iter9 of the panvk-bifrost campaign — operator-confirmed Vulkan output
on PineTab2 (Mali-G52 r1 MC1) 2026-05-20.
Patches Mesa 26.0.6's PanVk Vulkan driver:
- VK_KHR/EXT_robustness2 + nullDescriptor exposed for PAN_ARCH 6/7
- has_vk1_1 / has_vk1_2 = true on Bifrost
Patches applied via sed in PKGBUILD prepare() (cleaner than maintaining
a Mesa fork for two two-line tweaks; upstream context drifts between
Mesa releases would make a literal .patch brittle).
Co-installs at /usr/lib/panvk-bifrost/ — stock mesa untouched. Stock
libvulkan_panfrost.so and its ICD JSON keep working for everyone not
opting into the patched driver.
Ships /usr/bin/brave-vulkan that wires up:
VK_ICD_FILENAMES=/usr/lib/panvk-bifrost/icd.json
PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1
MESA_VK_VERSION_OVERRIDE=1.2 # ANGLE needs apiVersion>=1.1; the
# has_vk1_x flags don't move it, so
# the env-var override carries that
brave --use-gl=disabled --enable-features=Vulkan --use-vulkan=native
--ozone-platform=x11 --no-sandbox --disable-gpu-sandbox
--ignore-gpu-blocklist "$@"
Side-steps the stock "GLES3 is unsupported / GPU process exits" failure
documented in panvk-bifrost/README's "Consumer-side benefit" section.
Known limitation: WebGL/WebGL2 in-page won't work — ANGLE needs GLES3
which needs VK_EXT_transform_feedback, which PanVk-Bifrost doesn't
currently support. Browser chrome + standard page rendering work fine.
Gitea Actions job mesa-panvk-bifrost-aarch64 added to build.yml,
patterned on libva-v4l2-request-fourier-aarch64. Mesa build is slow
(~30-60min on actrunner-aarch64). Standalone (no needs:),
continue-on-error so it doesn't block other jobs.
Campaign artifacts: ~/src/panvk-bifrost/{README.md, phase8_iteration9_close.md,
phase0_evidence/iter9_brave_vulkan_breakthrough.txt}.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
481279c adds packaging/systemd/{daedalus-v4l2.service,modules-load} to
the upstream tree. This commit wires those into both the Arch
(PKGBUILD + .install) and Debian (build-deb.sh + postinst/prerm/postrm)
package layouts so that a fresh install of daedalus-v4l2 + daedalus-
v4l2-dkms on a Pi 5 leaves the kernel module loaded at next boot AND
the userspace broker daemon enabled — no manual modprobe / systemctl
enable dance needed.
arch/daedalus-v4l2:
* pkgver 0.1.0.r18.481279c, pkgrel reset to 1 (new upstream pin).
* Dropped 'systemd-libs' from depends — daemon doesn't link
libsystemd (no sd_notify); the .service unit is consumed by
systemd-the-init, no link-time dep required.
* package() now installs the .service to
/usr/lib/systemd/system/daedalus-v4l2.service and the modules-
load drop-in to /usr/lib/modules-load.d/daedalus-v4l2.conf.
* New .install file: post_install/post_upgrade run daemon-reload +
enable + systemd-modules-load + try-restart on upgrade; pre/post
remove tear down cleanly. No auto-start — operator decides.
arch/daedalus-v4l2-dkms:
* pkgver bump to 481279c, pkgrel reset to 1. Kernel module itself
is bit-identical to f0cd29a (commit only touches packaging/) but
bumping in lockstep keeps DKMS source-tree pkgver matched to the
userspace pkgver so /etc/modules-load.d points at a module that
actually exists.
debian/daedalus-v4l2:
* Same bump 481279c, PKGREL=1.
* build-deb.sh stages /lib/systemd/system/ + /usr/lib/modules-load.d/
and installs both files.
* Generates DEBIAN/postinst that runs daemon-reload, enables the
service, triggers systemd-modules-load, and conditionally starts
the service iff /dev/daedalus-v4l2 is already present (uses the
same ConditionPathExists= guard as the unit file so apt install
doesn't fail loudly on a host where dkms hasn't built yet).
* Generates DEBIAN/prerm (stop + disable on remove) and
DEBIAN/postrm (daemon-reload).
debian/daedalus-v4l2-dkms:
* Lockstep version bump, PKGREL=1. Postinst (loud-warn-on-missing-
headers) unchanged.
