tcp/dccp: fix lockdep issue when SYN is backlogged
authorEric Dumazet <edumazet@google.com>
Mon, 1 Oct 2018 22:02:26 +0000 (15:02 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 18 Oct 2018 07:16:20 +0000 (09:16 +0200)
commit17af5475aef350b78875646b2795381f452de756
tree5c5776d530109d71ef733aceb6be0072db5648f2
parent4b7b26024f52a5cdeca20b56657f90d74a0a1995
tcp/dccp: fix lockdep issue when SYN is backlogged

[ Upstream commit 1ad98e9d1bdf4724c0a8532fabd84bf3c457c2bc ]

In normal SYN processing, packets are handled without listener
lock and in RCU protected ingress path.

But syzkaller is known to be able to trick us and SYN
packets might be processed in process context, after being
queued into socket backlog.

In commit 06f877d613be ("tcp/dccp: fix other lockdep splats
accessing ireq_opt") I made a very stupid fix, that happened
to work mostly because of the regular path being RCU protected.

Really the thing protecting ireq->ireq_opt is RCU read lock,
and the pseudo request refcnt is not relevant.

This patch extends what I did in commit 449809a66c1d ("tcp/dccp:
block BH for SYN processing") by adding an extra rcu_read_{lock|unlock}
pair in the paths that might be taken when processing SYN from
socket backlog (thus possibly in process context)

Fixes: 06f877d613be ("tcp/dccp: fix other lockdep splats accessing ireq_opt")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
include/net/inet_sock.h
net/dccp/input.c
net/ipv4/tcp_input.c