Files
MobileGL/docs/Disaggregated/notes/recovered/wf2/p0-polish-findings.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

7.4 KiB

P0 polish package: findings from the final review of feat/disaggregated 8349babe..aa005720

All line numbers refer to the tree at aa005720 (~/w7/pipe). Fix every item below.

MAJOR

T-1 (transport, inproc doorbell has no death state -> Shutdown can never join a parked waiter)

InProcessChannel::Close() (MobileGL/MG_Remote/Transport/InProcessTransport.cpp:74-88) rings each CondVarDoorbell once and the comment at :83-84 claims this stops a peer parked mid-frame from hanging. It does not: Doorbell::Wait (Doorbell.h:136-176) consumes that one ring, re-tests ready() (false, nothing was published), tests Dead() - which CondVarDoorbell never overrides (Doorbell.h:113, :199-211; Doorbell.cpp:29-71 has no death state) - and with kWaitForever loops straight back into Park(), forever. Shutdown() of the inproc transport can therefore never join a server thread that is in the design's own steady state (consumer spins, sets consumerParked, then blocks). No test covers it (every Wait in MG_Test/Wire uses 5000/30/0 ms). Fix: give CondVarDoorbell an std::atomic<bool> m_dead + void Kill() (set under the mutex, notify_all), override Dead(), make Park return false immediately when dead, have Close() call Kill() (keep Notify() for raw Park callers), and add InProcessTransportTest.ShutdownUnparksAWaiterWithNoDeadline (a thread in Wait(parked, []{return false;}, kDefaultSpinUs, kWaitForever), Shutdown() from the test thread, join within a bound). Verify the test hangs (or times out) without the fix.

T-2 (transport, ring accepts records it can never place -> silent producer stall)

Ring.cpp:126 only refuses total > m_capacity, but a record with capacity/2 < total <= capacity is unreservable at any offset where neither spaceToEnd >= total nor offset >= total holds, even in an empty ring: Reserve (:137-148) needs spaceToEnd + total bytes when a pad is required, which exceeds capacity there. Concrete: 256-byte ring, one 16-byte record written, popped and retired (head offset 16, FreeBytes()==256), Reserve(1, kRecNone, 240) (total 248) returns nullptr forever; a producer waiting on FreeBytes() >= 248 deadlocks. For total <= capacity/2 the worst case is 2*total-8 <= capacity-8, so half is the correct bound. Fix: reject total > m_capacity / 2 at Ring.cpp:126 (same message, "half the ring"), state Capacity()/2 as the chunking bound in Ring.h:169-174 and in the G3 chunking comment in MobileGL/MG_Pipe generated header comments / gen_pipe.py where the "record larger than the segment -> chunk" rule is stated, and add RingTest.RecordLargerThanHalfTheRingIsRefused plus the offset-16/248 scenario as the negative control (today it spins on nullptr).

P-1 (interface hole: the six backend-owned compute limits have no carrier)

MGPCaps (MobileGL/MG_Pipe/MGPipeTypes.h:123-140) is DynamicBackendParameters + CallMask + two blobs, and DynamicBackendParameters (MobileGL/MG_Backend/BackendObject.h:316-530) carries MaxComputeWorkGroupInvocations but no per-axis MaxComputeWorkGroupCount[3] / MaxComputeWorkGroupSize[3]. Those six are exactly the backend-owned answers that survive the getter retirement (GL_Getter.cpp:1160-1177, MG_Util/ShaderTranspiler/CompileEnv.cpp:134-138, DirectVulkan.cpp:672-703 from VkPhysicalDeviceLimits), yet the .def footer (MobileGL/MG_Pipe/PipeCalls.def:143-145) repeats the disproved claim that "only GL_COMPUTE_WORK_GROUP_SIZE is a real backend answer and it lives in MGPCaps". Fix: add Int MaxComputeWorkGroupCount[3]; Int MaxComputeWorkGroupSize[3]; to DynamicBackendParameters and fill them in BOTH backends at capability init (DirectGLES from glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT/SIZE, i) where the other limits are probed; DirectVulkan from VkPhysicalDeviceLimits::maxComputeWorkGroupCount/Size), keep the GetIntegeri_v table path working exactly as today (do NOT change GL_Getter/CompileEnv behaviour in this package - that is P0.5), add the unit/integration assertion that the new fields equal what glGetIntegeri_v answers (AdvertisedLimitsScenario is the natural home), and rewrite the .def footer to say the six limits live in MGPCaps and GL_COMPUTE_WORK_GROUP_SIZE is a frontend link artifact. Update the MGPCaps size assertion accordingly (it is a composition of sizeof(DynamicBackendParameters)+...).

