On Tue, 2013-08-20 at 21:02 +0200, Johannes Berg wrote:

On Tue, 2013-08-20 at 10:28 +0200, Johannes Berg wrote:

quoted

The only way to fix this that I see right now (that doesn't rewrite the
locking completely) would be to make genetlink use parallel_ops itself,
thereby removing the genl_lock() in genl_rcv_msg() and breaking all
those lock chains that lockdep reported. After that, it should be safe
to use genl_lock() inside all the operations. Something like the patch
below, perhaps? Completely untested so far.

Tested now, and it still causes lockdep to complain, though that's a
lockdep issue I believe, it thinks that genl_mutex and nlk->cb_mutex can
be inverted although nlk->cb_mutex exists per family, so we need to
annotate lockdep there.

No, lockdep is correct - generic netlink uses the same cb_mutex for all
families, obviously, since it's all the same netlink family.

I'll just convert it to RCU.

johannes