OpenGL worker/RPC and render-bridge lease sweep — 2026-11-01
Run 20261101-gl-worker-lease-crash-sweep: both requested fixtures passed,
sequentially, with original assertions and 120-second command limits. No
unexpected native crash, unhandled Wasm trap, signal, assertion failure or
timeout occurred. The first fixture executes the native producer in real
Node worker_threads; the second uses deferred endpoints and VM/backend
doubles. These are distinct coverage results, not an end-to-end native
producer-to-render-worker/GPU result.
Revision, build and exact identity
Requested baseline: 6f1f57602e36c30fc45fdd755f7286b20d39cf4a, shared main.
Both commands ran at 81b757b5614dc495d31e7134826ac7b8b063497d; only
TODOS.md and ops/STATUS.md advanced, within the allowed documentation-only
exception. All 4,462 inventoried inputs match baseline tracked bytes and
accepted build scratch/runs/20261101-d3d-typed-stateblock-injection/inputs-after.json;
they remain unchanged before/after these tests. All 19 distinct loaded
CommonJS files match baseline and their exit-time hashes. VM-loaded endpoint
sources are separately preserved and hashed. The input inventory also covers
the sealed compiler graph, which is not represented fully by CommonJS cache.
Reused accepted build.log and outcomes.json from that build run:
timeout 120s flock /tmp/wine-build.lock bash tools/build.sh, exit 0,
46.479 seconds. No accepted test or build was rerun. Current primary artifact:
2,025,029 bytes, SHA-256
2be3cebaa1c08a8e12075de000bfb83fb8692841c1452d9e317203f5c2fce8cf.
Local environment: Linux x86-64, Node 24.18.1, V8 13.6.233.17-node.50.
No compiler/artifact or NODE_OPTIONS override was used.
Both real workers instantiated the same 2,024,992-byte module, SHA-256
ac9beb2ae3f746c621d0f258d4d3a2d1910207b0006e13ed96856bd0ac4975d9.
All nine non-custom sections equal the shipping primary module; only its
37-byte encoded wine-region-layout custom section is absent. This fixture
calls compileSrcWasm() without a transform. Exact module captures from
threads 1 and 2 and section hashes are retained in module-identity.json.
The bridge fixture compiles or instantiates no Wasm module.
Original fixture SHA-256 identities:
| Fixture | SHA-256 |
|---|---|
test/test-opengl-wat-worker.js |
a8747025d2b4ffba88b605d00dfbdfc61c91618aa91fb25a29887b02378cd926 |
test/test-gl-shared-render-worker.js |
a1f85520b08c1899c75f5d35d3e3b39e69b2145dd681b1d7baf4e5708721453b |
Commands and outcomes
Evidence is ignored under scratch/runs/20261101-gl-worker-lease-crash-sweep/.
Each test argument followed this exact prefix, with RUN the absolute path
of that directory:
SWEEP_CAPTURE_PREFIX="$RUN/<stem>" timeout 120s node --require "$RUN/capture-module.cjs"
The scratch-only hook is copied unchanged from the preceding native-encoder
sweep. It captures bytes before module construction/instantiation and hashes
loaded CommonJS files at exit; it does not replace source, compiler behavior,
fixture assertions or results. Workers inherit the preload. Full argument
arrays, HEADs, timings and logs are retained in outcomes.json and the logs.
| Test argument, in order | Exit | Seconds | Result |
|---|---|---|---|
test/test-opengl-wat-worker.js |
0 | 3.588 | Real-worker single-buffer and triple-pool cases both PASS |
test/test-gl-shared-render-worker.js |
0 | 0.068 | Entire shared-render protocol fixture PASS |
Real native worker/RPC execution
The first fixture creates two actual Node workers sequentially, each with a
real fixed 512 MiB shared Wasm memory and the full native module. It invokes
gl_wat_encoder_call in each worker through production
GuestRPC.createWorkerImports. Assertions reject runtime JS-encoder selector
exports. Process/thread bootstrap RPC identity is checked before GL work.
The main thread receives native shared-range descriptors rather than a
structured-cloned command buffer, observes STATUS_REQ, and deliberately
delays each GL broker service by 25 ms. The worker's query and finish calls
must each remain blocked for at least 15 ms. Production
createMainBroker().serveGlBatch consumes the request; query output is
written before response/wake, and the worker immediately observes value 4
plus return 0x2468. Finish returns 0x1357; the RPC slot finishes IDLE.