Verified the SHA via local rev-parse against ~/src/daedalus-v4l2 —
481279c is the "packaging/systemd: ship daedalus-v4l2.service +
modules-load drop-in" commit on main.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
DKMS build for daedalus_v4l2 fails against kernel 6.18+ with:
daedalus_v4l2_main.c:1049: error: too few arguments to function
'v4l2_fh_add'
v4l2-fh.h:97: void v4l2_fh_add(struct v4l2_fh *fh, struct file *filp);
(same for v4l2_fh_del). Signature changed exactly at v6.18 — verified
v6.13–v6.17 still use the one-arg form via raw.githubusercontent.com
tag walk.
Upstream commit f0cd29a wraps the calls with LINUX_VERSION_CODE so the
module keeps building against:
* 6.12 LTS / RPi 6.12.75 (one-arg) — hertz
* 6.12.88+deb13-arm64 (one-arg)
* 6.18.29+rpt-rpi-2712 (file* arg) — higgs running kernel
Higgs (Pi CM5) was hitting this: daedalus-v4l2-dkms 0.1.0+r16+gf55b2cd
showed 'installed' in dpkg but DKMS autoinstall failed for the running
6.18.29 kernel. Re-running 'dkms autoinstall' after this bump succeeds
+ /dev/daedalus-v4l2 appears.
Also widens debian/daedalus-v4l2-dkms Recommends from
linux-headers-generic | linux-headers
to
linux-headers-rpi-2712 | linux-headers-rpi | linux-headers-generic | linux-headers
so apt pulls the right metapackage on Raspberry Pi OS / RPi-2712
kernels by default.
Userspace pkgver bumps in lockstep (no userspace change in f0cd29a, but
keeps daedalus-v4l2 + daedalus-v4l2-dkms versions matching for
LIBVA_DRIVER_NAME selection sanity).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stock firefox.desktop disappears when our 'provides=firefox' replaces stock
firefox-arch, so installing firefox-fourier left the user without a Plasma/GNOME
start-menu entry. Add firefox-fourier.desktop with Categories=Network;WebBrowser;
(routes under Internet on KDE Plasma 6), MIME types for the usual web schemes,
new-window / private-window actions, and the 128px icon from the package's
internal browser/chrome/icons tree.
Reported on ohm (PineTab2 / Plasma 6) — manual /usr/share/applications shim
proved the file works; this commits it to the recipe so future installs Just Work.
Bumps pkgrel from 5 to 6. Cumulative regenerated from the redone
per-series reconstruction in kernel-agent (replacing the broken
PR #33 attempt with a proper rebase onto v7.0-danctnix1 baseline).
cumulative.patch: 162 716 -> 279 554 bytes
b2sum: 50397711a6a3... -> eb179c03f35a...
manifest.lock: 32 resolved patches (was 4 — c5x interim collapsed
the bes2600 driver scope into one cumulative blob;
per-series properly tracks each fix)
bes2600 series-dirs in kernel-agent: now 25 individual series-dirs
(one per cleanups commit + Patch H), each with the
proper in-tree drivers/staging/bes2600/* paths
Built + installed on ohm 2026-05-19 23:39. Functional verification:
uname: 7.0.0-danctnix1-6-pinetab2-danctnix-besser
srcversion: 1A919EED0E6DC2478559B17 (differs from pkgrel=5's
BEB625FA7443171EA8D55F7 — not byte-equivalent;
functional equivalence verified: wlan0 associates,
bes2600_btuart loads, Pattern A 0, no WARN/BUG)
Per-fix revertability now real: removing an include from
fleet/ohm.yaml drops that fix from the cumulative. Bisecting on
kernel-agent side becomes practical.
The recipes pinned `f55b2cdab8a8c0bc04e8c1bb1d0b6ca85e7d96d2` as the
"Phase 8.13: byte-exact end-to-end via libva" commit, but that SHA
does not exist in git.reauktion.de/reauktion/daedalus-v4l2.
The actual `main` tip (per gitea's for-each-ref) is
`f55b2cd002afdfd08f3c093627317f92e4929074` — same 7-char prefix
(`f55b2cd`), different full hash. Likely a manually-constructed SHA
based on a short prefix from a working copy that was never pushed.
git archive --format=tar.gz on the bad SHA fails with
fatal: not a tree object: f55b2cdab8a8...
which surfaces as 500 from Gitea's archive endpoint, which curl in
the CI build-deb.sh sees as `curl: (22) ... error: 500`.
Diagnosed by tailing gitea.log during a fresh archive request from
the CI runner; the underlying `git archive` command in the gitea
container is logged with the full failing SHA + error.
