Files
MobileGL/docs/Disaggregated/notes/p7/magma-c.md
T
swung0x48 f1b04edfc6 [MG_Backend, MG_IntegrationTest, docs] (P7): the magma vertex-layout refusal splits into one protocol fatal and seven masked attributes
CONTRACT-P7 §3.2's `vertex-layout` and `vertex-format-conversion` rows.
BuildWireVertexInput answered eight different conditions with one `false`, and the single
caller answered that with a P7-marked MagmaWireFatal. So a Magma split session that met a
GL_FIXED array, a dvec on a device with native fp64, or any format without
VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT did not decline the draw: it killed the process. The
monolith arm on the same device has always masked the attribute out and then declined the
draw loudly and once.

The split is by what the reason is ABOUT, not by severity. `buffer-window` is a statement
about the RECORD - an enabled attribute the program reads whose slot lies outside the
window set_vertex_buffers published - so it stays a named Fatal through the hook; it is
what caught the Redmi VAO hash collision fixed in 376c04be. Every other reason is a
statement about what Vulkan can express on this device, and those now set
unsupportedAttribMask and continue, with the monolith arm's own pre-flight in WireDraw.inc
declining the draw when a masked bit is one the program actually reads.

The contract says "the other six"; there are seven. `offset-overflow` post-dates the
audit's count and is an arithmetic guard rather than a shape one. It is masked with the
rest: an attribute whose base offsets cannot be added is one Vulkan cannot fetch, the
monolith arm never computes that sum at all, and masking is strictly safer than a Fatal
for a number no application can reach on purpose.

unsupportedAttribMask now enters the wire layout hash, for the reason the monolith entry's
hash already carries it: two VAOs can otherwise agree on every binding, attribute and
divisor and differ only in which enabled attribute was masked, and a blind hash would hand
one of them the other's pipeline. Omitting it was safe only while the mask was always zero.

vertex-format-conversion becomes a decline for a narrower reason: the converter is only
ever asked for a conversion BuildWireVertexInput already chose, so a refusal means the two
disagree - a defect on this side, not a peer's malformed record.

@P7 census: 15 sites / 4 files -> 13 / 3. Renderer/WireDraw.inc leaves the list entirely.
fatal_census.py unchanged at 79 abort sites / 43 family words / 0 unmarked, as expected:
wave 0 already funnelled these, so retiring a MagmaWireFatal CALL moves no abort count.

New MagmaVertexLayoutScenario, DirectVulkan only (Espryt hands GL_FIXED to a driver that
supports it, so the same name would assert a different fact there). DataType::Fixed32 has
no arm in ToVkVertexFormat at all, which makes it the one unmappable shape reachable on
host lavapipe - §2.5 records that no other P7-marked refusal is. One case has the program
READ the array and asserts the draw declines; the other leaves it enabled but unread and
asserts the draw lands, which is what stops the first being satisfied by "decline every
draw that mentions GL_FIXED". Registered on all three transports and picked up by the
ambient monolith DirectVulkan discovery, so the A/B is direct.

RED-ONCE: `format-map` put back on the protocol verdict reds the three split arms with
`Subprocess aborted` / `Fatal{UnmigratedVerb, "Magma:vertex-layout-buffer-window"}` while
the monolith control stays green - exactly the divergence this closes. Restored: 10/10.

The gating Magma lanes grow with the new entries: split 73 -> 75, spawn 52 -> 54, tcp
54 -> 56. Growth only.

Open for the integrator: the honest family word for buffer-window is ProtocolCorruption,
and MGPipeFatalFamily carries only UnmigratedVerb and RoleViolation. Adding a third costs
a row there plus an adapter arm, and §3.3 keeps the three Magma funnels on UnmigratedVerb
deliberately so the refusal census stays comparable to P5b. Flagged, not decided.
2026-09-22 16:07:59 -04:00

14 KiB
Raw Blame History

P7 wave 2 package C — Magma object death, resident sub-data, the reflection archive, the vertex-layout split

Base: p7/magma-c off feat/disaggregated @ a686131e (pipe HEAD at branch time). Build: build-split = -DMOBILEGL_BUILD_DISAGGREGATED=ON -DMOBILEGL_BUILD_DISAGGREGATED_INPROC=ON -DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_BUILD_INTEGRATION_TEST=ON, ICD pinned to /usr/share/vulkan/icd.d/lvp_icd.json (the same lavapipe ICD ~/w7/pipe/build-split pins; the worktree tool leaves MOBILEGL_ITEST_VK_ICD empty and CMake warns about it).

Contract rows: CONTRACT-P7 §5.5 (StateObjectDeathOps), §5.4 (OQ-10), §5.3 (OQ-8), §3.2 (vertex-layout, vertex-format-conversion). Discipline: rule I (no std::abort() on a Magma wire arm), rule J / R-16 (every fix red once on a two-process arm first).


