Follow-ups for the background worker stack

[nicolas-grekas]

Things left to do or decide after #2617 and its stack, so they don't get lost.

• Validate a configuration before stopping the running one: Start() calls Init() after the old app stopped, so any declaration error leaves the site down while Caddy rolls the configuration back. A side-effect-free preflight shared with Init(), called from Validate(). feat: declared background workers + frankenphp_get_worker_handle() #2617 (comment)

• Synchronize the startup-failure boundary: publishing ShuttingDown wakes the startup waiter before the thread finished its exit path, and a shutdown request is not distinguished from a thread that exited. Both worker handlers share that path. feat: declared background workers + frankenphp_get_worker_handle() #2617 (comment)

• Pull the rest of the common bookkeeping into workerLifecycle once that boundary has one shape. feat: declared background workers + frankenphp_get_worker_handle() #2617 (comment)

• num_regular_threads / max_regular_threads, so nobody computes the thread budget by hand. feat: declared background workers + frankenphp_get_worker_handle() #2617 (comment)

• Implement Io\Poll\Handle on FrankenPHP\WorkerHandle and the task handles, done in feat: declared background workers + frankenphp_get_worker_handle() #2617 and feat: SentTaskHandle and ReceivedTaskHandle #2636 on top of the 8.6 job of ci: run the tests against PHP 8.6 as an experimental job #2658

• Make the per-task php_stream of feat: SentTaskHandle and ReceivedTaskHandle #2636 optional: the state of each side moved onto its handle, the task settles in the handle's cleanup, and getStream() builds the stream on demand, so a script waiting through a context allocates neither a stream nor a resource per task. Done in feat: SentTaskHandle and ReceivedTaskHandle #2636.

• Recycle the task handles: measured and dropped. A per-thread pool of the per-side state benches flat against feat: SentTaskHandle and ReceivedTaskHandle #2636 (39.1 -> 38.7us a task, throughput within 1.4% either way, 15.4 syscalls on both sides, fork run 35466429028), since the engine already recycles the object slot and what is left is one 72-byte block per side.

• Give macOS the channel Linux gets from eventfd: done in feat: SentTaskHandle and ReceivedTaskHandle #2636 with a kqueue descriptor carrying an EVFILT_USER event, one per side. Measured on a macos-latest runner before writing it, per wake-up round trip: socketpair 16.1us, two pipes 14.0us, kqueue 11.7us, the same quarter eventfd takes off the pair on Linux (23.1 -> 17.5us on the same runner class). Windows keeps the socket pair for php_select().

[nicolas-grekas]

FTR for the last item, measured on a runner with the A/B workflow on my fork, alternating two refs over 4 rounds: a task round trip is ~38us, of which ~25us is the wait for the pickup, ie the two thread wake-ups. Dropping the per-task stream targets what's left, the fd syscalls already went with the pool. Careful with laptop-grade VMs for this, mine inverts the verdict on anything that moves wake-ups around.

Two things that gave exactly nothing, so nobody retries them: a state word written by Go in C memory instead of the await callback, and the park handshake with shared atomics. cgo callbacks aren't the cost, goroutine spawns are.