Fixed in all four recipes (arch + debian, daedalus-v4l2 + dkms).
pkgrel bumped to signal new build (PKGVER short-prefix `gf55b2cd`
stays the same — both bad and good SHA share that prefix).
fermi's makepkg.conf already has BUILDENV=(distcc ...) which prepends
/usr/lib/distcc/bin to PATH at build time. But Mozilla's mach
configure picks CC/CXX from the environment directly — it doesn't
treat the distcc shim in PATH as the default C compiler unless the
env vars are set explicitly. Without this guard the distcc workers
sat idle while mach drove a local-only build.
Mirrors the explicit-distcc pattern already used in
ffmpeg-v4l2-request-fourier/PKGBUILD.
Caveat documented in the comment: only the C/C++ portion
distributes. rustc and host-only build steps stay local. Empirical
~30-40% wall-clock improvement on a 4-worker pool (tesla, dcc1, dcc2,
ampere via Avahi zeroconf).
pkgrel bumped so a rebuild publishes a new package even though the
binary output for users is unchanged (build process speedup only).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`package_qt6-xcb-private-headers-fourier()` depended on
`qt6-base-fourier=$pkgver` (= literal "6.11.1"), but the package itself
ships with `epoch=1` so the installed version is `1:6.11.1-1`.
Pacman's strict-equality version compare treats "6.11.1" as epoch=0 and
"1:6.11.1" as epoch=1 — mismatch — and refuses every upgrade involving
qt6-xcb-private-headers-fourier with "unable to satisfy dependency".
Fix: include the epoch prefix in the dep string —
`qt6-base-fourier=$epoch:$pkgver`. pkgrel bumped to 2 so the rebuild
publishes a new package even though only the dep string changed.
Observed on ohm 2026-05-18 after the broader fleet pacman -Syu —
worked around there with --assume-installed; this fixes it properly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The PR branch contained the unpacked Linux 7.0 source tree
(~81k files, ~38M additions) under src/. These are makepkg build-dir
artifacts that should never be committed; PKGBUILD downloads + extracts
them at build time.
Their presence inflated the PR file-count display to 81030 changed
files. Removed src/ from tracking and added a stronger .gitignore
in the package dir to prevent regressions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bumps all 4 daedalus packages (arch + debian × userspace + dkms)
to pick up daedalus-v4l2 f55b2cd: "kernel: media_request_get/put
around inf->req (UAF safety)".
Closes the SHIP-WITH-EYES-OPEN concern Sonnet flagged in the
pre-deployment review — without explicit media_request_get on
capture + media_request_put on completion, a concurrent
MEDIA_IOC_REQUEST_REINIT or process-kill triggering
buf_request_complete from the cancel path could drop vb2's
reference before our completion handler ran, leaving inf->req
dangling through v4l2_ctrl_request_complete + buf_done.
Matches the cedrus / rkvdec refcount pattern. No protocol
change, no API change, no consumer-side adjustment required.
Same byte-exact output verified on hertz post-fix (libva path:
match; standalone test_m2m_stream: 30/30 frames).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sonnet pre-deployment review caught a BLOCKER: on a fresh higgs
(Debian 13 / Pi CM5) install without the RPi kernel headers
pre-installed, the postinst's `dkms autoinstall || true` silently
swallowed the build failure. Package appeared installed but the
.ko was absent; `modprobe daedalus_v4l2` then failed and the
entire stack was dead with no clear pointer to the cause.
Fix in both ecosystems:
debian/daedalus-v4l2-dkms/build-deb.sh:
- After `dkms autoinstall`, verify the post-condition with
`dkms status -m daedalus_v4l2 -v VER -k $(uname -r)`.
- If the module isn't 'installed' / 'loaded' for the running
kernel, emit a yellow-bolded ANSI warning naming the most
likely cause (kernel headers package missing) and the exact
recovery steps (linux-headers-rpi-2712 for RPi or
linux-headers-$KERNELVER for Debian generic, then
`dkms autoinstall` + `modprobe`).
- Colour only on TTY; the warning is unconditional regardless.
arch/daedalus-v4l2-dkms/:
- New daedalus-v4l2-dkms.install with post_install +
post_upgrade hooks that run the same `dkms status` check.
- post_upgrade catches the case where a kernel-headers package
was uninstalled / pruned between upgrades, silently
regressing the build.
- Wired into the PKGBUILD via install="${pkgname}.install".
Both versions point at the actual repair commands rather than
just saying "build failed", so the user is one apt/pacman away
from a working stack instead of debugging dkms internals.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PKGBUILD pinning the same daedalus-v4l2 @ f04d700 as the
userspace sibling. Installs kernel/ source to
/usr/src/daedalus_v4l2-<ver>/ with a generated dkms.conf;
AUTOINSTALL=yes builds the module against the running kernel.
