marfrit-packages/upstream-submissions/vb2-dma-resv/lkml-submission-notes.md

linux-media submission notes — vb2 dma_resv release-fence series

The 3-patch RFC series at marfrit-packages/kernel/vb2-dma-resv-rfc/ is ready to send. Cover letter is filled in; patches apply cleanly via git am against Linux 6.12 mainline; all kernel APIs verified.

Pre-flight gates (re-stated)

1. Full-tree compile-test — apply the 3 patches to a complete kernel checkout (not a sparse one), make defconfig + enable CONFIG_VIDEO_HANTRO=y and CONFIG_VIDEO_ROCKCHIP_RGA=y, then build. A focused make drivers/media/common/videobuf2/videobuf2-core.o drivers/media/platform/verisilicon/hantro_v4l2.o drivers/media/platform/rockchip/rga/rga-buf.o is the minimal safe-to-send threshold; full make is the higher confidence bar.
2. Boot-test on ohm — install the patched kernel, verify lsmod shows hantro_vpu and rockchip_rga loaded, run mpv --vo=drm --hwdec=v4l2request bbb_1080p30_h264.mp4 and confirm playback still works (smoke test that we haven't broken the existing buffer-state machine).
3. Runtime validation — install patched kernel, pacman -R kwin-fourier so KWin's stock watchDmaBuf is re-enabled, play bbb_1080p30_h264.mp4 in chromium-fourier under KDE Plasma. Should plays through end-to-end because the fence the compositor waits on is now real (signals when hantro completes). Compare CPU profile against the kwin-fourier-bypassed baseline — should be equivalent or marginally lower.
4. Capture syscall numbers — strace -c -p $(pgrep kwin_wayland) during a 60-second playback, before and after the kernel patch (with stock KWin, kwin-fourier uninstalled). Reports the ioctl reduction the patch enables (since EXPORT_SYNC_FILE will now signal real fences instead of stub ones, but the call itself remains). The reduction is greater if the KWin upstream-shape patch is also in, but the kernel patch alone is sufficient to make the existing KWin path correct.

Send command

After all four gates pass, format-patch and send:

cd <linux-tree-with-patches-applied>
git format-patch --cover-letter --base=auto -3 \
                 -o /tmp/vb2-dma-resv-out HEAD
# Verify cover letter has the right Subject and blurb (we already
# filled them in, but git format-patch will overwrite — so paste
# the canonical ones from kernel/vb2-dma-resv-rfc/0000-cover-letter.patch)
git send-email \
    --to=linux-media@vger.kernel.org \
    --cc='Mauro Carvalho Chehab <mchehab@kernel.org>' \
    --cc='Hans Verkuil <hverkuil-cisco@xs4all.nl>' \
    --cc='Ezequiel Garcia <ezequiel@vanguardiasur.com.ar>' \
    --cc='Tomasz Figa <tfiga@chromium.org>' \
    --cc='Sumit Semwal <sumit.semwal@linaro.org>' \
    --cc='Christian König <christian.koenig@amd.com>' \
    --cc='Jason Ekstrand <jason@jlekstrand.net>' \
    --cc='dri-devel@lists.freedesktop.org' \
    --cc='linux-rockchip@lists.infradead.org' \
    --cc='Heiko Stuebner <heiko@sntech.de>' \
    --cc='Jeffy Chen <jeffy.chen@rock-chips.com>' \
    /tmp/vb2-dma-resv-out/*.patch

Rationale for the CC list:

• linux-media@vger.kernel.org — primary list (V4L2 / videobuf2)
• Mauro / Hans / Ezequiel — V4L2 maintainers
• Tomasz Figa — hantro maintainer (Collabora / Chromium overlap)
• Sumit Semwal / Christian König / Jason Ekstrand — dma-buf reviewers (this patch touches dma_resv semantics, so the dma-buf folks need a look)
• dri-devel — graphics list, the consumer side of the fence
• linux-rockchip@lists.infradead.org — RK3xxx-specific feedback
• Heiko Stübner — Rockchip / arm64 maintainer
• Jeffy Chen — Rock-Chips media stack lead

Expected feedback

• "Why opt-in vs auto-on for all CAPTURE queues?" — answered in cover letter; willing to switch to auto-on if reviewers prefer the audit burden.
• "Does this need a vb2 selftest?" — videobuf2 doesn't have great in-tree dmabuf coverage; offer to add a minimal kselftest if reviewers ask.
• "Is vb2_buffer_done the right signal point for asynchronous drivers?" — for hantro and rga it is (they're synchronous m2m). For drivers with async post-processing, may need a separate hook; out of scope for this RFC.
• "Pick-to: stable trees?" — Maintainers will decide. The patch is invasive enough that LTS backport may not be wanted; the contract gap has been there since vb2's introduction so it's not a regression fix.

Withdraw conditions

If reviewers point out that:

• A driver in the wider tree already does this and we missed it (the auto-on path would just naturally pick that up; switch approach).
• vb2's dma_resv integration is being done in a different upstream effort already (search lkml for videobuf2 dma_resv before sending — should already be in the cover letter).
• The fence semantics are wrong for some driver class we didn't consider (most likely image-sensor → CSI → ISP → output, where the producer fence at the camera side and the user-visible fence at the ISP output may be different).

…drop the series, write up the lessons, retry next cycle.