Drafted but not yet compile-tested or runtime-validated. Draft
target: vb2 grows an opt-in dma_resv release-fence API; hantro and
rockchip-rga opt in as the demonstration drivers.
Series structure:
- 0000-cover-letter.patch — context, motivation, validation results
- 0001-media-videobuf2-add-dma_resv-release-fence-helper.patch
Adds vb2_buffer_attach_release_fence() that drivers call from
their buf_queue callback. Stores the fence on vb->release_fence;
vb2_buffer_done signals + puts. Per-queue fence context allocated
at vb2_core_queue_init.
- 0002-media-hantro-attach-dma_resv-release-fence-at-buf_queue.patch
Single call in hantro_buf_queue. ~5 lines.
- 0003-media-rockchip-rga-attach-dma_resv-release-fence-at-buf_queue.patch
Same shape in rga_buf_queue. ~5 lines.
Pre-flight before sending to linux-media (per kernel/README.md):
1. Compile the touched files against the kernel tree the patches
will land on (linux-next master as of 2026-04-28 was the source
of truth used for context-line generation).
2. Boot-test on ohm, smoke-test hantro + rga buffer flows.
3. Validate the fence semantics: install patched kernel, uninstall
kwin-fourier so KWin's watchDmaBuf is active, play 1080p30 H.264
under KDE Plasma — should plays through without the bypass
because the fence is now real.
4. Capture before/after dma_buf_export_sync_file timings.
5. Send via git format-patch --cover-letter to linux-media@,
CC dri-devel@ and the relevant maintainers.
This series is the kernel-correct fix for the architectural hole
that the chromium-fourier campaign's kwin-fourier package is
papering over. With this kernel side upstream, kwin-fourier
becomes either redundant (if KWin's existing wait works correctly)
or rewritten as a poll-fd-direct optimization.
Three patches landed on ohm in sequence today: qt6-base-fourier
(GL_ALPHA → GL_R8 on ES 3.x, 3 sites in qtbase), kwin-fourier
(watchDmaBuf no-op test fixture), chromium-fourier patch 4/4
(V4L2 capture pool floor at 16). Each unsticks one layer.
Together they produce smooth 1080p30 H.264 playback under KDE
Plasma 6.6.4 Wayland on RK3566 PineTab2 mainline, where stock
chromium previously stalled in 3 seconds. Combined chrome CPU
~81% steady, KWin ~9%, zero GL_INVALID_VALUE during playback.
Brave's YouTube on the same compositor session also feels snappier
independently — the kwin-fourier watchDmaBuf bypass is a
general-purpose latency reduction for every wp_linux_dmabuf client
on Mali-class hardware, not chrome-specific.
The kernel-side architectural hole is the right upstream-correct
fix: vb2 / hantro / rga don't populate dma_resv exclusive fences
for V4L2 producers, so dma_buf_export_sync_file substitutes a stub
fence representing nothing real. Per-driver fence wiring (3 small
commits: vb2 helper API, hantro opt-in, rga opt-in) plus a parallel
KWin commit using poll(POLLIN) directly on the dmabuf fd is the
planned MR pair. kwin-fourier as it stands is a working diagnostic,
not the upstream-bound shape.
Hypothesis under test: KWin's Transaction::watchDmaBuf calls
DMA_BUF_IOCTL_EXPORT_SYNC_FILE on every plane of every imported
dmabuf and parks the transaction on a QSocketNotifier(POLLIN)
waiting for that sync_file. On V4L2 hantro CAPTURE buffers (RK3566
mainline 6.19, panfrost mesa 26.0.5) the resulting fence either
never signals or signals so late that chrome's 6-buffer V4L2
capture pool exhausts at ~6s, hard-stalling the decoder. mpv with
gpu-next slideshows at 76% drop. weston A/B with same chrome v4
binary plays through clean — KWin's watchDmaBuf is the suspect.
This experiment patches watchDmaBuf to no-op. Wayland clients are
required by spec to ensure buffer contents are complete before
wl_surface.attach+commit, so the fence-wait is a defensive
optimization for misbehaving clients, not a correctness primitive.
