[inoway46]

Canceling a stream with queued outbound data leaves that data charged to maxSessionMemory. Subsequent requests on the same session can then fail even though the queued data has been discarded.

Release the remaining outbound memory accounting when cleaning up the stream's write queue, and add a regression test for session reuse after cancellation.

Found while investigating #66739. This does not change response-header admission or resolve the early-response issue itself.

Refs: #66739

[nodejs-github-bot]

Review requested:

• @nodejs/http2
• @nodejs/net

[mcollina]

lgtm

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/78712/

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 90.66%. Comparing base (98dbf60) to head (15bcc85).
⚠️ Report is 51 commits behind head on main.

Files with missing lines	Patch %	Lines
src/node_http2.cc	50.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##             main   #67113    +/-   ##
========================================
  Coverage   90.65%   90.66%            
========================================
  Files         797      797            
  Lines      278268   278793   +525     
  Branches    53517    53617   +100     
========================================
+ Hits       252261   252763   +502     
+ Misses      16386    16381     -5     
- Partials     9621     9649    +28     

Files with missing lines	Coverage Δ
src/node_http2.cc	82.05% <50.00%> (-0.08%)	⬇️

... and 42 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/78720/

[nodejs-github-bot]

Landed in 80a74e1

[mishushakov]

Could this get backport-requested-v24.x / backport-requested-v22.x (and land in v25.x)? We are hitting the leak this fixes in production through undici's HTTP/2 client.

Impact. undici ≥ 8 multiplexes up to 100 streams on one long-lived Http2Session. Every stream that is destroyed before its queued body is flushed (request timeout, AbortSignal, peer RST_STREAM, GOAWAY mid-upload) permanently loses its queued bytes from the session's maxSessionMemory budget. Once the budget is gone, Http2Stream::AddHeader fails for every later response on that session and Node resets each of them locally with NGHTTP2_ENHANCE_YOUR_CALM (surfaced by fetch as TypeError: fetch failed, cause ERR_HTTP2_STREAM_ERROR). A busy process never lets the session go idle, so it never recovers.

Measured on v24.22.2 (undici 8.12.3, maxSessionMemory: 1 to shorten the repro):

before abort : 50/50 small POSTs ok
one 16 MiB POST aborted after 300 ms
after abort  : 0/50 ok — every request RST(ENHANCE_YOUR_CALM), still 0/50 two seconds later

With the default 10 MB budget a single aborted 16 MiB upload loses 2–5 MB of it for the session's lifetime (checked by bisecting maxSessionMemory: 2 → poisoned, 5 → not).

Minimal repro (no undici, plain http2):

const http2 = require('node:http2')
const session = http2.connect('https://<any-h2-origin-accepting-uploads>', { maxSessionMemory: 1 })

function post (body, ms) {
  return new Promise((resolve) => {
    const req = session.request({ ':method': 'POST', ':path': '/upload' })
    req.on('response', (h) => resolve({ status: h[':status'] }))
    req.on('close', () => resolve({ rstCode: req.rstCode }))
    req.end(body)
    if (ms) setTimeout(() => req.destroy(), ms) // abort while body is still queued
  })
}

;(async () => {
  console.log(await post(Buffer.alloc(16 << 20), 300))      // { rstCode: 8 } CANCEL, expected
  console.log(await post(Buffer.alloc(4096)))                // { rstCode: 11 } ENHANCE_YOUR_CALM — leaked budget
  console.log(await post(Buffer.alloc(4096)))                // { rstCode: 11 } forever
})()

Context: e2b-dev/e2b SDK-425 — concurrent files.write() failing with ENHANCE_YOUR_CALM after moving from undici 7 (one stream per session) to undici 8 (100 streams per session).

[inoway46]

Thanks for the detailed report! Given the production impact, I agree this is worth backporting to the supported LTS lines.