netdev - [PATCH v2 net-next] udp: clear rps flow table for packets recv on UDP unconnected sockets

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <alpine.DEB.2.02.1408091013240.1087@tomh.mtv.corp.google.com>
Date:	Sat, 9 Aug 2014 10:15:30 -0700 (PDT)
From:	Tom Herbert <therbert@...gle.com>
To:	davem@...emloft.net, netdev@...r.kernel.org
cc:	eric.dumazet@...il.com
Subject: [PATCH v2 net-next] udp: clear rps flow table for packets recv on
 UDP unconnected sockets

This patch addresses a problem posed by Eric Dumazet in RPS/RFS
concerning the interaction between connected sockets and traffic on
unnconnected sockets.

On a server which has both traffic using connected connected sockets
and traffic that is going through unconnected UDP sockets, it is
very possible that the connected sockets could heavily populate the
RPS flow tables. Packets received on unconnected sockets would then
be steered based on unrelated entries in the flow tables which leads
to suboptimal steering. This happens even if the connections that
populate the table are inactive with no traffic, as long as the
connected sockets simply remain open unrelated traffic can be steered
using that information. This problem would further be exacerbated
if the packets on the unconnected UDP sockets are actually part of
long lived flows (which apparently would happen with QUIC in their
current implementation).

This patch clears the RPS flow hash table for packet recieved on
unnconnected UDP sockets. The effect is that the "flows" on unconnected
socekts will be steered using RPS. We don't do this for unconnected UDP
tunnels (encap_rcv) under the assumption that the flow table entries
will be adjusted during processing of inner packets.

Tested using super_netperf UDP_RR with 200 flows. Did not see any
noticeable regression in writing to flow table for every packet.

Before fix:
  76.99% CPU utilization
  112/158/230 90/95/99% latencies
  1.62776e+06 tps

After fix:
  76.03% CPU utilization
  115/162/239 90/95/99% latencies
  1.62257e+06 tps

Signed-off-by: Tom Herbert <therbert@...gle.com>
---
 net/ipv4/udp.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index f57c0e4..7ba764a 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -1444,6 +1444,8 @@ static int __udp_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
 	if (inet_sk(sk)->inet_daddr) {
 		sock_rps_save_rxhash(sk, skb);
 		sk_mark_napi_id(sk, skb);
+	} else {
+		sock_rps_reset_flow_hash(skb->hash);
 	}

 	rc = sock_queue_rcv_skb(sk, skb);
-- 
2.0.0.526.g5318336

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html