The kernel/ Makefile uses -I$(src)/../include for the shared
protocol header. In the userspace tree that's daedalus-v4l2/include/;
for DKMS we flatten by copying the header into kernel/include/
and patching the Makefile in package() to point there.
Sibling Debian package: debian/daedalus-v4l2-dkms/
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
kernel-agent corrected the @@ hunk counts in the SCS xor-neon patch
(the previous -9,7 +9,12 'fix' was an overcorrection; actual is
-9,6 +9,11). pkgrel=4 build #4 silently tolerated the wrong counts
because the trailer was stripped, but pkgrel=5 with besser#18
behind SCS exposed the inconsistency.
b2sum: ceec602afa8574c74354... -> 50397711a6a3ba522283...
Size unchanged 162 716.
Kernel-agent restored the trailer on the SCS patch (SCS is no longer
last in the includes ordering — besser#18 is). Cumulative grows
+12 bytes for the restored '-- \n2.54.0\n' sentinel.
b2sum: 0eb091ddaba4... -> ceec602afa85...
Size: 162 704 -> 162 716.
pkgrel=5 = pkgrel=4 + the besser#18 lockdep fix added to the
manifest's includes (in-tree mirror of bes2600-dkms d95453c).
Cumulative b2sum: 334c37b5d37067982bd9... -> 0eb091ddaba4a8f1c3c2a78...
Size: 157 446 -> 162 704 (+5 258 = besser#18 patch with trailer
stripped per the convention now established for the LAST patch in
ka-promote concatenation).
This single pkgrel satisfies all three deliverables of the ohm
migration goal: kernel-agent flow + new marfrit-packages pkg + the
besser#18 fix it contains.
After kernel-agent dropped the wrong git-format-patch trailer from
the SCS xor-neon patch (the trailer was a misdiagnosed fix; the
real problem was the @@ hunk counts).
Cumulative b2sum: ad9e2cb533957f218058... -> 334c37b5d37067982bd9...
Size: 157 458 -> 157 446 (12 byte trailer gone).
Re-promoted after kernel-agent fixed malformed @@ hunk counts in the
SCS xor-neon patch (kernel-agent commit ahead of this on
noether/migrate-pinetab2-pkg-and-patches).
b2sum: bd42cd39106298879eeb... -> ad9e2cb533957f218058...
size unchanged at 157 458.
Drops the f8f3ad9 baseline ("18 commits ahead, BLACK-SCREENS ampere"
per fleet/ampere.yaml bisect note) in favor of the kernel-agent-
managed 10-patch set produced by ka-promote from fleet/ampere.yaml.
Baseline: mainline v7.0-rc3 (3daa4f5dc6cc), plus the 10 scope-tagged
patches under marfrit/kernel-agent/patches/{soc,module,board,driver}/:
- 1 soc/rk3588 pwm15 pinctrl
- 6 board/coolpi-cm5-genbook DTS patches
- 3 driver/media Sarma VP9-on-VDPU381 patches (PR #24 closure)
New _commit 48a8c78 reflects this tree state in ~/src/linux-rockchip
(branch vp9-build on ampere, exactly v7.0-rc3 + 10 manifest patches).
End-to-end verified before this iteration was cut (hand-build of the
same tip on 2026-05-18 booted ampere via arch_vp9_test extlinux
label):
- VP9 decode bit-exact HW==SW==libva (sha c8624d7c42db66525f53a02a515bc38d0a17ef39f692660cc7bebb1e2d2e1b48)
- AV1 decode bit-exact HW==SW via kdirect (sha 30d2091158d92f3c5e0a807217c3e7307f873267673d92632e7fb147383e7dd1, av1-vpu-dec is mainline 7.0)
prebuild.sh canonical sha256 cleared — gzip-version-dependent, the
script warns-not-fails on mismatch. First successful kafr2 build can
pin a canonical sha here if a reproducibility audit ever needs it.
Cross-references:
- marfrit/kernel-agent#12 (VP9 enablement closure)
- marfrit/kernel-agent PR #24 (Sarma patch import + ampere.yaml bump)
The n8.1 pin's hwcontext_v4l2request.c deliberately blanks the
transfer-formats list for AV_PIX_FMT_YUV420P10 sw_format (the mapping
target for V4L2_PIX_FMT_NV15), so `ffmpeg -hwaccel v4l2request
-vf hwdownload,format=p010le` on a Hi10P / Main10 input failed at
filter-init with -22 EINVAL — even though kernel-side decode succeeded.
