We can make all dumps and lookups lockless.

Dumps currently only hold the nfnl mutex on the dump request itself.
Dumps can span multiple syscalls, dump continuation doesn't acquire the
nfnl mutex anywhere, i.e. the dump callbacks in nf_tables already use
rcu and never rely on nfnl mutex being held.

So, just switch all dumpers to rcu.

This requires taking a module reference before dropping the rcu lock
so rmmod is blocked, we also need to hold module reference over
the entire dump operation sequence. netlink already supports this
via the .module member in the netlink_dump_control struct.

For the non-dump case (i.e. lookup of a specific tables, chains, etc),
we need to swtich to _rcu list iteration primitive and make sure we
use GFP_ATOMIC.

This patch also adds the new nft_netlink_dump_start_rcu() helper that
takes care of the get_ref, drop-rcu-lock,start dump,
get-rcu-lock,put-ref sequence.

The helper will be reused for all dumps.

Rationale in all dump requests is:

 - use the nft_netlink_dump_start_rcu helper added in first patch
 - use GFP_ATOMIC and rcu list iteration
 - switch to .call_rcu

... thus making all dumps in nf_tables not depend on the
nfnl mutex anymore.

In the nf_tables_getgen: This callback just fetches the current base
sequence, there is no need to serialize this with nfnl nft mutex.

Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>