Re: [PATCH 02/14] tcp: fix mark propagation with fwmark_reflect enabled

On Thu, Jan 26, 2017 at 10:02:40AM -0800, Eric Dumazet wrote:

On Thu, 2017-01-26 at 17:37 +0100, Pablo Neira Ayuso wrote:

quoted

From: Pau Espin Pedrol <redacted>

Otherwise, RST packets generated by the TCP stack for non-existing
sockets always have mark 0.
The mark from the original packet is assigned to the netns_ipv4/6
socket used to send the response so that it can get copied into the
response skb when the socket sends it.

Fixes: e110861f8609 ("net: add a sysctl to reflect the fwmark on replies")
Cc: Lorenzo Colitti <redacted>
Signed-off-by: Pau Espin Pedrol <redacted>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
---
 net/ipv4/ip_output.c | 1 +
 net/ipv6/tcp_ipv6.c  | 1 +
 2 files changed, 2 insertions(+)

diff --git a/net/ipv4/ip_output.c b/net/ipv4/ip_output.c
index fac275c48108..b67719f45953 100644
--- a/net/ipv4/ip_output.c
+++ b/net/ipv4/ip_output.c

@@ -1629,6 +1629,7 @@ void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
 	sk->sk_protocol = ip_hdr(skb)->protocol;
 	sk->sk_bound_dev_if = arg->bound_dev_if;
 	sk->sk_sndbuf = sysctl_wmem_default;
+	sk->sk_mark = fl4.flowi4_mark;
 	err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
 			     len, 0, &ipc, &rt, MSG_DONTWAIT);
 	if (unlikely(err)) {

diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c
index 73bc8fc68acd..2b20622a5824 100644
--- a/net/ipv6/tcp_ipv6.c
+++ b/net/ipv6/tcp_ipv6.c

@@ -840,6 +840,7 @@ static void tcp_v6_send_response(const struct sock *sk, struct sk_buff *skb, u32
 	dst = ip6_dst_lookup_flow(ctl_sk, &fl6, NULL);
 	if (!IS_ERR(dst)) {
 		skb_dst_set(buff, dst);
+		ctl_sk->sk_mark = fl6.flowi6_mark;
 		ip6_xmit(ctl_sk, buff, &fl6, NULL, tclass);
 		TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
 		if (rst)

This patch is wrong.

ctl_sk is a shared socket, and tcp_v6_send_response() can be called from
many different cpus at the same time.

Right. This is not percpu as in IPv4.

I can send a follow up patch to get this in sync with the way we do it
in IPv4, ie. add percpu socket.

Fine with this approach? Thanks!