Files
MobileGL/docs/Disaggregated/notes/p4a/p4a-results/final-review-v1.md
T
swung0x48 aba329cc53 [Docs] (Disaggregated): P6's plan and contract draft; the working notes move from a local directory into the repository
Two things land together: the plan and contract draft P6 needs before it starts, and the
working notes that until now existed on exactly one machine.

1. P6's plan and contract draft (neither started; not one line of code)

     docs/Disaggregated/P6-SPAWN-PLAN.md      the package plan
     docs/Disaggregated/P6-CONTRACT-DRAFT.md  the contract draft (moves to
                                              MG_Remote/CONTRACT-P6.md when c6 lands)

   The three findings that mattered, all written into the documents:

   a. P6 is SMALLER than its roadmap cell suggests. The transport primitives are already in
      the tree and already tested: SocketDoorbell including hangup detection and the two
      lost-wakeup fences (Doorbell.cpp:211-360, four cases in FdPassingTest), SCM_RIGHTS fd
      passing, ShmSegment::Adopt, and the SurfaceOp / WindowKind schema already defined in
      protocol.fbs. inproc uses a real second shm mapping rather than an alias, and
      SessionRings.h:22-28 says that was done so today is not the first time it runs. P6 is
      assembly plus a control plane plus a process, not "write a socket transport".

   b. The evidence for "P6 is only a transport swap" has EXPIRED. It was bought by P5c, whose
      audit found 59 direct memory accesses in a tree that already looked clean. P5e changed
      the tree again. So the first package, a6, is a read-only audit that writes no code, and
      every contract row that depends on it is marked PENDING a6 rather than decided in
      advance.

   c. P6 does not get a game running on the phone. ANativeWindow* is meaningless in the child,
      and protocol.fbs:186-196's own comment says Android transfers the window out of band -
      that band is P12. P6 lands pbuffer / surfaceless / offscreen, which is what HeadlessGL,
      the split-equivalent lane and trace replay need, and which is exactly what P6's own exit
      gate measures. A real window reaching the server is a named refusal. This gets its own
      section because "P6 is done" would otherwise read as "the game runs now".

   The contract adds rule G: a role may not name a process-local handle. It extends P5c's
   rule E (a role may not name the other role's memory) from memory to identity - EGLDisplay /
   EGLSurface / EGLContext / ANativeWindow* / fd number / pid / mapped address either cross as
   a token this contract defines or they do not cross, and a descriptor crosses only through
   SCM_RIGHTS, where the kernel renumbers it.

   What P5e left for P6 gets its own section. The latch semantics are already designed
   (ARCHITECTURE.md:504) and are copied unchanged. What is new is that after run-ahead the
   client no longer waits while the peer process can DIE: the client holds records that were
   published and will never be applied, and Android's freezer suspends the child without
   closing anything. So slow and dead must be distinguishable WITHOUT a timeout threshold.
   That is P6's only new design; everything else is carriage. It shares a predicate with E1's
   outstanding re-specification (ID-122), and the two should be designed together.

   Two gates carried over from P5e's lessons: G2 requires integration-spawn's test-name set to
   match integration-split's name for name, because a differing set means the lane is quietly
   running less; and the arm-proof gate requires every entry to record the child pid and
   transport=spawn in its OWN private log - ID-124 established that this kind of false green
   really happens, and a spawn lane that actually ran monolith would pass every other gate.
   S2's shape must be settled before it is written, and the document says why: E1 is the
   cautionary case.

2. The working notes move into the repository (docs/Disaggregated/notes/, 415 .md files)

   ROADMAP / CURRENT_STAGE_PROGRESS / MEASUREMENTS cite ruling numbers, audit line numbers and
   the provenance of gate figures from ~/w7/notes/, a directory that exists on one machine and
   is under no version control. A record cited hundreds of times that lives only locally is not
   a record, so it moves in whole, with its directory structure intact.

   Only .md moves. Scripts, run logs, census and gate result JSON, patches, APKs and nm dumps
   stay where they are: they are evidence DATA rather than documents - large, regenerated per
   run, and mostly carrying absolute paths and device serials that mean something only on the
   machine that produced them. The mapping and this split are written in
   docs/Disaggregated/notes/README.md.

   Two exceptions are recorded there too: P6's BRIEF has been promoted to a first-class
   document and keeps no copy under notes/, so the two cannot drift apart; and recovered/ holds
   material salvaged from failed workflow runs, which may duplicate the like-named phase
   directories and is kept because a few of those were the only complete version at the time.

   CI's citation lint takes docs/Disaggregated/*.md only (non-recursive), and notes/ is
   deliberately outside it: those notes' file:line citations were written against their own
   code heads, and resolving them against today's HEAD would fail wholesale - which is not an
   error, it is their correct state as a historical record. To check one, use the revision it
   declares (--rev).

3. Inconsistencies fixed in passing

   - README.md's status line still said P5e was in progress while the stage table already said
     closed.
   - CURRENT_STAGE_PROGRESS's next-step item 1 still listed the exit gate as pending, though
     (a) and (b) were run in the previous commit; rewritten to closed plus the one item that
     actually remains (E1).
   - ROADMAP's P6 row still had the pre-phase plan in three of its cells; it now carries the
     package order, the scope boundary and the two new gates.

