We've had a series of bugs with resources that need
to be unique on the system across test runner
processes. Ports are used by neutron-server and the
OVS agent when run in native openflow mode. The function
that generates ports looks up random unused ports and
starts the service. However, it is raceful: By the time the
port is found to be unused and the service is started,
another test runner can pick the same random port.
With close to 65536 ports to choose from, the chance
for collision is low, but given enough test runs, it's
happened a non-trivial amount of times, and given that
a voting job needs a very low false-negative rate, we
need a more robust solution. The same applies to IP
addresses that are used by the OVS agent in tunneling
mode, and for the LB agent in all modes. With IP addresses,
we don't check if the IP address is used, we simply
pick a random address from a large pool, and again
we've seen a non-trivial amount of test failures.
The bugs referenced below had simple, short term solutions
applied but the bugs remain remain. This patch is a correct,
long term solution that doesn't rely on chance.
This patch adds a resource allocator that uses the disk
to persist allocations. Access to the disk is guarded
via a file lock. IP address, networks and ports fixtures
use an allocator internally.