net: rfs: add hash collision detection

Receive Flow Steering is a nice solution but suffers from
hash collisions when a mix of connected and unconnected traffic
is received on the host, when flow hash table is populated.

Also, clearing flow in inet_release() makes RFS not very good
for short lived flows, as many packets can follow close().
(FIN , ACK packets, ...)

This patch extends the information stored into global hash table
to not only include cpu number, but upper part of the hash value.

I use a 32bit value, and dynamically split it in two parts.

For host with less than 64 possible cpus, this gives 6 bits for the
cpu number, and 26 (32-6) bits for the upper part of the hash.

Since hash bucket selection use low order bits of the hash, we have
a full hash match, if /proc/sys/net/core/rps_sock_flow_entries is big
enough.

If the hash found in flow table does not match, we fallback to RPS (if
it is enabled for the rxqueue).

This means that a packet for an non connected flow can avoid the
IPI through a unrelated/victim CPU.

This also means we no longer have to clear the table at socket
close time, and this helps short lived flows performance.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 18 insertions, 40 deletions

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index ce784d5..ab3b7ce 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -644,39 +644,39 @@ struct rps_dev_flow_table {
 /*
  * The rps_sock_flow_table contains mappings of flows to the last CPU
  * on which they were processed by the application (set in recvmsg).
+ * Each entry is a 32bit value. Upper part is the high order bits
+ * of flow hash, lower part is cpu number.
+ * rps_cpu_mask is used to partition the space, depending on number of
+ * possible cpus : rps_cpu_mask = roundup_pow_of_two(nr_cpu_ids) - 1
+ * For example, if 64 cpus are possible, rps_cpu_mask = 0x3f,
+ * meaning we use 32-6=26 bits for the hash.
  */
 struct rps_sock_flow_table {
-	unsigned int mask;
-	u16 ents[0];
+	u32	mask;
+	u32	ents[0];
 };
-#define	RPS_SOCK_FLOW_TABLE_SIZE(_num) (sizeof(struct rps_sock_flow_table) + \
-    ((_num) * sizeof(u16)))
+#define	RPS_SOCK_FLOW_TABLE_SIZE(_num) (offsetof(struct rps_sock_flow_table, ents[_num]))
 
 #define RPS_NO_CPU 0xffff
 
+extern u32 rps_cpu_mask;
+extern struct rps_sock_flow_table __rcu *rps_sock_flow_table;
+
 static inline void rps_record_sock_flow(struct rps_sock_flow_table *table,
 					u32 hash)
 {
 	if (table && hash) {
-		unsigned int cpu, index = hash & table->mask;
+		unsigned int index = hash & table->mask;
+		u32 val = hash & ~rps_cpu_mask;
 
 		/* We only give a hint, preemption can change cpu under us */
-		cpu = raw_smp_processor_id();
+		val |= raw_smp_processor_id();
 
-		if (table->ents[index] != cpu)
-			table->ents[index] = cpu;
+		if (table->ents[index] != val)
+			table->ents[index] = val;
 	}
 }
 
-static inline void rps_reset_sock_flow(struct rps_sock_flow_table *table,
-				       u32 hash)
-{
-	if (table && hash)
-		table->ents[hash & table->mask] = RPS_NO_CPU;
-}
-
-extern struct rps_sock_flow_table __rcu *rps_sock_flow_table;
-
 #ifdef CONFIG_RFS_ACCEL
 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, u32 flow_id,
 			 u16 filter_id);
diff --git a/include/net/sock.h b/include/net/sock.h
index d28b8fe..e138245 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -857,18 +857,6 @@ static inline void sock_rps_record_flow_hash(__u32 hash)
 #endif
 }
 
-static inline void sock_rps_reset_flow_hash(__u32 hash)
-{
-#ifdef CONFIG_RPS
-	struct rps_sock_flow_table *sock_flow_table;
-
-	rcu_read_lock();
-	sock_flow_table = rcu_dereference(rps_sock_flow_table);
-	rps_reset_sock_flow(sock_flow_table, hash);
-	rcu_read_unlock();
-#endif
-}
-
 static inline void sock_rps_record_flow(const struct sock *sk)
 {
 #ifdef CONFIG_RPS
@@ -876,28 +864,18 @@ static inline void sock_rps_record_flow(const struct sock *sk)
 #endif
 }
 
-static inline void sock_rps_reset_flow(const struct sock *sk)
-{
-#ifdef CONFIG_RPS
-	sock_rps_reset_flow_hash(sk->sk_rxhash);
-#endif
-}
-
 static inline void sock_rps_save_rxhash(struct sock *sk,
 					const struct sk_buff *skb)
 {
 #ifdef CONFIG_RPS
-	if (unlikely(sk->sk_rxhash != skb->hash)) {
-		sock_rps_reset_flow(sk);
+	if (unlikely(sk->sk_rxhash != skb->hash))
 		sk->sk_rxhash = skb->hash;
-	}
 #endif
 }
 
 static inline void sock_rps_reset_rxhash(struct sock *sk)
 {
 #ifdef CONFIG_RPS
-	sock_rps_reset_flow(sk);
 	sk->sk_rxhash = 0;
 #endif
 }