unshare -n bash -c 'for i in $(seq 1 100);do ip link add foo$i type dummy;done'
Gives:

BUG: MAX_LOCK_DEPTH too low!
turning off the locking correctness validator.
depth: 48  max: 48!
48 locks held by kworker/u16:1/69:
 #0: ffff8880010b7148 ((wq_completion)netns){+.+.}-{0:0}, at: process_one_work+0x7ed/0x1350
 #1: ffffc900004a7d40 (net_cleanup_work){+.+.}-{0:0}, at: process_one_work+0xcf3/0x1350
 #2: ffffffff8bc6fbd0 (pernet_ops_rwsem){++++}-{4:4}, at: cleanup_net+0xab/0x7f0
 #3: ffffffff8bc8daa8 (rtnl_mutex){+.+.}-{4:4}, at: default_device_exit_batch+0x7e/0x2e0
 #4: ffff88800b5e9cb0 (&dev_instance_lock_key#3){+.+.}-{4:4}, at: unregister_netdevice_many_notify+0x1056/0x1b00
[..]

Work around this by splitting the list into lockdep-digestable sublists.
This patchset should have no effect whatsoever for non-lockdep builds.

This issue is a problem for me because of a recent test case added
to nftables userspace which will create/destroy 100 dummy net devices,
so when I run the tests on a debug kernel lockdep coverage is now lost.

Alternative suggestions welcome.

I did not yet encounter another code path that would take so many mutexes
in a row, so I don't see a reason to muck with task_struct->held_locks[].

Florian Westphal (2):
  net: core: move unregister_many inner loops to a helper
  net: core: split unregister_netdevice list into smaller chunks

 net/core/dev.c | 89 ++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 65 insertions(+), 24 deletions(-)

-- 
2.49.1