Checked with check_doc_citations.py: 125 citations across docs/Disaggregated/*.md, 0 problems.
2026-09-18 23:31:39 -04:00

36 KiB
Raw Blame History

P4a final whole-diff review — v1

Reviewer: final whole-diff review (fable). Tree reviewed: ~/w7/pipe at a38bdab4 = code head 6035c9d7 (+ the tools-only MOBILEGL_TRACE_SKIP_INSTALL commit). Phase diff read in full: git diff 37da3c3a 6035c9d7 (81 files, every hunk, including the tests, the harness, the scripts and the workflow). Experiments ran in the throwaway worktree ~/w7/p4a-finalreview (@ 6035c9d7, build-push / build-linux, base tree ~/w7/p4a-finalreview-base @ 37da3c3a) after ~/w7/p4a-bench.done appeared; the worktrees are removed at the end of this round (§9). Nothing was committed to pipe.

Reading order, as instructed: INTEGRATOR-DECISIONS.md (ID-1..49), BRIEF-P4A.md A/D/E/F, fable-seam-audit.md, fablefix-v1.md, then contract-v7, wire-v3 + review-v2, clientfb-v2 + review-v2, clientsp-v3 + review-v2, esprytobj-v3 + review-v2, esprytdraw-v3, gates-v3, and the three docs. Findings below were refuted against the code before being written; the refutation attempted is stated under each.


0. Verdict

REWORK — two critical defects in the texture-resource family's CLIENT half (package B's TextureEmit.h), both structural, both invisible to every gate that shipped, both with a small and well-bounded fix:

  • C-1 — TextureEmit.h never passes wire's MGPRespecifiedLevel, so every per-level glTexImage*D / glCopyTexImage2D / glGenerateMipmap grow clears the applier's WHOLE pending-upload set. A level accepted at one verb and defined-around at the next is uploaded by nobody; the client's flag is already clear (D-D5 step 1). Silent, no counter, no log line.
  • C-2 — a dead-but-not-recycled texture handle still resolves to the freed ITextureObject* inside the emitter, and the drain list is not emptied at death. glTexImage2D → glDeleteTextures → next verb walks freed memory (virtual GetStorageType() on it), and keeps walking it once per verb until the slot is recycled.

Both were confirmed by probes in the review worktree (§2), and C-1's shape was documented as B's obligation by wire-v3 §5 item 5 and never checked by clientfb-review-v2. Everything else in the phase is CLOSE-quality: the contract as landed is consistent (one encoding per field, six death helpers in one fixed order, c0f/c0g complete and non-transitive on both sides, the shutter table complete after the fable round), the applier is a per-object table with loud stale-generation refusals, Espryt's handle arms refuse loudly at every one of the 27 decline sites I re-walked, the exit-order rule holds, no debug or temporary-base artefact leaked, and every gate I mutated went red for the reason it exists (§4) except the one declared limitation in §4.3.

The rework is one package (B) plus one contract signature (A), ≈ 6 files + 2 tests, and does not reopen D/E/F. After it lands: re-run the list in §8. The recorded items (§5) and the D.5 docs checklist (§6) are unchanged by the rework.


1. Findings

CRITICAL

C-1 — the client never passes the level scope; a per-level respecify drops every pending upload

Where. MobileGL/MG_Impl/Pipe/TextureEmit.h:955-960:

static Bool ApplyRespecify(const MGPResourceDesc& desc) {
    if constexpr (MGPipeTextureRecordsReachTheApplier()) {
        return MGPipeApplyResourceRespecify(desc, nullptr);   // third parameter defaulted: level == nullptr
    }

MGPipeApplyResourceRespecify(desc, initialBytes, const MGPRespecifiedLevel* level) (MG_Pipe/PipeApply.cpp:1612-1613) then takes the whole-resource arm at :1684-1685: else if (level == nullptr) { record->PendingUploads.clear(); }. The non-test tree contains exactly three mentions of MGPRespecifiedLevel: PipeApply.h, PipeApply.cpp, and a comment in Managers.cpp. No emitter constructs one. The hook the frontend calls (MG_State/.../TextureObject.cpp:68-69 PipePublishDescriptor() → MGPipeEmitTextureResourceRespecify(*this), PipeMutation.h:320) carries no level at all, although every caller of it — AllocateStorage (TextureObject.cpp:527, TextureObject2DCube.cpp:68) — is per (uploadTarget, level).

Why the descriptor moves on a level-1 definition (so the metadata-only arm does not save it): TextureEmit.h:239 desc.Levels = mipmap->GetMipmapLevelCount(), and GetMipmapLevelCount() is m_textureStorage.GetLevelCount() (TextureObject.cpp:471-473), which grows as levels are defined. RespecifyRedefinesNoStorage (PipeApply.cpp:558-568) compares Levels, so the call is classified as a redefinition and the clear runs.

