unix: Wake sleeping operations immediately on scheduled callbacks.

[andrewleech]

Summary

On the unix port, time.sleep() and MICROPY_INTERNAL_WFE() were unresponsive to scheduled callbacks — time.sleep() used select() or sleep() with no wakeup mechanism, and WFE busy-polled with a 500us delay.

This adds a signal-based notification so that when something is scheduled (via MICROPY_SCHED_HOOK_SCHEDULED), a signal is sent to the process. A sig_atomic_t flag is set by both mp_hal_signal_event() and the signal handler. mp_unix_sched_sleep() uses pselect() to atomically unblock the signal and enter the wait, avoiding the TOCTOU race where signals for already-queued callbacks could be consumed before the sleep. The signal is blocked process-wide at init so only the thread inside pselect() can receive it, and pthread_sigmask is used on threaded builds for POSIX correctness.

time.sleep() is reworked to loop over mp_unix_sched_sleep() with elapsed-time tracking, processing pending callbacks each iteration. Negative values now raise ValueError to match CPython. WFE uses the same sleep primitive so it blocks properly and wakes on events rather than spinning.

Uses SIGRTMIN+7 where real-time signals are available, falling back to SIGURG.

Testing

Tested on the unix port (Linux), both standard and coverage variants. Windows codepath (_WIN32) is left unchanged — it keeps the existing select()/delay_us behaviour.

Trade-offs and Alternatives

Could have used a pipe/eventfd self-pipe pattern instead of signals, which would avoid any signal-number collision concerns. Signals are simpler and don't require managing an fd, and the chosen signal numbers avoid known conflicts with GC (SIGRTMIN+5) and thread terminate (SIGRTMIN+6).

The time.sleep() rewrite replaces the MICROPY_SELECT_REMAINING_TIME Linux-specific assumption about select() modifying the timeout struct with explicit mp_hal_ticks_ms() tracking, which is portable. Precision drops from microseconds to milliseconds which is acceptable for time.sleep().

Generative AI

I used generative AI tools when creating this PR, but a human has checked the code and is responsible for the description above.

[andrewleech]

/review

[mpy-reviewer]

The signal-based approach using pselect is the right design here — the atomic unblock-and-wait correctly eliminates the TOCTOU race. A few issues to address before merge.

[mpy-reviewer]

sa is a stack variable; sa_restorer (and any padding) is uninitialised. Use:

Suggested change

	struct sigaction sa;
	struct sigaction sa = {0};

[mpy-reviewer]

Use sigaction consistently rather than mixing with signal:

Suggested change

	signal(MP_SCHED_SIGNAL, SIG_DFL);
	struct sigaction sa = {0};
	sa.sa_handler = SIG_DFL;
	sigemptyset(&sa.sa_mask);
	sigaction(MP_SCHED_SIGNAL, &sa, NULL);

[mpy-reviewer]

"select()" → "pselect()".

[mpy-reviewer]

This truncates to milliseconds, so time.sleep(0.0001) becomes a no-op (total_ms=0 → loop exits immediately without sleeping). The old code had microsecond precision. At minimum, round up:

Suggested change

	uint64_t total_ms = (uint64_t)(val * MICROPY_FLOAT_CONST(1000.0));
	uint64_t total_ms = (uint64_t)MICROPY_FLOAT_C_FUN(ceil)(val * MICROPY_FLOAT_CONST(1000.0));

[github-actions]

Code size report:

Reference:  zephyr/mpconfigport: Remove duplicate builtins.open definition. [1ab9b66]
Comparison: unix: Use sched signal in MICROPY_INTERNAL_WFE. [merge of 3bd023c]
  mpy-cross:    +0 +0.000% 
   bare-arm:    +0 +0.000% 
minimal x86:    +0 +0.000% 
   unix x64: +1353 +0.158% standard[incl +40(data) +160(bss)]
      stm32:    +0 +0.000% PYBV10
      esp32:    +0 +0.000% ESP32_GENERIC
     mimxrt:    +0 +0.000% TEENSY40
        rp2:    +0 +0.000% RPI_PICO_W
       samd:    +0 +0.000% ADAFRUIT_ITSYBITSY_M4_EXPRESS
  qemu rv32:    +0 +0.000% VIRT_RV32