Fable review of the B3 rework, should-fix: run_trace_case.cmake read the
server log only `if(EXISTS ...)`. Under spawn the server process writes its
own file, and the lines the runner counts out of it - the applier Fatal{
lines and the MGWIRE-FLOOR unsound-serial-complete line, which only the
server's VulkanRenderer::OnSubmitsCompletedUpTo emits - are nowhere else,
so a server that died before its first line read as "0 lines" and went
green. The client log was already fail-closed, and so was the tcp arm's
forwarded server log; plain spawn was not.
The plain-spawn branch of the transport block now FATALs when
mobilegl.server.log does not exist, beside the tcp arm's identical check.
Demonstration: the SPAWN ctest command line with TRACE_REPLAY_EXE replaced
by a wrapper that runs the real replay and then deletes
output/mobilegl.server.log (541 lines):
runner at aab01275 -> rc 0, "Fatal{ lines across ... : 0",
"MGWIRE-FLOOR unsound-serial-complete lines: 0" (green, half unseen)
this runner -> rc 1:
CMake Error at run_trace_case.cmake:321 (message):
OpenRA DirectVulkan: MOBILEGL_TRANSPORT=spawn but the run wrote no
.../output/mobilegl.server.log, so the server role's Fatal{ and
MGWIRE-FLOOR lines cannot be counted. A spawn retrace with no
server log cannot be counted as a clean one.
MobileGLTraceReplay.OpenRA.DirectVulkan.SPLIT / .SPAWN unmodified: both
pass, ssim 1.000000, 0 px, 0 MGWIRE-FLOOR lines.
Review ID-P7-34, item 2. The primary B3 fix - the clamp in
VulkanRenderer::OnSubmitsCompletedUpTo that raises the completed-frame-
serial floor only to a serial no in-flight submission still carries - had
nothing behind it: deleting the clamp loop left every lane green and the
OpenRA picture gold on lavapipe (the host's barriers hide the tear). Its
only observable was the per-event MGLOG_W "MGWIRE-FLOOR
unsound-serial-complete".
run_trace_case.cmake now reads that line out of both roles' logs on every
split run, exactly as it already reads Fatal{: the count is printed pass
or fail and any nonzero count is a FATAL_ERROR. The R-16 counter (26 -> 0
on the OpenRA retrace) is thereby a gate.
Red/green on MobileGLTraceReplay.OpenRA.DirectVulkan.SPLIT / .SPAWN:
clamp loop deleted -> both FAIL at ssim 1.000000 / 0 px:
MGPipe split: OpenRA DirectVulkan transport=inproc, MGWIRE-FLOOR
unsound-serial-complete lines: 26
[mgl-srv-apply/WARN]: MGWIRE-FLOOR unsound-serial-complete #1: advancing
the completed frame-serial floor to 3, which submission 3 (serial 3)
still carries
(spawn: identical, 26)
clamp restored -> both PASS, 0 lines.
run_cts.py silently tolerates a missing --skip-file, so the known device-killers from
the skill (clip_distance.functional, the multisampled_to_singlesampled blit family,
the rgb565 repeat-mode case) would each have cost a reboot in the gate-5 window.
KHR-GL46 has no uniform_buffer_object group; the 90 UBO cases are
GTF-GL46.gtf31.GL3Tests.uniform_buffer_object*, present only when the deployed glcts
was built with GTF - the device probe decides. KHR-GL46.shader_storage_buffer_object
(124) is the block that exercises what cluster A actually refuses today: unaligned
and over-range SSBO bindings, not only UBO byte tails.
Cut from VK-GL-CTS 0b04c470ed78569b0b778ece1262d88455ab2b00 (main; the deployed glcts build's source) mustpass
gl46-main.txt and gl33-main.txt: texture_*, shader_image*, packed_pixels,
direct_state_access, uniform_buffer*. The gl33 variants are the fallback the runbook
names for an Android runner that cannot open a 4.6 core context; whichever version the
probe selects, BASE and AFTER read the same list. A 0.5 pp gate needs a committed
denominator, not a glob over someone's checkout.
wave 0 item 5. Three instruments P7's device window needs, and the CI steps that
can make each of them red.
run_tcp_matrix.py --backend DirectVulkan end to end. --backend already existed and
already filtered by each case's ci_backends; what did not exist is a per-backend
GOLDEN. The two backends shared `golden` plus an optional `alternate_golden`, and
alternate_golden is an OR across backends, not a per-backend key - so a
DirectVulkan run that matched the DirectGLES picture passed, which is the "went
green on the wrong arm" shape the exit gate exists to catch. trace_cases.json
gains `backend_overrides.<backend>` ({golden, alternate_golden, ssim_threshold}
only: restating target_call or a crop would stop the two backends retracing the
same frame through the same window). A per-backend golden REPLACES the shared
pair, matching another backend's declared image is refused BY NAME, and a --case
named explicitly that does not run a requested backend is an error instead of one
plan fewer and no message. The CMake emitter and the APK matrix refuse a case that
declares the key rather than silently emitting one golden for both arms.
DirectGLES is unchanged and that is asserted, not asserted-in-prose: a case with no
override resolves to itself, and the key is dropped from the resolved case even for
the backend that overrides nothing - so giving DirectVulkan its own golden does not
move the DirectGLES identity fingerprint and does not invalidate its checkpoints.
Each results.json row now records the golden, alternate golden and threshold it was
scored against.
run_android_retrace_local.py gets --matrix (the CI split subset, 39 cases, each
running only its own ci_backends - not --all, which adds the rd12 workload ID-P7-4
excludes), --repeat N, and --archive-dir. The archive is the point: the result root
is keyed by case and backend and by NOTHING else, so repeat 2 overwrites repeat 1
and the only artefact that survives a three-repeat run is the last one, which makes
"were the three repeats the same picture?" unanswerable after the fact - in the one
place where re-running to find out is expensive. It also refuses --benchmark with a
non-monolith --transport, which has always produced a monolith run under a spawn
label (run_benchmark_case passes args.env alone; p6/README.md records the pitfall).
compare_actuals.py computes the pair result.json cannot: SSIM between two actual
PNGs, plus a case x repeat x ssim-vs-golden x ssim-vs-first-repeat table over an
archive. It is the retrace gate's OWN SSIM - a transcription of ComputeChannelSsim
/ ComputeRgbSsim (trace_replay_core.cpp:630/:671), the only SSIM in the tree - and
the crop comes from each repeat's own result.json, so the two columns describe the
same rectangle. PNG decode is stdlib zlib, because an unavailable Pillow in the
device window would turn "the repeats diverged" into "the tool did not run".
TraceReplayActivity: resolveSpawnServerPath's two decisions were String.contains()
on the whole joined K=V;K=V list and neither was entry-aware. The one that bit:
`--env MOBILEGL_IPC_SERVER_PATH` with no '=' is the documented UNSET spelling
(trace_env_overrides_test.cpp:92) and the only way to ask for "spawn with no image"
and watch LaunchServer refuse by name - it did not match "MOBILEGL_IPC_SERVER_PATH="
so a resolved path was appended AFTER it and, because ApplyEnvOverrides applies in
order with setenv(...,1), the deliberate unset was silently undone. That is exactly
the "silently replacing it would make that knob untestable" the method's own comment
forbids. Two more: the key inside another entry's VALUE flipped both decisions the
wrong way, and the transport match was case-sensitive while ConfigLoader lower-cases
the value before matching (ConfigLoader.cpp:322). The parse moves to
SpawnServerPath, a class with no android.* import, so test_android_lifecycle.py
compiles and runs it with plain javac - the parse had no host-side coverage at all,
and its on-device failure mode is "the knob had no effect".