0002-nv15-to-p010-unpack.patch adds an inline NV15→P010 unpack
(5 bytes per 4 samples, little-endian → high-10-of-16) inside
v4l2request_transfer_data_from, exposes AV_PIX_FMT_P010 in
transfer_get_formats for that sw_format, and rejects non-P010
destinations explicitly with ENOSYS instead of silently corrupting
output via av_frame_copy on NV15-packed bytes.
Verified on fresnel (RK3399, linux-fresnel-fourier 7.0-14):
- 5-frame smoke test from issue #21 → exit 0, 13.8MB output
- 20-frame mid-fixture decode → bit-exact HW==SW
sha256 7d9b66d48d8f17b2281da1881c663ecc31722bb218aba1ae23bf28d07aa66b08
- 8-bit baseline (bbb_60s_720p.h264.mp4) still bit-exact HW==SW (no
regression in the existing NV12 path)
- Cross-device repro of original EINVAL on unpatched ampere (RK3588)
pkgrel=4, confirming the bug is upstream-FFmpeg-side, not RK3399-specific
Patch is upstream-able to Kwiboo's v4l2-request-n8.1 branch.
Closes#21.
Tracks upstream lmcp v1.1.1 (commit 9707f7a). Single-bug-fix
release: lmcp:tool() now normalises empty inputSchema.properties
by dropping the key, so Zod-strict MCP clients don't reject
tools/list with "expected: record, received: array".
Discovered live on a hertz-tools deployment where two custom
no-arg tools tripped the check and caused Claude Code to mark
the endpoint as disconnected.
New tarball sha256:
80c2e815aa61a2d3baab051c51cd247bdefa9dd03d72c4867b99c49b6eae9cb9
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PKGBUILD already renamed itself (pkgname=ffmpeg-v4l2-request-fourier,
replaces=(ffmpeg ffmpeg-v4l2-request-git)) but the containing
directory was never moved. This commit completes the rename to align
the path with the package identity and the rest of the -fourier
umbrella (libva, mpv, kwin, qt6-base, chromium, linux-*).
CI workflow path-trigger is wildcard (arch/**), unaffected. Step
names + cp source path updated to the new directory.
Predecessor experimental package (tarball + 18 stacked patches) is
superseded by libva-v4l2-request-fourier (git-tracked fork tip,
iter38 multi-device probe). The successor already declares
replaces=('libva-v4l2-request-ohm-gl-fix') so installs migrate
automatically; this commit just makes the deprecation visible in
README + pkgdesc.
Kept in-tree as historical reference for the ohm_gl_fix Phase 6
audit trail.
CI binary segfaulted on HEVC vaEndPicture even though
/etc/makepkg.conf has OPTIONS=(... !lto ...). Root cause: arch-meson's
wrapper hard-codes `-D b_lto=true` regardless of makepkg.conf, so
the binary still gets cross-function ICF (Identical Code Folding)
under -O2 + LTO.
HEVC is the only codec in the campaign that submits a per-frame chain
of 4 control structs (SPS + PPS + DECODE_PARAMS + SLICE_PARAMS); ICF
finds a near-duplicate per-codec helper across the codec dispatch and
merges them, then the wrong instance's local stack layout is invoked
on the HEVC path → SEGV. The other 4 codecs (H.264, VP8, VP9, MPEG-2)
submit fewer/simpler control structs and tolerate the folding by
accident.
Empirical confirmation from the issue body's binary bisection:
meson build (default debugoptimized) 485 KB HEVC ✓
arch-meson + --buildtype=release 145 KB HEVC ✓
arch-meson + release + -flto 76 KB HEVC SEGV
CI build (this package, 7ac934e-1) 133 KB HEVC SEGV
Fix: append `-Db_lto=false` to the arch-meson invocation.
pkgrel 1 -> 2.
The exec bit was missing on prebuild.sh as committed in PR #15
(it landed mode 0644). Running './prebuild.sh' fails with
'Permission denied'; bash ./prebuild.sh works as a workaround but
the shebang is intended to do its job.
Caught while running the first linux-ampere-fourier build on
boltzmann today. README.md examples show './prebuild.sh' so users
will hit the same error.
Single-bit fix via git update-index --chmod=+x. No content change.
Sibling to kernel-agent PR #8 which lands the scope-tagged board
patches + fleet/ampere.yaml manifest. This PR carries the
makepkg-side recipe.
