netfilter: nft_set_rbtree: use per-set rwlock to improve the scalability

netfilter: nft_set_rbtree: use per-set rwlock to improve the scalability

Karel Rericha reported that in his test case, ICMP packets going through
boxes had normally about 5ms latency. But when running nft, actually
listing the sets with interval flags, latency would go up to 30-100ms.
This was observed when router throughput is from 600Mbps to 2Gbps.

This is because we use a single global spinlock to protect the whole
rbtree sets, so "dumping sets" will race with the "key lookup" inevitably.
But actually they are all _readers_, so it's ok to convert the spinlock
to rwlock to avoid competition between them. Also use per-set rwlock since
each set is independent.

Reported-by: Karel Rericha <karel@unitednetworks.cz>
Tested-by: Karel Rericha <karel@unitednetworks.cz>
Signed-off-by: Liping Zhang <zlpnobody@gmail.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>