Only spawn gets a path, deliberately: it is the only transport that execs an image.
Red-once, all four: a stub case whose result names the DirectGLES golden fails a
--backend DirectVulkan run with "matched a golden another backend declares"; a
synthetic PNG pair with a closed-form SSIM (identical = 1.0 exactly, constant
black vs white = C1/(255^2+C1), inverted halves = (C2-2v)/(2v+C2)); the pre-fix
resolveSpawnServerPath body fails eight of SpawnServerPathTest's checks; and the
DirectVulkan TCP proof battery is the DirectGLES one under the other prefix,
because an empty prefix selection raises nothing and reads as a clean pass.
CI: test_run_tcp_matrix.py and test_compare_actuals.py join the R-16 step. The
matrix driver's stub regressions had NO CI step - they were run by hand and
recorded in a note - and they guard the tool that decides whether a device golden
run passed.
Counts in WSL: test_run_tcp_matrix 12 -> 18, tcp_lane_tools 4 -> 6,
test_compare_actuals 12 new, test_android_lifecycle green with the new Java unit.
Gate 8's last two numbers (SEG_STAGE bytes/frame, records/frame post-
chunking) had no producer at all: the summary line's byte classes are
backend-to-driver bytes and maxrec is a max, not a per-frame. Two push-
guarded counters answer them at their single choke points - ByteClass::
StageSegmentBytes at PipeWireCodec's StageAllocate (the only allocation
behind StageBytes, so every buffer walk, texture slab and CSO archive
lands there and emitters cannot double-count) and CallClass::WireRecords
on EncodeRecord's committed path (post-chunking an emitter no longer
knows how many records its one call produced). A staged-blob histogram
rides along. G1 holds: symbols 0/0/0/0, .text byte-identical, pull build
carries no MG_Remote symbols.
The spawn server never counted anything: PipeStats::Init() is a step of
MobileGL::Initialize(), which ServerMain never reaches, so every Enabled()
site in that process - including ServerLoop's own ServerWaits/ServerParks
publish - stayed false and the server log carried no summary line at all
(a zero that means 'another process' printed in the shape of one that
means 'never waited' is the milder misreading; the line was simply
absent). ServerMain now arms and closes PipeStats at its role boundary -
after MG_ConfigLoader::Init (Init latches from the parsed config), after
the server-role latch (the log sink's role), and before ::_exit(0).
Red-once: 0 server summary lines before, 29 after, SSIM 1.0 unchanged.
MOBILEGL_PIPE_STATS_FILE now derives per role through the log sink's
exported RoleLogPath rule - run.client.json / run.server.json, the bare
name resolves to no file - because both roles reach FormatJson() and ofstream
trunc made the second writer replace the first. G1 caught three separate
leaks on the way (a string temporary in the monolith body, one resized
Shutdown() from unequal #if arms, two rewritten MGLOG_W literals in
.rodata); the pull arm is statement-for-statement the original.
Device tooling for the paired A/B lands with it: pin_device.sh gains a
measured 2f7cbe2e entry (this board clamps the GPU at 1050 MHz -
thermal_pwrlevel writes 0 and reads back 1, max_gpuclk is read-only;
absolute times are not comparable with the 1100 MHz-era rows), and
tools/device_bench/p6/ carries the session collector and renderer beside
run_android_retrace_local.py's MOBILEGL_TRACE_PACKAGE override.
Unit 2367/2367 (PipeStats 25/25), spawn integration 102/102.
The spawn work had the two roles append to one MOBILEGL_LOG_FILE_PATH under
O_APPEND. It was correct - two processes at one offset would corrupt the file,
O_APPEND made each write atomic - but it put two sessions' lines in one place
and made every per-side assertion a search. This splits the sink: the client
role writes <base>.client.log and the server role writes <base>.server.log.
THE BASE NAME NAMES NO FILE ANY MORE, AND THAT IS THE POINT. Leaving the
client at the old unsuffixed path would have been the compatible choice and the
wrong one: a reader that was not updated would go on finding a file, go on
passing, and go on seeing only half the session - and a refusal census that
reports a confident zero for the half it cannot see is worse than one that
fails. With both names moved, an un-updated reader gets "file not found" and
says so. There were 34 such readers; every one is moved here.
THE ROLE IS PER-THREAD, not per-process, because under inproc both roles are
threads of one process. Log.cpp keys the sink on a thread-local role: the apply
thread declares itself server (ServerLoop.cpp), a spawned image is server for
its whole life (MOBILEGL_IPC_ROLE=server, read once), and everything else is
the client. The naming rule (RoleLogPath) and the two whole-log operations
(ReadRoleLogs / TruncateRoleLogs) are EXPORTED from the library, so no reader
reimplements the derivation - a second copy is the copy that falls behind.
Readers, by how they read:
- unit mains set MOBILEGL_LOG_FILE_PATH and read a role file: they keep the
base in the env and derive the client path from the library helper.
- the death-test child runners (ReadLog) now concatenate BOTH roles: a death
test asserts the CHILD said something, and the applier's refusals - the very
diagnostics these check - are written under the server role.
- the offset-based Emit readers derive the client path once after setenv.
- PipeStatsWindow reads both, because draws/vbs increment in the backend's
PrepareForDraw - the apply thread, the server role - so the summary line is
in the server file even under inproc.
- run_trace_case.cmake's split census reads both files; the client keeps the
transport markers.
- split_log_paths.py (and its self-test, and the control stub) derive both
role files from the declared base.
- the Android script pulls both role files off the device and concatenates
them, so every downstream check is unchanged - and it judges a role file by
CONTENT, because `run-as test -f` through `exec-out` returned 0 for a missing
file on the device and `exec-out` folds cat's own not-found line into stdout.
The monolith arm legitimately has no server file, and that is a silent skip
rather than a one-line server log.
Guarded so a pull build is untouched: RoleLogPath and the role sink are wholly
inside MOBILEGL_BUILD_DISAGGREGATED; a non-disaggregated build writes the base
path exactly as before. G1 re-measured - 0 symbol diff, .text unchanged at
0xa52203.
Verified on device 2f7cbe2e: monolith writes only the client log (no phantom
server log), inproc and spawn each write both and pass their transport proofs
(spawn: server role in pid 16125). Unit 2353 (one pre-existing environmental
failure), integration-split 186/186, integration-spawn 102/102, retrace 6/6,
control smoke 12/12, split-log-paths self-test 9/9.
The Android retrace matrix gains a third lane, spawn-acceptance, beside
monolith-control and inproc-acceptance. Until now nothing in CI had ever
started a second MobileGL process on a device at all.
The apk.yml retrace job already carried a `transport` matrix dimension and
already branched on it, so the lane itself is three edits: a row in
github_transport_matrix, an elif in the retrace step, and a summary to render.
What needed thought was the gate.
THE SPAWN LANE IS APK-ONLY, and that is a statement about where evidence is
missing rather than about what spawn can do. Linux already has
test.yml's `retrace-split`, which is {backend, case, transport} over inproc AND
spawn with the two negative controls that make it mean anything. A spawn row in
the Linux transport matrix would pay for a second full sweep proving nothing -
and test.yml's `retrace` job stages no server image, so the lane would not
merely be redundant, it would be red. self_test_transport_matrices now asserts
the absence in both directions, because an accidental removal of the `if apk`
has to be caught in the self-test rather than in CI.
