diff mbox series

[net,2/2] udp: Improve load balancing for SO_REUSEPORT.

Message ID 20200721061531.94236-3-kuniyu@amazon.co.jp
State Accepted
Delegated to: David Miller
Headers show
Series udp: Fix reuseport selection with connected sockets. | expand

Commit Message

Kuniyuki Iwashima July 21, 2020, 6:15 a.m. UTC 
Currently, SO_REUSEPORT does not work well if connected sockets are in a
UDP reuseport group.

Then reuseport_has_conns() returns true and the result of
reuseport_select_sock() is discarded. Also, unconnected sockets have the
same score, hence only does the first unconnected socket in udp_hslot
always receive all packets sent to unconnected sockets.

So, the result of reuseport_select_sock() should be used for load
balancing.

The noteworthy point is that the unconnected sockets placed after
connected sockets in sock_reuseport.socks will receive more packets than
others because of the algorithm in reuseport_select_sock().

    index | connected | reciprocal_scale | result
    ---------------------------------------------
    0     | no        | 20%              | 40%
    1     | no        | 20%              | 20%
    2     | yes       | 20%              | 0%
    3     | no        | 20%              | 40%
    4     | yes       | 20%              | 0%

If most of the sockets are connected, this can be a problem, but it still
works better than now.

Fixes: acdcecc61285 ("udp: correct reuseport selection with connected sockets")
CC: Willem de Bruijn <willemb@google.com>
Reviewed-by: Benjamin Herrenschmidt <benh@amazon.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
---
 net/ipv4/udp.c | 15 +++++++++------
 net/ipv6/udp.c | 15 +++++++++------
 2 files changed, 18 insertions(+), 12 deletions(-)

Comments

Willem de Bruijn July 21, 2020, 12:33 p.m. UTC | #1 
On Tue, Jul 21, 2020 at 2:17 AM Kuniyuki Iwashima <kuniyu@amazon.co.jp> wrote:
>
> Currently, SO_REUSEPORT does not work well if connected sockets are in a
> UDP reuseport group.
>
> Then reuseport_has_conns() returns true and the result of
> reuseport_select_sock() is discarded. Also, unconnected sockets have the
> same score, hence only does the first unconnected socket in udp_hslot
> always receive all packets sent to unconnected sockets.
>
> So, the result of reuseport_select_sock() should be used for load
> balancing.
>
> The noteworthy point is that the unconnected sockets placed after
> connected sockets in sock_reuseport.socks will receive more packets than
> others because of the algorithm in reuseport_select_sock().
>
>     index | connected | reciprocal_scale | result
>     ---------------------------------------------
>     0     | no        | 20%              | 40%
>     1     | no        | 20%              | 20%
>     2     | yes       | 20%              | 0%
>     3     | no        | 20%              | 40%
>     4     | yes       | 20%              | 0%
>
> If most of the sockets are connected, this can be a problem, but it still
> works better than now.
>
> Fixes: acdcecc61285 ("udp: correct reuseport selection with connected sockets")
> CC: Willem de Bruijn <willemb@google.com>
> Reviewed-by: Benjamin Herrenschmidt <benh@amazon.com>
> Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>

Acked-by: Willem de Bruijn <willemb@google.com>
diff mbox series

Patch

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 1b7ebbcae497..99251d3c70d0 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -416,7 +416,7 @@  static struct sock *udp4_lib_lookup2(struct net *net,
 				     struct udp_hslot *hslot2,
 				     struct sk_buff *skb)
 {
-	struct sock *sk, *result;
+	struct sock *sk, *result, *reuseport_result;
 	int score, badness;
 	u32 hash = 0;
 
@@ -426,17 +426,20 @@  static struct sock *udp4_lib_lookup2(struct net *net,
 		score = compute_score(sk, net, saddr, sport,
 				      daddr, hnum, dif, sdif);
 		if (score > badness) {
+			reuseport_result = NULL;
+
 			if (sk->sk_reuseport &&
 			    sk->sk_state != TCP_ESTABLISHED) {
 				hash = udp_ehashfn(net, daddr, hnum,
 						   saddr, sport);
-				result = reuseport_select_sock(sk, hash, skb,
-							sizeof(struct udphdr));
-				if (result && !reuseport_has_conns(sk, false))
-					return result;
+				reuseport_result = reuseport_select_sock(sk, hash, skb,
+									 sizeof(struct udphdr));
+				if (reuseport_result && !reuseport_has_conns(sk, false))
+					return reuseport_result;
 			}
+
+			result = reuseport_result ? : sk;
 			badness = score;
-			result = sk;
 		}
 	}
 	return result;
diff --git a/net/ipv6/udp.c b/net/ipv6/udp.c
index 7d4151747340..9503c87ac0b3 100644
--- a/net/ipv6/udp.c
+++ b/net/ipv6/udp.c
@@ -148,7 +148,7 @@  static struct sock *udp6_lib_lookup2(struct net *net,
 		int dif, int sdif, struct udp_hslot *hslot2,
 		struct sk_buff *skb)
 {
-	struct sock *sk, *result;
+	struct sock *sk, *result, *reuseport_result;
 	int score, badness;
 	u32 hash = 0;
 
@@ -158,17 +158,20 @@  static struct sock *udp6_lib_lookup2(struct net *net,
 		score = compute_score(sk, net, saddr, sport,
 				      daddr, hnum, dif, sdif);
 		if (score > badness) {
+			reuseport_result = NULL;
+
 			if (sk->sk_reuseport &&
 			    sk->sk_state != TCP_ESTABLISHED) {
 				hash = udp6_ehashfn(net, daddr, hnum,
 						    saddr, sport);
 
-				result = reuseport_select_sock(sk, hash, skb,
-							sizeof(struct udphdr));
-				if (result && !reuseport_has_conns(sk, false))
-					return result;
+				reuseport_result = reuseport_select_sock(sk, hash, skb,
+									 sizeof(struct udphdr));
+				if (reuseport_result && !reuseport_has_conns(sk, false))
+					return reuseport_result;
 			}
-			result = sk;
+
+			result = reuseport_result ? : sk;
 			badness = score;
 		}
 	}