The semantic JS consumer validates packed triangle mode, 3×19 vertex floats, first position/color, and exact batch opcode order. Single-buffer mode has two synchronous GL submissions; pooled mode has four. Import statistics require respectively two/four sync and local crossings after bootstrap. The released first range is demonstrably overwritten while its saved snapshot stays distinct. Triple-pool metadata must describe three distinct in-bounds 2 MiB slots, rotate in order, remain stable, and preserve the first two snapshots while the third is encoded before wrapping to slot zero.
This is real native producer execution, worker scheduling, shared memory and
blocking RPC. The consumer is a deterministic semantic callback, not an
OpenGL backend or GLRenderHost.Bridge. It does not stress asynchronous
bridge admission while the native worker rotates its pool. The fixture
terminates its own Node workers in finally; these are not managed boxes.
Bridge leases, ordering and backpressure through doubles
The second fixture executes production Bridge, SubmissionQueue,
SharedPacketPool, command replay and broker code with manually resolved
endpoint promises. It never constructs a real worker. Meaningful assertions
include:
- Safe void commands acknowledge early only after canonical GPU ownership is marked. Copied scalar and packed-vertex bytes survive producer overwrite. Two admitted batches bound the queue; a third producer waits, and a later query from another owner cannot overtake it. Query memory remains leased until completion and returns the actual result.
- Shared triple-pool transport sends offsets without command-buffer copies.
Third publication waits for oldest consumption before rotation; first and
third pending payloads survive the relevant writes. Four batches report
exactly four leases, zero copies and zero copied bytes. Overlapping pool
slots reject;
sharedPackets:falseuses the copy comparison path. - Presentation drain and
glFinishwait for replay. Borrowed pointers, unknown operations and software backend keep synchronous leases. Eight copied vector cases survive overwrite with a present at the tail; twelve malformed vector cases retain synchronous transport and actual result 37. Three supported small texture shapes, including the 65,536-byte threshold, own copied bytes; ten large/padded/unsupported/malformed upload cases retain leases and actual results. - GDI synchronization runs readback before later producer admission, with
ownership flag order
[1,0,1]. Cross-owner context deletion invalidates early-ACK eligibility. Async queue rejection propagates through drain, subsequent barriers and close without further submission. Close waits for pending replay, terminates the endpoint and rejects reopening/late frames. - Admission snapshots capture nested window rectangles and authoritative client sizes. Replaced HWND frames close their bitmaps, cannot attach to reused windows, and retire frame-owner references. Query writes, software producer configuration and surface copy-back precede lease completion. A live endpoint remains selected after its enable predicate changes.
- The broker leaves a promise-backed GL request in REQ, rejects duplicate GL/RPC service, and publishes result/output before RESP. This broker case uses direct shared-memory setup, without a waiting native worker.
VM execution of production gl-render-worker.js uses FakeGL to check owned
command-buffer selection, captured geometry overrides/live-export fallback,
and bitmap write-then-read coherence. D3DIM checks use deferred proxy ports,
FakeGPU cleanup and an actual JS D3DIMGpu wrapper with a deliberately
unavailable canvas and stub native exports. They check ready/fence results,
stop(), diagnosed fallback, state override cleanup, no per-triangle context
retry, and fatal propagation of an injected shader error. Those deliberately
rejected promises, overlap rejection and caught JS errors are successful
negative assertions, not unexpected native crashes or Wasm traps.
Limits and handoff
No source, test, helper, compiler, census, TODOS/STATUS or foreign WIP edits were made. No accepted fixture reruns, REP tests/diagnostics, held work, Heroes II, browser, GPU, heavy benchmark, boat, download, access retry, agent/successor or deployment was used. No crash reproduction was needed.
Remaining gaps: native producer connected to an actual render worker under consumer load, browser/GPU execution, broader concurrent scheduling, failed shared-packet lease quarantine/retirement, and application gameplay. These fixtures do not establish full-27 coverage or overall crash freedom.
environment.json, before/after input inventories, verification.json,
loaded-verification.json, exact original sources, modules and logs retain
reproduction identities. Post-test free disk: 2,427,588,608 bytes, above 2 GB.
Box ID/state/expiry and lifecycle: N/A; no managed box or ownership to hand
back. Next: coordinator reviews this single completion; this worker stops.