[gabedemesa]

This relates to...

Refs #6181, nodejs/node#67363

Rationale

abort() in lib/dispatcher/client-h2.js closes the stream with the default NO_ERROR code and then destroys it synchronously. When the request body is still queued, Node defers a NO_ERROR RST_STREAM until the body has been written, and the synchronous destroy means it is never sent. On affected Node versions this leaves the stream open in nghttp2, which can then loop on empty DATA frames until the process runs out of memory (nodejs/node#67363, fixed in Node by nodejs/node#67372).

NGHTTP2_CANCEL is the code for a stream that is no longer needed, and Node submits it immediately, so aborts no longer depend on the Node fix.

Changes

Features

N/A

Bug Fixes

• Close aborted HTTP/2 streams with NGHTTP2_CANCEL, and correct the comment that said close() had already queued the RST_STREAM.
• Add test/http2-abort-rst-cancel.js: aborting a request whose body is held by flow control must reset the stream with CANCEL. Without the change the peer sees code 0.
• Document in docs/docs/api/Client.md that an aborted HTTP/2 request resets its stream with CANCEL.

Breaking Changes and Deprecations

N/A

Testing

On Node v26.12.1, the reproduction from #6181 (concurrent fetch() uploads aborted mid-body) runs away with main and stays bounded with this change. ESLint is clean. Locally, test/h2-disconnect-guard.js, test/http2-max-concurrent-streams-zero-unref.js, test/http2-dispatcher.js and test/http2-connection.js fail or time out identically on main, so I could not use them to check this change; the other HTTP/2 tests I ran pass.

Benchmarks

The change only runs when a request is aborted.

Aborted uploads (32 KB fetch() POSTs over HTTP/2, aborted on the next tick, 20,000 requests, 32 concurrent, three runs each). This used Node main with nodejs/node#67372 applied, because main of undici runs away on released Node versions in this scenario:

	Run 1	Run 2	Run 3	RSS
main	11,669 req/s	12,131 req/s	11,406 req/s	267 MB
This PR	12,281 req/s	11,816 req/s	11,629 req/s	267 MB

benchmarks/benchmark-http2.js (SAMPLES=30 CONNECTIONS=50, Node v26.12.1), which does not abort requests:

Test	main	This PR
native - http2	6,627.83 req/s ± 4.30%	6,737.49 req/s ± 3.47%
undici - fetch	8,615.87 req/s ± 5.24%	8,840.41 req/s ± 1.80%
undici - pipeline	10,952.08 req/s ± 2.75%	11,245.60 req/s ± 2.80%
undici - dispatch	13,022.13 req/s ± 2.52%	13,398.57 req/s ± 2.64%
undici - stream	13,579.04 req/s ± 2.91%	14,894.53 req/s ± 2.96%
undici - request	13,630.19 req/s ± 0.97%	10,844.58 req/s ± 7.13%

The differences go in both directions and are run-to-run variation; the changed code is not reached in this benchmark. Both on macOS arm64.

Status

• I have read and agreed to the Developer's Certificate of Origin
• Tested
• Benchmarked (optional)
• Documented
• Review ready
• In review
• Merge ready

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.67%. Comparing base (691eb16) to head (3808fd1).
⚠️ Report is 7 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6201      +/-   ##
==========================================
- Coverage   91.68%   91.67%   -0.01%     
==========================================
  Files         113      113              
  Lines       44449    44442       -7     
==========================================
- Hits        40751    40744       -7     
  Misses       3698     3698              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 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.

[gabedemesa]

Thanks for the review! This is approved and CI is green. Could a collaborator land it when convenient?