modlwip: socket timeout ignored in send when out of memory #19746

Port, board and/or hardware

Not port-specific: generic lwip path in extmod/modlwip.c, every MICROPY_PY_LWIP port. Measured on OpenMV RT1062 (i.MX RT1062, cyw43 WiFi).

MicroPython version

master at a129b2f, by reading extmod/modlwip.c. Measured on OpenMV release firmware v5.0.0, MicroPython v1.28.0-49, built 2026-07-02, 8 KB lwIP heap.

Reproduction

1. settimeout(0.5) socket sends to a slow peer; a second socket to another slow peer holds the lwIP heap in unacked segments.
2. tcp_write returns ERR_MEM while tcp_sndbuf still reports room.
3. sock.send(buf) blocks in lwip_tcp_send past 0.5 s.

One-file script, board and host side: #19705, "Test", phase 2.

Expected behaviour

sock.send() raises OSError(ETIMEDOUT) at the socket timeout, as the tcp_sndbuf == 0 wait in the same function already does.

Observed behaviour

The ERR_MEM for (;;) loop in lwip_tcp_send never reads socket->timeout. It ends on tcp_write success or at 10 s with ENOMEM. On master it waits in poll_sockets(); on the v1.28.0 line in mp_hal_delay_ms(50).

Two readers at 20 KB/s, settimeout(0.5), 120 s: starved send() took 1651–3401 ms; 0 ETIMEDOUT, 0 ENOMEM.

Additional Information

Timed-socket case of #19704. Fixed by the second commit of #19705: ETIMEDOUT at the socket timeout, counted from call start; the 10 s limit stays for a socket with no timeout. Same test with the change (16 KB heap): 172 ETIMEDOUT, 0 ENOMEM, longest send() 504 ms.

Code of Conduct

Yes, I agree.