net: rfs: add sock_rps_delete_flow() helper

RFS can exhibit lower performance for workloads using short-lived
flows and a small set of 4-tuple.

This is often the case for load-testers, using a pair of hosts,
if the server has a single listener port.

Typical use case :

Server : tcp_crr -T128 -F1000 -6 -U -l30 -R 14250
Client : tcp_crr -T128 -F1000 -6 -U -l30 -c -H server | grep local_throughput

This is because RFS global hash table contains stale information,
when the same RSS key is recycled for another socket and another cpu.

Make sure to undo the changes and go back to initial state when
a flow is disconnected.

Performance of the above test is increased by 22 %,
going from 372604 transactions per second to 457773.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Octavian Purdila <tavip@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://patch.msgid.link/20250515100354.3339920-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

4 files changed, 31 insertions, 3 deletions

diff --git a/include/net/rps.h b/include/net/rps.h
index 507f4aa5d39b..d8ab3a08bcc4 100644
--- a/include/net/rps.h
+++ b/include/net/rps.h
@@ -123,6 +123,30 @@ static inline void sock_rps_record_flow(const struct sock *sk)
 #endif
 }
 
+static inline void sock_rps_delete_flow(const struct sock *sk)
+{
+#ifdef CONFIG_RPS
+	struct rps_sock_flow_table *table;
+	u32 hash, index;
+
+	if (!static_branch_unlikely(&rfs_needed))
+		return;
+
+	hash = READ_ONCE(sk->sk_rxhash);
+	if (!hash)
+		return;
+
+	rcu_read_lock();
+	table = rcu_dereference(net_hotdata.rps_sock_flow_table);
+	if (table) {
+		index = hash & table->mask;
+		if (READ_ONCE(table->ents[index]) != RPS_NO_CPU)
+			WRITE_ONCE(table->ents[index], RPS_NO_CPU);
+	}
+	rcu_read_unlock();
+#endif
+}
+
 static inline u32 rps_input_queue_tail_incr(struct softnet_data *sd)
 {
 #ifdef CONFIG_RPS
diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index da85cc30e382..77a0b52b2eab 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -23,11 +23,12 @@
 #if IS_ENABLED(CONFIG_IPV6)
 #include <net/inet6_hashtables.h>
 #endif
-#include <net/secure_seq.h>
 #include <net/hotdata.h>
 #include <net/ip.h>
-#include <net/tcp.h>
+#include <net/rps.h>
+#include <net/secure_seq.h>
 #include <net/sock_reuseport.h>
+#include <net/tcp.h>
 
 u32 inet_ehashfn(const struct net *net, const __be32 laddr,
 		 const __u16 lport, const __be32 faddr,
@@ -790,6 +791,7 @@ void inet_unhash(struct sock *sk)
 	if (sk_unhashed(sk))
 		return;
 
+	sock_rps_delete_flow(sk);
 	if (sk->sk_state == TCP_LISTEN) {
 		struct inet_listen_hashbucket *ilb2;
 
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 358b49caa7b9..dde52b8050b8 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -120,6 +120,7 @@
 #if IS_ENABLED(CONFIG_IPV6)
 #include <net/ipv6_stubs.h>
 #endif
+#include <net/rps.h>
 
 struct udp_table udp_table __read_mostly;
 
@@ -2200,6 +2201,7 @@ void udp_lib_unhash(struct sock *sk)
 		struct udp_table *udptable = udp_get_table_prot(sk);
 		struct udp_hslot *hslot, *hslot2;
 
+		sock_rps_delete_flow(sk);
 		hslot  = udp_hashslot(udptable, sock_net(sk),
 				      udp_sk(sk)->udp_port_hash);
 		hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash);
diff --git a/net/sctp/socket.c b/net/sctp/socket.c
index 53725ee7ba06..85a9dfeff4d6 100644
--- a/net/sctp/socket.c
+++ b/net/sctp/socket.c
@@ -8321,7 +8321,7 @@ static int sctp_hash(struct sock *sk)
 
 static void sctp_unhash(struct sock *sk)
 {
-	/* STUB */
+	sock_rps_delete_flow(sk);
 }
 
 /* Check if port is acceptable.  Possibly find first available port.