marfrit
84088141fd
build and publish packages / distcc-avahi-aarch64 (push) Successful in 35s
build and publish packages / lmcp-any (push) Successful in 7s
build and publish packages / lmcp-debian (push) Successful in 6s
build and publish packages / claude-his-any (push) Successful in 7s
build and publish packages / ffmpeg-v4l2-request-aarch64 (push) Successful in 12m2s
build and publish packages / claude-his-debian (push) Successful in 9s
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.
90 lines
3.7 KiB
Diff
90 lines
3.7 KiB
Diff
From: Markus Fritsche <mfritsche@reauktion.de>
|
|
Subject: [PATCH] transaction: bypass watchDmaBuf implicit-sync fence wait
|
|
Date: 2026-04-28
|
|
|
|
Background
|
|
----------
|
|
KWin's `Transaction::watchDmaBuf` (src/wayland/transaction.cpp) calls
|
|
`DMA_BUF_IOCTL_EXPORT_SYNC_FILE` on every plane of every imported
|
|
dmabuf and parks the transaction on a `QSocketNotifier(POLLIN)`
|
|
waiting for the resulting sync_file fd to become readable. The intent
|
|
is correct in principle — wait for the producer to finish writing
|
|
before sampling — but on V4L2 hantro CAPTURE buffers (RK3566 mainline
|
|
6.19, panfrost mesa 26.0.5) the resulting fence either signals so
|
|
late that chrome's 6-buffer V4L2 capture pool exhausts, or never
|
|
signals at all. Symptom (per chromium-fourier KWIN_PIVOT.md):
|
|
|
|
- chrome v4 attaches a video frame to a wp_subsurface, commits
|
|
- KWin's Transaction::commit calls watchDmaBuf, exports a sync_file,
|
|
parks on QSocketNotifier
|
|
- Sync_file never becomes readable
|
|
- Transaction never applies; old surface state never replaced
|
|
- wl_buffer.release for the previous video buffer never sent
|
|
- chrome's V4L2 capture pool starves at ~6 seconds, decoder blocks,
|
|
audio drains, hard stall
|
|
|
|
mpv with `--vo=gpu-next` on the same KWin session slideshows at 76%
|
|
drop rate but does not deadlock — its single-surface attach pattern
|
|
hits a different transaction shape than chrome's subsurface flow.
|
|
|
|
A clean weston A/B with the same chrome v4 binary plays through
|
|
end-to-end: the bug is specifically KWin's transaction fence-wait
|
|
path on this stack, not Wayland-as-a-protocol.
|
|
|
|
Fix
|
|
---
|
|
This experimental patch no-ops `watchDmaBuf` to test the hypothesis.
|
|
Implicit-sync correctness in this case is not lost: the V4L2
|
|
producer guarantees the buffer's contents are complete before
|
|
chrome sends `wl_surface.attach + commit`, and the wp_linux_dmabuf
|
|
client is required to do so by spec. The fence-wait was 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 under KWin with this patch, 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).
|
|
|
|
diff --git a/src/wayland/transaction.cpp b/src/wayland/transaction.cpp
|
|
index 967b22b..e3fbc06 100644
|
|
--- a/src/wayland/transaction.cpp
|
|
+++ b/src/wayland/transaction.cpp
|
|
@@ -263,27 +263,18 @@ static FileDescriptor exportWaitSyncFile(const FileDescriptor &fileDescriptor)
|
|
return FileDescriptor{};
|
|
}
|
|
#endif
|
|
|
|
void Transaction::watchDmaBuf(TransactionEntry *entry)
|
|
{
|
|
-#if defined(Q_OS_LINUX)
|
|
- const DmaBufAttributes *attributes = entry->buffer->dmabufAttributes();
|
|
- if (!attributes) {
|
|
- return;
|
|
- }
|
|
-
|
|
- for (int i = 0; i < attributes->planeCount; ++i) {
|
|
- const FileDescriptor &fileDescriptor = attributes->fd[i];
|
|
- if (fileDescriptor.isReadable()) {
|
|
- continue;
|
|
- }
|
|
-
|
|
- auto syncFile = exportWaitSyncFile(fileDescriptor);
|
|
- if (syncFile.isValid()) {
|
|
- entry->fences.emplace_back(std::make_unique<TransactionFence>(this, std::move(syncFile)));
|
|
- }
|
|
- }
|
|
-#endif
|
|
+ // kwin-fourier: no-op the implicit-sync fence wait. On V4L2
|
|
+ // hantro CAPTURE buffers (RK3566 mainline 6.19, panfrost mesa
|
|
+ // 26.0.5) the DMA_BUF_IOCTL_EXPORT_SYNC_FILE fence either never
|
|
+ // signals or signals so late that chrome's V4L2 capture pool
|
|
+ // exhausts at ~6s, hard-stalling the decoder. Wayland clients
|
|
+ // are required by spec to ensure the buffer's contents are
|
|
+ // complete before wl_surface.attach+commit, so this fence-wait
|
|
+ // is a belt-and-braces optimization, not a correctness primitive.
|
|
+ Q_UNUSED(entry);
|
|
}
|
|
|
|
} // namespace KWin
|