openvswitch: Per cpu flow stats.

With mega flow implementation ovs flow can be shared between
multiple CPUs which makes stats updates highly contended
operation. This patch uses per-CPU stats in cases where a flow
is likely to be shared (if there is a wildcard in the 5-tuple
and therefore likely to be spread by RSS). In other situations,
it uses the current strategy, saving memory and allocation time.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: Jesse Gross <jesse@nicira.com>

net/openvswitch/datapath.c

@@ -251,9 +251,9 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
 	OVS_CB(skb)->flow = flow;
 	OVS_CB(skb)->pkt_key = &key;
 
-	stats_counter = &stats->n_hit;
-	ovs_flow_used(OVS_CB(skb)->flow, skb);
+	ovs_flow_stats_update(OVS_CB(skb)->flow, skb);
 	ovs_execute_actions(dp, skb);
+	stats_counter = &stats->n_hit;
 
 out:
 	/* Update datapath statistics. */
@@ -459,14 +459,6 @@ static int queue_userspace_packet(struct net *net, int dp_ifindex,
 	return err;
 }
 
-static void clear_stats(struct sw_flow *flow)
-{
-	flow->used = 0;
-	flow->tcp_flags = 0;
-	flow->packet_count = 0;
-	flow->byte_count = 0;
-}
-
 static int ovs_packet_cmd_execute(struct sk_buff *skb, struct genl_info *info)
 {
 	struct ovs_header *ovs_header = info->userhdr;
@@ -505,7 +497,7 @@ static int ovs_packet_cmd_execute(struct sk_buff *skb, struct genl_info *info)
 		packet->protocol = htons(ETH_P_802_2);
 
 	/* Build an sw_flow for sending this packet. */
-	flow = ovs_flow_alloc();
+	flow = ovs_flow_alloc(false);
 	err = PTR_ERR(flow);
 	if (IS_ERR(flow))
 		goto err_kfree_skb;
@@ -641,10 +633,10 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow, struct datapath *dp,
 	const int skb_orig_len = skb->len;
 	struct nlattr *start;
 	struct ovs_flow_stats stats;
+	__be16 tcp_flags;
+	unsigned long used;
 	struct ovs_header *ovs_header;
 	struct nlattr *nla;
-	unsigned long used;
-	u8 tcp_flags;
 	int err;
 
 	ovs_header = genlmsg_put(skb, portid, seq, &dp_flow_genl_family, flags, cmd);
@@ -673,24 +665,17 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow, struct datapath *dp,
 
 	nla_nest_end(skb, nla);
 
-	spin_lock_bh(&flow->lock);
-	used = flow->used;
-	stats.n_packets = flow->packet_count;
-	stats.n_bytes = flow->byte_count;
-	tcp_flags = (u8)ntohs(flow->tcp_flags);
-	spin_unlock_bh(&flow->lock);
-
+	ovs_flow_stats_get(flow, &stats, &used, &tcp_flags);
 	if (used &&
 	    nla_put_u64(skb, OVS_FLOW_ATTR_USED, ovs_flow_used_time(used)))
 		goto nla_put_failure;
 
 	if (stats.n_packets &&
-	    nla_put(skb, OVS_FLOW_ATTR_STATS,
-		    sizeof(struct ovs_flow_stats), &stats))
+	    nla_put(skb, OVS_FLOW_ATTR_STATS, sizeof(struct ovs_flow_stats), &stats))
 		goto nla_put_failure;
 
-	if (tcp_flags &&
-	    nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS, tcp_flags))
+	if ((u8)ntohs(tcp_flags) &&
+	     nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS, (u8)ntohs(tcp_flags)))
 		goto nla_put_failure;
 
 	/* If OVS_FLOW_ATTR_ACTIONS doesn't fit, skip dumping the actions if
@@ -770,6 +755,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
 	struct datapath *dp;
 	struct sw_flow_actions *acts = NULL;
 	struct sw_flow_match match;
+	bool exact_5tuple;
 	int error;
 
 	/* Extract key. */
@@ -778,7 +764,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
 		goto error;
 
 	ovs_match_init(&match, &key, &mask);
-	error = ovs_nla_get_match(&match,
+	error = ovs_nla_get_match(&match, &exact_5tuple,
 				  a[OVS_FLOW_ATTR_KEY], a[OVS_FLOW_ATTR_MASK]);
 	if (error)
 		goto error;
@@ -817,12 +803,11 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
 			goto err_unlock_ovs;
 