--require-spawn REQUIRES TWO SENTENCES, and the second is the one that matters.
ConfigLoader's marker proves the VALUE RESOLVED - it exists only in the Spawn
arm, which exists only under MOBILEGL_BUILD_DISAGGREGATED - but that is a
statement about a parser. `spawn ARMED - the server role runs in pid N` is
written only after the launch, the connect and the handshake have all
succeeded, and the pid in it is the SERVER's. It is the only line in the log a
same-process run cannot produce.
Both halves red once, measured:
* a monolith run demanded to prove spawn: REFUSED, transport_resolution_marker
false, server_pid null - while the retrace itself scored ssim=1.000000,
which is exactly the trap. OpenRA matches its golden under monolith too, so
the picture has never been the evidence.
* the same run's real spawn log with the one ARMED line deleted: REFUSED,
transport_resolution_marker TRUE and server_pid null. That is the
"resolved the transport and then never left the process" case, isolated.
THE PROOF USES `any` WHERE THE inproc GATE USES `all`, and the difference is
load-bearing rather than a relaxation. Under spawn two processes append to one
MOBILEGL_LOG_FILE_PATH, and the launcher deliberately scrubs every
MOBILEGL_IPC_* from the child's environment (anti-recursion catch (b)) - so the
server's own `Config: IPC` line reads strict=0 role-split-state=0 BY
CONSTRUCTION. Demanding every line match would red every spawn run for the one
thing the design requires.
--use-pbuffer is not optional on this lane: until P12 a spawned server cannot be
handed a window, because an ANativeWindow* is a pointer into the CLIENT's
process (Rule H). MOBILEGL_IPC_SERVER_PATH is deliberately NOT passed -
nativeLibraryDir carries an install-time hash that nothing off the device can
spell, and the Activity resolves it.
The local runner now passes the same knobs and the same --require-* flag CI
does, so "it works locally" can no longer mean "it ran monolith locally". That
change immediately caught its own omission: the first local run with the gate
armed failed on client_ipc_line_present=false because the runner had been
sending MOBILEGL_TRANSPORT alone.
Verified on device 2f7cbe2e: OpenRA.DirectGLES spawn, ssim=1.000000,
mismatchPixels=0, "real spawn transport, server role in pid 9885, and zero
Fatal records confirmed".
OpenRA.DirectGLES, transport=spawn, on a Redmi: ssim=1.000000,
mismatchPixels=0, zero Fatal{ lines, three runs for three. The server is a
process the client launched out of the APK's lib/arm64/, and the client's own
log names its pid.
Four things stood between the desktop green and the device green, and every
one of them was invisible on desktop for the same reason: a build tree is not
a deployment, and Linux is not Android.
1. THE SERVER COULD NOT LINK. It links libMobileGL.so and is EXEC'd, not
dlopen'd, so it gets a fresh linker namespace with the DEFAULT search path
- and an APK's lib/<abi>/ is not on it. `CANNOT LINK EXECUTABLE ...:
library "libMobileGL.so" not found: needed by main executable`, before a
single line of MobileGL ran. On desktop CMake had been giving the target an
absolute build-tree RPATH, so the WSL server resolved its dependency
through a path no phone has. $ORIGIN is asked for via BUILD_RPATH (a raw
-Wl,-rpath,$ORIGIN reaches the generator unquoted and the SHELL eats it -
the device linker said `normalize_path - invalid input: "$"`), and because
the NDK drops RPATH from the link anyway, the LAUNCHER states
LD_LIBRARY_PATH: it is the one component that has already resolved where
the image lives, and a pair that is not co-located is not a pair.
2. THE RENDEZVOUS NEEDED A WRITABLE DIRECTORY. An Android app has no /tmp, so
bind() in the child had nowhere to put the node and the client refused by
name 20 s later. AF_UNIX endpoints now accept the ABSTRACT namespace with a
leading '@' - the systemd/D-Bus convention - which has no filesystem
presence at all: no directory, no mode bits, and no stale node to unlink
after a crash, which is the bind() failure Listen() has been working
around. Its reach is the network namespace, which for an app is the app.
StartSpawned mints '@' names; an explicit path still works everywhere, and
that is what P12's "connect to something somebody else is listening on"
will use.
3. PBUFFER WAS A DirectGLES-ONLY KNOB. Twice over: trace-replay-ci.sh gated
`--ez use_pbuffer true` on the backend name, and the JNI demanded a native
window for DirectVulkan unconditionally. A surface SHAPE is not a backend
property, and the coupling mattered because the Android GLWS makes a window
surface whenever gNativeWindow is non-null - so handing Magma a window was
the same as asking for one. Until P12 a spawned server cannot be given a
window at all: an ANativeWindow* is a pointer into the CLIENT's process and
SetWindowHandle is refused on the way across with Fatal{UnmigratedSurface,
"AndroidNativeWindow@P12"} (Rule H). Which is exactly what happened, one op
after the server had come up green.
4. THE RUNNER LOOKED IN ONE OF THE TWO OUTPUT DIRECTORIES. assembleTraceRelease
writes apk/trace/release; CI has staged the same artifact under
apk/trace/debug. A tree that has built locally has both, and picking the
wrong one retraces last month's APK and says nothing about it. Newest of
either, now.
MEASURED AND NOT FIXED: DirectVulkan on this device diverges under BOTH split
transports and does so nondeterministically - inproc scored ssim=0.988998 /
9836 px, spawn scored 0.976494 / 14658 px and then 0.988998 / 9836 px on a
repeat, while monolith with the same pbuffer surface scores 1.000000 / 0 px.
The two transports draw from the same set of outcomes, so this is the split
path on Magma against a real GPU, not the process boundary: P6 neither caused
it nor can fix it. DirectVulkan spawn IS green on desktop (lavapipe, 1.000000).
G14: one test name added, AnAbstractEndpointNeedsNoWritableDirectoryAndLeaves-
NoFile, which pins both halves - the handshake crosses an abstract name, and
nothing is left on the filesystem. A '@' mistaken for a path would have bound
a file of that literal name, passed here, and failed on the one platform this
exists for.
Two small pieces, so an Android spawn arm is a flag rather than a project.
THE RUNNER GAINS --transport. It already had a generic env passthrough - one
intent extra carries `K=V;K=V` all the way to setenv before the library is
dlopened - so the transport itself costs one entry and no new plumbing in the
five files between here and the device. monolith passes NOTHING rather than
MOBILEGL_TRANSPORT=monolith, so the default arm's environment stays
byte-identical to what it was before this option existed: a control that
changed when the option landed would not be one.
THE SERVER PATH CANNOT COME FROM THE HOST. Every other knob can, but this one
names a file that must be EXEC-ABLE, and on Android there is exactly one such
place: an APK's lib/<abi>/. W^X has forbidden the app's own data dirs since
API 29, and the packager only puts a file in lib/<abi>/ if it is named
lib*.so - which is why the server executable wears that name, and what P0
spike A proved on two devices. nativeLibraryDir contains an install-time
hash, so no host-side script can spell the path; the process about to launch
the server is the only one that can. TraceReplayActivity resolves it, and
only when the caller actually asked for spawn.
