child_process: serialize advanced IPC messages natively

[anonrig]

The advanced child_process IPC serialization codec was implemented in
JavaScript (ChildProcessSerializer / ChildProcessDeserializer in
lib/internal/child_process/serialization.js). It allocated a wrapper
serializer/deserializer per message and crossed the JS/C++ boundary several
times for every message (writeHeader, writeValue, releaseBuffer,
readHeader, readValue, …).

This moves the codec into a native ipc_serdes binding that drives the V8
ValueSerializer/ValueDeserializer with a C++ delegate. The wire format is
preserved byte-for-byte: a big-endian uint32 length prefix followed by the
V8 payload, with ArrayBufferViews tagged as host objects so that Node
Buffers round-trip as Buffers rather than plain Uint8Arrays. The json
codec is intentionally left unchanged (its hot path, JSON.stringify/parse,
is already native).

Performance

Measured A/B on identical built arm64 release binaries (only this change
differs), benchmark/child_process/child-process-ipc-roundtrip.js,
round-trips/sec, average of 3 runs:

Payload	Before (JS)	After (native)	Speedup
64 B	~300,000/s	~800,000/s	2.7× (+166%)
1 KiB	~272,000/s	~616,000/s	2.3× (+126%)
16 KiB	~91,000/s	~120,000/s	1.3× (+32%)
64 KiB	~30,000/s	~35,000/s	1.16× (+16%)

json mode is unchanged within noise (~543k → ~554k at 1 KiB).

The gain is largest for small messages, where the fixed per-message JavaScript
overhead (per-message serializer/deserializer allocation and the JS/C++
boundary crossings) dominated. It tapers for large messages, where the actual
serialization — already native in both versions — dominates. These are
codec/IPC-throughput numbers from a saturated round-trip; a real fork()
workload also pays for pipe I/O, the event loop and the user message handler,
so application-level gains will be smaller.

Verification

• test/parallel/test-child-process-* and test/parallel/test-cluster-*
  (190+ tests) pass, including advanced-serialization, -largebuffer,
  -splitted-length-field and fork-advanced-header-serialization.
• New cctest test/cctest/test_node_ipc_serdes.cc covers serialize/deserialize
  round-trips for primitives, objects, typed arrays and Buffers (including the
  Buffer-vs-Uint8Array distinction) and asserts the length-prefix framing; the
  full cctest suite passes.
• cpplint, git-clang-format, eslint and tsc --strict (typings) are clean.

There is no observable behavior change and the IPC wire format is unchanged.

[nodejs-github-bot]

Review requested:

• @nodejs/gyp

[anonrig]

Benchmark CI: https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1871/

[mcollina]

why can’t we use V8 serdes?

[panva]

There is no observable behavior change and the IPC wire format is unchanged.

There are observable changes tho:

repro.js

// repro.js// Run: node repro.js'use strict';const{ fork }=require('node:child_process');constfs=require('node:fs');constv8=require('node:v8');const{ MessageChannel }=require('node:worker_threads');if(process.argv[2]==='inspect'){process.on('message',(value)=>{process.send({isBuffer: Buffer.isBuffer(value),constructorName: value?.constructor?.name,keys: Object.keys(value),visible: value?.visible,bytes: ArrayBuffer.isView(value) ?
Buffer.from(value.buffer,value.byteOffset,value.byteLength).toString('hex') :
undefined,});});}elseif(process.argv[2]==='bad-tag'){classBadTagSerializerextendsv8.DefaultSerializer{_writeHostObject(value){this.writeUint32(2);// old child_process codec only accepts 0 or 1returnsuper._writeHostObject(value);}}constser=newBadTagSerializer();ser.writeHeader();ser.writeValue(Buffer.from('x'));constpayload=ser.releaseBuffer();constframed=Buffer.allocUnsafe(4+payload.length);framed.writeUInt32BE(payload.length,0);payload.copy(framed,4);fs.writeSync(process.channel.fd,framed);}else{main();}functioninspect(value){returnnewPromise((resolve)=>{constchild=fork(__filename,['inspect'],{serialization: 'advanced',stdio: ['ignore','ignore','inherit','ipc'],});child.once('message',(message)=>{child.disconnect();resolve({ok: true, message });});child.once('exit',(code,signal)=>{resolve({ok: false, code, signal });});try{child.send(value);}catch(err){child.disconnect();resolve({ok: false,threw: err.message});}});}functionsendBadTag(){returnnewPromise((resolve)=>{constchild=fork(__filename,['bad-tag'],{serialization: 'advanced',stdio: ['ignore','ignore','inherit','ipc'],});functiononUncaughtException(err){process.removeListener('uncaughtException',onUncaughtException);resolve({rejected: true,error: err.code||err.message});}process.once('uncaughtException',onUncaughtException);child.once('message',(value)=>{process.removeListener('uncaughtException',onUncaughtException);resolve({accepted: true,isBuffer: Buffer.isBuffer(value),value: Buffer.isBuffer(value) ? value.toString() : value,});});});}functioncheck(name,passed,details){console.log(`${passed ? 'PASS' : 'FAIL'}: ${name}`);console.log(details);console.log();if(!passed)process.exitCode=1;}asyncfunctionmain(){console.log('Expected on release/current Node: all PASS');console.log('Regression on PR build: one or more FAIL\n');const{ port1, port2 }=newMessageChannel();port1.visible=1;Object.defineProperty(Object.prototype,'visible',{configurable: true,set(){thrownewError('setter called');},});constsetter=awaitinspect(port1);deleteObject.prototype.visible;port1.close();port2.close();check('host-object spread must not invoke inherited setters',setter.ok&&setter.message.visible===1&&setter.message.keys[0]==='visible',setter.threw ? `regression: child.send() threw "${setter.threw}"` :
`result: ${JSON.stringify(setter)}`,);constbuf=Buffer.from('abc');buf.constructor=Uint8Array;constbufResult=awaitinspect(buf);check('Buffer with constructor = Uint8Array keeps old classification',bufResult.ok&&bufResult.message.isBuffer===false,`expected Buffer.isBuffer(received) === false\nresult: ${JSON.stringify(bufResult)}`,);constuint8=newUint8Array([1,2,3]);uint8.constructor=Buffer;constuint8Result=awaitinspect(uint8);check('Uint8Array with constructor = Buffer keeps old classification',uint8Result.ok&&uint8Result.message.isBuffer===true,`expected Buffer.isBuffer(received) === true\nresult: ${JSON.stringify(uint8Result)}`,);constbadTag=awaitsendBadTag();check('invalid child_process host-object tag is rejected',badTag.rejected===true,badTag.rejected ? `rejected with: ${badTag.error}` :
`regression: malformed tag accepted: ${JSON.stringify(badTag)}`,);}