 		/* Allocate flow. */
-		flow = ovs_flow_alloc();
+		flow = ovs_flow_alloc(!exact_5tuple);
 		if (IS_ERR(flow)) {
 			error = PTR_ERR(flow);
 			goto err_unlock_ovs;
 		}
-		clear_stats(flow);
 
 		flow->key = masked_key;
 		flow->unmasked_key = key;
@@ -866,11 +851,8 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
 		reply = ovs_flow_cmd_build_info(flow, dp, info, OVS_FLOW_CMD_NEW);
 
 		/* Clear stats. */
-		if (a[OVS_FLOW_ATTR_CLEAR]) {
-			spin_lock_bh(&flow->lock);
-			clear_stats(flow);
-			spin_unlock_bh(&flow->lock);
-		}
+		if (a[OVS_FLOW_ATTR_CLEAR])
+			ovs_flow_stats_clear(flow);
 	}
 	ovs_unlock();
 
@@ -908,7 +890,7 @@ static int ovs_flow_cmd_get(struct sk_buff *skb, struct genl_info *info)
 	}
 
 	ovs_match_init(&match, &key, NULL);
-	err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+	err = ovs_nla_get_match(&match, NULL, a[OVS_FLOW_ATTR_KEY], NULL);
 	if (err)
 		return err;
 
@@ -962,7 +944,7 @@ static int ovs_flow_cmd_del(struct sk_buff *skb, struct genl_info *info)
 	}
 
 	ovs_match_init(&match, &key, NULL);
-	err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+	err = ovs_nla_get_match(&match, NULL, a[OVS_FLOW_ATTR_KEY], NULL);
 	if (err)
 		goto unlock;
 

net/openvswitch/flow.c

@@ -35,6 +35,7 @@
 #include <linux/ip.h>
 #include <linux/ipv6.h>
 #include <linux/sctp.h>
+#include <linux/smp.h>
 #include <linux/tcp.h>
 #include <linux/udp.h>
 #include <linux/icmp.h>
@@ -60,23 +61,104 @@ u64 ovs_flow_used_time(unsigned long flow_jiffies)
 
 #define TCP_FLAGS_BE16(tp) (*(__be16 *)&tcp_flag_word(tp) & htons(0x0FFF))
 
-void ovs_flow_used(struct sw_flow *flow, struct sk_buff *skb)
+void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb)
 {
+	struct flow_stats *stats;
 	__be16 tcp_flags = 0;
 
+	if (!flow->stats.is_percpu)
+		stats = flow->stats.stat;
+	else
+		stats = this_cpu_ptr(flow->stats.cpu_stats);
+
 	if ((flow->key.eth.type == htons(ETH_P_IP) ||
 	     flow->key.eth.type == htons(ETH_P_IPV6)) &&
 	    flow->key.ip.proto == IPPROTO_TCP &&
 	    likely(skb->len >= skb_transport_offset(skb) + sizeof(struct tcphdr))) {
 		tcp_flags = TCP_FLAGS_BE16(tcp_hdr(skb));
 	}
 
