) {
-			ipprot->early_demux(skb);
+		if (ipprot && (edemux = READ_ONCE(ipprot->early_demux))) {
+			edemux(skb);
 			/* must reload iph, skb->head might have changed */
 			iph = ip_hdr(skb);
 		}

net/ipv4/protocol.c

@@ -28,7 +28,7 @@
 #include <linux/spinlock.h>
 #include <net/protocol.h>
 
-const struct net_protocol __rcu *inet_protos[MAX_INET_PROTOS] __read_mostly;
+struct net_protocol __rcu *inet_protos[MAX_INET_PROTOS] __read_mostly;
 const struct net_offload __rcu *inet_offloads[MAX_INET_PROTOS] __read_mostly;
 EXPORT_SYMBOL(inet_offloads);
 

net/ipv4/sysctl_net_ipv4.c

@@ -24,6 +24,7 @@
 #include <net/cipso_ipv4.h>
 #include <net/inet_frag.h>
 #include <net/ping.h>
+#include <net/protocol.h>
 
 static int zero;
 static int one = 1;
@@ -294,6 +295,58 @@ static int proc_tcp_fastopen_key(struct ctl_table *ctl, int write,
 	return ret;
 }
 
+static void proc_configure_early_demux(int enabled, int protocol)
+{
+	struct net_protocol *ipprot;
+#if IS_ENABLED(CONFIG_IPV6)
+	struct inet6_protocol *ip6prot;
+#endif
+
+	ipprot = rcu_dereference(inet_protos[protocol]);
+	if (ipprot)
+		ipprot->early_demux = enabled ? ipprot->early_demux_handler :
+						NULL;
+
+#if IS_ENABLED(CONFIG_IPV6)
+	ip6prot = rcu_dereference(inet6_protos[protocol]);
+	if (ip6prot)
+		ip6prot->early_demux = enabled ? ip6prot->early_demux_handler :
+						 NULL;
+#endif
+}
+
+static int proc_tcp_early_demux(struct ctl_table *table, int write,
+				void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+	int ret = 0;
+
+	ret = proc_dointvec(table, write, buffer, lenp, ppos);
+
+	if (write && !ret) {
+		int enabled = init_net.ipv4.sysctl_tcp_early_demux;
+
+		proc_configure_early_demux(enabled, IPPROTO_TCP);
+	}
+
+	return ret;
+}
+
+static int proc_udp_early_demux(struct ctl_table *table, int write,
+				void __user *buffer, size_t *lenp, loff_t *ppos)
+{
+	int ret = 0;
+
+	ret = proc_dointvec(table, write, buffer, lenp, ppos);
+
+	if (write && !ret) {
+		int enabled = init_net.ipv4.sysctl_udp_early_demux;
+
+		proc_configure_early_demux(enabled, IPPROTO_UDP);
+	}
+
+	return ret;
+}
+
 static struct ctl_table ipv4_table[] = {
 	{
 		.procname	= "tcp_timestamps",
@@ -749,6 +802,20 @@ static struct ctl_table ipv4_net_table[] = {
 		.mode		= 0644,
 		.proc_handler	= proc_dointvec
 	},
+	{
+		.procname       = "udp_early_demux",
+		.data           = &init_net.ipv4.sysctl_udp_early_demux,
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = proc_udp_early_demux
+	},
+	{
+		.procname       = "tcp_early_demux",
+		.data           = &init_net.ipv4.sysctl_tcp_early_demux,
+		.maxlen         = sizeof(int),
+		.mode           = 0644,
+		.proc_handler   = proc_tcp_early_demux
+	},
 	{
 		.procname	= "ip_default_ttl",
 		.data		= &init_net.ipv4.sysctl_ip_default_ttl,

net/ipv6/ip6_input.c

@@ -49,6 +49,8 @@
 
 int ip6_rcv_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
 {
+	void (*edemux)(struct sk_buff *skb);
+
 	/* if ingress device is enslaved to an L3 master device pass the
 	 * skb to its handler for processing
 	 */
@@ -60,8 +62,8 @@ int ip6_rcv_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
 		const struct inet6_protocol *ipprot;
 
 		ipprot = rcu_dereference(inet6_protos[ipv6_hdr(skb)->nexthdr]);
-		if (ipprot && ipprot->early_demux)
-			ipprot->early_demux(skb);
+		if (ipprot && (edemux = READ_ONCE(ipprot->early_demux)))
+			edemux(skb);
 	}
 	if (!skb_valid_dst(skb))
 		ip6_route_input(skb);

net/ipv6/protocol.c

@@ -26,7 +26,7 @@
 #include <net/protocol.h>
 
 #if IS_ENABLED(CONFIG_IPV6)
-const struct inet6_protocol __rcu *inet6_protos[MAX_INET_PROTOS] __read_mostly;
+struct inet6_protocol __rcu *inet6_protos[MAX_INET_PROTOS] __read_mostly;
 EXPORT_SYMBOL(inet6_protos);
 
 int inet6_add_protocol(const struct inet6_protocol *prot, unsigned char protocol)

net/ipv6/tcp_ipv6.c

@@ -1925,8 +1925,9 @@ struct proto tcpv6_prot = {
 	.diag_destroy		= tcp_abort,
 };
 
-static const struct inet6_protocol tcpv6_protocol = {
+static struct inet6_protocol tcpv6_protocol = {
 	.early_demux	=	tcp_v6_early_demux,
+	.early_demux_handler =  tcp_v6_early_demux,
 	.handler	=	tcp_v6_rcv,
 	.err_handler	=	tcp_v6_err,
 	.flags		=	INET6_PROTO_NOPOLICY|INET6_PROTO_FINAL,

net/ipv6/udp.c

@@ -1436,8 +1436,9 @@ int compat_udpv6_getsockopt(struct sock *sk, int level, int optname,
 }
 #endif
 
-static const struct inet6_protocol udpv6_protocol = {
+static struct inet6_protocol udpv6_protocol = {
 	.early_demux	=	udp_v6_early_demux,
+	.early_demux_handler =  udp_v6_early_demux,
 	.handler	=	udpv6_rcv,
 	.err_handler	=	udpv6_err,
 	.flags		=	INET6_PROTO_NOPOLICY|INET6_PROTO_FINAL,