AN EXPLICIT MOBILEGL_IPC_SERVER_PATH STILL WINS. A caller that set one is
pointing somewhere deliberately - at a pushed build, say - and silently
replacing it would make that knob untestable on the one platform where it
matters most.
The APK must be built with -Pmobilegl.buildDisaggregated=ON or none of this
is reachable: without it ConfigLoader has no transport parser, accepts the
value, ignores it (CONTRACT-P5 rule 5), and the run is monolith under a name
that says spawn. The --transport help says so where somebody will read it.
It was one case on one backend under one transport: OpenRA, DirectGLES,
inproc. It is now every CI case, both backends, inproc AND spawn - 154 CI
entries, 156 local ctest entries - because every defect the spawn work turned
up was a question about the SERVER's process and not about any one trace or
either backend.
THE DirectVulkan ARM. tools/trace_replay/CMakeLists.txt carried "a
DirectVulkan arm would be measuring a server nobody has written yet". True
when P5 wrote it; false since P5f/P7 landed Magma's server consumers and its
run-ahead readiness gate. Leaving it would have shipped a split shape whose
second backend nothing ever ran - the same class of hole as a lane that runs
monolith under a name that says split. Magma now brings up a server process
of its own and OpenRA retraces across it at ssim=1.000000, which is only
possible because this branch also taught the child to load its config: before
that, a DirectVulkan session brought up a DirectGLES server and the two
halves disagreed about which backend the capability mask describes.
`split` BECOMES AN OPT-OUT. It began as an opt-in naming one case, because
one case was the work. Marking 39 cases `"split": true` would carry no
information and would reinstate the hole the key was written to close, only
mirrored: a case added later would get the monolith arms and be silently
absent from the split ones, and absent is the one state nothing reds on. A CI
case is now in the subset unless it says `"split": false`, which is a
reviewable claim that a trace cannot cross a role boundary.
THE CI MATRIX GAINS A THIRD DIMENSION. The job took {backend, case} and
hard-coded MOBILEGL_TRANSPORT=inproc in its own run block, so `spawn` had no
way into CI at all; the .SPAWN ctest entries existed only for a local build.
It is {backend, case, transport} now, the backends are the case's own
ci_backends rather than a hard-coded DirectGLES, and the artifact names carry
the transport or the two arms of one {backend, case} would overwrite each
other's uploads.
inproc STAYS AS THE CONTROL. Same library, same codec, same applier, same
picture; the only difference is which process the server role runs in.
Running only spawn would leave "did the transport break it?" and "did the
phase break it?" answerable only together - which is the question every
failure in this phase has turned on, including the SIGSEGV that D1 explained.
BOTH NEGATIVE CONTROLS FOLLOW THE ARM. retrace_pull_library_control.sh and
retrace_drop_draw_control.sh each exported MOBILEGL_TRANSPORT=inproc
themselves, so on a spawn job they would have gone on proving something about
the other arm while spawn ran unguarded. They read SPLIT_TRANSPORT now, and
default to inproc so a caller that sets nothing behaves exactly as before.
scripts/ci/control_smoke_test.sh: 22 passed, 0 failed against the stubbed
ctest, including the two modes that matter most - a red whose SSIM never fell
and a knob that armed and dropped nothing.
The spawn arm needs a second binary and the inproc arm needs none, so the job
places libMobileGLServer.so only on the spawn arm and a missing one is a
named failure there rather than a LaunchServer refusal 78 jobs deep.
Measured locally as this lands: 39 of the 78 SPAWN entries complete, all
green. The rest are still running.
OpenRA.DirectGLES.SPAWN is green: ssim=1.000000, mismatchPixels=0, zero
Fatal{ lines, against a server in a process the client started by name. The
inproc SPLIT arm stays beside it as the control, so "did the transport break
it" and "did the phase break it" remain separately answerable.
Five defects stood between the handshake and that picture. Each one was
silent in its own way, and none of them looked like what it was.
1. THE DOORBELL COULD NOT WAKE ITSELF. SocketDoorbell's one-fd form sends on
the end it polls, so a write is delivered to the PEER - correct for waking
the other process, useless for waking a thread in this one. The server's
control pump posts into the apply thread's mailbox and must ring the bell
that thread parks on. Symptom: every eglInitialize crossed, landed in the
mailbox, and timed out at 5 s. The two-fd form (park on [0], notify on [1])
is restored here, and the bell handover goes from two descriptors to THREE
- the client needs the same split for the same reason. Seven offers now.
2. InitCapabilities HAD NO WIRE KIND. CONTRACT-P6 §2 predicted this exactly
("the request has a live client caller and the codec refuses to encode
it"). SurfaceOpKind gains 11, append-only; the decode range's upper bound
moves with it, or the new tag reads as out-of-range and surfaces as
Fatal{ProtocolCorruption} rather than the missing row it is. The op's name
loses its InprocOnly suffix, which stopped being true here.
3. THE CHILD LOADED NO CONFIG AT ALL. mobilegl_server_main calls
InitServerRoleForSpawn directly, and MG_ConfigLoader::Init is a step of
MobileGL::Initialize - so every MG_Config value in the server was its
static default. MOBILEGL_BACKEND_TYPE survives the launcher's scrub on
purpose and was read by nobody: a DirectVulkan session would have brought
up a DirectGLES server and then disagreed with its client about which
backend the capability mask describes.
4. AND THEREFORE Transport READ AS Monolith, which CONTRACT-P6 §3.1 (D1)
forbids in as many words. a6 measured the cost: 226 live lines test
`Transport != Monolith` to select the SERVER arm, 148 in MG_Backend.
Monolith flips all of them to frontend glue in the one process with no
frontend. Observed precisely as predicted - UnitTexturesByHandle() answered
false, Clear took the legacy arm that dereferences borrowed FRONTEND slot
pointers, and the apply thread took a SIGSEGV inside SyncNeccessaryTextures
on the first Clear record of the trace. It is pinned in-process rather than
through the environment because the environment is where it must NOT be:
LooksLikeAClientEnvironment exits 65 on a surviving MOBILEGL_TRANSPORT, and
that check is anti-recursion catch (b). Dial, not Transport, is that axis.
5. THE LOG WAS CORRUPT, NOT MERELY INTERLEAVED. Two processes opened the same
MOBILEGL_LOG_FILE_PATH with "w" and each carried its OWN file offset, so
they overwrote each other's bytes - half-lines like `ote client: CapsMirror
generation 1 adopted`, and the server's whole startup block vanishing under
the client's later writes. This cost real diagnosis time twice, because a
missing line read as a code path not taken. O_APPEND now; the client still
truncates once so a fresh run is not a pile of old ones, and the server
never does - it would delete the very startup lines a failed session needs.
run_trace_case.cmake's Fatal{ census reads exactly this file, so it was
also reading a lie.
The retrace runner's split assertions become transport-aware: one row per
transport it admits, and a transport with no row is a FATAL rather than a
fall-through, because the fall-through is the accident the block exists to
prevent. The spawn row carries a SECOND sentence and it is the stronger half
- ConfigLoader's marker proves the value RESOLVED, while `spawn ARMED - the
server role runs in pid N` is written only after the launch, the connect and
the handshake all succeeded, and carries a pid that is not this process's.
inproc resolves its own marker while running the server role on a thread
here; nothing on that path can write this one.