Failure scenario (any of these, application-side, all legal GL):

  1. glTexImage2D(L0, data) → any verb that does not reach the texture (a draw with another texture, a glClear, glReadPixels, glGetTexImage on another texture, glBindImageTexture — every MGP_FILL point drains) → glTexImage2D(L1, data) → draw sampling the texture. At the first verb the drain emits L0, the applier accepts it, the client clears L0's flag (TextureEmit.h:1097). Espryt does not sync the texture (bound nowhere / not sampled), so the applier's entry is the only owner of those texels. The L1 definition respecifies with level == nullptr → PendingUploads.clear(). At the sampled draw the handle arm allocates from the descriptor (MGB_STORAGE_*, Managers.cpp:6315-6333) and uploads only what FindPipeTextureUpload returns (MGB_LEVEL_NEEDS_UPLOAD, :6262-6265 — no frontend fallback on the handle arm): L1 only. Level 0 is allocated undefined. No refusal, no counter, no line.
  2. The same with glGenerateMipmap: GL_Texture.cpp:6782-6783 runs EnsureGeneratedMipmapStorageAllocated (one AllocateStorage → one respecify per level, :505-550) BEFORE GenerateMipmap_Backend (MGP_FILL(GenerateMipmap) at :1675). A texture whose level 0 was accepted-but-unconsumed loses it inside the generate; the driver builds the chain from garbage.
  3. Cube maps: six faces of level 0 defined across verbs — each face's AllocateStorage respecifies and clears the other faces' pending uploads.
  4. Texture views: TextureObjectView::AllocateStorage forwards to the owner (TextureObjectView.cpp:272), so a per-level definition through a view clears the OWNER's set.

Evidence.

  • Applier+client probe (worktree, TextureEmitTest with the real frontend AllocateStorage / MarkStorageDirtyRegion and DrainTextureSubData): FinalReviewALevelDefinedAfterAnEmittedButUnconsumedUploadKeepsThatUpload → RED: after the level-1 definition, level 0 is no longer pending and its client flag is CLEAR; after the next drain only level 1 is pending (~/w7/p4a-finalreview-exp.log).
  • End-to-end (worktree itest FinalReviewC1Scenario, DirectGLES, llvmpipe): cases 1 and 4 read a 100 % black level 0 on the handle arm (0x1fff) and red on the legacy arm (0x1ff); the three controls are red on both (§2.1).

Refutation attempted.

  • "Espryt's incomplete-texture bail keeps the level pending until the chain is complete": the bail at Managers.cpp:6480 is !IsComplete() && !HasAnyDefinedMipmapLevel(...) — it only bails when NO level is defined. So when the texture IS reached, L0 is consumed at the first sync and the hazard closes; the hazard needs the texture to be unreached between the two definitions. That is the ordinary loader shape (progress bars, one-level-per-tick uploaders, atlas stitchers that upload level 0 and mip later).
  • "The client re-dirties": nothing in MG_Impl re-marks a level; the only re-dirty is RearmPipeTextureLevelUpload on a backend bail, which needs the set to still hold the entry.
  • "glTexStorage / glBufferData are whole-resource anyway": yes, and those are the calls that SHOULD pass nullptr (wire-v3 §5 item 5 lists them). The bug is that everything passes nullptr.
  • "The applier could infer the scope": it cannot — a Levels move from 1 to 2 and a level-0 redefinition at a new format both change the descriptor; only the caller knows which level moved.

Why no gate saw it. TextureEmitTest.cpp:1567-1645 and :1647-… (wire's C1 cases and gates' MINOR-1 cube case) drive MGPipeApplyResourceRespecify(..., &levelZero) DIRECTLY — they prove the applier's arm and assume B passes the pointer (:1584-1591 even says so in prose). No unit case drives AllocateStorage twice around a drain. clientfb-review-v2 verified M1 (the region tail) at TextureEmit.h:1068-1069 and did not verify wire-v3 §5 item 5. No itest or retrace fixture defines two levels of one texture around a verb without sampling it in between (the itest textures are drawn right after their upload; MC defines every level with NULL data in one burst and sub-images later).

Fix path (B + A, one commit, ≈ 6 files): PipeMutation.h:320 MGPipeEmitTextureResourceRespecify(ITextureObject&, const MGPRespecifiedLevel* = nullptr); TextureObject.h:235 / TextureObject.cpp:68 PipePublishDescriptor(TextureUploadTarget, Uint level) called from AllocateStorage with the level and from TruncateMipmapLevels / the view / immutable paths with none; PipeFill.cpp:1258 forwards; TextureEmit.h:955 builds the key with MGPipePackSubDataTarget(MGPipeResourceTargetForTextureTarget(target), uploadTarget) — the SAME packed Target the drain uses at :1037-1039 (wire-v3 §5 item 5's exact requirement), and RepublishMask keeps nullptr. Tests: the applier+client probe above becomes a permanent TextureEmitTest case; the itest in §2.1 (cases 1 and 4 with their controls) becomes a lane.

C-2 — a dead texture handle resolves to freed memory in the drain; the drain list survives the death

Where. TextureEmit.h:446-453:

ITextureObject* ResolveTexture(MGPipeHandle handle) const {
    ... if (entry.Texture == nullptr || entry.Gen != handle.Gen) return nullptr;
    if (MGPipeSlots().GenOfSlot(MGPipeKind::Texture, handle.Slot) != handle.Gen) return nullptr;
    return entry.Texture;

MGPipeSlotAllocator::Free (SlotAllocator.cpp:184-207) sets Live = false and pushes the slot on the free list; the generation moves only in AllocateFor (:90 ++entry.Gen) at the NEXT acquire. So between the death and the recycle, GenOfSlot == handle.Gen, entry.Texture is the freed object, and ResolveTexture hands it back. The death helper (PipeFill.cpp:1391-1395 MGPipeEmitTextureDestroyAndFree) does not forward to the emitter — TextureEmit.h:927-929 says so in its own words ("the death helper is A's ... and does not forward to this emitter") — and RetireIfRecycled (:931-935) clears the entry only at re-acquire. Nothing removes the slot's DrainKeys / m_drain entries at death.