Slice 1 — Magma StateObjectDeathOps (§5.5)

What landed. MG_Backend/DirectVulkan/DirectVulkan.cpp gains g_magmaStateObjectDeathOps and InstallStateObjectDeathOps(), installed from BackendObject_DirectVulkan::Initialize() — i.e. from ServerLoop::CreateBackend, step 1 of InitServerRoleCommon, which both Magma server roles (inproc and the spawn/TCP child) run and which is before any client session exists. Guarded #if MOBILEGL_BUILD_DISAGGREGATED && MOBILEGL_PIPE_PUSH, so the pull build compiles nothing new (G1).

The one arm, and why it is only one. Espryt's g_glesStateObjectDeathOps has two jobs: emit the wire object_death record off the apply thread, and otherwise destroy a backend twin keyed by lifetime id. Magma has no lifetime-id-keyed twin at all — its two client-minted kinds (VertexElementsCso, Buffer) are keyed on {slot, gen} by MagmaPipeIdentityTables and the other six are still reached from their frontend objects — so the Magma table is exactly the EMIT arm and nothing else.

Correction to the plan's wording (worth the integrator's attention). Plan §1.3 and the task brief describe the gap as a client SLOT leak provable with PipeSlotPeek. It is not, and has not been since P4a: MG_Impl/Pipe/PipeFill.cpp's NotifyAndFree frees the client slot unconditionally and backend-neutrally, so PipeSlotPeek never saw a Magma leak and does not see a change now. What leaked was the server's twin. A framebuffer has no wire delete opcode at all (BRIEF-P4A D-I2), so object_death is its only death delivery; with no consumer installed under Magma inproc, the server's record stayed Live for the life of the session. The CtWireScenario framebuffer case now reads BOTH numbers — the server's deaths tally and the client's PipeSlotKind::Framebuffer live count — so a future regression in either half is attributable.

Cleanup of the Magma special cases. The two executable ones are the CtWireScenario skip guards (:153 / :209), both deleted. Everything else the audit counted is comment text, and it is reworded rather than deleted because the underlying argument survives: PipeSlotPeek.h, MagmaPipeArms.h (the age-sweep-instead-of-death-notice paragraph — still true, because the new table never touches the per-renderer identity tables), PipeMutation.h, PipeFill.cpp, VertexArrayObject.cpp, DirectGLES/Managers.cpp, HandleRecycleScenario.cpp, WireTables.cpp.

The install-order question, answered rather than assumed. SetStateObjectDeathOps is a bare last-writer-wins pointer with no stacking, so two installers could in principle race. They cannot meet: Magma's runs only in a process that OWNS a DirectVulkan backend, and InstallClientWireTables' arm runs only under Transport == Spawn, which is by construction the process that owns no backend (InitSplitRoles builds a BackendObject_Remote there). The red-once below measures exactly that split of responsibility.

Proof (green). ctest -R "DirectVulkan\.(Split|Spawn|Tcp)\.Ct\..*Death" — 6/6 pass (integration-magma-all-{split,spawn,tcp}), plus the two TcpServer fixtures. DirectGLES control ctest -R "DirectGLES\.(Split|Spawn|Tcp)\.Ct\..*Death" — 6/6 pass, unchanged.

RED-ONCE (executed). BackendObject_DirectVulkan::Initialize()'s call changed to if (false) InstallStateObjectDeathOps();, rebuilt, lane re-run:

1/8 DirectVulkan.Split.Ct.CtWireScenario.TextureDeath...        ***Failed
2/8 DirectVulkan.Spawn.Ct.CtWireScenario.TextureDeath...            Passed
4/8 DirectVulkan.Tcp.Ct.CtWireScenario.TextureDeath...              Passed
5/8 DirectVulkan.Split.Ct.CtWireScenario.FramebufferDeath...    ***Failed
6/8 DirectVulkan.Spawn.Ct.CtWireScenario.FramebufferDeath...        Passed
7/8 DirectVulkan.Tcp.Ct.CtWireScenario.FramebufferDeath...          Passed
75% tests passed, 2 tests failed out of 8

CtWireScenario.cpp:264: Failure
Expected: (afterCounters.deaths) > (deathsBefore), actual: 0 vs 0

Restored and rebuilt: 8/8 green again. This is precisely the shape package L's temporary probe predicted (notes/p7/magma-two-process-first-run.md §6): spawn and tcp were already carried by InstallClientWireTables' emitter, and inproc was the only hole — which is the hole this table closes.