G14: one test name added (TheEleventhFramedOpIsInitCapabilitiesAndItCrosses-
TheWire), none removed. The ten-op table keeps its name and its scope rather
than being renamed to eleven.
Two unit failures predate this change and are environmental in this tree:
PipeCatalogue.FrontendNeverTakesAnApplierAddress (the frontend source root
resolves empty) and SessionHandshakeTest.TheAbiFingerprintChangesWhenAnyOf-
ItsInputsDoes (GIT_COMMIT_HASH_SHORT is empty here, which a6 already recorded
as red in this configuration).
Stage updates to adopted GPU-resident buffers on the command timeline instead of overwriting memory that earlier draws still read. Wait for submitted copies as well as pending commands before CPU readback.
Commit coherent mapped writes before Android trace buffer copies, and retire buffer shadows on deletion so copies see current vertices and name reuse cannot leave dangling dirty-shadow entries.
Validated draw ordering, submitted-copy readback, traced buffer copies, and mapped-buffer deletion/name reuse on Redmi Adreno 830. Minecraft 26.3-rc-3 Magma passed movement, HUD, and menu checks and exited normally; the new trace parsed all 64,410 frames without warnings.
- bench.sh's pin_freqs writes /proc/ppm and /proc/gpufreq, which exist on neither 35d0befa
(SM8750) nor 3B159D009VZ00000 (MT6993 dropped both for /proc/gpufreqv2); the guard the
PROFILE_VERIFIED key exists for was about to be defeated on the device its own comment
guessed was safe
- pin_device.sh pins big/little/GPU through the nodes each device actually has (Xiaomi:
policy6 1958400 / policy0 1555200 / kgsl pwrlevel 0; Oppo: policy4 2000000 + policy7
2000000 / policy0 1600000 / gpufreqv2 fix_target_opp_index 0), checks only against its own
pins because ColorOS moves stock maxima by itself, and refuses to call a zero-key read a verdict
- both profiles are PROFILE_VERIFIED=1 on 30 s pinned windows with zero drift; the evidence and
the two things not verified (Oppo GPU held at DVFS-request level only, no full bench window)
are in docs/Disaggregated/devices/pin-verification-2026-09-07.md
- bench.sh and session.sh are invoked as ./tools/device_bench/bench.sh in their own README
and are the harness for the paired two-device A/B the GO/NO-GO asks for; a previous commit
in this package dropped the bit and every check invoked them through bash, so nothing saw it
- The verified-profile guard read the process environment as well as the profile: the test ran
after the source, so PROFILE_VERIFIED=1 exported in an operator's shell re-opened the fail-open
hole for every profile that says nothing. Both scripts now set PROFILE_VERIFIED=0 immediately
before sourcing, so the file is the only thing that can answer.
- A --device path that cannot be sourced was diagnosed as an unverified profile, because both
scripts cd to their own directory first and neither checked readability. The path is now also
tried relative to the directory the script was invoked from (which is what a repo-root-relative
--device means), and an unreadable one is reported as unreadable, naming both places tried.
- Verified: exported PROFILE_VERIFIED=1 + an unverified profile -> rc 2; exported 1 + a profile
with no key -> rc 2; a repo-root-relative path -> resolved, then refused for its own reason;
a missing file -> "cannot read the device profile"; odinlite.env -> past the guard;
--allow-unverified-profile -> the three warnings, then proceeds.
- test.yml's new step described a mechanism the tree does not have. G6's and G10's entries are
registered in the pull build too - they must be, for G2's name-for-name comparison - and skip
inside their bodies. The step's value is unchanged and its comment now says the true thing: the
`test` job runs those names as a column of skips, and this is the first CI job that unpacks a
build which compiled the assertions.
- trace_benchmark takes the wall baseline before the CPU baseline, the order OnFrameBoundary
already reads them in, so frame 0 stops reporting a CPU delta biased upward against its own
wall delta; and it includes <time.h> rather than <ctime> for the POSIX names it uses.
- format_benchmark printed a p50 taken with the nearest-rank rule beside a medianFrameCpuMs the
device computes as the average of the two middle frames, and documented the two as one rule; on
an even window they differ (the pre-flight printed p50=8.261ms next to medianCpuMs=271.766).
p50 now goes through series_median, which is SummarizeSeries' rule transcribed; p95 and p99 stay
nearest rank, which is the device's rule for p95 and the honest extension of it for the p99 the
device does not compute at all
- require_verified_profile treated a profile that simply omits PROFILE_VERIFIED as verified, which
is the fail-open default a profile written by copying another one inherits - exactly the case the
guard exists for. It defaults to unverified now, odinlite.env carries PROFILE_VERIFIED=1
explicitly (it is the one profile that earned it), and the refusal says "says 0, or says nothing"
- the two new profiles claimed profile.sh refuses an unverified profile; it has no such check and
needs none - it records a simpleperf profile and pins nothing. The claim is corrected in both
profiles and in the README rather than a guard added where there is nothing to guard
- the handle-ABA / CSO control step in test.yml set only MOBILEGL_ITEST_REQUIRE_GPU while its
sibling verify step sets the three MOBILEGL_MAGMA_* fixes and arms core dumps. It runs the same
DirectVulkan binary on the same runner, so a crash there left no core; it now carries both
- DriverBenchStateToggle runs mc_state_toggle as its own ctest entry. The case has been in
kBenchCases since P0 and nothing executed it, so nothing would have noticed it rotting - and it
is the exact enable/draw/disable/draw shape the microbenchmark P2 owes the GO/NO-GO measures.
About 1.2 s inside an existing three-minute job.
- scripts/g7_negative_control.sh breaks the pipeline/dynamic split on purpose: it inserts two
boundaries so ColorMasks becomes a dynamic chunk of its own, which keeps the partition sorted,
non-overlapping and complete - so it still COMPILES - while making glColorMask bump
m_pipelineStateVersion without moving the pipeline-subset hash. A non-zero ctest is the pass.
- Everything that could make that control lie is refused rather than reported: a missing
SetterConsistency test exits 2 instead of reading "no tests matched" as a failure; a tree that
is already red or already broken exits 2; a patched table that does not compile exits 2, since
a build break would prove the static_asserts work rather than that the test still checks; and
the restore is from byte-for-byte copies (never from git, so a dirty tree is given back
intact), followed by a rebuild and a re-run that must be green. --verify-patch-only exercises
the mechanism where the test does not exist yet and says explicitly that it is not a pass.
- Profiles for the two campaign devices, and the guard that stops them being trusted early. Both
carry PROFILE_VERIFIED=0 and every device-specific field is TODO_VERIFY_ON_DEVICE rather than a
guess: the harness pins through MediaTek nodes and 35d0befa is a Qualcomm part, where
`su -c 'echo ... > /proc/ppm/...'` fails with a zero exit and the run would report numbers it
believes were pinned. bench.sh and session.sh now refuse an unverified profile unless
--allow-unverified-profile is passed, which warns that the run is not comparable with a pinned
one. The README records what earns PROFILE_VERIFIED=1.
- The metric the disaggregation GO/NO-GO hangs on is per-thread CPU p50/p99, and the tree had no
first-party collector for it: no CLOCK_THREAD_CPUTIME_ID, no getrusage, no RUSAGE_THREAD, no
/proc/self/task anywhere under MobileGL/, tools/, android-plugin/ or scripts/.
