http2: async iteration reports premature close with buffered responses in v26.3.0+ and v24.20.0

Version

v26.8.1 and v24.20.0 (official binaries). The regression starts between v26.2.0 and v26.3.0; v24.18.0 also passes. See the comparison below.

Platform

macOS, Darwin 25.6.0, arm64.

Subsystem

http2, stream (eos() and Readable async iteration)

What steps will reproduce the bug?

Save this as repro-http2-premature-close.mjs and run it with node repro-http2-premature-close.mjs. It uses only Node built-in modules, localhost, and an ephemeral port. No proxy or third-party dependency is needed.

// Run with an unmodified Node binary: node repro-http2-premature-close.mjs
// Uses only built-in modules, localhost, and an ephemeral port.
import assert from 'node:assert/strict';
import { once } from 'node:events';
import http2 from 'node:http2';
import { setImmediate } from 'node:timers/promises';

const server = http2.createServer();
server.on('stream', (stream) => {
  stream.resume();
  stream.on('end', () => {
    stream.respond({ ':status': 200 }, { waitForTrailers: true });
    stream.on('wantTrailers', () => stream.sendTrailers({ 'grpc-status': '0' }));
    stream.end('complete response');
  });
});
server.listen(0, '127.0.0.1');
await once(server, 'listening');
const session = http2.connect(`http://127.0.0.1:${server.address().port}`);

try {
  await once(session, 'connect');
  const request = session.request({ ':method': 'POST' });
  const response = once(request, 'response');
  request.end('request');
  await response;

  // Delay consuming the response until HTTP/2 reports a normal stream close.
  // The complete response body is still buffered, and the stream is not destroyed.
  const deadline = Date.now() + 2_000;
  while (!request.closed && Date.now() < deadline) await setImmediate();
  assert.equal(request.closed, true);
  assert.equal(request.rstCode, 0);
  assert.equal(request.destroyed, false);
  assert.equal(request.readableLength, Buffer.byteLength('complete response'));

  let body = '';
  for await (const chunk of request) body += chunk.toString();
  assert.equal(body, 'complete response');
  console.log(`${process.version}: PASS`);
} finally {
  session.destroy();
  await once(session, 'close');
  server.close();
  await once(server, 'close');
}

How often does it reproduce? Is there a required condition?

The equivalent direct-connection test failed on all 20 attempts in each affected version listed below. The required condition is that iteration starts after the HTTP/2 stream closes normally, while the response body is still buffered.

Before reading, the reproduction asserts closed === true, rstCode === 0, destroyed === false, and that all 17 response bytes remain in readableLength. It does not destroy the stream before consuming it; session destruction is only in cleanup.

Without intentionally waiting for this state, the failure depends on scheduling. Downstream gRPC calls through proxy/Duplex transports exposed it, but the reproduction removes both the proxy and the gRPC library.

What is the expected behavior? Why is that the expected behavior?

for await consumes complete response successfully and finishes without an error. The process exits with status 0, as it does on v24.18.0 and v26.2.0. The HTTP/2 stream closed normally and its complete response is still readable.

What do you see instead?

for await throws ERR_STREAM_PREMATURE_CLOSE; the process exits with status 1. On v26.8.1:

Error [ERR_STREAM_PREMATURE_CLOSE]: Premature close
    at getEosOnCloseError (node:internal/streams/end-of-stream:98:14)
    at eos (node:internal/streams/end-of-stream:159:23)
    at start (node:internal/streams/readable:1425:15)
    at Object.next (node:internal/streams/readable:1574:23)

v24.20.0 fails with the same error through getEosOnCloseError / eos / createAsyncIterator.

Additional information

Version comparison for the equivalent direct-connection reproduction, delaying consumption until request.closed:

Node version	Successful responses / attempts
24.18.0	20 / 20
24.20.0	0 / 20
26.2.0	20 / 20
26.3.0	0 / 20
26.7.0	0 / 20
26.8.1	0 / 20

The single-case script above was also verified separately with v24.18.0 and v26.2.0 (exit 0), and with v24.20.0, v26.3.0 and v26.8.1 (exit 1).

Suspected regression mechanism: the version boundary and source comparison point to #62394, cbee0de1cb, released in v26.3.0. The same change was backported as 5658c631fa in v24.20.0.

Http2Stream.closed can be true after a normal protocol close while the readable body is still buffered. eos() now computes immediateResult before that buffer is consumed, then reports the captured premature-close error in a subsequent process.nextTick() callback. The previous implementation evaluated the close error on the next tick instead.

A diagnostic comparison on v26.8.1, keeping its async iterator unchanged and only recomputing the close result inside the next-tick callback, changes the direct-connection test from 0/20 to 20/20 successful responses. That is evidence for the timing regression, not a proposed production monkey patch.

Related reports checked: #63989 and #64098 concern HTTP/1.1 keep-alive / node-fetch. This reproduction uses core HTTP/2 with a normal protocol close, and also fails on v26.8.1 after those reports were resolved. #44866 is the older Web Stream termination hang addressed by #62394, rather than this new async-iteration failure.

記載されている Node バージョンで repro-http2-premature-close.mjs を実行し、その後、eos() と createAsyncIterator の周辺にある node:internal/streams/end-of-stream および node:internal/streams/readable を調査します。#62394 で導入された動作と、そのバックポートを比較します。バッファーされたデータを含む正常に閉じられた HTTP/2 ストリームが、ERR_STREAM_PREMATURE_CLOSE なしで for await によって完全に消費されれば完了です。