Files
MobileGL/docs/archive/notes/p5b/p5b-results/wait-codex-v1.md
T

3.2 KiB

P5b closing-review fixes: FenceWait and ReadPixels status

2026-09-16. Branch p5b/wait-codex, dedicated tree /home/swung/w7/p5b-wait-codex. Base: 348d22a4b9c15cc571af840dedc4b4edc8c53dbb. Commit: 6388035f809861fd9b45cd1ba67c458a14b8b836.

Changes

Closing review Major #2: FenceWait applied/reply wait now allows ceil(TimeoutNs / 1,000,000) + 30,000 milliseconds. Quotient/remainder rounding avoids overflow at UINT64_MAX. The finite 64-bit budget is consumed in transport chunks of at most UINT32_MAX - 1, avoiding both narrowing and the transport's forever sentinel, and avoiding overflow from one enormous chrono deadline. Shutdown still returns the original ShutDown result immediately; finite exhaustion still reports Fatal{BarrierTimeout} with the full effective budget. Only CPU FenceWait expands its budget. All other verb barriers and command-space retirement retain 30 seconds; FenceWaitServer with GL_TIMEOUT_IGNORED remains an ordinary bounded CPU submission.

Closing review Minor #3: both tight and scattered ReadPixels paths now abort Fatal{ReplyStatusInvalid, "ReadPixels"} for unknown reply status. Existing ERROR/DECLINED diagnostics remain intact. Correctly sized pixels with status 3 can no longer return as successful readback.

Directed verification

Built only RemoteClientTest in this dedicated split/inproc tree with Clang Release and -j4. Selected 7, passed 7, failed 0, skipped 0 (80 ms runtime).

New runtime controls use an installed test peer and a wrapper of the real transport doorbell to observe the deadline actually selected by EmitAndWait; no 30/60-second sleep or production test switch is needed. Coverage:

  • FenceWait timeouts 0 ns, 1 ns, 60 seconds and UINT64_MAX: requested timeout is respected and huge waits use a finite transport chunk.
  • FenceWaitServer + GL_TIMEOUT_IGNORED: ordinary 30-second budget.
  • Huge FenceWait interrupted by a dead doorbell: DECLINED, no reply, prompt return.
  • Status 3 with a complete readback payload: named abort on both tight and bounce paths.
  • Existing complete readback, short tight reply, ERROR reply and short bounce reply controls remain green.

Command:

build-split/MobileGL/MG_Test/Wire/RemoteClientTest --gtest_filter='RemoteClientControls.FenceWaitBudgetHonorsGlTimeoutAndFiniteTransportChunks:RemoteClientControls.LongFenceWaitBudgetStillWakesOnShutdown:RemoteClientControls.UnknownReadPixelsReplyStatusRefusesTightAndBounce:RemoteClientControls.ReadPixelsPutsTheTightExtentOnTheWire:RemoteClientControls.ShortOkReadPixelsReplyRefusesByName:RemoteClientControls.ErrorReadPixelsReplyRefusesByName:RemoteClientControls.BounceReadPixelsShortReplyRefusesBeforeScatter'

Evidence: /home/swung/w7/p5b-wait-{configure,build,focused}.log. git diff --check passed. Commit tree is clean; dependency links are off. Old r1 tree/commit and all existing evidence remain intact. No repeated phase gate, additional adversarial review, device operation or push was performed.

The host full gate remains pinned to 348d22a4; this directed result is explicitly a later review-fix delta. APKs must be built from the final integrated head containing both closing-review fixes.