Failure scenario. glGenTextures; glTexImage2D(data) (or glTexSubImage2D) — the level is on the drain list — glDeleteTextures (the last reference; the frontend destructor runs the helper) — any verb. DrainTextureSubData → EmitOneLevel (:998-1005) → ResolveTexture returns the dangling pointer → AsMipmapTexture(texture) calls virtual GetStorageType() (TextureObject.h:342-346) on freed memory → IsStorageDirty reads freed MipmapStorage. If the allocator has not scribbled the block the read succeeds, a resource_subdata is emitted for a record the wire delete has already dropped, the applier refuses it (a counted no-op, :1082-1083), EmitOneLevel returns false and the entry STAYS on the list (:1090-1095, "retried at the next validate point") — so the freed object is dereferenced once per verb until a new texture recycles the slot. In a process with many texture creates this is short; in one that deletes a temporary texture and then draws for a long time it is a per-draw UAF read plus one MGLOG_E_ONCE line. MALLOC_PERTURB_ / ASan / a reused block turns it into a crash.

Evidence.

  • Emitter probe (worktree, TextureEmitTest): FinalReviewADeadTexturesHandleStillResolvesAndTheDrainWalksIt → IsLive == false passes, ResolveTexture(handle) == 0x5a88f04261f0 (expected nullptr), and DrainListSize() stays 1 after the death (~/w7/p4a-finalreview-exp2.log).
  • End-to-end (worktree itest, case 6 of FinalReviewC1Scenario, 8 create-upload-delete-draw rounds, plain and under MALLOC_PERTURB_=165, both arms): SIGABRT pure virtual method called on the handle arm in the first round, with and without the perturb; the legacy arm passes (§2.1). This is a hard crash on glTexImage2D; glDeleteTextures; <any verb> at the shipping default mask.

Refutation attempted.

  • "The GenOfSlot compare is the guard": it guards a RECYCLED slot, not a dead one — the design comment at :442-445 ("the entry exists only between the create the constructor emits and the destroy the destructor emits") describes an invariant no code establishes.
  • "The applier refuses, so nothing reaches the server": true for the server — the client-side deref is the hole, and the refusal path is what keeps the entry alive.
  • "No lane saw it, so the sequence is rare": no itest or retrace deletes a texture with a level still on the drain list (every itest draws first). CTS negative tests and packed_pixels (upload, check errors, delete, next case draws) do exactly this; G15 has not run yet.
  • The renderbuffer table has the same stale-Gen entry but no drain and no pointer, so no deref; the framebuffer NamedEntry table resolves the live object at the DSA call; sampler CSOs are content-addressed with reference counts; sampler views and shader CSOs are re-acquired from live objects at the validate point. Only the texture drain defers by handle.

Fix path (B + A, same commit as C-1): the texture death helper — or the emitter's own hook from NotifyAndFree's notice — nulls entry.Texture, drops entry.DrainKeys and the slot's entries from m_drain; ResolveTexture tests MGPipeSlots().IsLive(kind, handle) rather than GenOfSlot. The emitter probe above becomes a permanent case; the itest case 6 (with MALLOC_PERTURB_ in its lane environment) becomes the negative control.

MAJOR

M-A — no producer of kMGPipeBindSampler / kMGPipeBindShaderImage; ImageBindableHint is dead

MGPipeTypes.h:256-259 says "P4a is that producer: a texture sets kMGPipeBindSampler, kMGPipeBindShaderImage, kMGPipeBindRenderTarget and kMGPipeBindDepthStencil". The only caller of NoteTextureBoundAs (TextureEmit.h:474) in the non-test tree is FramebufferEmit.h:577, with RenderTarget / DepthStencil. SamplerEmit.h and ImageEmit.h never call it. Consequences:

  • desc.ImageBindableHint = (BindMask & kMGPipeBindShaderImage) != 0 (TextureEmit.h:230, :520, :982) is always 0; the ShaderImage-transition ForceParamsResync at :486 is dead.
  • Espryt's pushedWantsImageBindableStorage (Managers.cpp:6408, read from Desc.BindMask) never fires from the record; the widening still comes from the legacy trigger (SyncTextureObjectToBackend(texture, /*imageBindable=*/true)), so the PICTURE is right — but D-A4 / D-M's mitigation (the hint arriving BEFORE the first image bind so the level is not re-uploaded) is not delivered, and ROADMAP.md:85 open question 2 (the remint pull rate that decides MOBILEGL_PIPE_TEXEL_RETAIN_MB) cannot be answered from this tree — nothing counts a remint.
  • The metadata-respecify arm (ID-18 M4, RepublishMask :975-986) has exactly one live trigger (framebuffer attachment) and none of the two the ruling was written for.

Refutation attempted: ResourceTracker.h:97 returns kMGPipeBindSampler for BufferTarget::Texture — that is the BUFFER family (a TBO's backing store), not a texture. clientsp-review-v2 §"sampler views" checks the view's Target/Res and not the mask.