## Baseline
marfrit/linux-rk3588-marfrit @ f8f3ad9 (mainline v7.0-rc3 + 18 commits):
10 by Markus Fritsche (board DTS + Kconfig + suspend/wakeup + bt)
4 by Shawn Lin (phy: rockchip-snps-pcie3 stability)
2 by Cristian Ciocaltea (clk + dw-dp bridge — Collabora)
1 by Sebastian Reichel (Rock 5 ITX hdmirx — unused by ampere)
1 by Pedro Alves (CLK_SET_RATE_PARENT VOP2 fix)
The 6 board-relevant patches (pwm15 pinctrl + RK806 power-off +
genbook pwm-fan/speaker/USB-C/lid) are scope-tagged in
kernel-agent. The 12 others are 'cherry-picks-in-baseline'
currently — splitting them into scope-tagged patches is a
follow-up.
## Source distribution
The kernel tree isn't pushable to Gitea (boltzmann's working clone
is shallow). Tarball (260MB) doesn't belong in marfrit-packages
either. Pragmatic shortcut: ship prebuild.sh that produces the
tarball locally from $LINUX_RK3588_MARFRIT_TREE (default
~/src/linux-rockchip) just before makepkg runs. PKGBUILD source
array references the tarball by name; makepkg picks it up from
the PKGBUILD dir.
Build flow:
cd arch/linux-ampere-fourier
./prebuild.sh # produces 260MB tarball; idempotent w/ sha check
makepkg -s --noconfirm
Future iteration: host the tarball on Gitea release assets or
packages.reauktion.de/sources/ and switch source to URL.
## What the PR ships
PKGBUILD — the recipe (linux-rk3588-marfrit @ f8f3ad9)
config — snapshot of ampere's /proc/config.gz
linux-ampere-fourier.preset — mkinitcpio preset, /boot/firmware/ paths
extlinux-add.hook — alpm hook, fires on /boot/firmware/Image* changes
extlinux-add.sh — idempotent managed-label injector for
/boot/firmware/extlinux/extlinux.conf
(vfat on mmcblk0p1, ampere-specific path)
prebuild.sh — source-staging helper (see above)
README.md — build + install runbook
## ampere-specific boot path differences vs fresnel
ampere boots from /boot/firmware/ (vfat partition), not /boot/ (root
partition) like fresnel. The PKGBUILD installs Image + initramfs + DTB
under /boot/firmware/ directly. The extlinux-add hook fires on
/boot/firmware/<X> path changes and edits
/boot/firmware/extlinux/extlinux.conf. The mkinitcpio preset writes
initramfs to /boot/firmware/.
Only one PRESET ('default') — no fallback image — because
/boot/firmware is ~1G total and two ~20MB initramfs files plus the
DTB and a couple of kernel-image backups would squeeze it tight.
## Out of scope this PR
- Ask 2 (VP9 enablement) and Ask 3 (AV1 dec) from kernel-agent issue #6
- Build + publish + install (sequenced after this PR + kernel-agent PR #8 merge)
- Splitting the 12 non-board commits into scope-tagged kernel-agent patches
Sibling to kernel-agent PR #7 (which lands the vb2_dma_resv RFC v2
series under patches/subsystem/media/videobuf2/ and updates
fleet/fresnel.yaml to include them). This PR carries the same three
patches in the build-tree-ready form for the linux-fresnel-fourier
PKGBUILD.
Background: Markus iterated v2 locally on boltzmann reaching pkgrel=14
since the 2026-05-09 bootstrap published 7.0-1. v2 attaches the V4L2
producer fence at device_run in slept-OK context per Dufresne's v1
review on linux-media. Drivers opted in here: hantro + rockchip-rga.
Changes:
- pkgrel 1 -> 14
- pkgdesc adds '+ vb2_dma_resv RFC v2'
- source array gains 3 new local patch files (scope-tagged comments
point to where they live in marfrit/kernel-agent)
- sha256sums extended to 11 SKIP entries
No prepare() edit needed — the existing for-loop applies every *.patch
in $srcdir, so the 4/5/6 patches get picked up automatically. config
unchanged (verified diff vs boltzmann's local pkgrel=14 tree).
Pre-built artifacts: linux-fresnel-fourier-7.0-14-aarch64.pkg.tar.zst
and linux-fresnel-fourier-headers-7.0-14-aarch64.pkg.tar.zst already
exist on boltzmann at /home/mfritsche/src/kernel-agent-bootstrap/
build/marfrit-packages/arch/linux-fresnel-fourier/ — will be signed
and published via marfrit-publish-arch after this PR merges (no fresh
compile needed; the artifacts were built locally during the iteration
loop that reached pkgrel=14).
