net: non-blocking host netdev I/O with an epoll poller

[0pcom]

The host netdev now uses non-blocking sockets and a small epoll-based poller (netpoll_native.go): when a syscall would block (EAGAIN), the goroutine parks on a channel until the fd is ready, its deadline passes, or the fd is closed. The last case is the important one — closing an fd now wakes every goroutine parked on it (real cancellation), so a blocked Accept/Recv/Send returns promptly on shutdown instead of pinning a thread and hanging teardown.

This replaces the previous blocking-syscall + SO_*TIMEO approach so that:

• a blocked read/accept parks the goroutine instead of pinning an OS thread;
• closing an fd unblocks every goroutine parked on it, which is what lets a graceful shutdown actually complete.

Follow-on to the host netdev that landed via #60.

Verified against tinygo dev (0.42.0): blocked Accept unblocks on Close; 4 concurrent connections doing echo round-trips; 20 sequential request/response round trips on one connection; and a full software-TLS handshake + HTTPS GET served through it (with the crypto/tls host change in the tinygo repo). Also validated by a real skycoin daemon serving its API and syncing blocks.

🤖 Generated with Claude Code

https://claude.ai/code/session_01C3X3gvq8ZggMRvzVzWm66i

[0pcom]

Pushed a second commit: deadline changes now interrupt I/O that is already parked in the poller.

The first version captured the deadline when Recv/Send entered the wait, so a concurrent SetReadDeadline had no effect until the next wakeup. The visible symptom was in net/http: abortPendingRead interrupts the server's background read by setting a past read deadline, and with that inert, an HTTP/1.1 response with Connection: close did not actually close the connection until the client gave up.

Mechanism: the conn deadline setters call the netdev's PollInterrupt (optional interface — netdevs without a poller are unaffected), which wakes parked waiters with an internal sentinel; Read/Write retry under the fresh deadline, a partially-sent Write continues from where it left off, and an interrupt that lands between deadline-capture and parking is remembered on the pollDesc so it cannot be missed.

Verified: an HTTP/1.1 Connection: close request now closes promptly (plain and TLS), and the earlier tests (accept-unblock-on-close, concurrent echo, 20 sequential round trips, TLS handshake + HTTPS GET) still pass.

[yohimik]

I audited the local Crier shutdown patch. Repeated Close on a TCP listener, TCP connection, or UDP connection can close a new socket that has reused the descriptor. A test now closes the old socket, installs a replacement at the same descriptor, calls Close again, and checks the replacement. All three cases failed before the guard. A separate test makes 32 concurrent Close calls and checks that only one reaches the netdev.

The guard and tests are available in a0c3ff5. The combined fork passes these tests and blocked Accept, Read, and Write tests on Linux arm64 and amd64 under emulation, using released TinyGo 0.43.0-net.1 with an isolated source root.

This is a limited Close guard, not an equivalent to this poller. It does not solve stale descriptor use by other operations or changes to deadlines during blocked I/O. The shutdown call stays a fork stopgap. I opened no duplicate shutdown PR. Please keep cancellation and deadline work here. The guard and regression tests are available to combine with your lifecycle work. #82 only implements ListenConfig and is independent of #78 and #80.