net: loopback: fix a dst refcounting issue
authorEric Dumazet <edumazet@google.com>
Fri, 25 Jan 2013 07:44:41 +0000 (07:44 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 14 Feb 2013 18:48:15 +0000 (10:48 -0800)
commitbe730f3d2049d1161ea22ada338dd1aeaef20065
treec0a9ba58ec906046e3c4d9390840e06757b94224
parentd432256090663bd155fad2c0649cc2d9e0c56c99
net: loopback: fix a dst refcounting issue

[ Upstream commit 794ed393b707f01858f5ebe2ae5eabaf89d00022 ]

Ben Greear reported crashes in ip_rcv_finish() on a stress
test involving many macvlans.

We tracked the bug to a dst use after free. ip_rcv_finish()
was calling dst->input() and got garbage for dst->input value.

It appears the bug is in loopback driver, lacking
a skb_dst_force() before calling netif_rx().

As a result, a non refcounted dst, normally protected by a
RCU read_lock section, was escaping this section and could
be freed before the packet being processed.

  [<ffffffff813a3c4d>] loopback_xmit+0x64/0x83
  [<ffffffff81477364>] dev_hard_start_xmit+0x26c/0x35e
  [<ffffffff8147771a>] dev_queue_xmit+0x2c4/0x37c
  [<ffffffff81477456>] ? dev_hard_start_xmit+0x35e/0x35e
  [<ffffffff8148cfa6>] ? eth_header+0x28/0xb6
  [<ffffffff81480f09>] neigh_resolve_output+0x176/0x1a7
  [<ffffffff814ad835>] ip_finish_output2+0x297/0x30d
  [<ffffffff814ad6d5>] ? ip_finish_output2+0x137/0x30d
  [<ffffffff814ad90e>] ip_finish_output+0x63/0x68
  [<ffffffff814ae412>] ip_output+0x61/0x67
  [<ffffffff814ab904>] dst_output+0x17/0x1b
  [<ffffffff814adb6d>] ip_local_out+0x1e/0x23
  [<ffffffff814ae1c4>] ip_queue_xmit+0x315/0x353
  [<ffffffff814adeaf>] ? ip_send_unicast_reply+0x2cc/0x2cc
  [<ffffffff814c018f>] tcp_transmit_skb+0x7ca/0x80b
  [<ffffffff814c3571>] tcp_connect+0x53c/0x587
  [<ffffffff810c2f0c>] ? getnstimeofday+0x44/0x7d
  [<ffffffff810c2f56>] ? ktime_get_real+0x11/0x3e
  [<ffffffff814c6f9b>] tcp_v4_connect+0x3c2/0x431
  [<ffffffff814d6913>] __inet_stream_connect+0x84/0x287
  [<ffffffff814d6b38>] ? inet_stream_connect+0x22/0x49
  [<ffffffff8108d695>] ? _local_bh_enable_ip+0x84/0x9f
  [<ffffffff8108d6c8>] ? local_bh_enable+0xd/0x11
  [<ffffffff8146763c>] ? lock_sock_nested+0x6e/0x79
  [<ffffffff814d6b38>] ? inet_stream_connect+0x22/0x49
  [<ffffffff814d6b49>] inet_stream_connect+0x33/0x49
  [<ffffffff814632c6>] sys_connect+0x75/0x98

This bug was introduced in linux-2.6.35, in commit
7fee226ad2397b (net: add a noref bit on skb dst)

skb_dst_force() is enforced in dev_queue_xmit() for devices having a
qdisc.

Reported-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Ben Greear <greearb@candelatech.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/net/loopback.c