Empirical follow-up to #9. After packaging 150.0.1-4 and installing on
fresnel, MOZ_LOG showed the failure pattern was still present:
D/FFmpegVideo FFMPEG: Using preferred software codec h264
No VAAPI_VLD engagement, no dmabuf surface locking — Gecko bailed
before reaching the patched VAAPI path. Same symptom as issue #8
pre-fix, despite the prefs file being on disk at the expected path.
Root cause: /usr/lib/firefox-fourier/browser/defaults/preferences/ is
NOT a vendor-prefs scan location in Firefox 150. Mozilla's canonical
scan dir is /usr/lib/<app>/defaults/preferences/ — without the
'browser/' prefix.
Verified by hand-copying the same file to /usr/lib/firefox-fourier/
defaults/preferences/ and re-running the H.264 playback test:
D/FFmpegVideo FFMPEG: Requesting pixel format VAAPI_VLD
D/Dmabuf VideoFrameSurface: VAAPI locking dmabuf surface UID 26267
FFMPEG ID 0x4000000 mAVHWFrameContext ...
D/Dmabuf VideoFrameSurface: VAAPI locking dmabuf surface UID 26268
FFMPEG ID 0x4000001 ...
(15+ surface locks)
End-to-end zero-copy DMABUF path engaged, hantro/rkvdec dekodes the
H.264 stream via libva-v4l2-request-fourier iter38b.
pkgrel 4 -> 5. No other PKGBUILD changes.
Triggered by 'pacman -Syu' on fresnel showing local 1:6.11.0-2 newer
than (then-)extra 6.11.0-4. By the time of this commit Arch has moved
on to 6.11.1-1; rebase straight to that.
Diff against Arch's qt6-base 6.11.1 PKGBUILD reduced:
- _pkgver 6.11.0 -> 6.11.1
- pkgrel 3 -> 1 (reset on pkgver bump)
- sha256 of git tag updated to 2eafe504... (matches Arch's 6.11.1 PKGBUILD)
- Drop cherry-pick of 8b54513cdcf6 (qdbus crash fix landed upstream
in 6.11.1; Arch dropped it from theirs as well)
- depends: gcc-libs replaced by libgcc + libstdc++ (Arch split these
into separate packages between 6.11.0 and 6.11.1)
- Build flag: -DFEATURE_sql_ibase=OFF removed, -DFEATURE_mimetype_database=OFF
added (Arch toggled in 6.11.1)
Kept (the reason this fork exists):
- 0001/0002/0003 GL_R8-on-ES3 patches (Mali GLES3 KWin compositor stall fix)
- qt6-base-cflags.patch + qt6-base-nostrip.patch (system CFLAGS + skip strip)
- qt6-xcb-private-headers-fourier sub-package (depended on by other fork pkgs)
- epoch=1 (precedence guarantee until upstream lands GL_R8/ES3 fix)
Patches NOT dry-run-verified against 6.11.1 source (qt clone is multi-GB
and impractical to fetch for a quick check). They touch glyph cache and
RHI gles2 paths that have been stable since Qt 6.10; expected to apply
cleanly. Will revert pkgrel and re-pin if first build hits hunk failures.
Triggered by 'pacman -Syu' on fresnel showing local 1:6.6.4-1 newer
than extra 6.6.5-1 (epoch wins, but our base lags upstream).
- pkgver 6.6.4 -> 6.6.5
- pkgrel 3 -> 1 (reset on pkgver bump)
- sha256sum updated for new tarball
- 0001-transaction-bypass-watchDmaBuf-fence-wait.patch dry-run verified
clean against 6.6.5/src/wayland/transaction.cpp
Epoch=1 retained until upstream lands a proper V4L2 implicit-sync
fence-wait bypass (the patch-header hypothesis is still being tested
through ohm/fresnel HW-decode validation; not upstreamable yet).
Follow-up to #10. The subdir layout I assumed worked turned out to not
work — per 'man PKGBUILD':
The filename in the array must NOT include any path components like ./
makepkg's source-staging only honors basenames; the patches/ subdir
references staged the files but the basename-lookup at apply time
failed with the same 'not found in build directory' error that #9 hit.
I copied the chromium-fourier convention thinking it was proven, but
chromium-fourier has no CI either and likely has the same latent
breakage (separate issue, not in this PR's scope).
The actually-working pattern is the mpv-fourier one: patches at the
top level next to the PKGBUILD. mpv-fourier iter2 just built green
on CI (#92) with that layout.
This commit:
- Moves all 8 patches (5 fourier + 3 arch upstream) from patches/
to arch/firefox-fourier/ via git mv (preserves blame).
- Removes the now-empty patches/ subdir.
- Strips 'patches/' prefix from source array entries.
- Strips '/patches' from prepare() patch -i paths.