- The retrace loop is the cheapest honest place to take it. Retrace runs --singlethread and
trace_benchmark states that Begin/OnFrameBoundary/End are only ever reached from that one
thread, so that thread's CPU time IS the client-side CPU cost: one extra clock_gettime per
frame, no root, no profiler, no debuggable build, no sampling - and no instrumentation
committed to a hot path inside the library.
- The CPU reading is taken before the wall reading at each boundary, so the syscall lands in the
wall delta rather than hiding inside the CPU delta. An inflated wall number is visible; a
deflated CPU number would not be.
- Report::frameCpuMs is the same length as frameMs or it is EMPTY. A clock that started failing
mid-run would otherwise be silently re-indexed and put frame N next to frame N+k, and an empty
series and a series of zeroes are different claims about the platform.
- SummarizeBenchmark is split into SummarizeSeries and reused verbatim for the CPU series rather
than duplicated: same tail window, same even-count median rule, same nearest-rank p95, or the
delta between the two series stops meaning anything.
- benchmark.json gains meanFrameCpuMs / medianFrameCpuMs / p95FrameCpuMs and the WHOLE
frameCpuTimesMs[] array beside frameTimesMs[]; result.json and the completion line gain the
three headline numbers. p99 therefore needs no device change - it is a host-side reduction over
an artefact that already exists, and run_android_retrace_local.py prints p50/p95/p99 off the
same trailing window the device summarised.
- Pre-flighted on the desktop CLI, which shares the same core: 2-frame run, cpu series aligned
with the wall series and strictly below it (wall 1290.787/20.225 ms, cpu 535.271/8.261 ms).
- the block looked for Fatal{PipeVerifyDiffer and Fatal{UnmigratedPipeInput only. A
misspelt MOBILEGL_PIPE_VERIFY_CORRUPT / MOBILEGL_PIPE_POISON_OMIT reports
Fatal{PipeVerifyBadKnob, and it was caught only because D2 makes that one abort the
process - which is precisely what MOBILEGL_PIPE_VERIFY_FATAL=0, the supported triage
configuration, takes away. A typo'd knob would have left a negative-control run
looking healthy
- the FATAL_ERROR text now says what each of the three means and where the two
vocabularies live, because that message is the whole diagnosis a `cmake -P` step gets
- A retrace that exports MOBILEGL_PIPE_VERIFY=1 at a library configured without
-DMOBILEGL_PIPE_VERIFY=ON is a no-op: the frames still match their goldens and the
case reports green having compared nothing. run_trace_case.cmake now demands the
evidence whenever the variable is set to anything but 0/false - mobilegl.log must
exist, must carry "MGPipe: verify armed", and must carry neither
Fatal{PipeVerifyDiffer nor Fatal{UnmigratedPipeInput. With the variable unset the
script is byte-for-byte the old one.
- The Fatal scan is not redundant with the replay's exit status:
MOBILEGL_PIPE_VERIFY_FATAL=0 is the supported triage configuration, and there a
divergence is logged and counted rather than aborted, so the run would finish 0
with its own report sitting unread in the log.
- LABELS retrace on both registrations: the lane was selectable only by regex, so
`ctest -L retrace --no-tests=error` - the spelling that reds a lane which
registered nothing - could not be written at all.
- "verify": true on eight cases (the five 180s cases plus minecraft-1.21.4-in-world,
minecraft-1.21.4-fabric-sodium-in-world and improved-transparency-minecraft-26.3),
and --format github-verify-matrix over that subset: 16 entries, against the full
lane's 77. The verify build compares at every verb boundary and again at every
accessor read, which the design budgets at 5-10x, so the per-push job runs the
subset and the full sweep is a phase-exit / workflow_dispatch run.
- The flag is validated in the manifest loader, not at the matrix, so a non-boolean
or a verify case excluded from CI is a loud error in every consumer instead of a
subset that is quietly one case short.
- brings the write-map landing into GPU-resident stores, the open-ended fp64 storage block flattening, the idle-based CTS chunk timeout and the MSVC /WHOLEARCHIVE test link
- verified on the merged tree: unit 1458, integration 868, Wire 54, retrace 79/79
- plan-B §8.3 asks which tier `AcquirePersistentMap` lands in, but the probe only
asked Vulkan. DirectGLES ("Espryt") reaches a persistent map through
`glBufferStorageEXT` + `glMapBufferRange(PERSISTENT|COHERENT)`, not a
`VkDeviceMemory` map, so a Vulkan-only answer decides nothing for that backend.
Add a GLES leg to T1: the exported fd imported with `glCreateMemoryObjectsEXT`
+ `glImportMemoryFdEXT` + `glBufferStorageMemEXT`, then mapped
PERSISTENT|COHERENT -- in-process first (isolates "GL can import this fd" from
"the fd survives a process boundary"), then cross-process (new `t1gl` child).
Drivers disagree about how the import must be phrased, so each attempt walks a
ladder over {dedicated flag} x {import size = memory requirement or the fd's
own size} x {buffer size} and reports the rung the driver accepted plus every
rejected rung with its GL error -- a driver *preference* must never be reported
as a missing capability. A driver that backs the storage but refuses
PERSISTENT|COHERENT is reported separately from one that refuses the storage:
that distinction is exactly T1 vs T2 for DirectGLES. The T0 GLES leg
(`EGL_ANDROID_get_native_client_buffer` + `glBufferStorageExternalEXT` +
persistent map, verified by `AHardwareBuffer_lock` on the client side) now
reports every step's GL enum and requires the persistent flags for OK.
- the verdict was unfalsifiable: T1 reported PARTIAL when neither leg had moved a
byte. Replace it with an explicit decisive-leg model -- OK only when every
decisive leg round-tripped in both directions, PARTIAL when at least one did,
FAIL otherwise with the failing step and its driver error named in `why:`.
Every row now opens with a per-leg trace (`vkimport[D]=rt gpu[D]=rt`). The raw
`mmap` leg is informational for opaque-fd (Vulkan forbids interpreting that
payload outside the driver, so a refusal is conformant) and decisive for
dma-buf, where a CPU mapping is the point of the handle type.
- T3 never ran the direction that would make it a tier: both ends were the
importing process. Add `T3-client-memfd-server-import` (new `t3c` child) -- the
client creates and writes the memfd, the server mmaps the received fd, imports
the client's host pointer into a `VkDeviceMemory`, reads what the client wrote,
writes back, and takes a GPU access on the client's memory, which the client
then verifies through its own mapping.
- no route touched the GPU, so an OK proved only that a map call returned a
pointer. Every tier row now takes a real GPU access before it can be OK:
`vkCmdCopyBuffer` out of the shared allocation into private staging (mismatch =
the GPU could not read what the peer wrote) plus `vkCmdFillBuffer` into it,
queue-idle and an explicit host-read barrier, with the peer checking the filled
region through its own mapping. VkCtx grows a queue and command pool for it.
- the device run executes in the `shell` SELinux domain, not the `untrusted_app`
domain MobileGL runs in, and the two do not share dmabuf/gralloc rules. Print
uid/pid/`/proc/self/attr/current` in a run-context header, repeat the caveat in
the summary, and document in README.md how to answer it for the real domain
later (exec the same binary from the trace app's spike hook, spike-A package)
without implementing that here.
- `vkStr()` returned a pointer into one static buffer while several results
routinely appear in one format call, so all of them showed the last one; it
returns std::string now, `memFlagStr` likewise, and `fmt`/`pr` carry
`format(printf)` so a missed `.c_str()` is a compile error rather than UB.