[anonrig]

@panva thanks for the thorough repro — you're right, those were real regressions, and they're fixed in ef73d8f:

1. Inherited setters — non-view host objects are now copied with CreateDataProperty instead of Set, so inherited setters are no longer invoked (matching the old { ...object } spread).
2. Buffer classification — a Node Buffer is now identified by value.constructor === Buffer (matching v8.DefaultSerializer._writeHostObject) instead of by its prototype, so reassigning .constructor classifies exactly as before: a Buffer with constructor = Uint8Array round-trips as a Uint8Array, and a Uint8Array with constructor = Buffer round-trips as a Buffer.
3. Invalid tag — the deserializer now rejects any host-object tag other than 0/1 with ERR_INVALID_STATE, matching the previous assert(tag === kNotArrayBufferViewTag).

I added test/parallel/test-child-process-advanced-serialization-host-objects.js covering all four cases from your repro; it fails on the previous build and passes now.

[anonrig]

@mcollina it does use V8 serdes — the binding wraps v8::ValueSerializer/v8::ValueDeserializer, the same way the previous code used v8.DefaultSerializer/DefaultDeserializer. What's custom (and all this PR moves from JS into C++) is the thin child_process layer around them: the kArrayBufferViewTag/kNotArrayBufferViewTag host-object tagging, the JSON-like shallow copy for non-view host objects, and the 4-byte big-endian length framing — all needed to keep the wire format byte-identical for cross-version IPC. The motivation is removing the per-message JS allocations and JS↔C++ boundary crossings (benchmark numbers in the description), not replacing V8's serializer.

[panva]

@anonrig The fix should compare against a stable original Buffer constructor reference, not read it back from Buffer.prototype. This is still a regression, I'll let you be the judge of how big one.

'use strict';const{ fork }=require('node:child_process');if(process.argv[2]==='child'){process.on('message',(value)=>{process.send({isBuffer: Buffer.isBuffer(value),constructorName: value.constructor.name,keys: Object.keys(value),text: ArrayBuffer.isView(value) ?
Buffer.from(value.buffer,value.byteOffset,value.byteLength).toString() :
undefined,});});}else{console.log('Expected behavior: Buffer.prototype.constructor tampering affects advanced IPC classification.');console.log('On Node v26.3.0, the received value is a Uint8Array, not a Buffer.');console.log('Regression on the PR: the received value is still a Buffer.\n');constoriginal=Buffer.prototype.constructor;Buffer.prototype.constructor=Uint8Array;constchild=fork(__filename,['child'],{serialization: 'advanced',stdio: ['ignore','ignore','inherit','ipc'],});child.once('message',(message)=>{Buffer.prototype.constructor=original;constpassed=message.isBuffer===false&&message.constructorName==='Uint8Array'&&message.text==='abc';console.log(`${passed ? 'PASS' : 'FAIL'}: received classification`);console.log(message);if(!passed)process.exitCode=1;child.disconnect();});child.once('exit',()=>{Buffer.prototype.constructor=original;});child.send(Buffer.from('abc'));}

[anonrig]

@panva good catch — fixed in 557a596. The serializer now compares the view's constructor against the stable require('buffer').Buffer reference captured in serialization.js (passed into the native serialize()), instead of reading it back from Buffer.prototype.constructor. So tampering Buffer.prototype.constructor no longer fools the classifier: a genuine Buffer is then sent as a Uint8Array, matching v8.DefaultSerializer (and Node v26.x).

Your repro now passes (received Uint8Array, not Buffer), and I added that case to test/parallel/test-child-process-advanced-serialization-host-objects.js.

[mcollina]

lgtm

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74214/

[nodejs-github-bot]

CI: https://ci.nodejs.org/job/node-test-pull-request/74234/

[nodejs-github-bot]

Landed in a1074b8