Add Async Handlers Support

[g41797]

Async handlers ("split" handlers) are based on the already existing async body reading flow

Split Handler Pattern

One handler proc, called twice:

handle :: proc(h: ^http.Handler, req: ^http.Request, res: ^http.Response) {
    if res.work_data == nil { // first call runs inside nbio.tick()
        work := new(My_Work, alloc)
        http.mark_async(h, res, work)
        // start background work .....
        return
    }

    // Second call - after exit from tick()
    work := (^My_Work)(res.work_data)
    defer {
        // free work struct
        res.work_data = nil
    }
    http.respond_plain(res, work.result)
}

res.work_data - shared data between two calls

Flow (includes body reading)

        IO thread                          Background thread
─────────────────────────────────────   ─────────────────────────
handler (first call)
  res.async_handler = h
  http.body(req, -1, res, body_callback)
  return

body_callback (IO thread)
  mark_async(res.async_handler, res, work)
  thread.start(background_proc) ----------------
  return                                       |
                                               V
                                        background processing
                                        result => work
                                        resume(res):
        [nbio] <---------------------(2)
          |        [mpsc queue]<-----(1)
          V             |
[after tick() returns]  |
                        V
              handler (second call)
                       respond

"Async" Changes

New:

• resume.odin — public API: mark_async, cancel_async, resume
• internal/mpsc/ — lock-free MPSC queue, used for resumed handlers

Modified:

• response.odin — async fields:

Response :: struct {
	// ------ async "header"
	using node:       list.Node,
	async_handler:    ^Handler, // handler to resume
	work_data:        rawptr, // non-nil means async is pending
	// ------

• server.odin

  • resume loop after nbio.tick() {gets handler from queue + runs it}
  • shutdown waits for pending async handlers (max 5s)

Non-Async Changes

conn_handle_req does nothing when closing is set, it allows safe close and prevents SIGBUS (macOS) and memory leaks (Windows):

conn_handle_req :: proc(c: ^Connection, ...) {
    if atomic_load(&c.server.closing) {return}
    ...
}

Sometimes the server hung during shutdown — added 1ms timeout to nbio.tick().

Examples

examples/async/ — three examples

All use thread.create per request (for illustration only)

Production code should use a worker pool or pipeline

Known problems

Docs generation on GitHub failed - Odin version dev-2026-04-nightly:0c9ea9b

Run ./docs/generate.sh
+ cd docs
+ rm -rf build
+ mkdir build
+ odin doc . -all-packages -doc-format
./docs/generate.sh: line 10:  2496 Segmentation fault      (core dumped) odin doc . -all-packages -doc-format
Error: Process completed with exit code 139

Local generation - ok - Odin version dev-2026-03-nightly