- `advertisedExportable` decided the status at the allocate site but not at the
`vkGetMemoryFdKHR` site. One rule at every export failure now
(`exportFailStatus`): advertised EXPORTABLE and then declining is FAIL, never
advertised is UNSUPPORTED. Export + map + fd is factored into `exportHostVisible`.
- `T0-ahb-blob-transfer` was recorded OK on the socket handoff alone. The handoff
keeps its own informational row; the tier row is now composed at the end from
the full import+map+compare+writeback chain over the Vulkan, GL and GPU legs.
- `mmapErrno` kept the first attempt's errno after the second-chance mmap
succeeded, so a working mapping carried a failure code; it is cleared on
success and the first errno moves into the note.
- a failed `glImportMemoryFdEXT` no longer closes the fd: EXT_memory_object_fd
does not say whether ownership still transfers on failure and Mesa closes it
either way, so closing risks a double close landing on the socket. Leaking a
handful of dups in a short-lived probe is the safe side of that trade.
- validated end to end on the host harness (lavapipe + llvmpipe,
`VK_DRIVER_FILES=lvp_icd.json EGL_PLATFORM=surfaceless`): T1-opaque-fd OK,
T3-external-memory-host OK, T3-memfd-cross-process OK,
T3-client-memfd-server-import OK, T1-dma-buf UNSUPPORTED (not advertised
exportable). The two T1-gles rows FAIL there with GL_OUT_OF_MEMORY on every
ladder rung although GL_DEVICE_UUID_EXT matches the Vulkan deviceUUID --
llvmpipe's GL does not implement importing a lavapipe opaque-fd allocation,
a Mesa interop gap recorded in README.md so a device FAIL stays attributable.
Rebuilt for arm64-v8a with NDK r27d (PIE, android-30); the device run is
pending, both device locks are held by another campaign.
- Plan B §11 P0 requires spike B ("external memory 导出，两台设备") to run before the
persistent-map decision of §8.3 can be taken: T0 (server imports a client
allocation), T1 (server exports its own HOST_VISIBLE|HOST_COHERENT allocation)
or T2 (AcquirePersistentMap returns nullptr, making the §5.10 client-side block
push mandatory). §8.3 says the answer must be measured on the two campaign
devices, and that a platform unknown must not block the interface work.
- tools/spikes/extmem_probe/ is a self-contained NDK command-line program: it
links vulkan/EGL/GLESv3/android/log and nothing from MobileGL, is configured by
its own CMakeLists with the android toolchain file, and is deliberately absent
from the project's build graph (the root CMakeLists only pulls in
tools/trace_replay), so the default ALL target is untouched.
- Phase A enumerates VK_KHR_external_memory_fd, VK_EXT_external_memory_dma_buf,
VK_EXT_external_memory_host, VK_ANDROID_external_memory_android_hardware_buffer
and, through a headless EGL pbuffer context, GL_EXT_memory_object{,_fd},
GL_EXT_external_buffer, GL_EXT_buffer_storage, GL_OES_EGL_image_external{,_essl3}
and EGL_ANDROID_get_native_client_buffer, plus the memory-type table and the
vkGetPhysicalDeviceExternalBufferProperties verdict per handle type for the exact
buffer usage set MobileGL needs.
- Route T1 allocates a HOST_VISIBLE|HOST_COHERENT buffer memory with
VkExportMemoryAllocateInfo, writes a pattern through vkMapMemory, exports an fd
with vkGetMemoryFdKHR (opaque-fd and, where advertised, dma-buf), hands it to a
second process over SCM_RIGHTS, and has that process both mmap() the fd and
import it into its own VkDeviceMemory + vkMapMemory. Both sides write and both
sides compare, so a copy-on-import or one-directional mapping is reported as
PARTIAL rather than as success.
- Route T0 has the second process allocate an AHardwareBuffer BLOB
(CPU_READ_OFTEN|CPU_WRITE_OFTEN|GPU_DATA_BUFFER), send it with
AHardwareBuffer_sendHandleToUnixSocket, and the first process import it three
ways -- AHardwareBuffer_lock, VkDeviceMemory via
VK_ANDROID_external_memory_android_hardware_buffer, and a GL buffer via
eglGetNativeClientBufferANDROID + glBufferStorageExternalEXT mapped
persistent/coherent -- with a write-back leg the allocating process verifies.
- Route T3 covers VK_EXT_external_memory_host: a memfd-backed mmap region aligned
to minImportedHostPointerAlignment, imported through
VkImportMemoryHostPointerInfoEXT, plus the same memfd handed to another process.
- The second process is /proc/self/exe re-exec'd with --child=<route> and one end
of a socketpair on fd 3. A bare fork() is not usable: neither side's Vulkan
driver survives fork, and both sides need live Vulkan. It is also the topology
the transport actually has (§8.1, inheriting PLAN.md §11.1-§11.6: the client
spawns the server), so the probe measures the arrangement that would ship.
- The probe also builds for the host with T0 compiled out. That is not scope
creep: a negative device result is only worth something if the harness is known
to report a working route as working. Running it on lavapipe did that, and paid
for itself immediately by exposing two harness bugs that would have produced
false negatives on the devices -- (a) the child wrote through its plain mmap
before reading through the Vulkan import, so on a driver whose exported fd maps
at an offset the probe overwrote the very payload the second read compares
(lavapipe reports payloadAt=4096); reads through both mappings now precede
writes through either, and the offset is searched for and reported; (b) an
export failure on a handle type the driver never advertised as EXPORTABLE was
classified FAIL instead of UNSUPPORTED (lavapipe's dma-buf answer).
- Output is a RESULT/summary table carrying the raw driver verdicts (VkResult
names, errno, GL enums) because those codes -- not a pass/fail bit -- are what
§8.3 needs in order to pick the tier.
- Built with NDK 27.3.13750724 for arm64-v8a / android-30, RelWithDebInfo, PIE,
warning-clean; host build clang/RelWithDebInfo, warning-clean.
- PLAN-B.md §8.2 and appendix B add a batch of new runtime switches
(MOBILEGL_PIPE_PUSH / _VERIFY / _STATS / _LEGACY_MEMOS / _TEXEL_RETAIN_MB /
_INDEX_MIRROR_MB, plus MOBILEGL_IPC_* later), and §11 P0 wants them parsed beside
the existing ones. Today every knob that has to reach an Android replay costs an
edit in five files - run_android_retrace_local.py, trace-replay-ci.sh,
TraceReplayActivity's request record, the JNI marshalling, and the setenv block in
trace_replay_core.cpp. That per-knob tax is what this replaces: one extra,
`--es mobilegl_env "K=V;K=V"`, carries all of them.
- Applied last, immediately before dlopen(libMobileGL.so), so it can also override
the dedicated fields above it - MobileGL's config is read during the load, and an
escape hatch that cannot beat the defaults is not one. An entry with no '=' unsets
the variable, which is the only way to clear a default the marshalling sets.
- The existing per-knob flags stay: they carry semantics beyond a setenv (use_angle
also selects a variant, the dump lists are joined, DirectVulkan forces the
R11G11B10F fallback), and rewriting them as env strings would move that logic into
the callers.
- Surface: --env / MOBILEGL_TRACE_ENV in trace-replay-ci.sh, repeatable --env
KEY=VALUE in run_android_retrace_local.py, `mobilegl_env` intent extra,
Request::envOverrides.
