Summary

#472 fixed the primary port-reservation race in libtest::Server::start(): get_free_port() now keeps its reservation socket bound until release_port() is called, and Server::start() releases its own _port once the server is confirmed up (or definitively failed).

That fix only releases the Server's own _port. Several call sites reserve a port for something other than a Server object's primary port and never released it — the reservation socket now leaks (stays bound) for the rest of that test binary's process lifetime:

• tests/context.h's Context, tests/cycle.cc's cycle_context_st, and tests/round_robin.cc's Context all overwrite their stored port in reset() with a fresh get_free_port() value without releasing the old one first, and never released the final port on destruction.
• tests/httpd_test.cc reserves http_port, passed only as --http-port= to a gearmand instance whose own primary port is default_port() — the Server never knows about http_port.
• tests/gearmand.cc's long_keepalive_start_TEST and config_file_SIMPLE_TEST reserve a port for --check-args validation runs that never actually bind it (no Server involved at all).
• libtest/unittest.cc's get_free_port_TEST exercises get_free_port() directly and never releases what it reserves.

Fix

Release the port explicitly wherever the above patterns discard/overwrite/finish with a reservation:

• reset()-style methods now call release_port() on the old port before generating a new one; destructors now release the current port directly instead of wastefully fetching a new one just to abandon it (round_robin.cc's destructor previously called reset(), which did exactly that).
• httpd_test.cc releases http_port right after server_startup() returns (success or failure) — by then gearmand has either bound it or failed outright.
• gearmand.cc's two --check-args tests release the port immediately after formatting it into the CLI/config argument, since --check-args never binds.
• get_free_port_TEST releases each of its throwaway reservations at the end.

default_port()'s single process-wide reservation is intentionally left untouched — callers throughout the test suite rely on it staying valid for the whole process lifetime, so releasing it would be a behavior change, not a leak fix.

Test plan

• Full rebuild of all affected binaries (t/unittest, t/cycle, t/round_robin, t/gearmand, t/httpd, plus everything else that includes tests/context.h: t/drizzle, t/memcached, t/redis, t/mysql, t/tokyocabinet, t/postgres, t/ephemeral) — all compile clean.
• Ran every affected binary individually — all pass (backends without a local install, e.g. drizzle/memcached/mysql/postgres, skip as expected; t/unittest's only failures are pre-existing environment artifacts unrelated to this change — getenv() checks that only pass under make check's harness, and a /bin/echo-path issue specific to this NixOS dev machine).
• Compared 4x-concurrent t/cycle and a heavier 16-way mix (protocol/worker/round_robin/cycle) against the pre-fix baseline on the same machine: an occasional t/cycle flake under the heaviest combination reproduces identically with or without this change, confirming it's pre-existing high-load contention in this environment, not a regression from this PR.

Fixes #474

🤖 Generated with Claude Code