[Impact]
Occasionally starting new containers or creating new net namespaces may soft lockup because of improper refcounting of conntrack entires.
In the issue that I face, I can find a kworker thread using up an entire core, and when I cat /proc/$pid/stack I see this:
<ffffffffbe01e9b6>] ___preempt_schedule+0x56/0xb0
[<ffffffffc02223e4>] nf_ct_iterate_cleanup+0x134/0x160 [nf_conntrack]
[<ffffffffc0223dae>] nf_conntrack_cleanup_net_list+0x4e/0x170
[nf_conntrack]
[<ffffffffc022436d>] nf_conntrack_pernet_exit+0x4d/0x60 [nf_conntrack]
[<ffffffffbe6040d3>] ops_exit_list.isra.1+0x53/0x60
[<ffffffffbe6048d0>] cleanup_net+0x100/0x1d0
[<ffffffffbe084991>] process_one_work+0x171/0x470
[<ffffffffbe08563b>] worker_thread+0x11b/0x3a0
[<ffffffffbe08bb82>] kthread+0xd2/0xf0
[<ffffffffbe71757c>] ret_from_fork+0x7c/0xb0
[<ffffffffffffffff>] 0xffffffffffffffff
The kworker is looping forever and failing to clean up conntrack state.
All the while, it holds the global netns lock. Given that I've bisected
to commit e53376bef2cd97d3e3f61fdc677fb8da7d03d0da which is to do with refcounting, I suspect that borked refcounting on conntrack entries makes them impossible to properly free/destroy, which prevents this worker from cleaning up the namespace, which then goes on to prevent anything else from interacting with namespaces (add/delete/etc).
[Test Case]
bug 1403152 has a testcase which can occasionally hit this issue
[Fix]
$ git describe --contains e53376bef2cd97d3e3f61fdc677fb8da7d03d0da
v3.14-rc3~36^2~28^2~12
SRU Patch sent to Ubuntu kernel-team ML.