upstream-submissions/kwin-fourier: validation findings + revised MR body

Captured stock kwin 6.6.4 ioctl baselines on ohm (PineTab2 / RK3566 / mainline 6.19.10 / Plasma 6.6.4 Wayland): - Brave + bbb 1080p30 (sw decode), 60 s: 96,120 ioctls, 0 EXPORT_SYNC_FILE - chromium-fourier + bbb 1080p30, 30 s: 29,128 ioctls, 0 EXPORT_SYNC_FILE, ~7,800 SYNCOBJ_* (explicit sync) Finding: KWin 6.6.4 + drm-syncobj-aware clients negotiate wp_linux_drm_syncobj_v1 explicit sync, leaving Transaction::watchDmaBuf on the legacy implicit-sync path that fires only for clients that don't advertise drm-syncobj support. Updated kde-mr-body.md to reflect honest scope: structural cleanup of the legacy path (still relevant for older clients / V4L2 pipelines that don't wrap dmabufs in syncobjs) rather than the originally- hypothesized per-frame ioctl reduction. Patch is correct by construction; no regression on the no-targeted-path workload. Evidence files saved under measurements/. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 19:10:55 +00:00
parent 00aa186b99
commit 70a01bd41a
3 changed files with 29212 additions and 10 deletions
@@ -71,17 +71,62 @@ documents as the recommended primitive for this case.

 ## Validation

-Tested on PineTab2 (RK3566 / Mali-G52 / mainline kernel 6.19.10 /
-panfrost mesa 26.0.5 / KDE Plasma 6.6.4 Wayland) playing 1080p30
-H.264 in chromium under chromium-fourier. Frame rate and CPU
-profile equivalent to the previous code path; the savings are in
-compositor-loop microseconds rather than user-visible fps.
-`strace -c` on `kwin_wayland` during a 60-second playback shows
-**X% fewer `ioctl` calls** with this patch versus stock.
+Built and run on PineTab2 (RK3566 / Mali-G52 / mainline kernel
+6.19.10 / panfrost mesa 26.0.5 / KDE Plasma 6.6.4 Wayland). KWin
+session starts cleanly, multi-window OpenGL workloads (Brave,
+plasma-overview, kate, dolphin) render normally — no regression.

-(Will fill in the actual `strace -c` numbers once the kwin-fourier
-package built with this patch is validated end-to-end on ohm —
-work in progress.)
+### Honest finding from `strace -e trace=ioctl` on `kwin_wayland`
+
+In 30 s of 1080p30 H.264 playback under chromium-fourier
+(wp_linux_dmabuf_v1 client, V4L2-stateless decode capable):
+
+- **Zero `DMA_BUF_IOCTL_EXPORT_SYNC_FILE`** calls.
+- 29,128 ioctl calls total, dominated by:
+  - `DRM_IOCTL_PANFROST_*` (rendering)
+  - `DRM_IOCTL_SYNCOBJ_*` (≈ 7,800 — explicit-sync syncobj operations)
+
+The same shape holds with stock Brave + 60 s playback (96,120 total
+ioctls, 0 EXPORT_SYNC_FILE).
+
+KWin 6.6.4 with a modern client negotiates `wp_linux_drm_syncobj_v1`
+**explicit sync** — `Transaction::watchDmaBuf` is not on the hot path
+in this configuration. The function still exists and runs for clients
+that don't advertise drm-syncobj support (older Wayland clients,
+some video pipelines that import implicit-sync producer dmabufs
+without wrapping them), but on a current Plasma + drm-syncobj-aware
+client stack, the call frequency we initially expected (≈ 60
+ioctls/sec/client at 1080p30) does not materialize.
+
+### What this means for the patch
+
+The patch is still correct and worth taking, but the *value
+proposition* shifts from "measurable per-frame win on V4L2 video
+playback" to "remove a kernel round-trip on the legacy implicit-sync
+path that still services older Wayland clients and V4L2 producers
+that don't go through drm-syncobj-aware compositor paths".
+
+The structural improvement is real wherever `Transaction::watchDmaBuf`
+fires; we just couldn't construct a benchmark on this hardware that
+fires it heavily enough to put a percentage on. Reviewers with
+older client baselines (Plasma 5 era apps, GTK on older Wayland,
+xwayland passing implicit-sync buffers in some configurations) may
+see the call site fire more often.
+
+### What is reproducible
+
+- No regression: 60-second video playback, no fps loss, no rendering
+  artifacts, no compositor stalls.
+- Strace evidence available at
+  https://git.reauktion.de/marfrit/marfrit-packages
+  under `upstream-submissions/kwin-fourier/measurements/`:
+  - `stock-6.6.4-bbb1080p30-60s.txt` — `strace -c -f` summary
+  - `stock-cr-fourier-ioctls-30s.log` — full per-ioctl trace under
+    chromium-fourier playback
+- Patch correctness: the dmabuf fd's `poll()` semantics are
+  documented in `Documentation/driver-api/dma-buf.rst` "Implicit
+  Fence Poll Support" and match what `EXPORT_SYNC_FILE`+poll
+  observes.

 ## Future work (out of scope here)

@@ -0,0 +1,29 @@
+strace: Process 3891 attached with 6 threads
+strace: Process 3891 detached
+strace: Process 3931 detached
+strace: Process 3941 detached
+strace: Process 3932 detached
+strace: Process 3930 detached
+strace: Process 3928 detached
+% time     seconds  usecs/call     calls    errors syscall
+------ ----------- ----------- --------- --------- ----------------
+ 44.28    8.740874         801     10908           ppoll
+ 35.51    7.011272        1221      5740       459 futex
+  8.65    1.708574          17     96120           ioctl
+  3.41    0.673061         427      1576           clock_nanosleep
+  2.57    0.506512          48     10361           mmap
+  1.94    0.383496          37     10358           munmap
+  1.08    0.212343          16     13241           close
+  0.73    0.145039          42      3426           sendmsg
+  0.59    0.116589           7     16232           getpid
+  0.31    0.061258          18      3337       173 read
+  0.28    0.055864          19      2888      1444 recvmsg
+  0.23    0.045432          10      4428           fcntl
+  0.19    0.037662          24      1514           write
+  0.15    0.029347          16      1749           epoll_pwait
+  0.06    0.011514           8      1432           dup
+  0.01    0.002820        2820         1           restart_syscall
+  0.00    0.000272          34         8           timerfd_settime
+  0.00    0.000054          27         2           sysinfo
+------ ----------- ----------- --------- --------- ----------------
+100.00   19.741983         107    183321      2076 total