Fix (C, ≈ 2 lines + 1 unit case): ImageEmit.h's window walk notes each image-bound texture as kMGPipeBindShaderImage; SamplerEmit.h's view acquisition notes kMGPipeBindSampler; ImageEmitTest: "a texture bound to an image unit carries ImageBindableHint = 1 on its next descriptor". Not blocking on its own — recorded as MAJOR because a design ruling (D-A4) is silently undelivered and a ROADMAP number depends on it.

MINOR

  • m-1 F-7 (recorded by the audit, still open): set_texture_params latches BEFORE acceptance. TextureEmit.h:641-643 sets HasParamsLatch / ParamsVersion / SamplerVersion and :676 clears ForceParamsResync before the void MGPipeApplySetTextureParams(params) at :679. A refused record (no applier record — the SD-1/SD-3 shape) leaves the latch advanced; the params are not re-sent until the next glTexParameter* moves the version. PublishCreate (:941-953) and NoteRespecified (:962-968) got the acceptance rule; params did not. Owner B; make MGPipeApplySetTextureParams return Bool (A) and latch on true.
  • m-2 Config.h:336-340 lists 0x200→0x400, 0x400→0x80, 0x800→0x400 and omits the fourth row kMGPipeP4aFamilyDependencies actually enforces (PipeFill.cpp:961-…: bit 10 also requires bit 11). The scenario ASamplerBitWithoutTheTextureBitIsRefusedAndNamed pins the code; the comment misleads a reader of the mask.
  • m-3 MGPipeTypes.h:1107-1108 says the stored BindMask "is sticky and therefore ORed, never replaced"; PipeApply.cpp:1644 replaces the descriptor whole. Equivalent today only because the client's entry.BindMask is itself sticky; fix the comment (or OR in the applier).
  • m-4 F-6 (audit class 1, recorded): encodings stated in package headers, not the contract. MGPipeEncodeImageAccess lives in ImageEmit.h; MGPSamplerView::Target's meaning and the DrawBuffers narrowing are stated at their emitters. Espryt reads them back with its own copies. Owner: contract, P4b.
  • m-5 DirectGLES.cpp:8687-8754 (GenerateThreeChannelFloatMipmapOnCpu, called at :8780): marks frontend levels dirty from inside the backend without RearmPipeTextureLevelUpload; on the handle arm the CPU-built levels reach the driver only at the NEXT verb's drain (self-heals, one verb late). The storage grow beside it (EnsureGenerateMipmapStorageAllocated, :8038 / :8777) is fine — and note it is a SECOND per-level AllocateStorage → respecify path inside the backend that C-1's fix must pass the level through as well.
  • m-6 G9's dirty-surface check is blind to the two P4a bits on the UseProgram row when row AND mark are dropped together (declared in DirtySurface.def:300-306; measured in §4.3). The behaviour is pinned by TrackerTest.cpp:833-849 and the F-1/F-2 itests, not by the gate.
  • m-7 pull-build text change: DirectGLES.cpp:4137/4174 adds a Bool broadcastCountResolved local and an if on the pull path (the only non-#if code change the preprocess diff of Managers.cpp / DirectGLES.cpp / TextureObject.cpp shows; everything else is macro parentheses, __LINE__ immediates and one static_assert string). Benign; the per-symbol size comparison in §3 is the evidence.
  • m-8 the review worktree's include/ska submodule prints fatal: not a git repository on every git status — a wsl_tree.sh-shaped tree does not get the nested-worktree .git pointer right for include/ska; git commit -a fails silently there (it cost this review one false gate reading, §4.1). Tooling note for ~/w7/notes/tools/.

2. Experiments (review worktree, after the bench)

2.1 FinalReviewC1Scenario (DirectGLES, llvmpipe, itest binary run directly with the lane's env)

case sequence 0x1fff handle arm 0x1ff legacy arm
1 PerLevelDefinitionAcrossAnUnrelatedDraw L0(red); draw(other); L1(red); draw(T) magnified BLACK rgba(0,0,0,0), 11625/11625 px — FAIL red — pass
2 ConsecutiveDefinitionsNoVerbBetween (control) L0; L1; draw(T) red — pass red — pass
3 LevelZeroConsumedBeforeLevelOne (control) L0; draw(T); L1; draw(T) red — pass red — pass
4 GenerateMipmapAcrossAnUnrelatedDraw L0(red); draw(other); glGenerateMipmap; draw(T) BLACK 11625/11625 px — FAIL red — pass
5 GenerateMipmapImmediately (control) L0(red); glGenerateMipmap; draw(T) red — pass red — pass
6 ADirtyTextureDeletedBeforeAnyVerbIsWalkedByTheNextDrain 8 × (L0(red); delete; draw(other)) SIGABRT in round 1: pure virtual method called / terminate called without an active exception (rc 134) white, 8/8 rounds — pass
6 under MALLOC_PERTURB_=165 same SIGABRT, same message pass

Notes. 128×96 pbuffer, llvmpipe (LLVM 22.1.6), MOBILEGL_BACKEND_TYPE=DirectGLES, __EGL_VENDOR_LIBRARY_FILENAMES=…/50_mesa.json, the itest binary run directly with --gtest_filter; logs ~/w7/p4a-finalreview/.finrev-logs/c1-{handles,legacy,perturb,perturb-legacy}.{log,mglog}. The three controls (2, 3, 5) are green on the handle arm, so the mip path and the consumed-first path are fine and the two black pictures are exactly the accepted-but-unconsumed window C-1 describes. Case 6's pure virtual method called is the dangling-vtable signature of a virtual call on an object whose destructor has run — AsMipmapTexture → GetStorageType() from EmitOneLevel — and it fires on the FIRST round without any allocator help; the library log carries no was refused line because the process dies inside that first drain. The legacy arm (no drain, no handle table) passes all six.

