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.
15 KiB
Adversarial review — P1 package fix-tests (F3), branch p1/fix-tests @ c67c47f7
Reviewer ran independently in WSL ~/w7/p1-fix-tests (the tree; the task text's ~/w7/fix-tests
does not exist — the implementer declared this, deviation 1, and it is correct).
Tree state at start and at end: git status --porcelain shows only the three untracked build
directories; git log -1 is c67c47f7. Nothing was left changed (two temporary mutations were
made for negative controls and reverted with git checkout --; verified clean afterwards).
Verdict: approved. 0 majors, 6 minors.
I set out to refute "correct and complete" and could not. Every claim in
fix-tests-v1.md reproduced, and two of them reproduce stronger than claimed.
1. What was re-run independently (command + observed output)
1.1 The diff is what it says it is
$ git -C ~/w7/p1-fix-tests diff --name-only 97b997d5..HEAD
MobileGL/MG_Test/Framebuffer/FramebufferTest.cpp
MobileGL/MG_Test/SanityTest.cpp
MobileGL/MG_Test/ScopedPipeVerb.h
MobileGL/MG_Test/Texture/TextureTest.cpp
128 insertions / 5 deletions, all inside MobileGL/MG_Test/**. No production source, no
scripts/, no CMake, no .def/generated file, no fixture, no MG_Test/Pipe/** (the other
package's area), no MG_IntegrationTest/**. The full diff was read line by line.
Commit message shape checked against the lane rule (subject + blank line + - bullets, no
Co-Authored-By): conforms.
1.2 Build configurations are the reference ones (deviation 2/3 verified, not taken on trust)
Diffing every MOBILEGL_* / CMAKE_BUILD_TYPE / CMAKE_CXX_COMPILER cache entry against
~/w7/pipe/<dir>/CMakeCache.txt:
--- build-linux --- IDENTICAL to pipe/build-linux
--- build-push --- only CMAKE_CXX_COMPILER:STRING vs :UNINITIALIZED (same value)
--- build-verify --- only CMAKE_CXX_COMPILER:STRING vs :UNINITIALIZED (same value)
MOBILEGL_BUILD_BENCHMARK:BOOL=OFF is present in ~/w7/pipe's three caches too, so
deviation 3 reproduces the reference trees rather than changing them. Compile definitions
confirmed from build.ninja: build-linux has neither, build-push has
-DMOBILEGL_PIPE_PUSH=1, build-verify has -DMOBILEGL_PIPE_PUSH=1 -DMOBILEGL_PIPE_VERIFY=1.
cmake --build {build-linux,build-push,build-verify} -j 12 → all three ninja: no work to do,
i.e. the committed tree is exactly what was tested.
1.3 Pull-build symbol report, --threshold 0
$ python3 scripts/symbol_report.py --before ~/w7/p1-before-libMobileGL.so \
--after build-linux/libMobileGL.so --threshold 0
symbol-report: .text 10792579 -> 10792579 (+0, +0.000%)
symbol-report: .data 76824 -> 76824 (+0) .bss 1296872 -> 1296872 (+0) .rodata 1534938 -> 1534938 (+0)
symbol-report: 27799 -> 27799 defined symbols: 0 added, 0 removed, 0 resized, 0 renamed
G1 holds. (Trivially, by construction — the helper is header-only test code — but re-run.)
1.4 gen_pipe.py --check / --self-test
gen_pipe: 71 calls (11 screen, 60 context), 69 verify payloads, 63 PipeInputs fields (7 sticky), 69 verbs, 9 classes
gen_pipe: generated files are up to date rc=0
gen_pipe: self-test: 6 negative-control trip(s), positive control OK rc=0
scripts/check_include_closure.py also clean (3 probes, 0 forbidden) — the new header does not
break the layering gate.
1.5 ctest -L unit in all three build directories
build-linux rc=0 100% tests passed, 0 tests failed out of 1482
build-push rc=0 100% tests passed, 0 tests failed out of 1482
build-verify rc=0 100% tests passed, 0 tests failed out of 1482
1.6 ctest -N name diff
build-linux 2360 names, build-push 2360, build-verify 3178
removed vs ~/w7/p1-before-ctest-names.txt: 0 added: 26
vs the integrated tree ~/w7/pipe/build-linux (2360): 0 removed, 0 added <-- exact match
The 26 additions are the other P1 packages' (already at the branch point); this package adds
none, removes none, renames none. The build-verify surplus is confirmed by label census:
integration-verify 818 + integration-gpu 1696; ~/w7/pipe/build-verify also lists 3178.
So the surplus is the configuration, not a name change.
