commit 03a3ca37e4 upstream.
Under extremely rare conditions TCP early demux will retrieve the wrong
socket.
1. local machine establishes a connection to a remote server, S, on port
p.
This gives:
laddr:lport -> S:p
... both in tcp and conntrack.
2. local machine establishes a connection to host H, on port p2.
2a. TCP stack choses same laddr:lport, so we have
laddr:lport -> H:p2 from TCP point of view.
2b). There is a destination NAT rewrite in place, translating
H:p2 to S:p. This results in following conntrack entries:
I) laddr:lport -> S:p (origin) S:p -> laddr:lport (reply)
II) laddr:lport -> H:p2 (origin) S:p -> laddr:lport2 (reply)
NAT engine has rewritten laddr:lport to laddr:lport2 to map
the reply packet to the correct origin.
When server sends SYN/ACK to laddr:lport2, the PREROUTING hook
will undo-the SNAT transformation, rewriting IP header to
S:p -> laddr:lport
This causes TCP early demux to associate the skb with the TCP socket
of the first connection.
The INPUT hook will then reverse the DNAT transformation, rewriting
the IP header to H:p2 -> laddr:lport.
Because packet ends up with the wrong socket, the new connection
never completes: originator stays in SYN_SENT and conntrack entry
remains in SYN_RECV until timeout, and responder retransmits SYN/ACK
until it gives up.
To resolve this, orphan the skb after the input rewrite:
Because the source IP address changed, the socket must be incorrect.
We can't move the DNAT undo to prerouting due to backwards
compatibility, doing so will make iptables/nftables rules to no longer
match the way they did.
After orphan, the packet will be handed to the next protocol layer
(tcp, udp, ...) and that will repeat the socket lookup just like as if
early demux was disabled.
Fixes: 41063e9dd1 ("ipv4: Early TCP socket demux.")
Closes: https://bugzilla.netfilter.org/show_bug.cgi?id=1427
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
f66b8e7381