2.2 First-attempt probes (same worktree, 21:43–22:00, logs ~/w7/p4a-finalreview-exp*.log)

  • TextureEmit.FinalReviewALevelDefinedAfterAnEmittedButUnconsumedUploadKeepsThatUpload — RED (C-1 at the emitter+applier level, driven through the real AllocateStorage).
  • TextureEmit.FinalReviewADeadTexturesHandleStillResolvesAndTheDrainWalksIt — RED (C-2 at the emitter level).

3. G1 / G5 mechanics

  • G5 as landed: p4a_untouched_regions.sh 37da3c3a HEAD rc 0; --self-test 8/8 controls tripped. Mutation (§4.1): a (void)0; statement committed inside RingAvailable → rc 1, the row's sha moves and the region is named. (A comment-only mutation also moves the sha by the script's own rule at :63-72; my first two tries read rc 0 only because git commit -a failed on the broken submodule — m-8.)
  • G1 spot-check at TU granularity (pull configuration, compile_commands.json of both build-linux trees, clang++ -E -P and nm -S -C on the objects):
    • Managers.cpp.o: 1410 defined symbols at base and at tree, 0 differing (size, type, name) rows; DirectGLES.cpp.o: 1004 and 1004, 0 differing. The symbol SET and every symbol's SIZE are unchanged in the pull configuration of the two files the phase touched most.
    • Preprocessed bodies: BackendSamplerObject::SyncToBackend identical (the widened signature is #if MOBILEGL_PIPE_PUSH-guarded at Managers.h:2552-2564, so the pull symbol keeps its name); SyncMipmapsToBackend and SyncBuiltinSamplerToBackend differ only in __LINE__ immediates inside ErrorLopper::Loop lambdas and in the extra parentheses the MGB_* macros add (no codegen effect — the nm rows above are the proof); the TextureObject.cpp TU (first attempt) differs only in a static_assert string. The one pull-path CODE change found is m-7.
  • D-N regions: the 17 rows are the same 17 the brief lists (15 Managers.cpp functions, DepthStencilSamplingReadImpl namespace, ShouldUseCaveatTextureFormat), FlushPendingRangesFrom consulted against the pinned sha unconditionally (D-N/2, declared in the script header).

4. Gates — can each go red?

4.1 Mutations this round (worktree; every edit restored, tree back at 6035c9d7, git status clean of tracked changes)

gate mutation result
G5 p4a_untouched_regions.sh (void)0; inside RingAvailable, committed rc 1, RingAvailable named, sha row moves
G9 gen_pipe_dirty_surface.py --check SetBlendFunc row drops NEW_PIPELINE_STATE rc 1 "MISSING publisher NEW_PIPELINE_STATE for SetBlendFunc"
G9 same UseProgram row drops `NEW_SAMPLER_VIEWS NEW_SHADER_IMAGES`, marks kept
G9 same (first attempt) row AND undecided marks dropped rc 0 — blind (m-6, declared)
G7 p4a_descriptor_negative_control.sh build-push its own two field drops rc 0: Layered (FramebufferEmit.h, 3 assignments neutralised) tripped, borderColorForm (SamplerEmit.h) tripped, both suites green again after restore; tree clean

4.2 Mutations from the first attempt (same worktree, ~/w7/p4a-finalreview-exp.log)

gate mutation result
P4aSeamAudit F-3 delete MGP_NOTE_AGGREGATE(FramebufferAttachment) at TextureObject.cpp:83 ATextureRespecifiedWhileAttachedReachesTheFramebufferRecord red
G9 white-box delete SyncReadFramebufferTextureAttachments(...) at DirectGLES.cpp:2309 AReadAttachmentOnlyTexturesDepthStencilModeReachesTheDriver red (6402 ≠ 6401)
G12 / D-K2 PipeSubsystemDependencyMissing returns false ObjectSubsystemControl.Refused.…ASamplerBitWithoutTheTextureBitIsRefusedAndNamed red, names 0x9ff and the two bits

4.3 Standing observations on the gates

  • .github/workflows/test.yml:9-12 still carries the TEMPORARY feat/disaggregated trigger — left in place as instructed; the integrator removes it at the dev merge. BASELINE: "37da3c3a" at :1599; --expect-probes 4 at :802 pinned with check_include_closure.py's PROBES list.
  • G2 (pull == push test names): not re-taken here (the review's pull tree has no test binaries); gates-v3 §6 row 2 is the standing evidence.
  • The subdata-refusal line (resource_subdata for texture {…} level N was refused, TextureEmit.h:1092) is not one of the ten census shapes esprytdraw §6.1 greps, and no itest reads RefusedSubDataCount. Add it to the post-merge census (it is C-2's tell).
  • ctest re-runs on the final tree this round (build-push, llvmpipe/lavapipe): -R 'ObjectSubsystemControl|TextureParamsWithoutASamplerView|TextureUploadShape|P4aSeamAudit|HandleRecycle' 181/181; -L unit -R 'TextureEmit|FramebufferEmit|SamplerEmit|ImageEmit|ProgramEmit|Tracker|CompositeResolver|PipeCatalogue|ResourceEmit|Sanity|ProgramArtifacts' 304/304. Green — and, as §1 shows, green over both criticals: none of the 181 defines a level around a verb or deletes a texture with a level on the drain list.