If chrome plays through end-to-end at the recorded 34.7% combined
CPU number with this patched KWin, the bug is confirmed and the
upstream fix can be refined (timeout, V4L2-source skip, or use the
dmabuf fd directly in the QSocketNotifier instead of an extra
exported sync_file).
KWIN_PIVOT.md (in chromium-fourier/) carries the discovery thread.
Today's deltas:
- qt6-base-fourier built, installed, validated on ohm. Static-idle
journal shows zero GL_INVALID_VALUE post-relogin; the Qt 6
GL_ALPHA bug is genuinely fixed.
- chrome v4 under KWin still stalls — at ~6s vs ~3s pre-Qt-fix, so
the GL_ALPHA churn was contributing some load but wasn't the
primary cause.
- Clean A/B with weston: same chrome v4 binary, same panfrost,
same V4L2, same hardware → swapping KWin → weston turns the
stall off entirely. Chrome plays through with elevated CPU
(~96 % vs KWin's ~50 % when it isn't stalled) because weston
falls back to LINEAR composite vs KWin's fast-tile path.
- mpv triangulation:
--vo=null --hwdec=v4l2request: clean (decode only)
--vo=drm --hwdec=v4l2request: 0.7 % drops in 19 s (KMS scanout)
--vo=gpu-next --hwdec=v4l2request under KWin: 76 % drops, slideshow
Decode + display hardware path is fully capable. The wall is
specifically KWin's compositor scheduling/presentation pipeline on
this stack — panfrost ES 3.2 + V4L2 stateless NV12 dmabuf clients.
KWIN_PIVOT.md rewritten:
- Phase 4 (qt6 patch, ship, upstream) marked done.
- New Phase 5 (find the KWin culprit): WAYLAND_DEBUG on chrome +
KWin to capture the missing wl_buffer.release / wp_presentation
exchange around the 6 s stall, plus strace-on-kwin and
effects-disable bisection.
- New Phase 6 (fix and ship): kwin-fourier package pattern, ohm
validation, bugs.kde.org filing.
Three small runtime checks in qtbase 6.11.0:
1. src/opengl/qopengltextureglyphcache.cpp createTextureData() and
load_glyph_image_to_texture() — the `#else` branch (active when
qtbase is built with QT_CONFIG(opengles2), every aarch64 distro)
hard-codes GL_ALPHA without checking the runtime context's ES
version. Replace with `useR8 = ctx->format().majorVersion() >= 3`
so ES 3+ contexts get GL_R8 + GL_RED.
2. src/gui/rhi/qrhigles2.cpp toGlTextureFormat() RED_OR_ALPHA8 case
— gated only on caps.coreProfile. Extend with
`caps.gles && caps.ctxMajor >= 3`.
3. src/opengl/qopengltextureuploader.cpp Format_Alpha8 and
Format_Grayscale8 cases — short-circuit on `context->isOpenGLES()`
before reaching the swizzle fallback. Restrict to ES 2 only so
ES 3+ falls through to the existing TextureSwizzle path.
Discovered while validating chromium-fourier patch 3/3 (NV12
EXTERNAL_OES) on ohm. The chrome stall, VLC vout init failure and
mpv "could not initialize video chain" all share this Qt 6 root
cause: every Qt application running on Mali / panfrost / panthor
under a KWin Wayland session emits glTexImage2D(GL_ALPHA) on every
text-glyph cache upload, mesa returns GL_INVALID_VALUE, the
compositor's frame-callback path stalls, and every wayland video
client deadlocks behind it.
PKGBUILD inherits from extra/qt6-base 6.11.0-2 with arch+=aarch64,
epoch=1 to dominate Arch's pkgrel until upstream lands the fix.
qt6-base-cflags.patch and qt6-base-nostrip.patch carried verbatim
from upstream packaging.
KWIN_PIVOT.md (in chromium-fourier/) carries the full diagnosis
thread.
