Kernel BUG on network namespace deletion

* Kernel BUG on network namespace deletion
@ 2024-01-15 13:59 Марк Коренберг
  2024-01-19  0:49 ` Jakub Kicinski
  0 siblings, 1 reply; 6+ messages in thread
From: Марк Коренберг @ 2024-01-15 13:59 UTC (permalink / raw)
  To: netdev

Kernel 6.6.9-200.fc39.x86_64

The following bash script demonstrates the problem (run under root):

```
#!/bin/bash

set -e -u -x

# Some cleanups
ip netns delete myspace || :
ip link del qweqwe1 || :

# The bug happens only with physical interfaces, not with, say, dummy one
ip link property add dev enp0s20f0u2 altname myname
ip netns add myspace
ip link set enp0s20f0u2 netns myspace

# add dummy interface + set the same altname as in background namespace.
ip link add name qweqwe1 type dummy
ip link property add dev qweqwe1 altname myname

# Trigger the bug. The kernel will try to return ethernet interface
# back to root namespace, but it can not, because of conflicting
# altnames.
ip netns delete myspace

# now `ip link` will hang forever !!!!!
```

I think, the problem is obvious. Although I don't know how to fix.
Remove conflicting altnames for interfaces that returns from killed
namespaces ?

On kernel 6.3.8 (at least) was another bug, that allows duplicate
altnames, and it was fixed mainline somewhere. I have another script
to trigger the bug on these old kernels. I did not bisect.

````
[  494.473906] default_device_exit_net: failed to move enp0s20f0u2 to
init_net: -17
[  494.473926] ------------[ cut here ]------------
[  494.473927] kernel BUG at net/core/dev.c:11520!
[  494.473932] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
[  494.473935] CPU: 3 PID: 3852 Comm: kworker/u32:17 Not tainted
6.6.9-200.fc39.x86_64 #1
[  494.473938] Workqueue: netns cleanup_net
[  494.473944] RIP: 0010:default_device_exit_batch+0x295/0x2a0
```

-- 
Segmentation fault

^ permalink raw reply	[flat|nested] 6+ messages in thread