No semantic change to the patch content or apply order.
pkgrel 3 -> 4.
The PKGBUILD has been structurally broken since at least the 150.0.1
import: source array referenced patches at top level ('0001-...patch')
but the actual files live in patches/ — makepkg searched for them in
the package root and failed before doing any work. The arch-0002/3/4
upstream toolchain patches (clang 22, glibc 2.43, Rust 1.95) weren't
shipped in the repo at all.
Caught when boltzmann tried to build 150.0.1-2 (PR #9 follow-up); fails
at the source-staging phase with:
==> ERROR: arch-0002-Bug-2033279-Make-enable-rust-simd-work-with-Rust-1.9.patch
was not found in the build directory and is not a URL.
Why nobody noticed: firefox-fourier has no Gitea Actions job (too heavy
for the runner pool). The shipped 150.0.1-1 .pkg.tar.zst was produced
by hand-staging files; the PKGBUILD itself never reproduced from a
clean checkout.
This commit:
- Moves all source-array patch references under patches/ to match the
on-disk layout (same convention as chromium-fourier).
- Updates prepare() paths accordingly.
- Adds arch-0002/3/4 patches fetched verbatim from
https://gitlab.archlinux.org/archlinux/packaging/packages/firefox
(raw/main/{0002,0003,0004}-...patch). Renamed with arch- prefix in
the repo to avoid namespace collision with the local 0001..0005.
- Adds the previously-omitted 0005-rdd-sandbox-v4l2-media-ctl.patch
to the source array + prepare(); this is the patch that broadens
the RDD sandbox so /dev/media* and /dev/video* ioctls reach the
decoder child (README §1 enumerates it as load-bearing for the HW
path). It was present in patches/ but not wired up.
Dry-run verified all 8 patches apply cleanly against firefox-150.0.1
source (some hunks offset by ≤3 lines, no conflicts).
pkgrel 2 -> 3.
Three prefs that gate the patched VAAPI / V4L2-request code path are
'false' upstream and need to be 'true' for the libva-v4l2-request
backend to actually engage on RK3399 / panfrost EGL:
widget.dmabuf.force-enabled
media.hardware-video-decoding.force-enabled
media.ffvpx-hw.enabled
Without them, Gecko's DMA-BUF probe silently fails (panfrost EGL
doesn't trip the right detection), the (Intel-tuned) HW-vs-SW cost
heuristic picks SW, and the FFvpx PDM never selects its HW-capable
variant for VP8/VP9. Net result: fresh profiles SW-decode despite the
0001..0004 unlock patches being applied.
Shipped as a vendor-default file at:
/usr/lib/firefox-fourier/browser/defaults/preferences/rockchip-fourier-defaults.js
Lower precedence than user.js / about:config — power users debugging
HW decode can still flip them off without touching the package.
pkgrel 1 -> 2.
Closes: #8
Carries the fork from libva-multiplanar iter8 close (2026-05-06) to
fresnel-fourier iter38b (2026-05-14): multi-device probe so a single
libva session serves rkvdec (H.264 + HEVC + VP9) + hantro (MPEG-2 +
VP8) without LIBVA_V4L2_REQUEST_VIDEO_PATH overrides, plus the
MAX_PROFILES bounds-check fix from iter38b.
Required by the fresnel-fourier README's HW-decode stack (vainfo
exposes all 10 profiles in one session, byte-exact bench verified
across 5 codecs).
68 commits between the old pin (65969da3) and the new (7ac934e0).
pkgver() recomputes the rev count at build time; the static placeholder
is bumped from r280 to r348 to reflect the new pin.
Adds Kwiboo / Langdale's 2024-08 patches (0001-meson + 0002-vo-hwdec-drmprime)
so AV_HWDEVICE_TYPE_V4L2REQUEST flows through mpv's drmprime VO hwdec and
`--hwdec=v4l2request` engages against the dmabuf-wayland VO instead of failing
with "Could not create device". The matcher previously filtered on
AV_HWDEVICE_TYPE_DRM and rejected v4l2request even when libavcodec's hwaccel
returned a valid PRIME descriptor.
PKGBUILD: pkgrel 9 → 10, source[] picks up both new patches, prepare() applies
them before the existing 0001-vo_dmabuf_wayland cache-sync workaround,
meson build flag -Dv4l2request=enabled added so the new option compiles in.
Stacks on top of iter1 (713a856): the dma_buf cache-sync workaround stays
load-bearing for the existing vaapi_dmabuf and drmprime_dmabuf import paths;
this iter only adds the wiring for the drmprime VO's hwdec selector.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
marfrit
claude-noether
test0r
claude-noether