feat: frankenphp_send_task(), frankenphp_receive_task() and friends

[nicolas-grekas]

Stacked on #2617 and #2635: the first two commits are those PRs', review the third one (4c5af56).

The task half of #2319, rewritten on the background workers of #2617 and the shared vars of #2635. A request, an HTTP worker or another background worker hands work to a named background worker with frankenphp_send_task(string $name, array $payload, ?float $timeout = 30.0): resource and reads what comes back with frankenphp_read_task($stream): ?array; the worker dequeues with frankenphp_receive_task(): ?array and answers with frankenphp_update_task($stream, array $data): void, fclose() completing the task. Names resolve like frankenphp_get_vars() does, payloads and updates follow the set_vars() whitelist and travel as persistent tables through the Go side. Roughly 280 lines of Go and 330 of C.

The handle of #2617 is the only stream the worker needs: each task sent writes a task\n line to the handle of every thread of the worker, so the loop is fgets() on the handle, EOF still meaning drain, and stream_select() users keep a single stream. The line is a wake-up, not a count: in a pool the first thread back in its loop takes the task and the others get null from receive_task(), so the documented loop drains the queue on each wake-up.

send_task() blocks until a thread has picked the task up and throws on timeout, so a busy worker pushes back on its senders instead of queueing without bounds; null waits forever. Tasks queued while a thread restarts are signaled again on its next run, and the wait ends when the sender's own thread is drained for a restart or the shutdown, since the target's threads are drained too. The sender gets a socket stream over a pair opened per task (the Windows php_select() rule of #2617), one byte per update and EOF at completion, so stream_select() bounds the wait or multiplexes tasks and a blocking read parks as well; closing it abandons the task, which the worker's stream reports as EOF to stream_select() and feof(), and the worker's next update throws. The receiver's stream completes the task on close, unless that close is the resource cleanup of request shutdown, which means the script ended with the task open: the sender's next read throws instead of returning null. Sixteen updates are buffered per task, past that update_task() waits for the sender to read.

Waking threads is what a task costs, so wake-ups are kept to a minimum. A send wakes one parked thread of the worker, round-robin, with the line on its handle; a thread that reads its handle while tasks are queued gets the line from the read op itself, so no wake-up is lost whichever loop shape the script uses, and after 10ms without pickup every thread is woken as a fallback. The sender waits for the pickup in the kernel, on its end of the task's pair, rather than in a Go select: waking a PHP thread parked inside a cgo callback costs Go a P hand-off, a byte on a socket does not. The queue mutex is never held across a syscall and taken once per wake-up, since a thread inside a cgo callback that loses it parks the same expensive way. Benchmarked in the Docker builder image (the harness stayed out of the PR): a task went from 549 to 285µs with one thread, a pool of 8 from 1745 to 320µs, and 8 senders on 8 threads from 2k to 32k tasks/s.

Each side of a task waits on its own descriptor of the task's channel, an eventfd on Linux, one end of a socket pair elsewhere for Windows's php_select(): the streams carry no data, they are what stream_select() waits on and what fclose() ends, the Go side holds the state the functions report. The descriptors belong to the task until both sides closed, then the pair is drained and pooled, so a task costs no socketpair, fcntl or close: about 12 syscalls instead of 18, 13% off the latency and up to a third more throughput under load in the same measurement.

Tasks are referenced from the streams through cgo handles and freed once both sides closed, or by the sender when nobody picked the task up. The stop sockets of a worker's threads are now guarded by its task-queue mutex, since senders write to them.

Compared to #2319: no queue ahead of pickup and no cancellation before it, no dedicated signaling stream, no global task table, no metrics yet.

[nicolas-grekas]

Pushed 23d3094 after benchmarking the task path in the Docker builder image (harness kept out of the PR). Three changes: a send wakes one parked thread instead of every thread of the pool; the sender waits for the pickup in the kernel, on the task's socket, instead of in a Go select (waking a thread parked inside a cgo callback costs the scheduler a P hand-off); and the queue mutex is no longer held across a syscall nor taken twice per wake-up. One thread: 549 to 285µs per task. A pool of 8: 1745 to 320µs. 8 senders on 8 threads: 2k to 32k tasks/s. Absolute numbers are that VM's, the ratios are the point.