Skip to content

fix: close review issues #6-#13 - #25694

Closed
enochgroot wants to merge 17 commits into
msgpack:masterfrom
enochgroot:fix/review-open-6-13
Closed

enochgroot wants to merge 17 commits into
msgpack:masterfrom
enochgroot:fix/review-open-6-13

Conversation

@enochgroot

Copy link
Copy Markdown
Collaborator

Closes #6 #7 #8 #9 #10 #11 #12 #13

One PR for the eight OPEN review findings on the 2.0 modernization tree.

Changes

Tests

npm test: 135 pass.
JS coverage gate 95% (c8 --check-coverage).
Native coverage: 96.0% lines, 99.5% branches, 100% functions (gcovr --fail-under-line 95 --fail-under-branch 95).

Sabotage: Stream cap tests fail without the allocate check; pack sparse-array test fails without kMaxContainer; CLI tests fail without the stdin cap.

Coverage gate in package.json / COVERAGE.md is unchanged.

Opened from fork enochgroot/msgpack-node branch fix/review-open-6-13.

enochgroot and others added 17 commits September 10, 2026 00:02
Vendor msgpack-c c-7.0.2, fail-closed unpack limits, and fix the
sbuffer leak on pack throw (nodejs/node#25686). Replace nodeunit with
node:test, require Node 18+, and run GitHub Actions on 18/20/22.
Dates and toJSON apply at every nesting level. Numeric object keys
are packed instead of dropped. Cycle marks use V8 private symbols
so a user key named _msgpack_stack is no longer stripped. Integral
doubles outside uint64/int64 range (e.g. 1e30) pack as float64.
Pack recursion is capped at 512 so 8000-deep input throws instead
of SIGSEGV. The vendored C unpacker is built with
MSGPACK_EMBED_STACK_SIZE=512 so advertised unpack depth matches
the walker. Stream emits error on unpack throw and treats packed
nil as a message. The CLI bins run on Node 18+.
DefineOwnProperty on map keys so a wire __proto__ cannot replace
the decoded object's prototype. Property reads during pack go
through Nan::TryCatch so a throwing getter or Proxy ownKeys
raises a catchable error instead of aborting on ToLocalChecked.
Stream snapshots bytes_remaining and advances the buffer before
emit('msg'), so a listener that unpacks or throws cannot desync
or replay a frame.
pack() ran a JS pre-pass that called toJSON() on any top-level object.
Buffer.prototype.toJSON exists, so msgpack.pack(Buffer.from([1,2,3]))
emitted a {type:'Buffer',data:[...]} map instead of msgpack bin, and
unpack() no longer returned a Buffer. Nested Buffers were unaffected
because they went straight to the binding.

The pre-pass is redundant: the native JsToMsgpack checks
node::Buffer::HasInstance before the generic object path, PackObject
applies toJSON at every level, and Dates pack as ISO strings natively.
Drop it and forward arguments to the binding unchanged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
NODE_MODULE is not context-aware, so requiring msgpack in the main
thread then in a Worker throws "Module did not self-register".

Register with NODE_MODULE_CONTEXT_AWARE and keep the sbuffer pool,
unpack remainder, and cycle-detection key thread_local so workers
do not race the main isolate.

Fixes msgpack#60
Packed integer 0 must emit on Stream (msgpack-node#44). A map written
by python-msgpack with a bin8 payload must unpack Payload as Buffer
(msgpack-node#10). Both already work on 2.0.0; these tests keep them so.

Refs msgpack#10
Refs msgpack#44
Adds index.d.ts matching the 2.0.0 runtime: pack(...values) returns
Buffer, unpack.bytes_remaining is per-thread, Stream wraps a duplex.

Fixes msgpack#39
Address review on PR #2 for worker_threads (msgpack#60):

- Mark sbuf_pool thread_local so a worker pack cannot share the pool
- Register with NAN_MODULE_WORKER_ENABLED instead of NODE_MODULE_CONTEXT_AWARE
- Return non-pooled sbuffers to the pool so it actually fills
- Add concurrent pack (workers + main) and sequential pool-reuse tests
lib/ and bin/ reach 100% statements/branches/functions/lines under c8;
src/msgpack.cc reaches 95.9% lines and 99.5% branches under gcovr.

