soreuseport: fix initialization race
authorCraig Gallek <kraig@google.com>
Thu, 19 Oct 2017 19:00:29 +0000 (15:00 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 18 Nov 2017 10:30:37 +0000 (11:30 +0100)
commit3212d077c679f189ac41c186f66f38591807d6f4
tree76e332a4aa9ad2724459123c85a6a9e933f273d3
parent318eb5ca5bbba6a7fb05fde69b0bcf9f67bfd45c
soreuseport: fix initialization race

[ Upstream commit 1b5f962e71bfad6284574655c406597535c3ea7a ]

Syzkaller stumbled upon a way to trigger
WARNING: CPU: 1 PID: 13881 at net/core/sock_reuseport.c:41
reuseport_alloc+0x306/0x3b0 net/core/sock_reuseport.c:39

There are two initialization paths for the sock_reuseport structure in a
socket: Through the udp/tcp bind paths of SO_REUSEPORT sockets or through
SO_ATTACH_REUSEPORT_[CE]BPF before bind.  The existing implementation
assumedthat the socket lock protected both of these paths when it actually
only protects the SO_ATTACH_REUSEPORT path.  Syzkaller triggered this
double allocation by running these paths concurrently.

This patch moves the check for double allocation into the reuseport_alloc
function which is protected by a global spin lock.

Fixes: e32ea7e74727 ("soreuseport: fast reuseport UDP socket selection")
Fixes: c125e80b8868 ("soreuseport: fast reuseport TCP socket selection")
Signed-off-by: Craig Gallek <kraig@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
net/core/sock_reuseport.c
net/ipv4/inet_hashtables.c
net/ipv4/udp.c