63ed1f3c1f
Squash of experiment/ff-big-modules vs main.
Big-module routing removed: native-EH shrank kicad_editor below
SpiderMonkey's x86-64 code budget (runs 29355049705/29356152413 green on
stock Firefox), so BIG_MODULE_SPECS routing and the baseline-only-JIT
crutch are gone — kicad-firefox and kicad-chromium both run the full
suite, with the module compiled the way real users' browsers compile it.
Per-engine screenshots end to end: stableShot/shotPath write
test-results/<engine>/<name>.png; baselines move to
baseline-screenshots/{chromium,firefox}/ and the whole tools/screenshots
pipeline (compare/promote/manifest/spec-map/changelog/Discord) keys on
<engine>/<name>. Previously Firefox and Chromium renders of one spec
overwrote each other and Firefox renders were never actually gated.
Seeded from CI run 29421380806 (92 new firefox baselines, +24 chromium
web-suite shots); manifest generated from the baseline tree.
One merged playwright.config.ts (kicad/asyncify/coroutine/perf as
projects); ~25 dead npm scripts dropped. The web suite is gated in CI for
the first time ever (4 rotted specs fixed, 5 broken lib-bridge specs
triaged as fixme in docs/features/web-e2e-rot/); cheap lint step after
npm ci; last 26 blind-sleep violations fixed.
SwiftShader retired: CI Chromium renders WebGL on ANGLE → Mesa llvmpipe
(--use-gl=angle --use-angle=gl --ignore-gpu-blocklist; the blocklist flag
is mandatory — llvmpipe is blocklisted and WebGL is silently unavailable
without it) in BOTH configs. Under WORKERS=4 congestion SwiftShader
transiently failed the first post-board-load draw and the recovery
cascade ended in a silent permanent Cairo fallback — that engine flip was
the "~1.2% changedRatio both directions" occ-export baseline flake.
Validated 160/160 across two 80-repeat rigs; full analysis in
docs/features/wx-parity-bugs/occ-export-context-eviction.md. Chromium
baselines shift slightly on llvmpipe — promote once from the first green
run. Deflakes the new coverage exposed: presence baselines settle before
capture; presence fixtures declare current file formats; perf gets its
own outputDir so CI evidence survives; occ-export settles the board paint
before the export dialog; menu-item waits (waitForRenderedByLabel before
clickMenuItem) in 4 specs + the TESTING.md rule.
Web suite runs the PROD build, in parallel: webServer becomes backend
`start` + the standalone's e2e:preview (build-preview.mjs: link-wasm →
stash the public/wasm symlink aside during vite build, build-demo.mjs's
move — then vite preview as the persistent server). The wasm middleware
serves /wasm/* in preview and emits COOP/COEP/CORP itself (a pthread
worker script's own response must carry COEP or Chrome kills it with
ERR_BLOCKED_BY_RESPONSE). VITE_* flags bake at build time;
VITE_ALLOW_USER_OVERRIDE joins turbo globalEnv. fullyParallel + default
workers: 5.2m → 1.4m. Determinism fixes the parallel run exposed:
shared-page specs become serial groups; locks.spec grabs alice's exact
item via the new kicadCollabTestSelectByUuid hook (cross-tab "first
footprint" order is not a ysync invariant); quit specs poll page.url()
(quit supersedes its own navigation — NS_BINDING_ABORTED on Firefox).
Suite: 51 passed / 12 skipped / 0 failed in 1.6m.
CI-coverage gate (lint:ci-coverage): every tests/**/*.spec.ts must be
reachable from the npm scripts the workflows invoke — scraped from
.github/workflows/, resolved through package.json, coverage asked from
playwright --list itself. Rules: uncovered-spec + orphan-project (with a
documented LOCAL_ONLY_PROJECTS allowlist). Gating next to
lint:determinism; 138 spec files / 13 projects accounted for.
