Jesse Brandeburg reported that commit acaf4e70 caused a panic
when adding a network namespace while vxlan module was present in
the system:

[<ffffffff814d0865>] vxlan_lowerdev_event+0xf5/0x100
[<ffffffff816e9e5d>] notifier_call_chain+0x4d/0x70
[<ffffffff810912be>] __raw_notifier_call_chain+0xe/0x10
[<ffffffff810912d6>] raw_notifier_call_chain+0x16/0x20
[<ffffffff815d9610>] call_netdevice_notifiers_info+0x40/0x70
[<ffffffff815d9656>] call_netdevice_notifiers+0x16/0x20
[<ffffffff815e1bce>] register_netdevice+0x1be/0x3a0
[<ffffffff815e1dce>] register_netdev+0x1e/0x30
[<ffffffff814cb94a>] loopback_net_init+0x4a/0xb0
[<ffffffffa016ed6e>] ? lockd_init_net+0x6e/0xb0 [lockd]
[<ffffffff815d6bac>] ops_init+0x4c/0x150
[<ffffffff815d6d23>] setup_net+0x73/0x110
[<ffffffff815d725b>] copy_net_ns+0x7b/0x100
[<ffffffff81090e11>] create_new_namespaces+0x101/0x1b0
[<ffffffff81090f45>] copy_namespaces+0x85/0xb0
[<ffffffff810693d5>] copy_process.part.26+0x935/0x1500
[<ffffffff811d5186>] ? mntput+0x26/0x40
[<ffffffff8106a15c>] do_fork+0xbc/0x2e0
[<ffffffff811b7f2e>] ? ____fput+0xe/0x10
[<ffffffff81089c5c>] ? task_work_run+0xac/0xe0
[<ffffffff8106a406>] SyS_clone+0x16/0x20
[<ffffffff816ee689>] stub_clone+0x69/0x90
[<ffffffff816ee329>] ? system_call_fastpath+0x16/0x1b

Apparently loopback device is being registered first and thus we
receive an event notification when vxlan_net is not ready. Hence,
when we call net_generic() and request vxlan_net_id, we seem to
access garbage at that point in time. In setup_net() where we set
up a newly allocated network namespace, we traverse the list of
pernet ops ...

list_for_each_entry(ops, &pernet_list, list) {
	error = ops_init(ops, net);
	if (error < 0)
		goto out_undo;
}

... and loopback_net_init() is invoked first here, so in the middle
of setup_net() we get this notification in vxlan. As currently we
only care about devices that unregister, move access through
net_generic() there. Fix is based on Cong Wang's proposal, but
only changes what is needed here. It sucks a bit as we only work
around the actual cure: right now it seems the only way to check if
a netns actually finished traversing all init ops would be to check
if it's part of net_namespace_list. But that I find quite expensive
each time we go through a notifier callback. Anyway, did a couple
of tests and it seems good for now.

Fixes: acaf4e70 ("net: vxlan: when lower dev unregisters remove vxlan dev as well")
Reported-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Tested-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>