ipv4: percpu nh_rth_output cache

Input path is mostly run under RCU and doesnt touch dst refcnt

But output path on forwarding or UDP workloads hits
badly dst refcount, and we have lot of false sharing, for example
in ipv4_mtu() when reading rt->rt_pmtu

Using a percpu cache for nh_rth_output gives a nice performance
increase at a small cost.

24 udpflood test on my 24 cpu machine (dummy0 output device)
(each process sends 1.000.000 udp frames, 24 processes are started)

before : 5.24 s
after : 2.06 s
For reference, time on linux-3.5 : 6.60 s
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/net/ip_fib.h

@@ -21,6 +21,7 @@
 #include <linux/rcupdate.h>
 #include <net/fib_rules.h>
 #include <net/inetpeer.h>
+#include <linux/percpu.h>
 
 struct fib_config {
 	u8			fc_dst_len;
@@ -81,7 +82,7 @@ struct fib_nh {
 	__be32			nh_gw;
 	__be32			nh_saddr;
 	int			nh_saddr_genid;
-	struct rtable __rcu	*nh_rth_output;
+	struct rtable __rcu * __percpu *nh_pcpu_rth_output;
 	struct rtable __rcu	*nh_rth_input;
 	struct fnhe_hash_bucket	*nh_exceptions;
 };

net/ipv4/fib_semantics.c

@@ -176,6 +176,23 @@ static void rt_nexthop_free(struct rtable __rcu **rtp)
 	dst_free(&rt->dst);
 }
 
+static void rt_nexthop_free_cpus(struct rtable __rcu * __percpu *rtp)
+{
+	int cpu;
+
+	if (!rtp)
+		return;
+
+	for_each_possible_cpu(cpu) {
+		struct rtable *rt;
+
+		rt = rcu_dereference_protected(*per_cpu_ptr(rtp, cpu), 1);
+		if (rt)
+			dst_free(&rt->dst);
+	}
+	free_percpu(rtp);
+}
+
 /* Release a nexthop info record */
 static void free_fib_info_rcu(struct rcu_head *head)
 {
@@ -186,7 +203,7 @@ static void free_fib_info_rcu(struct rcu_head *head)
 			dev_put(nexthop_nh->nh_dev);
 		if (nexthop_nh->nh_exceptions)
 			free_nh_exceptions(nexthop_nh);
-		rt_nexthop_free(&nexthop_nh->nh_rth_output);
+		rt_nexthop_free_cpus(nexthop_nh->nh_pcpu_rth_output);
 		rt_nexthop_free(&nexthop_nh->nh_rth_input);
 	} endfor_nexthops(fi);
 
@@ -817,6 +834,7 @@ struct fib_info *fib_create_info(struct fib_config *cfg)
 	fi->fib_nhs = nhs;
 	change_nexthops(fi) {
 		nexthop_nh->nh_parent = fi;
+		nexthop_nh->nh_pcpu_rth_output = alloc_percpu(struct rtable __rcu *);
 	} endfor_nexthops(fi)
 
 	if (cfg->fc_mx) {

net/ipv4/route.c

@@ -1206,11 +1206,15 @@ static inline void rt_free(struct rtable *rt)
 
 static void rt_cache_route(struct fib_nh *nh, struct rtable *rt)
 {
-	struct rtable *orig, *prev, **p = (struct rtable **)&nh->nh_rth_output;
+	struct rtable *orig, *prev, **p;
 
-	if (rt_is_input_route(rt))
+	if (rt_is_input_route(rt)) {
 		p = (struct rtable **)&nh->nh_rth_input;
-
+	} else {
+		if (!nh->nh_pcpu_rth_output)
+			goto nocache;
+		p = (struct rtable **)__this_cpu_ptr(nh->nh_pcpu_rth_output);
+	}
 	orig = *p;
 
 	prev = cmpxchg(p, orig, rt);
@@ -1223,6 +1227,7 @@ static void rt_cache_route(struct fib_nh *nh, struct rtable *rt)
 		 * unsuccessful at storing this route into the cache
 		 * we really need to set it.
 		 */
+nocache:
 		rt->dst.flags |= DST_NOCACHE;
 	}
 }
@@ -1749,8 +1754,11 @@ static struct rtable *__mkroute_output(const struct fib_result *res,
 	fnhe = NULL;
 	if (fi) {
 		fnhe = find_exception(&FIB_RES_NH(*res), fl4->daddr);
-		if (!fnhe) {
-			rth = rcu_dereference(FIB_RES_NH(*res).nh_rth_output);
+		if (!fnhe && FIB_RES_NH(*res).nh_pcpu_rth_output) {
+			struct rtable __rcu **prth;
+
+			prth = __this_cpu_ptr(FIB_RES_NH(*res).nh_pcpu_rth_output);
+			rth = rcu_dereference(*prth);
 			if (rt_cache_valid(rth)) {
 				dst_hold(&rth->dst);
 				return rth;