net_sched: act_gact: remove spinlock in fast path

Final step for gact RCU operation :

1) Use percpu stats
2) update lastuse only every clock tick to avoid false sharing
3) Remove spinlock acquisition, as it is no longer needed.

Since this is the last contended lock in packet RX when tc gact is used,
this gives impressive gain.

My host with 8 RX queues was handling 5 Mpps before the patch,
and more than 11 Mpps after patch.

Tested:

On receiver :

dev=eth0
tc qdisc del dev $dev ingress 2>/dev/null
tc qdisc add dev $dev ingress
tc filter del dev $dev root pref 10 2>/dev/null
tc filter del dev $dev pref 10 2>/dev/null
tc filter add dev $dev est 1sec 4sec parent ffff: protocol ip prio 1 \
	u32 match ip src 7.0.0.0/8 flowid 1:15 action drop

Sender sends packets flood from 7/8 network
Signed-off-by: NEric Dumazet <edumazet@google.com>
Acked-by: NAlexei Starovoitov <ast@plumgrid.com>
Acked-by: NJamal Hadi Salim <jhs@mojatatu.com>
Acked-by: NJohn Fastabend <john.fastabend@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

include/net/act_api.h

@@ -70,6 +70,17 @@ static inline void tcf_hashinfo_destroy(struct tcf_hashinfo *hf)
 	kfree(hf->htab);
 }
 
+/* Update lastuse only if needed, to avoid dirtying a cache line.
+ * We use a temp variable to avoid fetching jiffies twice.
+ */
+static inline void tcf_lastuse_update(struct tcf_t *tm)
+{
+	unsigned long now = jiffies;
+
+	if (tm->lastuse != now)
+		tm->lastuse = now;
+}
+
 #ifdef CONFIG_NET_CLS_ACT
 
 #define ACT_P_CREATED 1

net/sched/act_gact.c

@@ -90,7 +90,7 @@ static int tcf_gact_init(struct net *net, struct nlattr *nla,
 
 	if (!tcf_hash_check(parm->index, a, bind)) {
 		ret = tcf_hash_create(parm->index, est, a, sizeof(*gact),
-				      bind, false);
+				      bind, true);
 		if (ret)
 			return ret;
 		ret = ACT_P_CREATED;
@@ -104,7 +104,7 @@ static int tcf_gact_init(struct net *net, struct nlattr *nla,
 
 	gact = to_gact(a);
 
-	spin_lock_bh(&gact->tcf_lock);
+	ASSERT_RTNL();
 	gact->tcf_action = parm->action;
 #ifdef CONFIG_GACT_PROB
 	if (p_parm) {
@@ -117,7 +117,6 @@ static int tcf_gact_init(struct net *net, struct nlattr *nla,
 		gact->tcfg_ptype   = p_parm->ptype;
 	}
 #endif
-	spin_unlock_bh(&gact->tcf_lock);
 	if (ret == ACT_P_CREATED)
 		tcf_hash_insert(a);
 	return ret;
@@ -127,9 +126,8 @@ static int tcf_gact(struct sk_buff *skb, const struct tc_action *a,
 		    struct tcf_result *res)
 {
 	struct tcf_gact *gact = a->priv;
-	int action = gact->tcf_action;
+	int action = READ_ONCE(gact->tcf_action);
 
-	spin_lock(&gact->tcf_lock);
 #ifdef CONFIG_GACT_PROB
 	{
 	u32 ptype = READ_ONCE(gact->tcfg_ptype);
@@ -138,12 +136,11 @@ static int tcf_gact(struct sk_buff *skb, const struct tc_action *a,
 		action = gact_rand[ptype](gact);
 	}
 #endif
-	gact->tcf_bstats.bytes += qdisc_pkt_len(skb);
-	gact->tcf_bstats.packets++;
+	bstats_cpu_update(this_cpu_ptr(gact->common.cpu_bstats), skb);
 	if (action == TC_ACT_SHOT)
-		gact->tcf_qstats.drops++;
-	gact->tcf_tm.lastuse = jiffies;
-	spin_unlock(&gact->tcf_lock);
+		qstats_drop_inc(this_cpu_ptr(gact->common.cpu_qstats));
+
+	tcf_lastuse_update(&gact->tcf_tm);
 
 	return action;
 }