Source-grep collapsed Phase 1+2 onto a single pass. KWin's own GL paths
use GL_R8 correctly (gltexture.cpp:61, shadowitem.cpp:494). The
glTexImage2D(GL_ALPHA) calls observed in the journal originate from
Qt 6:
- qtbase/src/opengl/qopengltextureglyphcache.cpp:111-117 — text glyph
cache upload path. The #else branch (active when qtbase is built
with QT_CONFIG(opengles2)) unconditionally uses GL_ALPHA, with no
runtime check for ES context major version. Correct on ES 2.x;
broken on ES 3.x where GL_ALPHA is no longer a valid glTexImage2D
internalFormat.
- qtbase/src/gui/rhi/qrhigles2.cpp:1373-1378 — Qt-Quick-RHI sibling.
Same logic, gated only on caps.coreProfile, missing the ES≥3 case.
- qtbase/src/opengl/qopengltextureuploader.cpp:253-257 — QImage→GL
upload path; same shape.
KWin runs an ES 3.2 context on Mali-G52 panfrost (RK3566), Qt picks
GL_ALPHA, mesa returns GL_INVALID_VALUE, every dependent draw errors
at level 0, the compositor's frame-callback path stalls. KWin is the
visible victim because it's the compositor, but the bug is in Qt.
KWIN_PIVOT.md rewritten: the patch series and packaging now target
qt6-base-fourier instead of kwin-fourier. Three small hunks (~3 lines
each), runtime-safe via existing caps.gles + caps.ctxMajor / surface
format majorVersion checks. Upstream landing path: bugreports.qt.io
+ Gerrit change against qtbase dev branch.
Earlier framing was wrong — the wall isn't 'Arch ARM clang 22 vs Arch
x86_64 clang 23'. Arch x86_64 is also on 22.1.3; LLVM 23 isn't anywhere
in extra/staging. The flags chromium 147 emits come from chromium's
clang fork (Google maintains an LLVM fork with chromium-specific
passes), not upstream LLVM 23. PKGBUILD'ing clang 23 is the wrong tree.
Right tree: cross-compile from x86_64 so chromium's bundled clang
prebuilt is reachable. CIPD has full linux-amd64 prebuilts, gclient
sync works cleanly, no qemu-x86_64-static dance needed.
his provisioned CT 220 chromium-builder-x86 on data (Ryzen 7 1700,
14 cores, 32 GiB RAM, 200 GiB ZFS). data is normally asleep — woke
via /opt/herding/bin/wake-data. Reach pattern: hertz -> ssh data ->
pct exec 220.
Source fetch running as chromium-fetch.service transient unit on
CT 220. Once src is in, plan: tools/clang/scripts/update.py for
chromium's bundled clang + arm64 sysroot, gn gen with target_cpu=arm64,
build, transfer aarch64 binary to ohm/fresnel/ampere.
boltzmann chromium-builder LXD container preserved as fallback; can
be torn down if cross-compile pans out.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First-build summary on chromium-builder@boltzmann. gn gen succeeds
with our V4L2VDA-unlock args. ninja fails immediately on:
1. chromium 147 emits clang flags (-fno-lifetime-dse,
-fsanitize-ignore-for-ubsan-feature=array-bounds) that clang 22
doesn't know. Arch Linux ARM is on clang 22; clang 23 hasn't
landed in extra yet.
2. Bundled x86_64 esbuild is invoked via qemu-x86_64-static but
/lib64/ld-linux-x86-64.so.2 isn't installed — same shape as the
bundled node-linux-x64 issue we already fixed by symlinking to
system node. Smaller wall.
Documents 5 paths forward (grind patches / pin chromium 132 (7Ji's
known-good) / pin 138-141 middle ground / use chromium's bundled
clang / wait for Arch ARM clang 23) with estimated effort and trade-
offs. Recommends pinning to a chromium version that compiles clean
against clang 22 as the fastest path to a working browser, then
bumping as Arch ARM bumps clang.
Build host state preserved — container running, source extracted,
gn-gen'd, no compile artifacts. Easy to resume from any of the five
paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Initial chromium-fourier shape on the chromium-builder@boltzmann LXD
container we provisioned today. Approach is the 7Ji-style "tarball +
system tools" pattern (no gclient/CIPD, the linux-arm64 dependency
binaries don't exist anyway) but stripped of the MPP/X11/panfork
specifics — chromium-fourier targets mainline kernel + Wayland +
panfrost/panthor + V4L2 stateless on /dev/video0, not the vendor
stack 7Ji's chromium-mpp targets.
