On Wed, Oct 30, 2013 at 12:07:11AM +0000, Patrick McHardy wrote:

The problem is that the reassembled packet is referenced by the individual
fragments, so we trigger the BUG_ON in pskb_expand_head(). In this
particular case the case we BUG() on is actually OK, but I'm looking at
a way we can fix this without special casing. Hope to have a patch for
testing in the next hours.

Just for the record. I'm observing similar, quite reproducable crashes when
receiving fragmented icmp echo request packets on an IPsec gateway with
nf_conntrack_ipv6.

Since git commit 58a317f10 ("netfilter: ipv6: add IPv6 NAT support")
netfilter might insert a reassembled ipv6 packet with a shared skb and
local_df = 1 to the ok function. In case of xfrm, __xfrm6_output()
fragments the packet again and when adjusting the headroom later, we
crash because of a shared skb.

I can fix it by checking for a shared skb in ip6_fragment() and do
slow path fragmentation then. But we never needed such a check in
ip6_fragment(), so it's maybe better to fix it in netfilter.