MINOR (one hygiene commit)

  • S-1 stats: PipeStats.cpp OnPresent() does store(0) on g_frameBytes/Calls/Gate* which races with a concurrent Bump from another thread (buffer/texture staging is multi-threaded per the header): a frame-scoped add can be lost from the Tracy plot. Fix: exchange(0, relaxed) and plot the exchanged value.
  • T-3 FdPassing.cpp:207-209: when MSG_CMSG_CLOEXEC is absent (macOS/BSD) received fds have no CLOEXEC; add fcntl(fd, F_SETFD, FD_CLOEXEC) for every received fd under #if !defined(MSG_CMSG_CLOEXEC). Same platform: MSG_NOSIGNAL (FdPassing.cpp:136, Doorbell.cpp:94) is undefined there; #ifndef MSG_NOSIGNAL -> 0 plus SO_NOSIGPIPE on the socket.
  • T-4 CMakeLists.txt:444-445: set(... CACHE BOOL ... FORCE) writes OFF into the cache, so after the operator runs git submodule update a plain re-configure still yields OFF with no message. Use a normal-variable set(MOBILEGL_BUILD_DISAGGREGATED OFF) (shadows the cache for this configure only) with the same message(WARNING).
  • P-2 PipeCalls.def: three live GLFunctionsTable entries have no carrier: GetGpuTimestampNs (GL_Getter.cpp:1355/2265, glGetInteger64v(GL_TIMESTAMP), synchronous server answer), QueryCounterTimestamp (glQueryCounter, not a begin/end pair), WaitSync (GPU-side wait, distinct from FenceWait's client wait). Records are append-only: APPEND QueryTimestamp (kCtxQuery, kReplySlot), QueryCounter (kCtxQuery) and FenceWaitServer (kScreen) at the end of their groups' natural place BUT without renumbering any existing opcode (append to the end of the list if opcodes are positional), regenerate with gen_pipe.py, update the catalogue counts in the .def header and the PipeCatalogueTest expectations.
  • P-3 MGPipeTypes.h:411-417 inlines a 32 B MGHostSpan in every MGPBufferRange of every class, and MGPStreamOutputTargets (:429-436) reuses it: 32 B/entry of dead space for SSBO/atomic/XFB records, and the plan's D-B8 says not to freeze this payload's shape before the stage-ubo-named numbers exist. Keep the range at 24 B and carry the named-UBO host span as an optional var-tail (flag kHostSpan on set_shader_buffers already exists) - adjust the payload, its field list, the size assertion and the test.
  • P-4 (accepted in review 1, never reworked): ConfigLoader.cpp:164-183 QueryEnvUint64 accepts -1 and silently reads octal via strtoull(...,0) - parse base 10 (and 0x prefix explicitly if wanted), reject a leading '-' with the warning; test.yml pipe-gates stdio grep misses fprintf(stdout - extend the alternation; MGPSubData has no buffer offset/size convention - document (and if needed add) how the buffer half of resource_subdata expresses [offset,size) (e.g. UnionBox.x/w with target==Buffer) and assert it in the test; gen_pipe.py does not check that every call payload has a verify field list - make it fail when a payload named in PipeCalls.def has no entry in PipeFields.def (except the documented memcmp-fallback types), and add the check to --check.
  • The first-review minor "MGPPixelPackState size assertion is tautological" (it asserts sizeof against itself): assert against the literal 28 that PixelStoreParameters measured.