Slice 2 — OQ-10 kCapResidentSubData (§5.4)

What landed. InitServerRoleCommon ORs kCapResidentSubData when g_resourceOpsAtStep2->SubDataResident != nullptr — the pointer that function already pinned at step 2 and that step 5 already refuses to see swapped. Never from ActiveBackendType (ID-39). rsd= joins the emit[] bracket on the PipeStats summary line, counted at MGPipeEmitBufferSubDataResident once per record. SplitRuntimePeek gains residentSubDataCap, read through the same CapsMirrorInstance().HasCap() accessor the frontend's MGPipeResourceOpsHaveSubDataResident reads.

BLOCKER FOR THE INTEGRATOR: §5.4's white-box case is not achievable at this tree

§5.4 asks for a case in integration-magma-buffers and one in integration-split proving that the emitted record is BufferSubDataResident (opcode 49). Neither can exist today, and the reason is R-6 rather than anything this package did or did not do. The chain:

step site verdict under split
the resident emitter runs only from a resident store BufferObject.cpp UploadSubData → if (m_resource.IsGpuResident()) —
residency comes only from AdoptPersistentMap PipeResource.h:181, the only writer of m_gpuMapped —
adoption needs a non-null map_persistent answer BufferObject.cpp TryAdoptLargeStorage / EnsureGpuResidentStorage / AcquireMemoryRange —
the answer is a decline PipeApply.cpp:2331 "A SPLIT BUILD RUNS AT TIER T2 AND DECLINES EVERY ACQUISITION, ALWAYS" always
the tier gate has no other implemented value PersistentMapTracker.cpp:1033 AdoptTierIsEmulate — T0/T1 are P11's and are a named Fatal{UnimplementedAdoptTier} always

So under any transport no buffer is ever GPU-resident, LandBytesIntoResidentStore is unreachable, and opcode 49 cannot cross. The capability is therefore wired ahead of its consumer, deliberately: when P11 lands a tier that adopts, the bit is already correct and already derived from the table, and rsd= is the counter that will show the records. The plan's phrasing ("both backends fall back to the in-place memcpy at BufferObject.cpp:~711-716") is one step optimistic — under split the code does not reach that fallback either, it takes the plain shadow Memcpy in UploadSubData above it.

What is proved instead, and it is both halves of the bit.

  • Server half — CapsMirrorTest.MagmaTransportPublishesRealBufferConsumersWithoutRunAhead (RemoteClientTest.cpp) now asserts (mask & kCapResidentSubData) != 0 == (ops->SubDataResident != nullptr). The two sides are compared to each other rather than both to true, so the case also covers a backend WITHOUT the arm on the day one exists.
  • Client half — the new CtWireScenario.TheServerPublishesTheResidentSubDataCapabilityFromIts OwnTable, registered through the ctCase list so tier 2 replays it: green on DirectGLES.{Split,Spawn,Tcp} (integration-{split,spawn,tcp}) and DirectVulkan.{Split,Spawn,Tcp} (integration-magma-all-*), 6/6.
  • Negative half — CapsMirrorTest.APlaceholderMirrorConsumesNothing keeps the no-snapshot-withholds-the-bit reading, which stopped being vacuous the moment anything set it.

RED-ONCE (executed, both levels). capBits |= MG_Pipe::kCapResidentSubData; → capBits |= 0;:

RemoteClientTest.cpp:692: Failure
Expected equality of these values:
  (server.CallMask() & static_cast<Uint64>(kCapResidentSubData)) != 0
    Which is: false
    Which is: true

25% tests passed, 6 tests failed out of 8
  DirectGLES.{Split,Spawn,Tcp}.Ct....ResidentSubDataCapability...   ***Failed
  DirectVulkan.{Split,Spawn,Tcp}.Ct....ResidentSubDataCapability... ***Failed
CtWireScenario.cpp:262: Failure
Value of: runtime.residentSubDataCap
  Actual: false

Restored: 11/11 unit and 6/6 integration green again.

One incidental finding, recorded because it cost a cycle. The first cut of the capability case did no GL work and failed on ScenarioFixture.h:90 — an armed split lane refuses a case whose encoder ordinal did not move, because "emitted nothing" and "resolved the transport and then put nothing on the wire" are the same observation. The case now clears and reads back first. This is the same F1 arming rule package L hit on tier 3.