5. Recorded for later (owner, phase)

# item owner phase
R-1 M-A producers for kMGPipeBindSampler / kMGPipeBindShaderImage + the remint counter ROADMAP.md:85 asks for C before D.5 docs if cheap, else P3b/P4b
R-2 m-1 F-7: MGPipeApplySetTextureParams returns Bool; latch on acceptance A + B with the C-1/C-2 rework (same files)
R-3 m-4 F-6: image-access / sampler-view target / draw-buffer encodings stated once in the contract contract P4b
R-4 m-6: G9 derivation for UseProgram / BindVertexArray (the attachedShaders / Access undeclared-member taint) A P4b
R-5 SD-5: 0x7ff half-run — c0g's table now refuses it on the client; the server's table (ResolveTextureResourceSubsystemArm) is the same four rows; keep both lists in one place D P4b
R-6 g_rawDepthFetchSamplerState (a frontend SamplerObject inside MG_Backend, synced through the "object is the authority" branch Managers.cpp:12361-12372) — the one place the handle arm reads a frontend object on purpose (S-1, audit §B) D P3b/P4b (ROADMAP.md:23)
R-7 m-5: GenerateThreeChannelFloatMipmapOnCpu rearm D P4b
R-8 ProgramArtifacts.h:563-565 NDK/libc++ size pin still inert (brief D.5 chore) integrator D.5
R-9 Post-merge census must include the retrace logs (grep -c 'no applier record|does not describe|declining the built-in|cannot be pushed|was refused' retrace-out/*/*/DirectGLES/output/mobilegl.log) — audit §B's correction plus C-2's line integrator D.3 re-run
R-10 HandleRecycle.AbaControl for the six P4a kinds is inert on both backends (no PipeHandleAbaControl consumer over the object slot tables; the cases say so and assert the correct pixels) — a real control needs one if in GetOrCreateByHandle D P4b
R-11 TextureUploadShapeScenario recorded, not gated (D-D4) — becomes a gate with the Mali delta F P3b/P4b
R-12 m-8 tooling: nested-worktree include/ska .git pointer in wsl_tree.sh-shaped trees tools now

6. D.5 docs checklist (state at 6035c9d7 / a38bdab4)

All D.5 placeholders are still unfilled; scripts/check_doc_citations.py docs/Disaggregated/*.md must be run after each.

doc line brief asks state
README.md:3 status → P4a landed + hash; next P3b/P4b, P5 not updated
ROADMAP.md:3 — still "P3a 已落地 … 下一步 P4a"
ROADMAP.md:20 ✅ + the real numbers (15 calls, 6 kinds, memos retired, two red-before logs, CTS blocks, T3) not marked
ROADMAP.md:23 what P4a left to P3b/P4b (D-Q table) not updated
ROADMAP.md:67 39-day checkpoint: actual elapsed days per package "P4a > 39 天
ROADMAP.md:85 (Q2) texture-remint pull rate not updated — and NOT MEASURABLE on this tree (M-A)
ROADMAP.md:91 (Q8) one reflection archive, Espryt half not updated
ROADMAP.md:100 (Q17) create-indirect state at exit + fix the :7→:98 citation not updated (SD-0 fixed it; say so)
ARCHITECTURE.md:63 [deviation] D-F2 identity-addressed sampler views for Espryt not updated
ARCHITECTURE.md:100 read-attachment-only parameter gap closed, scenario red-first not updated
ARCHITECTURE.md:134 [deviation] D-E1 MGPTextureParams 40 bytes + built-in sampler CSO not updated
ARCHITECTURE.md:136 MGPFramebufferState::Target, Complete = CheckCompleteness() (D-C3); add Named (ID-27/D-I2) not updated
ARCHITECTURE.md:212 CompositeResolver file + two release paths not updated
ARCHITECTURE.md:249-254 drain list is P4a's, cursor is P3b/P4b's, TextureUploadShapeScenario built/gated not updated
ARCHITECTURE.md:261 ProgramArtifactsCodec.{h,cpp}, verify-only not updated
ARCHITECTURE.md:294-295 kNeedsAck narrowed to the buffer target (D-A2) not updated
ARCHITECTURE.md:318/:321 P4a's six byte-identical rows beside P3a's eleven (now 17 rows in p4a_untouched_regions.sh) not updated
ARCHITECTURE.md:323 g_rawDepthFetchSamplerState still a frontend object inside the backend not updated
ARCHITECTURE.md:369 MOBILEGL_PIPE_LEGACY_MEMOS P4a arm + the +1 day the only P4a mention in the file; the day is not recorded
ARCHITECTURE.md:598 push default 0x1fff, bits 9-12, three dependency refusals (there are four: add 10→11) not updated
MEASUREMENTS.md new numbered P4a section (D.4.2 table with the acc/draw caveat above it, D.4.3 headline numbers, T1/T2/T3 + mc_state_toggle, six counters, four suppressors, per-bit fire rates, two red-before logs, Track H census, CTS blocks, peak-RSS hedge) no P4a section; DriverBench rows exist in ~/w7/p4a-bench.log (e.g. magma mc_pass_switch pull 518785.6 / push 513177.1 / native 475307.8) — transcribe with the two-device A/B
MEASUREMENTS.md:465 reading (5) discharged: CSO blob byte class numbers not updated
MEASUREMENTS.md:519 g_uploadRing reset asymmetry still open + ScopedDefaultUnpackState::s_synced (D-O) beside it not updated
ProgramArtifacts.h:563-565 NDK size pin (chore) inert

