[v26.x backport] http2: reduce per-request allocations by mcollina · Pull Request #64663 · nodejs/node

mcollina · 2026-07-21T19:58:32Z

Backport the v26-compatible parts of #64265.

Included:

header construction and compat timeout-handler allocation reductions;
write-path closure reductions; and
response-option allocation reductions.

Excluded changes require the HTTP/2 stream-lifecycle work in #63249 or fail v26 regression tests: stream lifecycle/pending-stream changes, trailer-round-trip changes, native auto-empty trailers, and scheduled-callback reductions.

Tests:

make -j4
tools/test.py http2
test-h2-large-header-cause-client-to-hangup
test-h2leak-destroy-session-on-socket-ended
test-http2-many-writes-and-destroy

Cut several sources of per-stream/per-request overhead on the hot path: - Track 'priority'/'frameError' stream listeners by overriding the EventEmitter methods on Http2Stream instead of subscribing to 'newListener'/'removeListener', which made every listener add and remove on every stream emit an extra tracking event. - Replace the per-call SafeSet and sensitive-header mapping in buildNgHeaderString with a lazily allocated array and an empty-array fast path, and skip the HTTP token regex and connection-specific header checks for well-known single-value header names. - Replace per-call closures with shared named handlers in onStreamClose, afterShutdown and Http2Stream._destroy. - Skip the pendingStreams Set add/delete for streams that are created with their native handle already available (all server streams). - Hoist the per-request onStreamTimeout closure factories in the compat layer to module-level handlers, and avoid a once() wrapper allocation per server stream. h2load, 1 KiB response payload, -c 4 -m 100, mean of 6 alternating runs: core API 60.2k -> 69.3k req/s (+15%), compat API 43.6k -> 46.2k req/s (+5.9%). Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: nodejs#64265 Reviewed-By: James M Snell <jasnell@gmail.com> Reviewed-By: Gürgün Dayıoğlu <hey@gurgun.day> Reviewed-By: Tim Perry <pimterry@gmail.com>

Every _write()/_writev() on an Http2Stream allocated four closures and an anonymous nextTick callback to coordinate the write callback with the end-of-stream check. Since the stream machinery dispatches at most one write at a time, that coordination state can live on the stream's kState object instead, with shared named functions for the end check and completion logic. When trailers are pending the writable side cannot be shut down early anyway, so the end-of-stream check tick is now skipped entirely for those writes. Also pre-initialize the kState fields that used to be added dynamically (shutdownWritableCalled, fd) so hot-path stores no longer transition the object shape. h2load, 1 KiB response payload, -c 4 -m 100, mean of 6 alternating runs vs main: core API 61.0k -> 70.7k req/s (+15.9% cumulative), compat API 43.7k -> 50.4k req/s (+15.3% cumulative). Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: nodejs#64265 Reviewed-By: James M Snell <jasnell@gmail.com> Reviewed-By: Gürgün Dayıoğlu <hey@gurgun.day> Reviewed-By: Tim Perry <pimterry@gmail.com>

respond() copied the user-provided options object on every call just so it could normalize and locally flip options.endStream, and prepareResponseHeadersObject() then looked the :status and date fields up again on the dictionary-mode null-prototype headers copy it had just built. Use a local variable for endStream and pick up :status/date while copying the headers instead. No measurable throughput change on its own; this removes an object clone and several dictionary-mode property lookups per response. Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: nodejs#64265 Reviewed-By: James M Snell <jasnell@gmail.com> Reviewed-By: Gürgün Dayıoğlu <hey@gurgun.day> Reviewed-By: Tim Perry <pimterry@gmail.com>

nodejs-github-bot · 2026-07-21T19:58:37Z

Review requested:

@nodejs/http
@nodejs/http2
@nodejs/net

aduh95 · 2026-07-21T22:01:25Z

The commit message does not match the ones that landed on main, the dates are also different, can you fix?

mcollina · 2026-07-21T22:16:44Z

Updated the branch: the retained commits now preserve the messages and timestamps from main. I also reduced the backport to the v26-compatible hunks; the previously included lifecycle and trailer changes regress v26 HTTP/2 tests.