Slice 4 — the vertex-layout split and vertex-format-conversion (§3.2)

What landed. BuildWireVertexInput answered eight conditions with one false, and the caller answered that with a P7-marked MagmaWireFatal — a session kill where the monolith arm masks the attribute out and declines the draw. Split by what the reason is ABOUT:

reason before after
buffer-window false → P7-marked Fatal false → named Fatal Magma:vertex-layout-buffer-window (no P7 marker: it is a permanent protocol error, not an unmigrated shape)
attribute-shape, native-fp64-format, format-map, native-format-feature, converted-format-feature, element-size, offset-overflow false → Fatal MGLOG_E_ONCE + unsupportedAttribMask |= 1<<loc + continue, then the monolith arm's own pre-flight in WireDraw.inc declines the draw when a masked bit is one the program reads
vertex-format-conversion P7-marked Fatal MGLOG_E_ONCE + return false (decline, rule I (a))

unsupportedAttribMask now enters the wire layout hash, for the reason the monolith entry's hash already carries it: two VAOs can otherwise produce identical bindings/attributes/divisors and differ only in which enabled attribute was masked, and a blind hash would serve one the other's pipeline. It was safe to omit only while the mask was always zero.

Count correction. CONTRACT-P7 §3.2 says "the other six". There are SEVEN non-buffer-window return sites in the tree: offset-overflow post-dates the audit's count and is an arithmetic guard rather than a shape one. It is masked with the rest — an attribute whose base offsets cannot be added is one Vulkan cannot fetch, the monolith arm never computes that sum at all, and masking is strictly safer than a Fatal for a number no application can reach on purpose.

Family-word question for the integrator. buffer-window now goes through MagmaWireFatal, whose family is UnmigratedVerb. The honest word for "the record's own numbers are impossible" is ProtocolCorruption. MG_Pipe::MGPipeFatalFamily carries only UnmigratedVerb and RoleViolation and its own comment says a third "costs a row here and an arm in the adapter - which is the point, because that arm is where the choice gets argued". Not taken here: it is a cross-package edit (MG_Pipe/PipeSessionFail.h + the MG_Remote adapter) and §3.3 records that the three Magma funnels keep UnmigratedVerb deliberately, because renaming invalidates the refusal census kept since P5b. Flagged rather than decided.

Census. grep -rn "@P7" MobileGL/ --include=*.cpp --include=*.inc --include=*.h: 15 sites / 4 files → 13 sites / 3 files. Renderer/WireDraw.inc leaves the list entirely (both of its marked refusals are retired). scripts/ci/fatal_census.py is unchanged at 79 abort sites over 20 files, 43 distinct family words, 3 refusal words, 0 unmarked — the funnels were already converted in wave 0, so retiring a MagmaWireFatal call moves no std::abort() count. (Note for anyone editing near this row: the census counts the marker literally, so a COMMENT that spells it raises the number. The first cut of this slice pushed 15 → 16 that way.)

New scenario. Scenarios/MagmaVertexLayoutScenario.cpp, DirectVulkan only — Espryt hands GL_FIXED to a driver that supports it, so the same case under the same name would assert a different fact there. DataType::Fixed32 has no arm in ToVkVertexFormat at all, which is what makes it the one unmappable shape reachable on host lavapipe; CONTRACT-P7 §2.5 records that no other P7-marked refusal is (they need real hardware), so without this shape the retirement would have had no host gate.

  • AnUnmappableVertexFormatDeclinesTheDrawInsteadOfEndingTheSession — the program READS the GL_FIXED array, so the draw declines and the framebuffer keeps the clear colour.
  • AnUnreadUnmappableArrayDoesNotStopTheRestOfTheDraw — the array is enabled and unmappable but unread, so the draw lands. This is what stops the first case being satisfied by "decline every draw that mentions GL_FIXED".

Registered on DirectVulkan.{Split,Spawn,Tcp}.VertexLayout. AND picked up by the ambient monolith DirectVulkan. discovery, so the A/B is direct: the same two assertions on the wire arm and on the monolith arm.

Proof (green). 8/8 — monolith DirectVulkan. ×2, Split ×2, Spawn ×2, Tcp ×2.

RED-ONCE (executed). format-map put back on the protocol verdict (return reject("format-map")):

70% tests passed, 3 tests failed out of 10
#4238 DirectVulkan.Split.VertexLayout...Unmappable...  Subprocess aborted***Exception
  MGPipe: Fatal{UnmigratedVerb, "Magma:vertex-layout-buffer-window"}
#4240 DirectVulkan.Spawn.VertexLayout...Unmappable...  Subprocess aborted
#4242 DirectVulkan.Tcp.VertexLayout...Unmappable...    Subprocess aborted
#3408 DirectVulkan.MagmaVertexLayoutScenario...Unmappable...   Passed   <- monolith control

The three split arms KILL THE PROCESS and the monolith arm stays green, which is exactly the divergence this slice closes. Restored: 10/10.

Lane arithmetic. The three new Split entries carry integration-magma-split, so that gating lane grows 73 → 75 (and -spawn 52 → 54, -tcp 54 → 56). Growth only; G14's name sets are grow-only by rule.