* Fix data race for max connection limiting in proxy directive.
The Conns and Unhealthy fields are updated concurrently across all active
requests. Because of this, they must use atomic operations for reads and
writes.
Prior to this change, Conns was incremented atomically, but read unsafely.
Unhealthly was updated & read unsafely. The new test
TestReverseProxyMaxConnLimit exposes this race when run with -race.
Switching to atomic operations makes the race detector happy.
* oops, remove leftover dead code.
* Balance round robin evenly when some hosts are down
Before, when load balancing across multiple hosts, if a host went down
then the next host in line would be sent a double share of requests.
This is because the round robin counter was only incremented once per
request, regardless of the health of the selection. If current
selection was unhealthy then the policy would advance to the next host,
but this would not be reflected in the policy counter. To fix this, the
counter is now incremented for every attempted host.
This commit adds a test case that identifies the issue, and a fix.
* Make robin counter private
* Use a mutex to sync round robin selection
