Files
MobileGL/docs/archive/notes/p5b/p5b-results/r1-codex-v1.md
T

7.2 KiB
Raw Blame History

P5b r1 — server ownership, readback byte order, retirement back-pressure

Completed 2026-09-16. Branch p5b/r1-codex, isolated tree /home/swung/w7/p5b-r1-codex, commit db8b3479520862cd47c546e96506b7332158abcf. Base is r2 d50183cb, so merge r2 first or merge this branch including its ancestry. This implements ID-73 findings 1–4; it is not another adversarial review or a phase gate.

The inherited Windows tree /mnt/c/Users/geekerwan/p5b-r1 was at 37fc4fdb with three uncommitted files: PersistentMapTracker.cpp, .h, and BufferObject.cpp. Their complete file contents were copied into /home/swung/w7/p5b-r1-codex-backup/ before continuing in this new tree. The old tree and any old Claude process were left alone.

Changes and evidence

Finding Implementation Focused evidence
1: coherent map bypass Under split, retained backend SyncPersistentMappedRange calls return on the apply thread. PushAllMembers, PushBlocksFor, and the final PushMappedSpanBlock producer refuse server-role entry with Fatal{RoleViolation,...}, before reading map bytes or modifying frontend content serials. Server map-state teardown keeps membership valid but does not emit client live-write state. Real GPU store readback test now holds a coherently mapped frontend object with bytes A and a server shadow with bytes B; Ensure uploads B. A live client session also asserts the producer's named apply-thread refusal. Separately, removing only client pre-verb push calls makes both coherent-pointer pixel cases fail while backend sync sites remain intact; restoring the calls makes both pass.
2: framebuffer death on client GLES' shared frontend-death callback forwards synchronously through the existing apply-thread control mailbox before any twin destruction or binding-cache changes. This covers every kind in that callback; framebuffer still has no invented wire opcode. Native glDeleteFramebuffers interception observes one deletion on the apply thread and verifies that the native draw binding becomes zero. The previously failing DirectGLES.P4aFinalFixScenario.ARenderbufferAndItsFramebufferDeletedAfterAClearLeaveTheNextDrawIntact now passes.
3: PACK_SWAP_BYTES After validating the tight reply, swap each requested component/packed word on the client. Apply to both direct and bounce paths, before scatter; byte types remain unchanged. FLOAT_32_UNSIGNED_INT_24_8_REV swaps two separate 32-bit words. Live emit-table peer covers ushort, float, packed uint, mixed depth/stencil, and byte data with swap off/on and tight/scattered layouts; sentinel gaps are preserved. Actual GL test reads a non-symmetric ushort from RGBA16F with tight and skipped/strided pack destinations: both monolith and inproc pass. RGBA8 was deliberately avoided because expanded ushort components repeat the same byte and cannot expose a missing swap.
4: applied before retired Preserve r2's staging allocator wait and truthful wait counter unchanged. ClientSession::EmitAndWait now treats command-ring capacity exhaustion as back-pressure after the encoder's individual-record size check. It waits for enough retired bytes including the wrap pad, then retries the same record without changing its sequence. Shutdown/timeout gets a named RetirementWaitFailed instead of a silent drop or indefinite wait. Two live-session cases hold the server between applied publication and retirement using r2's scheduling-only hook. The staging case fills 3/4 of 1 MiB twice and observes one wait. The command case positions a legal 393 KiB draw record so its successor needs a wrap pad, observes the producer parked, releases retirement, and verifies the next ordinal is exactly prior+1.

The death callback uses the existing inproc control mailbox, not a new wire lifetime record. P6's separate address space still needs the already-planned lifetime/control transport work. Staging's r2 five-second shutdown-aware bound and once-per-blocked-allocation counter are intact. The command wait uses the session's existing kBarrierTimeoutMs (30 seconds); individually oversized records still fail at PipeWireEncoder::EncodeRecord.

Executed acceptance

All paths below are under /home/swung/w7/.

  • Split build completed, CTestTestfile.cmake verified by the build driver. Logs: p5b-r1-codex-config-split.log, p5b-r1-codex-build-split.log, p5b-r1-codex-build-fixes.log. Build concurrency was at most 8.
  • Final targeted run: 10/10 passed, zero skips, MOBILEGL_TRANSPORT=inproc, MOBILEGL_ITEST_REQUIRE_GPU=1, p5b-r1-codex-final-targeted.log. Six unit entries plus four integration entries: the six ownership/wait/swap controls, actual ushort swap pixels, FBO/RBO deletion pixels, and two coherent-pointer draw cases.
  • Actual GL ushort swap: 1/1 monolith and 1/1 inproc, zero skips, p5b-r1-codex-swap-{monolith,inproc}.log.
  • Client-only push suppression: two own pixel failures (CTest exit 8), plus the existing map-arm metadata case's one intentional skip. The skip is not counted as a red. p5b-r1-codex-coherent-red.log; driver p5b-r1-codex-coherent-control.sh. It restores the source through an EXIT trap and rebuilds the restored integration binary. The restored two pixel cases are included in the final 10/10 run.
  • git diff --check passed. Final worktree clean, submodule links off. Commit is a single line without trailers, author/committer Swung0x48 <swung0x48@outlook.com>.

The primary negative control removes only PushPersistentMapsBeforeVerb() from the client pre-verb hooks. It does not set the block-size-zero knob, alter the backend's retained sync sites, or remove the new server-role guard. It therefore proves client omission is no longer repaired by server producer re-entry.

No new all-lane census, four-flavor unit run, G1 binary comparison, full integration lane, device operation, push, or phase review was run here: the integrator explicitly owns the combined quick gate. All production changes outside MG_Remote are inside #if MOBILEGL_BUILD_DISAGGREGATED, so pull code is structurally unaffected; this is not reported as a new measured G1 result. No .def, generated catalogue, contract, or opcode changed.

Integration notes

  1. Keep r2's StageAllocate implementation, its session doorbell hookup, telemetry semantics, and before-retire hook. This commit changes command wait policy in ClientSession only; it does not replace r2's codec implementation.
  2. Five new unit entries; the existing staged-shadow unit is strengthened to use a persistent coherent mapping. One new ordinary integration scenario, registered for both backends; integration-split remains the same 22 selected entries. The new scenario is PixelStoreSweepScenario.ReadPixelsSwapsUnsignedShortComponentsAndPreservesPackGaps.
  3. The known 27 inproc ordinary failures should no longer be relabeled wholesale as unchanged: the GLES FBO/RBO deletion case is now green on this isolated tree. The merged census must determine the new aggregate; no count was extrapolated from this single case.

Tools retained: p5b-r1-codex-submodlinks.sh, p5b-r1-codex-build.sh, p5b-r1-codex-coherent-control.sh; external backup directory above. No integration worktree, other package source/build directory, adb device, or remote ref was modified.