When we receive a SYN flood from a limited set of hosts, the TCP hostcache
is the bottleneck. This is hash row lock contention just to read hostcache
The core of the change is that only update requests do mtx_lock. The read
requests use SMR section. This required some changes to writers as well,
e.g. directly reusing the least used entry in a row is no longer possible,
also updates need to be atomic.