2. The four things I tried to catch it on
2.1 "A test helper that makes tests pass unconditionally" — refuted, twice
Control A (mine, stronger than the implementer's): neutralise the helper entirely.
sed -i 's/^#if MOBILEGL_PIPE_PUSH$/#if defined(MGL_REVIEW_NEUTRALISE) \&\& MOBILEGL_PIPE_PUSH/'
on MobileGL/MG_Test/ScopedPipeVerb.h (all four guards, so ctor, Renew() and dtor become
no-ops in every build), rebuild build-verify, ctest -L unit -j 8:
99% tests passed, 11 tests failed out of 1482
5 DirectGLESSanity.BindsAMultisampleTextureDespiteTheDefaultMipmapFilter (Subprocess aborted)
6 DirectGLESSanity.BindingZeroClearsPreviousNativeTextureBinding
82 DirectGLESTextureSync.UnitMemoRefusesToDriveATwinFromAnotherTexture
310 FramebufferTest.DrawIntoAWidenedDrawBufferReachesTheDriverWithAlphaWritesMaskedOff
311 FramebufferTest.ClearIntoAWidenedDrawBufferKeepsAlphaWritableAndSubstitutesOne
312 FramebufferTest.ApplicationAlphaMaskOffIsStillHonouredOnANativeDrawBuffer
313 FramebufferTest.DualSourceBlendFactorsReachTheDriverWhenTheExtensionIsThere
314 FramebufferTest.DualSourceBlendIsDeclinedRatherThanThrownWhenTheExtensionIsMissing
315 FramebufferTest.DualSourceFactorsAreDeclinedEvenWithBlendingDisabled
316 FramebufferTest.DualSourceFactorsWithBlendingDisabledStillReachACapableDriver
471 TextureTest.StorePackedWordsToClientCopiesWordsVerbatimUnderPackParams
Exactly eleven, and exactly F3's eleven — every declaration in the diff is load-bearing (none is decorative), and no other test in the 1482 depends on the helper. So the helper cannot be making anything pass that was not already failing for the F3 reason, and none of these eleven is a test that cannot fail.
Control B (mine): substitute the WIDEST verb class for a narrow one. In
MobileGL/MG_Test/Texture/TextureTest.cpp:4481 change MGPipeVerb::ReadPixels →
MGPipeVerb::DrawArrays (kDraw, 47 fields, the largest mask there is) and run that one test in
build-verify with MOBILEGL_LOG_FILE_PATH set:
exit=134 (SIGABRT)
[Linux TextureTest/FATAL]: MGPipe: Fatal{UnmigratedPipeInput, "GetPixelStoreParameters@DrawArrays"}
This is the decisive anti-rubber-stamp result: the ctor runs the production
MGPipeFillForVerb (MobileGL/MG_Impl/Pipe/PipeFill.cpp:528-575), which stamps only
kMGPipeClassFieldMask[class of verb] — so a read outside the declared verb's may-read set is
still Fatal{UnmigratedPipeInput}, naming the field and the declared verb, even when the
declared verb is the broadest one. FillPoints.def remains the only thing that says what a verb
may read. It also proves the test binary and the backend share one gPipeInputs: the abort came
out of the backend's accessor, on the block the test's ctor stamped.
Both were reverted with git checkout -- and the affected targets rebuilt; the restored tests
pass and git status --porcelain -- MobileGL is empty.
2.2 "Fields / call sites the enumeration missed" — refuted
Enumerated every direct backend-helper call in the unit tree myself
(awk over the three files, plus a repo-wide grep):
MobileGL/MG_Test/Framebuffer/FramebufferTest.cpp— 11SyncRenderState(calls at:1238,1271,1283,1323,1356,1377,1395,1422,1440,1453,1471; every one sits inside aScopedPipeVerbscope or after aRenew().MobileGL/MG_Test/SanityTest.cpp— 4BindCurrentTextures()(:329,372,380,391) and 2SyncNeccessaryTextures()(:3025,3046); all covered.MobileGL/MG_Test/Texture/TextureTest.cpp— 2StorePackedWordsToClient((:4482,4508); both covered.grep -rlE 'MG_Backend::(DirectGLES|DirectVulkan)::[A-Za-z_]+::[A-Za-z_]+\(' MobileGL/MG_Test --include=*.cppreturns exactly those three files and no others.- The same grep over
MobileGL/MG_IntegrationTestreturns nothing — integration scenarios drive GL entry points only, soMGP_FILLalways fires there. The implementer's stated reason for not touchingMG_IntegrationTest/**is factually right, not an excuse.
2.3 "A mechanism that silences the comparator" — refuted, and the package under-claims
The unit lane in build-verify runs with Features.PipeVerify off unless
MOBILEGL_PIPE_VERIFY=1 is in the environment, so the implementer's green ctest -L unit only
exercised the poison, not the compare-at-read hook. That was the one place a wrongly placed
declaration (fill before a mutation the helper must see) could hide: the push value would go
stale against the live context and nobody would look. So I armed it:
$ MOBILEGL_PIPE_VERIFY=1 ctest --test-dir build-verify -L unit --no-tests=error -j 8
rc=0 100% tests passed, 0 tests failed out of 1482
and proved the arming was real rather than a silent no-op:
$ MOBILEGL_PIPE_VERIFY=1 MOBILEGL_LOG_FILE_PATH=/tmp/mgl-verify.log \
./build-verify/.../FramebufferTest --gtest_filter='*DrawIntoAWidened*'
[Linux FramebufferTest/INFO]: MGPipe: verify armed - 63 fields, 69 verbs, fatal=1
# same run without the env var: 0 occurrences of "MGPipe: verify armed"
With the comparator genuinely armed, every accessor read in all eleven tests returned exactly
what a live pull would have returned. That is a stronger statement than the result file makes,
and it closes the placement question completely: no ScopedPipeVerb in this diff is placed
before a mutation its helper is meant to see.
