net: add a default host netdev backed by raw Linux syscalls

[0pcom]

Summary

On the native linux target the net package defaults to a NOP netdev, so every
operation returns Netdev not set and net.Dial/Listen, DNS, and anything built on
them (net/http, tls) fail at runtime. This adds a default netdev for native Linux that
implements the netdever interface directly on top of raw kernel sockets, so the net
package works on a regular Linux host with no network driver.

Addresses #28.

Why this approach

The earlier attempt in #28 tried to drop TinyGo's net package and pull in the upstream
stdlib net, which stalled on the runtime netpoller:

ld.lld: undefined symbol: internal/poll.runtime_pollReset
ld.lld: undefined symbol: internal/poll.runtime_pollWait
ld.lld: undefined symbol: internal/poll.runtime_pollSetDeadline

This PR keeps the existing lightweight netdev-based net package and instead supplies the
one missing piece — a netdev for the host — which sidesteps that blocker entirely. It works
because:

• Native Linux does not override the syscall package (loader/goroot.go), and the
  TinyGo compiler lowers syscall.Syscall/RawSyscall into real system calls
  (compiler/syscall.go), so the stdlib socket functions
  (syscall.Socket/Connect/Bind/Listen/Accept/...) are available and work directly. musl's
  omitted src/network module is not needed.
• The native Linux scheduler is threads, so blocking socket syscalls per goroutine are
  safe. This deliberately avoids the internal/poll netpoller dependency above.

What's included

• netdev_native.go — hostNetdev implementing the full netdever interface
  (Socket/Bind/Connect/Listen/Accept/Send/Recv/Close/SetSockOpt/Addr) over the syscall
  package, registered as the default via init(). Deadlines map to
  SO_RCVTIMEO/SO_SNDTIMEO (re-armed per write iteration so a deadline bounds the whole
  Send); Send loops over short writes; EOF and timeout errors are surfaced
  (timeoutError satisfies net.Error); Connect retries on EINTR.
  GetHostByName resolves IP literals, then /etc/hosts, then a small UDP DNS resolver
  driven by /etc/resolv.conf (the reply is validated against the query ID and the
  response/RCODE bits). Build-tagged
  linux && !baremetal && !nintendoswitch && !wasm_unknown && !tinygo.wasm, so only the
  native Linux target is affected — embedded/WASM netdevs are untouched.
• lookup.go — restores the standard net.DNSError type used by resolution errors.

Verification

Built and run with tinygo on linux/amd64:

• net.Listen + Accept + net.Dial on 127.0.0.1 round-trips (the case from
  TinyGo build of the default binary: ~3.5× smaller, offline commands work; blockers documented skycoin/skycoin#2902).
• DNS resolution + outbound TCP to example.com:80 with deadlines — 200 OK.
• http.Get("http://example.com/") — 200 OK.
• An IPv6 literal ([::1]) returns a clean "only IPv4 is supported" error rather than
  panicking.
• A stock build (no netdev), same program, reproduces ... failed: Netdev not set.

Scope / caveats

• Blocking sockets, one OS thread per blocked connection (no epoll netpoller). Correct, and
  fine for modest connection counts; not intended to match upstream Go's scalability at very
  high connection counts.
• IPv4 only, matching the rest of the TinyGo net package. IPv6 would require broadening
  the package's address handling (e.g. DialTCP's len(IP) != 4 guard) and is left as a
  follow-up.
• The DNS resolver is a minimal stub (A records, first answer, no search-domain handling);
  sufficient for typical client/daemon use, not a full resolver.

Happy to add tests or adjust the default-registration approach / build constraints to taste.

[b0ch3nski]

Also see: https://github.com/niemeyer/muslnet

[deadprogram]

Also see: https://github.com/niemeyer/muslnet

OK that looks very interesting indeed cc @soypat

[soypat]

DNSError and a lot of other logic has been added. Please rebase branch

net/dnserror.go

Line 10 in d5da3dd

type DNSError struct {

[0pcom]

Rebased on main. You were right that #63 landed on top of this: it added DNSError in dnserror.go, so the copy this branch was carrying in lookup.go is now redundant and has been dropped in favour of yours — the netdev only constructs *DNSError values, and the field set it uses is identical. errNoSuchHost also comes from main now rather than from here. The diff is back to two files: lookup.go and the new netdev_native.go.

On niemeyer/muslnet — thanks, that is genuinely relevant prior art and I had not seen it. It reaches the same place from the other direction: linking musl to get a real libc resolver and socket layer, where this uses raw syscalls plus a small /etc/hosts and /etc/resolv.conf resolver, and so needs no C at all. Happy to be told the musl route is the one you would rather carry — it would mean deleting most of this — but the two do not conflict, so I have left this as-is pending your call.

Worth noting these sources are not a standalone Go module, so I could not build them outside TinyGo. The rebase was mechanical (dropping a duplicate type), and I checked for duplicate top-level declarations across the package and for the identifiers the dropped block used.

[0pcom]

This landed already: #60 was stacked on this branch, so merging it brought 7bf471a ("net: add default host netdev backed by raw Linux syscalls") onto main along with the IPv6 commit on 2026-08-20.

Rebasing confirms it — the commit is detected as already applied and the branch comes out empty against main. Closing rather than leaving a conflicting duplicate open. Thanks for the reviews.