Phil Sutter <phil@xxxxxx> wrote:
> So it is possible that ops are still used somewhere after
> nf_unregister_net_hook() returns? I don't quite get that code, with all
> the RCU-bells and ONCE-whistles.
Not ops, but other core could still run the hookfn that
is being unregistered, i.e.:
cpu 0
unreg hookfn nf_hook_slow
runs hookfn
if nf_hook() fetched the hook blob (struct nf_hook_entries *)
before unreg path called its rcu_assign_pointer with the updated
incarnation.
That said, this might be a special case as entire nf_hook_entries blob
is tied to device that is going down, so packet flow might have stopped
already.