Product fixes kept from the investigations (reachable on real GPUs too):
wx 7799fd1be5 — paint flags clear before dispatch + Invalidate always
propagates; kicad 3dcfea5e45 — SwiftShader pass-boundary flush +
per-instance font texture + first-frame GL-error drain (GAL recovery
recovers instead of falling back to Cairo) + the user-facing eeschema
switch navigates again under __EMSCRIPTEN__ (project-sync's
FaceRegistered gate had rerouted it into the hidden sync player; caught
by the newly-gated web suite).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018eUxiPApHgGiu9NFyQfhAq
119 lines
6.5 KiB
Markdown
119 lines
6.5 KiB
Markdown
# occ-export flake: SwiftShader congestion → GAL recovery cascade (fixed by GL-stack swap)
|
||
|
||
Status 2026-07-16: fully root-caused with page-level canvas/context telemetry +
|
||
temp C++ instrumentation in a local docker rig. **Final remedy: CI Chromium
|
||
now runs WebGL on ANGLE → Mesa llvmpipe instead of SwiftShader**
|
||
(`tests/playwright.config.ts`, `CHROMIUM_CI_ARGS`) — the same software-GL
|
||
stack the Firefox projects use, where none of this ever happened. Four small
|
||
product fixes found along the way were kept (below); everything else was
|
||
deliberately **not** hardened — the trigger cannot occur outside congested
|
||
SwiftShader.
|
||
|
||
## Symptom
|
||
|
||
`chromium/occ-export-{dialog,done}.png` bistable on CI only: the board behind
|
||
the modal STEP-export dialog flips between runs (~1.2% changedRatio both
|
||
directions). Firefox (llvmpipe) never affected.
|
||
|
||
## The causal chain (every step observed in the rig)
|
||
|
||
1. **Transient SwiftShader failure at the first post-board-load draw**
|
||
(~10-15% of instances at `REPEAT=40-80 WORKERS=4`): `DoRePaint threw:
|
||
Requested render buffer size is not supported`, plus LRU context evictions
|
||
the moment a 3rd context appears (21/80 runs; the usual victim is a leaked
|
||
hidden 881×159 startup GAL canvas, so the `CONTEXT_LOST_WEBGL` line alone
|
||
was harmless). No flag stops it — `--max-active-webgl-contexts=64`,
|
||
`--force-gpu-mem-available-mb=4096`, `--disable-low-end-device-mode`,
|
||
`--disable-gpu-driver-bug-workarounds` all verified present and useless.
|
||
|
||
2. **`recoverFromGalError` attempt 1** (`common/draw_panel_gal.cpp:205`)
|
||
destroys + recreates the WebGL GAL (board canvas `glcanvas-2` →
|
||
`glcanvas-3`).
|
||
|
||
3. **Attempt 1 then always failed** on stale cross-context GL state:
|
||
- static `g_fontTexture` bound on the new context → `INVALID_OPERATION:
|
||
bindTexture: object does not belong to this context` → `checkGlError`
|
||
throw (**fixed — kept**: per-instance `m_fontTexture`,
|
||
`common/gal/webgl/webgl_gal.{cpp,h}`);
|
||
- sticky GL error flags left by the old GAL's teardown (its `LockCtx`
|
||
targets a lost context, so the deletes land on the current one) →
|
||
the fresh GAL's first `checkGlError` ("generating vertices buffer:
|
||
invalid enum") misattributed the leftover and threw (**fixed — kept**:
|
||
drain `glGetError()` before first-frame `init()` in
|
||
`WEBGL_GAL::BeginDrawing`). With both fixes, reinit-on-WebGL succeeds
|
||
(validated: 0 Cairo fallbacks / 60 after, vs 100% of recoveries before).
|
||
|
||
4. **Attempt 2 = silent permanent Cairo fallback** (`GAL_FALLBACK ==
|
||
GAL_TYPE_CAIRO` on emscripten, dialog suppressed). Cairo renders the board
|
||
through the wx software path — shots LOOK painted but are a different
|
||
engine → subtle whole-board diffs vs the WebGL baseline. **This engine
|
||
flip IS the original "~1.2% changedRatio both directions" flake.**
|
||
`FULLSCREEN_QUAD` (singleton VAOs) is the same static-GL-state disease and
|
||
produced `bindVertexArray: object does not belong` + one wasm abort in
|
||
rare runs — NOT fixed (unreachable once recovery never triggers; noted
|
||
here in case in-app recovery robustness is ever wanted).