nodejs-github-bot · 2026-07-22T07:57:48Z

CI: https://ci.nodejs.org/job/node-test-pull-request/75001/

nodejs-github-bot · 2026-07-22T11:30:37Z

CI: https://ci.nodejs.org/job/node-test-pull-request/75008/

Ethan-Arrowood · 2026-07-23T21:11:14Z

~~@mcollina it seems like your commits are missing the verified signatures. That will prevent merging if I'm not mistaken.~~ nvm I think its fine as long as we use commit-queue

This reverts commit 3329647. PR-URL: #64663 Reviewed-By: Antoine du Hamel <duhamelantoine1995@gmail.com> Reviewed-By: Tim Perry <pimterry@gmail.com> Reviewed-By: Ethan Arrowood <ethan@arrowood.dev>

This reverts commit a44fca5. PR-URL: #64663 Reviewed-By: Antoine du Hamel <duhamelantoine1995@gmail.com> Reviewed-By: Tim Perry <pimterry@gmail.com> Reviewed-By: Ethan Arrowood <ethan@arrowood.dev>

Cut several sources of per-stream/per-request overhead on the hot path: - Track 'priority'/'frameError' stream listeners by overriding the EventEmitter methods on Http2Stream instead of subscribing to 'newListener'/'removeListener', which made every listener add and remove on every stream emit an extra tracking event. - Replace the per-call SafeSet and sensitive-header mapping in buildNgHeaderString with a lazily allocated array and an empty-array fast path, and skip the HTTP token regex and connection-specific header checks for well-known single-value header names. - Replace per-call closures with shared named handlers in onStreamClose, afterShutdown and Http2Stream._destroy. - Skip the pendingStreams Set add/delete for streams that are created with their native handle already available (all server streams). - Hoist the per-request onStreamTimeout closure factories in the compat layer to module-level handlers, and avoid a once() wrapper allocation per server stream. h2load, 1 KiB response payload, -c 4 -m 100, mean of 6 alternating runs: core API 60.2k -> 69.3k req/s (+15%), compat API 43.6k -> 46.2k req/s (+5.9%). Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: #64265 Backport-PR-URL: #64663 Reviewed-By: Antoine du Hamel <duhamelantoine1995@gmail.com> Reviewed-By: Tim Perry <pimterry@gmail.com> Reviewed-By: Ethan Arrowood <ethan@arrowood.dev>

Every _write()/_writev() on an Http2Stream allocated four closures and an anonymous nextTick callback to coordinate the write callback with the end-of-stream check. Since the stream machinery dispatches at most one write at a time, that coordination state can live on the stream's kState object instead, with shared named functions for the end check and completion logic. When trailers are pending the writable side cannot be shut down early anyway, so the end-of-stream check tick is now skipped entirely for those writes. Also pre-initialize the kState fields that used to be added dynamically (shutdownWritableCalled, fd) so hot-path stores no longer transition the object shape. h2load, 1 KiB response payload, -c 4 -m 100, mean of 6 alternating runs vs main: core API 61.0k -> 70.7k req/s (+15.9% cumulative), compat API 43.7k -> 50.4k req/s (+15.3% cumulative). Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: #64265 Backport-PR-URL: #64663 Reviewed-By: Antoine du Hamel <duhamelantoine1995@gmail.com> Reviewed-By: Tim Perry <pimterry@gmail.com> Reviewed-By: Ethan Arrowood <ethan@arrowood.dev>

respond() copied the user-provided options object on every call just so it could normalize and locally flip options.endStream, and prepareResponseHeadersObject() then looked the :status and date fields up again on the dictionary-mode null-prototype headers copy it had just built. Use a local variable for endStream and pick up :status/date while copying the headers instead. No measurable throughput change on its own; this removes an object clone and several dictionary-mode property lookups per response. Signed-off-by: Matteo Collina <hello@matteocollina.com> PR-URL: #64265 Backport-PR-URL: #64663 Reviewed-By: Antoine du Hamel <duhamelantoine1995@gmail.com> Reviewed-By: Tim Perry <pimterry@gmail.com> Reviewed-By: Ethan Arrowood <ethan@arrowood.dev>

aduh95 · 2026-08-04T08:23:46Z

Landed in 6fa2b2d...e443fde