-	spin_lock(&flow->lock);
-	flow->used = jiffies;
-	flow->packet_count++;
-	flow->byte_count += skb->len;
-	flow->tcp_flags |= tcp_flags;
-	spin_unlock(&flow->lock);
+	spin_lock(&stats->lock);
+	stats->used = jiffies;
+	stats->packet_count++;
+	stats->byte_count += skb->len;
+	stats->tcp_flags |= tcp_flags;
+	spin_unlock(&stats->lock);
+}
+
+static void stats_read(struct flow_stats *stats,
+		       struct ovs_flow_stats *ovs_stats,
+		       unsigned long *used, __be16 *tcp_flags)
+{
+	spin_lock(&stats->lock);
+	if (time_after(stats->used, *used))
+		*used = stats->used;
+	*tcp_flags |= stats->tcp_flags;
+	ovs_stats->n_packets += stats->packet_count;
+	ovs_stats->n_bytes += stats->byte_count;
+	spin_unlock(&stats->lock);
+}
+
+void ovs_flow_stats_get(struct sw_flow *flow, struct ovs_flow_stats *ovs_stats,
+			unsigned long *used, __be16 *tcp_flags)
+{
+	int cpu, cur_cpu;
+
+	*used = 0;
+	*tcp_flags = 0;
+	memset(ovs_stats, 0, sizeof(*ovs_stats));
+
+	if (!flow->stats.is_percpu) {
+		stats_read(flow->stats.stat, ovs_stats, used, tcp_flags);
+	} else {
+		cur_cpu = get_cpu();
+		for_each_possible_cpu(cpu) {
+			struct flow_stats *stats;
+
+			if (cpu == cur_cpu)
+				local_bh_disable();
+
+			stats = per_cpu_ptr(flow->stats.cpu_stats, cpu);
+			stats_read(stats, ovs_stats, used, tcp_flags);
+
+			if (cpu == cur_cpu)
+				local_bh_enable();
+		}
+		put_cpu();
+	}
+}
+
+static void stats_reset(struct flow_stats *stats)
+{
+	spin_lock(&stats->lock);
+	stats->used = 0;
+	stats->packet_count = 0;
+	stats->byte_count = 0;
+	stats->tcp_flags = 0;
+	spin_unlock(&stats->lock);
+}
+
+void ovs_flow_stats_clear(struct sw_flow *flow)
+{
+	int cpu, cur_cpu;
+
+	if (!flow->stats.is_percpu) {
+		stats_reset(flow->stats.stat);
+	} else {
+		cur_cpu = get_cpu();
+
+		for_each_possible_cpu(cpu) {
+
+			if (cpu == cur_cpu)
+				local_bh_disable();
+
+			stats_reset(per_cpu_ptr(flow->stats.cpu_stats, cpu));
+
+			if (cpu == cur_cpu)
+				local_bh_enable();
+		}
+		put_cpu();
+	}
 }
 
 static int check_header(struct sk_buff *skb, int len)

net/openvswitch/flow.h

@@ -19,6 +19,7 @@
 #ifndef FLOW_H
 #define FLOW_H 1
 
+#include <linux/cache.h>
 #include <linux/kernel.h>
 #include <linux/netlink.h>
 #include <linux/openvswitch.h>
@@ -146,6 +147,22 @@ struct sw_flow_actions {
 	struct nlattr actions[];
 };
 
+struct flow_stats {
+	u64 packet_count;		/* Number of packets matched. */
+	u64 byte_count;			/* Number of bytes matched. */
+	unsigned long used;		/* Last used time (in jiffies). */
+	spinlock_t lock;		/* Lock for atomic stats update. */
+	__be16 tcp_flags;		/* Union of seen TCP flags. */
+};
+
+struct sw_flow_stats {
+	bool is_percpu;
+	union {
+		struct flow_stats *stat;
+		struct flow_stats __percpu *cpu_stats;
+	};
+};
+
 struct sw_flow {
 	struct rcu_head rcu;
 	struct hlist_node hash_node[2];
@@ -155,12 +172,7 @@ struct sw_flow {
 	struct sw_flow_key unmasked_key;
 	struct sw_flow_mask *mask;
 	struct sw_flow_actions __rcu *sf_acts;
-
-	spinlock_t lock;	/* Lock for values below. */
-	unsigned long used;	/* Last used time (in jiffies). */
-	u64 packet_count;	/* Number of packets matched. */
-	u64 byte_count;		/* Number of bytes matched. */
-	__be16 tcp_flags;	/* Union of seen TCP flags. */
+	struct sw_flow_stats stats;
 };
 
 struct arp_eth_header {
@@ -177,7 +189,10 @@ struct arp_eth_header {
 	unsigned char       ar_tip[4];		/* target IP address        */
 } __packed;
 
-void ovs_flow_used(struct sw_flow *, struct sk_buff *);
+void ovs_flow_stats_update(struct sw_flow *flow, struct sk_buff *skb);
+void ovs_flow_stats_get(struct sw_flow *flow, struct ovs_flow_stats *stats,
+			unsigned long *used, __be16 *tcp_flags);
+void ovs_flow_stats_clear(struct sw_flow *flow);
 u64 ovs_flow_used_time(unsigned long flow_jiffies);
 
 int ovs_flow_extract(struct sk_buff *, u16 in_port, struct sw_flow_key *);