Two additions the brief did not foresee: (a) the C-1/C-2 rework and the MGPRespecifiedLevel rule belong in ARCHITECTURE.md:249-254's texture paragraph ("the client passes the level it redefines; whole-resource calls pass none"); (b) ARCHITECTURE.md:598 should list the FOUR dependency rows.


7. Public-GL sequences still uncovered by the retrace / itest lanes (audit classes extended), ranked

  1. Per-level definition around a verb (C-1): glTexImage2D(L0); verb; glTexImage2D(L1) — also the glCopyTexImage2D, glCompressedTexImage2D, cube-face and texture-view forms. No fixture, no itest.
  2. Upload-then-generate around a verb (C-1/4): glTexImage2D(L0); verb; glGenerateMipmap.
  3. Dirty-then-delete (C-2): a level on the drain list when the last reference drops. CTS negative tests and packed_pixels do this constantly; G15 is the first place it will run.
  4. Image-unit-only textures across program switches (M-A/F-2): bound with glBindImageTexture, never sampled, program changes the window — F-2's itest covers the window; nothing measures the remint.
  5. Sampler-object binding churn (glBindSampler on registered samplers across many units) — SD-1's two scenarios are the only coverage of the registered-sampler CSO path; no retrace fixture binds application samplers (audit §B: zero occurrences in the 20 traces).
  6. Framebuffer death while bound / recycled name with a Named record pending — HandleRecycle covers the bound case; DSA-configured-then-deleted-then-recycled (Named record on a slot whose Gen moved) is only covered by FramebufferRecordFor's loud refusal, not by a scenario.
  7. Level-scoped respecify of a cube face while the other faces are pending — gates' MINOR-1 unit case pins the applier; no GL-level case.
  8. Texture views over a mutable owner redefined per level (class of 1).
  9. glTexParameter* refused-then-repeated (m-1): the latch-before-acceptance shape has no lane because the refusal needs a texture without an applier record (S-1/SD-3 shape), which the fixes removed; a MOBILEGL_PIPE_PUSH mask that arms bit 10 without a consumer (NoConsumer lane) is the place to add it.
  10. The 0x7ff / 0x5ff half-masks with a REGISTERED sampler bound (SD-5 was measured without one).

8. What the integrator must re-run

After the C-1/C-2 rework lands (B + A; and R-2 if taken with it):

  1. -L unit on pull, push and verify; the two probe cases from §2.2 and the two itest cases from §2.1 (cases 1, 4, 6) added as permanent tests, with case 6's lane carrying MALLOC_PERTURB_=165.
  2. ctest -L integration-gpu on build-push at 0x1fff, 0x1ff, 0x9ff, 0x5ff, consumer / no-consumer, plus -L integration-verify — the 1079/1079 answer gates-v3 §6 names.
  3. The full retrace at the default mask (79 cases) AND the retrace-log census of R-9 (the ten shapes + was refused must all be zero).
  4. p4a_untouched_regions.sh 37da3c3a HEAD + --self-test; p4a_descriptor_negative_control.sh; gen_pipe_dirty_surface.py --check; the include-closure step.
  5. G1 symbol report (symbol_report.py --fail-on-symbol-set-change --fail-on-added-bytes 0) — the rework touches PipeMutation.h / TextureObject.h / TextureObject.cpp, all of which must stay #if MOBILEGL_PIPE_PUSH-guarded on the pull side.
  6. The device A/B and DriverBench rows are unaffected by the rework's mechanism but the two-device Release A/B must be re-taken on the rework head before the numbers go into MEASUREMENTS.md.
  7. Remove the TEMPORARY trigger at test.yml:9-12 in the dev merge commit; D.5 docs (§6); check_doc_citations.py; G15 scheduled after the verdict.

9. Artefacts and cleanup

  • Kept beside this report in ~/w7/notes/p4a/p4a-results/final-review-v1-logs/: the probe scenario FinalReviewC1Scenario.cpp (drop-in for MG_IntegrationTest/Scenarios/, add one line to the CMake list), the two unit probes finrev-repro.py / finrev-repro2.py (they splice a TEST into TextureEmitTest.cpp), the itest run logs c1-{handles,legacy,perturb,perturb-legacy}.log, the gate-mutation outputs g5-*.err / g9-*.out / g7.log, the ctest subset logs, the four nm -S symbol tables of §3, and the first attempt's p4a-finalreview-exp*.log.
  • Removed: the worktrees ~/w7/p4a-finalreview and ~/w7/p4a-finalreview-base (git worktree remove --force + prune from ~/w7/pipe), their build dirs, ~/w7/finrev-g1/ (42 MB of .i) and the first attempt's ~/w7/p4a-finalreview-*.log. pipe itself was never checked out, edited or committed to (HEAD a38bdab4, no tracked change).