48f65f6
From 0d3f6d297bfb7af24d0508460fdb3d1ec4903fa3 Mon Sep 17 00:00:00 2001
48f65f6
From: Martin KaFai Lau <kafai@fb.com>
48f65f6
Date: Wed, 11 Nov 2015 11:51:06 -0800
48f65f6
Subject: [PATCH] ipv6: Avoid creating RTF_CACHE from a rt that is not managed
48f65f6
 by fib6 tree
48f65f6
48f65f6
The original bug report:
48f65f6
https://bugzilla.redhat.com/show_bug.cgi?id=1272571
48f65f6
48f65f6
The setup has a IPv4 GRE tunnel running in a IPSec.  The bug
48f65f6
happens when ndisc starts sending router solicitation at the gre
48f65f6
interface.  The simplified oops stack is like:
48f65f6
48f65f6
__lock_acquire+0x1b2/0x1c30
48f65f6
lock_acquire+0xb9/0x140
48f65f6
_raw_write_lock_bh+0x3f/0x50
48f65f6
__ip6_ins_rt+0x2e/0x60
48f65f6
ip6_ins_rt+0x49/0x50
48f65f6
~~~~~~~~
48f65f6
__ip6_rt_update_pmtu.part.54+0x145/0x250
48f65f6
ip6_rt_update_pmtu+0x2e/0x40
48f65f6
~~~~~~~~
48f65f6
ip_tunnel_xmit+0x1f1/0xf40
48f65f6
__gre_xmit+0x7a/0x90
48f65f6
ipgre_xmit+0x15a/0x220
48f65f6
dev_hard_start_xmit+0x2bd/0x480
48f65f6
__dev_queue_xmit+0x696/0x730
48f65f6
dev_queue_xmit+0x10/0x20
48f65f6
neigh_direct_output+0x11/0x20
48f65f6
ip6_finish_output2+0x21f/0x770
48f65f6
ip6_finish_output+0xa7/0x1d0
48f65f6
ip6_output+0x56/0x190
48f65f6
~~~~~~~~
48f65f6
ndisc_send_skb+0x1d9/0x400
48f65f6
ndisc_send_rs+0x88/0xc0
48f65f6
~~~~~~~~
48f65f6
48f65f6
The rt passed to ip6_rt_update_pmtu() is created by
48f65f6
icmp6_dst_alloc() and it is not managed by the fib6 tree,
48f65f6
so its rt6i_table == NULL.  When __ip6_rt_update_pmtu() creates
48f65f6
a RTF_CACHE clone, the newly created clone also has rt6i_table == NULL
48f65f6
and it causes the ip6_ins_rt() oops.
48f65f6
48f65f6
During pmtu update, we only want to create a RTF_CACHE clone
48f65f6
from a rt which is currently managed (or owned) by the
48f65f6
fib6 tree.  It means either rt->rt6i_node != NULL or
48f65f6
rt is a RTF_PCPU clone.
48f65f6
48f65f6
It is worth to note that rt6i_table may not be NULL even it is
48f65f6
not (yet) managed by the fib6 tree (e.g. addrconf_dst_alloc()).
48f65f6
Hence, rt6i_node is a better check instead of rt6i_table.
48f65f6
48f65f6
Fixes: 45e4fd26683c ("ipv6: Only create RTF_CACHE routes after encountering pmtu")
48f65f6
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
48f65f6
Reported-by: Chris Siebenmann <cks-rhbugzilla@cs.toronto.edu>
48f65f6
Cc: Chris Siebenmann <cks-rhbugzilla@cs.toronto.edu>
48f65f6
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
48f65f6
Signed-off-by: David S. Miller <davem@davemloft.net>
48f65f6
---
48f65f6
 net/ipv6/route.c | 8 +++++++-
48f65f6
 1 file changed, 7 insertions(+), 1 deletion(-)
48f65f6
48f65f6
diff --git a/net/ipv6/route.c b/net/ipv6/route.c
48f65f6
index c8bc9b4..74907c5 100644
48f65f6
--- a/net/ipv6/route.c
48f65f6
+++ b/net/ipv6/route.c
48f65f6
@@ -1322,6 +1322,12 @@ static void rt6_do_update_pmtu(struct rt6_info *rt, u32 mtu)
48f65f6
 	rt6_update_expires(rt, net->ipv6.sysctl.ip6_rt_mtu_expires);
48f65f6
 }
48f65f6
48f65f6
+static bool rt6_cache_allowed_for_pmtu(const struct rt6_info *rt)
48f65f6
+{
48f65f6
+	return !(rt->rt6i_flags & RTF_CACHE) &&
48f65f6
+		(rt->rt6i_flags & RTF_PCPU || rt->rt6i_node);
48f65f6
+}
48f65f6
+
48f65f6
 static void __ip6_rt_update_pmtu(struct dst_entry *dst, const struct sock *sk,
48f65f6
 				 const struct ipv6hdr *iph, u32 mtu)
48f65f6
 {
48f65f6
@@ -1335,7 +1341,7 @@ static void __ip6_rt_update_pmtu(struct dst_entry *dst, const struct sock *sk,
48f65f6
 	if (mtu >= dst_mtu(dst))
48f65f6
 		return;
48f65f6
48f65f6
-	if (rt6->rt6i_flags & RTF_CACHE) {
48f65f6
+	if (!rt6_cache_allowed_for_pmtu(rt6)) {
48f65f6
 		rt6_do_update_pmtu(rt6, mtu);
48f65f6
 	} else {
48f65f6
 		const struct in6_addr *daddr, *saddr;
48f65f6
-- 
48f65f6
2.5.0
48f65f6