2.4 "The pull build moves" / "a global escape hatch" / "the poison is weakened" — refuted
- Symbol report 0/0/0/0,
.text+0 (§1.3). The class body is#if MOBILEGL_PIPE_PUSH-gated to nothing (MobileGL/MG_Test/ScopedPipeVerb.h:46,62,67), the ctor signature is identical in both arms, so call sites cannot rot. - No global escape hatch: no
#define, no env knob, no weakening ofMGP_INPUT_CHECK, no change toPipeInputs.h/PipeFill.cpp/FillPoints.def. The whole mechanism is "call the production filler for a named verb". - Whole-binary single-process runs reproduce (the destructor re-arm's stated purpose):
SanityTestexit=0 82 tests,FramebufferTestexit=0 61 tests,TextureTestexit=0 192 tests, 0[ FAILED ]lines each. With the helper neutralised the same three binaries are 134/1/1 with 11 and 3 failures — i.e. the re-arm is not covering anything up.
3. Minors
-
Nested scopes silently invalidate the enclosing declaration, and the header does not say so.
MobileGL/MG_Test/ScopedPipeVerb.h:68— the destructor bumpsCurrentVerbSerial, so closing an inner scope makes the outer scope's stamps stale. The two current sites survive only because adraw.Renew()happens to follow (FramebufferTest.cpp:1453after theClearscope at:1439-1441). The header's guidance at:38-42documentsRenew()as "a second verb of the same kind begins" and nested scopes as the way to handle another class, but never states the interaction. A future author who nests a scope and then calls the outer helper again gets an abort. One sentence in theRenew()comment would fix it. -
The exit fill makes the next poison Fatal name a misleading verb.
ScopedPipeVerb.h:74picksMGPipeVerb::GetGpuTimestampNsas the leave-verb, so a test that forgets a declaration after another scope closed aborts withFatal{UnmigratedPipeInput, "<Field>@GetGpuTimestampNs"}— reading as "a timestamp query tried to read this", which is not what happened. The implementer's own demo A produced the clearer@<none>only because that test declared nothing at all. A dedicated "left the verb" sentinel (or documenting the string) would keep the diagnostic honest. -
Only one of the verb classes that reach each helper is exercised.
SanityTest.cpp:3025,3046declareDrawArraysforSyncNeccessaryTextures(), but that helper is reached in production fromkReadbacktoo —DirectGLES.cpp:9020(ReadPixels) and:9191(GetTexImage), among 14 call sites. Choosing the class the helper "belongs to" is exactly what F3 asks for, and the retrace / integration lanes are the oracle for the other classes (that is F1's job), so this is not a defect — but the result file's table reads as if the declared verb were the only class involved, and it is not. -
The known
GetFramebufferBindingSlotFastpoison gap is what produced the pre-fix push SEGFAULT, and it is worth recording.DirectGLES.cpp:147-157: when the block is unfilled,MGB_CTX_IDENTITYisnullptrandg_fbSlotCacheContextis alsonullptr, so the cache is treated as warm and*g_fbSlotCache[target]dereferences null. This is the documented D4 gap (brief line 151, closed in P2), not this package's, but it means a push-build test that reaches that path can crash or read a stale slot without the poison ever firing — the one place where "green" is not proof. No such test exists today (§2.2). -
A brief-vs-tree fact, recorded per the standing instruction. The brief (G3) calls
integration-verify"the 742-entry lane" andP1-LANE-FINDINGS.mdsays 878 integration entries; the tree says 818integration-verifyand 1696integration-gpu(ctest --test-dir build-verify -N -L <label>), identical in~/w7/pipe/build-verify. The result file's 818 is right; the brief's number is stale. Tree wins. -
The result file under-reports its own strongest evidence. It does not mention that the verify build's unit lane is also green with the comparator actually armed (
MOBILEGL_PIPE_VERIFY=1, §2.3). That is the result that rules out mis-placed declarations, and it belongs in the record.
4. Declared deviations — all six check out
- tree
~/w7/p1-fix-tests(confirmed:~/w7/fix-testsdoes not exist; branch, branch point97b997d5and ownership are as briefed); - the three build dirs did not exist and were configured by the implementer (confirmed
equivalent to
~/w7/pipe's, §1.2); -DMOBILEGL_BUILD_BENCHMARK=OFF(confirmed present in all three reference caches);~/w7/p1-before-ctest-names.txtas the-Nbaseline (I additionally diffed against the integrated tree: 0/0);- the extra wrong-verb demonstration (I reproduced an independent, stronger variant);
- the destructor's re-arm as an unbriefed design choice (reproduced; see minors 1 and 2).
No undeclared deviation found. F1, F2 and F4 are untouched, as stated; the
integration-verify lane and the 79-case retrace were not run here, which is correct — they are
red for F1/F2 reasons this package does not own, and nothing in this diff can reach them
(no MG_IntegrationTest/**, no production source, no fill table).