PKGBUILD highlights:
- pkgver=147.0.7727.116 (current Chrome stable as of 2026-04-25)
- gn args: use_v4l2_codec=true, use_v4lplugin=true, use_linux_v4l2_only=true,
use_vaapi=true. The first three are the magic that unlocks V4L2VDA on
Linux non-ChromeOS without source patches; if they're sufficient on
their own, the chromeos-pipeline-bypass patch stays a no-op.
- ffmpeg_branding="Chrome" + proprietary_codecs=true for H.264.
- enable_widevine=false, enable_nacl=false to keep the tree small.
- Currently development-shaped: prepare()/build() operate on a
pre-extracted /build/chromium/src rather than makepkg-fetched
source. Will switch to canonical source=(...tarball.xz) shape once
the patches stabilise.
patches/chromeos-pipeline-bypass.patch is a placeholder; the actual
patch (if any) gets developed once we see what 7Ji's gn args do or
don't unlock for us.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stage the next side-project: patch upstream Chromium to do HW video
decode through VaapiVideoDecoder + our marfrit/libva-v4l2-request-fourier
backend on a mainline Linux Wayland system, instead of going through
the chromeos pipeline that fails on Brave today.
STUDY.md captures:
- The exact failure stack we're fixing (PickDecoderOutputFormat ->
ImageProcessor init failure in media/gpu/chromeos/video_decoder_pipeline.cc)
- Three candidate patches (chromeos pipeline bypass, V4L2VideoDecoder
factory un-gate, libva backend default)
- Reference forks (JeffyCN, igel-oss, 7Ji-PKGBUILDs/chromium-mpp,
amazingfate/chromium-debian-build) — all use the older V4L2VDA path
with vendor MPP, not VAAPI; useful for PKGBUILD shape and
factory-un-gating patterns but not directly applicable
- Build plan on fermi (depot_tools, ~30 GB fetch, 6-10 h initial build,
distcc-avahi acceleration through CT108 + tesla)
- Phase order — workspace done now, build env next session,
patches after that, package after that, brave-fourier rebase last
No PKGBUILD added yet; one will land when there's something to actually
package. Build artifacts intentionally not in repo (chromium tree is
~100 GB).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First carrier package for the Fourier umbrella. Tracks Kwiboo's
v4l2-request-n8.1 branch (ffmpeg 8.1 base) pinned at commit b57fbbe —
bump _commit to rebase onto newer tip. CI job chained after
claude-his-any so it serializes on the shared aarch64 repo-db;
continue-on-error: true so long ffmpeg builds don't block the debian
downstream.
Deliberately diverges from the AUR package:
- AUR pins 6.1.1 with epoch=2 → would downgrade stock Arch ffmpeg 2:8.1-3.
We track 8.1 so install is a sideways swap, not a regression.
- AUR pulls X11/AMF/CUDA/FireWire/AviSynth/OpenMPT/Bluray/OpenMAX/JPEG-XL/
Theora/XVid/rsvg/soxr/ssh/vidstab/modplug/SDL2/Vulkan/JACK/GSM/Speex —
dropped here; none are needed on a Wayland ARM video-decode fleet.
- AUR uses #branch=…, sha256sums=(SKIP) → every build is tip-of-branch,
not reproducible. We pin via #commit=<sha>.
Kept: encoders (libx264/libx265/libvpx/libdav1d), VAAPI, libdrm, libv4l2,
neon, OpenGL, PulseAudio, subtitle/font stack, gnutls TLS. SDL2 dropped
means no ffplay binary (mpv covers interactive playback).
provides=(ffmpeg) conflicts=(ffmpeg) so it replaces stock ffmpeg on the
target host deliberately. Primary consumers: ohm (Fourier step 5),
fresnel, ampere.
pkgrel 16 builds ship an upstream distcc 3.4 bug: src/compile.c sizes
the rewritten-compiler-name buffer with strlen(argv[0] + 1) — pointer
arithmetic applied before strlen — under-allocating by 2 bytes. glibc
FORTIFY_SOURCE=2 catches the resulting overflow in strcat and aborts,
so every "distcc gcc ..." invocation dies on a modern Arch.