|
||
|
||
5. **wx paint-flag races turned it into a stuck-blank board** (both
|
||
**fixed — kept**, `wxwidgets/src/wasm/window.cpp`):
|
||
- `Invalidate` early-out skipped the upward walk while the sweep can
|
||
strand a window "flags set, ancestors clear" → every later `Refresh()`
|
||
swallowed. Fix: always propagate to the top.
|
||
- `PaintSelf`/`PaintChildren` cleared the dirty flags AFTER dispatching
|
||
paint, so an `Invalidate` issued reentrantly from inside a paint handler
|
||
(canvas recreated mid-frame, board-load refresh mid-sweep) was clobbered
|
||
with nothing re-issuing it. Fix: `DoPaint` snapshots + clears both flags
|
||
BEFORE dispatching. (Pre-fix rigs showed stuck half-painted toolbars —
|
||
this is a real production-reachable bug, not a CI artifact.)
|
||
|
||
One rare mode ("grid-mode": board data loaded, UI painted, canvas alive,
|
||
items never drawn, zoom-to-fit doesn't help; VIEW had all 464 items and
|
||
the same Clear/DisplayBoard ordering as healthy runs) was still being
|
||
diagnosed when the GL-stack swap made it — and everything above —
|
||
unobservable: **80/80 runs clean on llvmpipe**, all dialog shots pixel-
|
||
identical (single 55.65% ratio value across all 80; SwiftShader scattered
|
||
over 55.5–56.8%).
|
||
|
||
## The fix that ships
|
||
|
||
`CHROMIUM_CI_ARGS` = `--use-gl=angle --use-angle=gl --ignore-gpu-blocklist`
|
||
(CI-gated). Requirements, all already present on CI: an X display
|
||
(`xvfb-run -a` wraps the whole e2e step) and Mesa (installed by
|
||
`playwright install --with-deps`). `--ignore-gpu-blocklist` is mandatory —
|
||
llvmpipe is on Chromium's software-GL blocklist and WebGL silently reports
|
||
unavailable without it (the exact analog of the Firefox projects'
|
||
`webgl.force-enabled`). Headless works; headed is NOT needed.
|
||
Renderer-string ground truth (Chrome falls back silently — always verify):
|
||
`gl.getExtension('WEBGL_debug_renderer_info')` must report
|
||
`ANGLE (Mesa, llvmpipe …)`, not SwiftShader.
|
||
|
||
Consequence: chromium renders shift slightly → the chromium baseline set
|
||
must be re-promoted once from the first green CI run
|
||
(`cd tests && npm run screenshots:promote -- --run <id>`).
|
||
|
||
## Test-side hardening (kept, independent)
|
||
|
||
Popup menu items register in the element registry progressively while the
|
||
menu paints; a coarse ">3 menuitems" gate + single-shot `clickMenuItem` races
|
||
(4–8/20 under contention, pre-existing). Rule (tests/TESTING.md): wait for the
|
||
specific item via `waitForRenderedByLabel(page, label, { elementType:
|
||
'menuitem' })` before every `clickMenuItem`. Applied to `occ-export`,
|
||
`occ-export-models`, `3d-viewer-models`, `load-pcb`; `pl_editor` already did
|
||
this; `clickMenuItemByText` waits internally.
|
||
|
||
## Open leads (not blocking)
|
||
|
||
- The leaked hidden 881×159 startup GAL canvas: destroying it would free a
|
||
context + memory.
|
||
- If in-app GL-error recovery ever matters on real GPUs: fix the
|
||
`FULLSCREEN_QUAD` singleton VAOs (per-instance like the font texture) and
|
||
audit remaining cross-context statics.
|
||
|
||
## Repro rig (for revalidation)
|
||
|
||
Docker `mcr.microsoft.com/playwright:v1.57.0-jammy`, worktree bind-mount,
|
||
container-local node_modules volume, `CI=1`, `xvfb-run -a npx playwright test
|
||
--project=kicad-chromium --repeat-each=N --workers=4` with a temp spec that
|
||
loads `pic_programmer`, screenshots board-load / settled / export-dialog /
|
||
post-poke, and logs the WebGL renderer string + canvas/context telemetry per
|
||
instance. Canaries: `CONTEXT_LOST_WEBGL`, `does not belong to this context`,
|
||
board-strip pixel ratio of the shots.
|