New tests:

- test/coverage-native.test.js walks every MessagePack format family from
  hand-built wire bytes, including the ones pack() never emits (float32,
  str8/16/32, bin16/32, array32, map16/32, all eight ext forms, negative
  fixint); a truncation point for every header and payload; the kMaxBytes,
  kMaxContainer and kMaxDepth rejections; 0xc1; the pack-side type dispatch
  (Symbol, BigInt, non-finite numbers, integer edges, undefined, zero- and
  multi-argument pack); Date and toJSON failure modes with mark cleanup; and
  a worker nesting 600 packs deep to saturate the thread-local sbuffer pool
  and reach the "pool is full, free it" arm of ~PackBuffer.
- test/cli.test.js covers the exit-1 paths of both CLIs. The pack-failure arm
  of json2msgpack is reachable from real JSON only through nesting deeper
  than the 512-level pack cap; every other pack error needs a value
  JSON.parse cannot produce.

Infrastructure:

- binding.gyp grows an msgpack_coverage variable, default 0. Only when it is
  set to 1 does the addon compile and link with --coverage -O0 -g, so
  npm install and node-gyp rebuild stay uninstrumented.
- npm run coverage runs coverage:js (c8, gated at 95% on all four metrics)
  then coverage:native (scripts/coverage-native.js). The native script
  rebuilds instrumented, runs the suite, gates on gcovr --fail-under-line 95
  --fail-under-branch 95 over src/ excluding deps/, and always rebuilds
  uninstrumented afterwards -- including when the gate fails.

src/msgpack.cc gains comments only. GCOVR_EXCL_BR markers, each with its
reason inline, mark branches unreachable without stubbing malloc or V8:
allocation-failure arms of msgpack_pack_*, empty-MaybeLocal guards, Skip()
calls a preceding CheckBytes has already proved safe, and the post-ScanOne
error tail of Unpack. No production code was removed. COVERAGE.md lists all
45 marked lines, the 20 still-uncovered lines, the 2 still-uncovered branches
and the un-gated numbers (70.6% raw, 86.3% throw-excluded, 99.5% as shipped).
Adds an ubuntu-latest / Node 20 job that installs gcovr via pip when it is
not already present and runs npm run coverage. Both halves are threshold-
gated, so the job fails when JS or native coverage drops below 95%. The
existing node 18/20/22 x ubuntu/macos test matrix is unchanged.
Acceptance requires the coverage script in the Building section
and a pointer to COVERAGE.md for gates and remainder.
Strip personal-fork branding so this reads as the upstream 2.0.0
tree. Document that vendored msgpack-c c-7.0.2 is Boost Software
License 1.0 (relicensed from Apache-2.0 in msgpack-c 1.3.0). Restore
Command Line Utilities and Benchmarks, refresh the bench runner for
Node 18+, and record current numbers.
Restore the master contributors list in package.json. Point SECURITY.md,
src/msgpack.cc, and the leak regression at msgpack#25686
instead of nodejs/node#25686. Correct the Stream comment: bytesRemaining
is thread_local. Add windows-latest to the CI matrix.

Closes msgpack#25686
Pin the Windows matrix to windows-2022 so node-gyp 10/11 (Node 18/20/22)
can find Visual Studio. windows-latest currently ships VS 2026, which
those node-gyp versions report as unknown version "undefined".

Run install and test as one bash command so a failed rebuild cannot
continue into npm test. List test files explicitly so npm test works on
Windows Node 18/20 (cmd.exe does not expand *, and node --test globs
need Node 21+).

The Stream test comment now matches thread_local bytesRemaining.
Pin nan 2.28.0 and c8 12.0.0; CI uses npm ci.
Raise engines.node to >=22 and drop Node 18/20 from CI.
Pin GitHub Actions to full SHAs.
Cap Stream receive before allocate and drop buf on close.
Apply kMaxContainer to pack array/map walks.
Cap CLI stdin concat at MAX_STDIN_BYTES.
Document a risk-based bump window in SECURITY.md.

Closes #6 #7 #8 #9 #10 #11 #12 #13
@enochgroot

Copy link
Copy Markdown
Collaborator Author

Closing. This branch still carries the pre-squash 2.0 history, so GitHub is showing 17 commits / 116 files against master. The scoped #6–#13 fix is enochgroot#14

@enochgroot enochgroot closed this Sep 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

npm install issue

1 participant