soft lockup - may be related to wireguard (backported)

[Jason A. Donenfeld]

On 5/4/20 6:49 AM, Wang Jian wrote:
 > Jason A. Donenfeld <Jason at zx2c4.com> 于2020年5月4日周一 下午1:26写道：
 >>
 >> Are you routing wireguard over wireguard, or something interesting 
like that?
 >>
 >> Is ipsec being used?
 >
 > I don't think I have any fancy use cases. But wireguard over pppoe?
 >
 > Other details are
 > - nftable's iptables compatible mode is used, along with ipset
 > - pppoe link is default route, and wg-quick is configured to install
 > additional default route into new created routing table (2000)
 > - ipset matches are used to MARK traffic to specific destinations in
 > mangle table, PREROUTING & OUTPUT, for both v4 and v6
 > - ip rules are added to match the fwmark and lookup new routing table
 > (2000, so go out via wireguard interface) for forwarded traffic and
 > output traffic
Can you send full networking configuration in enough detail that I'll be 
able to reliably reproduce this problem? If I can't reproduce it, it's 
unlikely I'll be able to fix it.

Meanwhile, it really really looks from your stacktrace that you have one 
wireguard interface going over another wireguard interface:

[27929.506367]  wg_packet_send_staged_packets+0x320/0x5d0 [wireguard]
[27929.506426]  wg_xmit+0x324/0x490 [wireguard]
[27929.506469]  dev_hard_start_xmit+0x8d/0x1e0
[27929.506508]  __dev_queue_xmit+0x721/0x8e0
[27929.506549]  ip_finish_output2+0x19b/0x590
[27929.506604]  ? nf_confirm+0xcb/0xf0 [nf_conntrack]
[27929.506648]  ip_output+0x76/0xf0
[27929.506681]  ? __ip_finish_output+0x1c0/0x1c0
[27929.506720]  iptunnel_xmit+0x174/0x210
[27929.506761]  send4+0x120/0x390 [wireguard]
[27929.506806]  wg_socket_send_skb_to_peer+0x98/0xb0 [wireguard]
[27929.506860]  wg_packet_tx_worker+0xa9/0x210 [wireguard]

Here, a wireguard encrypted udp packet is being sent to an endpoint that 
then is being routed to a wireguard interface. What in your network 
config would make that possible?

Jason