Patch moves the +1 outside the strlen, as intended.
Closesmarfrit/marfrit-packages#3.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Observed two flakes in 24h — distcc-avahi-aarch64 at 66a7050 and lmcp-any
at 22060e0. Journal for both showed the job starting then failing silently
before any sudo/makepkg line. Theory: transient pacman mirror sync
failures on \"pacman -Syu --needed\" in the bootstrap steps, with the
rsync-to-nc step as secondary exposure (IPv6 prefix rotations from the
Fritz reconnect cron are visible in the runner journal).
Wrap pacman -Syu (5 sites), rsync to nc and hertz (5 sites), and ssh
publish-deb (2 sites) in a tiny 3-try exponential-backoff retry helper
defined per-step. No workflow restructure, no job-graph changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
distcc-avahi build failure was breaking the entire chain (lmcp-* and
claude-his-* depend transitively on it). These jobs only share a runner,
not real build outputs — one failing package should not block the
others.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Tag v0.1.3 added the scripts + quickref, v0.1.4 fixed doc paths.
Package both /usr/bin/repo-inventory.sh and /usr/bin/repo-inventory-nosudo.sh.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Claude Code Edit semantics — literal string replace with uniqueness
check. replace_all flag for bulk edits. Fixes missing edit tool on
lmcp-backed boltzmann / tesla / broglie.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- arch/claude-his-agent/PKGBUILD: fetches v0.1.0 tarball from
git.reauktion.de/marfrit/claude-his-agent, installs agent+skill+helper
to /usr/share/claude-agents/, /usr/share/claude-skills/his/, /usr/bin/
- debian/claude-his-agent/: control+changelog+copyright + build-deb.sh
mirroring the lmcp-debian pattern (dpkg-deb, reproducible mtimes)
- .gitea/workflows/build.yml: two new serialized jobs (claude-his-any +
claude-his-debian) after lmcp-debian; same publish flow as lmcp.
makepkg output isn't byte-deterministic across runs; same pkgver-pkgrel
rebuilds produce different .pkg.tar.xz bytes. repo-add no-ops on
same-version entries, leaving the db's CSIZE/SHA256 pointing at the
previous build while the nc copy is the new one — alpm then refuses
download with 'Maximum file size exceeded'. Force-remove first.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
reprepro refuses to re-include a .deb if the bytes differ from one
already in pool/. Fixed mtimes via SOURCE_DATE_EPOCH make subsequent
builds byte-identical, so CI re-runs of the same pkgver/pkgrel are
accepted as no-ops instead of failing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- debian/lmcp/build-deb.sh fetches the v0.3.0 tarball, lays out the
filetree, and uses dpkg-deb to assemble lmcp_0.3.0-1_all.deb directly
on the Arch aarch64 runner (no debhelper needed for a pure-Lua pkg).
- workflow job 'lmcp-debian' rsyncs the .deb to hertz's marfritrepo
incoming dir, then ssh-triggers 'publish-deb <suite>' for both
bookworm and trixie. publish-deb wraps 'reprepro includedeb' and
rsyncs dists/+pool/ to nc.
- New secret MARFRIT_REPO_HERTZ_KEY uploaded to Gitea repo. Forced
command on hertz routes rsync uploads vs publish-deb triggers.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
lmcp is arch=any (pure Lua). One build on the aarch64 runner serves
all pacman targets — the pure-Lua package drops into both repo dbs.
Depends on lua + lua-socket from the target distro's base repos.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes install conflict with filesystem pkg on merged-usr systems where
/usr/sbin is a symlink to /usr/bin.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- distcc-3.4 tarball hash + local support files hashed (no more SKIP)
- .gitea/workflows/build.yml now builds, signs, repo-adds and rsyncs
to nc via the marfrit-repo-deploy key. Triggers on push to arch/**
or manual workflow_dispatch.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
marfrit
test0r