- The two-level parse now lives in trace_env_overrides.hpp, beside the semicolon
splitter it shares with the texture and FBO dump lists, and
tools/trace_replay/trace_env_overrides_test.cpp pins it: the empty entries a
trailing ';' leaves behind must not become unsetenv(""), `K=` must stay a Set of
the empty string rather than an Unset (a knob read with getenv() != nullptr sees
those as opposite answers), and only the FIRST '=' may separate, or a value
carrying '=' is truncated without a word of warning. The whole MOBILEGL_PIPE_*
batch rides on this parse, and the only lane that exercised it end to end was an
on-device retrace, which would have reported a splitting bug as "the knob had no
effect".
- The check is built and RUN at build time and mobilegl_trace_replay depends on it,
so `cmake --build ... --target mobilegl_trace_replay` - the exact command of
test.yml's "Build trace replay" job, which never invokes ctest - runs it. It is
assert-free on purpose: that lane configures Release, and <cassert> under NDEBUG
would compile every check into a green run that checked nothing. Negative control:
swapping find('=') for rfind('=') fails 2 checks, and keeping the splitter's empty
entries fails 2 more.
- PLAN-B.md §11 P0 lists spike A (the Android delivery chain) as a P0 deliverable,
inherited verbatim from PLAN.md §15 P0; §8.1 inherits PLAN.md §11.1-§11.6, whose
Android path needs a second process. Android gives an application no writable
exec-able directory, so the only supported route is to name the binary lib*.so, let
the packager put it in lib/<abi>/, and exec it out of
getApplicationInfo().nativeLibraryDir. This builds that route end to end so the
spike can be answered with evidence instead of folklore.
- New root option MOBILEGL_BUILD_SERVER_SPIKE (OFF, ANDROID-only) adds the
MobileGLServer target from tools/spikes/server_stub/main.cpp with PREFIX "lib" /
SUFFIX ".so" and -fPIE/-pie: an .so name does not exempt the file from Android's
PIE requirement. Its RUNTIME_OUTPUT_DIRECTORY is pointed at
CMAKE_LIBRARY_OUTPUT_DIRECTORY, because AGP packages what lands in the per-ABI
library output directory and CMake would otherwise put an executable elsewhere.
- The option is opt-in on both sides. The plugin flavour cannot turn it on at all,
and the trace flavour builds it only when asked, with
`-Pmobilegl.buildServerSpike=ON` or MOBILEGL_BUILD_SERVER_SPIKE=ON in the
environment; a flavour that silently carries an executable nothing loads is the
kind of thing nobody notices until it ships. Verified both ways:
assembleTraceDebug -Pmobilegl.buildServerSpike=ON packages
lib/arm64-v8a/libMobileGLServer.so and `file` reports "ELF 64-bit LSB pie
executable, ARM aarch64 ... interpreter /system/bin/linker64, for Android 26";
the same task with no property packages only libMobileGL.so and
libtrace_replay_runner.so.
- The stub prints one line to stdout and writes the same line to the file named by
argv[1], then exits 0. The line carries pid/ppid/uid/gid and, decisively, the
child's own /proc/self/attr/current: only `u:r:untrusted_app:...` proves an
ordinary app process did the exec. An `adb run-as` shell runs in a different
SELinux domain, so a success there would prove nothing.
- RunSpawnSpike() starts the stub with argv [serverPath, markerPath], redirects the
child's stdout/stderr into a captured file (an app process has stdout on
/dev/null, so a printed line would otherwise vanish), waits for it, and reports
exit status, signal, the exec errno, the parent's own SELinux context, the marker
content and the captured stdout - to logcat, to the returned string, and to a
<marker>.report file, because the Activity finishes immediately afterwards.
- The child reports the errno of a REFUSED execve through a close-on-exec pipe.
Without it the one datum the spike exists to produce is lost: the parent only ever
sees a wait status, in which every reason has already been flattened into one exit
code, and EACCES (SELinux, or a noexec mount) versus ENOEXEC (a packager that
mangled the file) are opposite verdicts for the design. A successful exec closes
the write end for free, so the parent reads EOF and reports execErrno=0.
- fork/execve only. The earlier draft also carried a posix_spawn arm behind
`__ANDROID_API__ >= 28`, which was dead code in every configuration this repo can
build - bionic declares posix_spawn from API 28 and the root CMakeLists.txt pins
MOBILEGL_ANDROID_API_LEVEL to 26 and refuses to configure lower - and would have
silently become the production path, untested, on a minSdk bump. Keeping the arm
that actually ships means the spike measures the code the server would really use.
Nothing happens between fork and execve except open/dup2/execve/write/_exit, all
async-signal-safe, because the parent is a multi-threaded JVM process.
- The spike lives in its own TU, spawn_spike.cpp/.hpp, listed only by the trace
APK's CMakeLists. Its sibling trace_replay_core.cpp is compiled verbatim by the
DESKTOP mobilegl_trace_replay runner (tools/trace_replay/CMakeLists.txt names the
same file), where <android/log.h> does not exist, so nothing Android-only may live
there; spawn_spike.cpp carries an #error for anyone who adds it to that list.
- The Activity runs the spike, and nothing else, when launched with the
`mobilegl_spike_spawn` intent extra; that mode needs no trace, no golden and no
render surface. It is a separate JNI entry point rather than another parameter on
the 30-argument replay call, which it shares nothing with.
- Not yet run on a device: both device locks are held by another campaign. The
on-device verdict is the coordinator's step.
Initialize the desktop replay Request from the three iterationRP environment flags before loading MobileGL. Without this, the Request defaults caused the replay core to unset CI's exported flags, leaving all repairs disabled despite the workflow configuration.
iterationRP's Program 203 declares shared vec2 prefixSumCache[32] for a
512-invocation workgroup indexed by gl_SubgroupID; any device narrower
than 16 lanes partitions into more than 32 subgroups and the pack writes
shared memory out of bounds (heap corruption on lavapipe's CPU
rasterizer, ssim 0.028 on the CI retrace). Fix it where the fault lies -
in the fixture - and keep the GL contract sound everywhere else:
- FixIterationRPSubgroupScratchPass: fingerprint-gated SPIR-V pass that
grows exactly that array to ceil(invocations/width) entries on sub-16-lane devices; every other module passes through byte-identical.
- DeriveNumSubgroupsPass stays default-on for the Adreno topology bug
and is made spec-sound: pipelines request REQUIRE_FULL_SUBGROUPS
whenever the workgroup shape makes the flag legal (computeFullSubgroups
enabled, local_size_x a multiple of the native width, subgroup count
within maxComputeWorkgroupSubgroups).
- EmulateSubgroupsPass: 32-lane virtual-subgroup lowering kept in-tree
as a last resort, enabled only by MOBILEGL_MAGMA_EMULATE_SUBGROUP=1 on
devices with no native subgroup support; fails closed on extended
subgroup instructions and on modules whose added scratch would exceed
maxComputeSharedMemorySize.
- IterationRPFirstReductionScenario skips gracefully outside the pack's
16..256-lane source domain; the new IterationRPScratchFixScenario runs
the fixture-shaped reduction on any width and asserts the exact
width-independent total. DriverPost keeps reporting FAIL on
out-of-domain devices.
- Program203 -> IterationRP rename throughout; the per-trace
num_subgroups_quirk plumbing is removed from the trace replayer, JNI
chain, and CI workflows.
swung0x48