https://bugs.kde.org/show_bug.cgi?id=527582
Bug ID: 527582
Summary: Native Wayland GPU clients cause kwin_wayland to
constantly runtime-resume an idle secondary AMD dGPU
(~600 wakeups/s)
Classification: Plasma
Product: kwin
Version First 6.8.0
Reported In:
Platform: Other
OS: Linux
Status: REPORTED
Severity: normal
Priority: NOR
Component: platform-drm
Assignee: [email protected]
Reporter: [email protected]
CC: [email protected]
Target Milestone: ---
On a hybrid AMD laptop, the secondary dGPU (no outputs connected) is kept
from runtime-suspending as soon as a native Wayland client that submits
GPU-rendered buffers is active. Kernel tracing shows the wakeups come from
kwin_wayland itself, not from the client. The client renders on the iGPU only.
When the same client runs through Xwayland, the wakeups disappear entirely.
--- HARDWARE / SOFTWARE ---
- Laptop: ASUS TUF A16 FA617NS
- iGPU: AMD Radeon 680M (0000:77:00.0) - drives the internal panel
- dGPU: AMD Radeon RX 7600S (0000:03:00.0) - only HDMI/DP are wired to it,
nothing connected during tests
- OS: EndeavourOS (Arch), kernel 7.2.11-arch1-1
- Plasma: 6.8.0, KWin: 6.8.0, Wayland session
- Mesa: 26.2.5
- supergfxctl in Hybrid mode
- dGPU runtime PM: power/control=auto, autosuspend_delay_ms=100
(lowered because amdgpu itself wakes the GPU every ~2 s)
--- STEPS TO REPRODUCE ---
1. Leave nothing connected to the dGPU outputs. Check it is idle:
cat /sys/bus/pci/devices/0000:03:00.0/power/runtime_status -> suspended
2. Start a Chromium-based browser (Brave 155 / Chromium 155.0.8059.85) natively
on Wayland, pinned to the iGPU:
brave --ozone-platform=wayland --render-node-override=/dev/dri/renderD129
(chrome://gpu confirms: Ozone platform = wayland, active GPU = 0x1681 /
680M)
3. Scroll a page while recording runtime-PM resumes of the dGPU:
T=/sys/kernel/tracing
echo 1 > $T/events/rpm/rpm_resume/enable
timeout 15 cat $T/trace_pipe | grep "0000:03:00.0" > /tmp/rpm.log
echo 0 > $T/events/rpm/rpm_resume/enable
awk '{print $1}' /tmp/rpm.log | sort | uniq -c | sort -rn
--- OBSERVED (15 s capture) ---
9492 kwin_wayland-1780
~14 kworker/* (amdgpu internal, expected)
runtime_status flips between active/suspended continuously while the client
draws. It stops as soon as the window is idle. The browser feels laggy
because of the repeated resume latency.
--- EXPECTED ---
No dGPU activity from KWin while no output is enabled on the dGPU and the
client renders on the iGPU. The dGPU should stay suspended.
--- WHAT WAS RULED OUT ---
- The client does not touch the dGPU: no fd open on the dGPU nodes
(fuser on renderD128/card1 only shows kwin_wayland, Xwayland, systemd,
logind), and strace -e openat on the browser GPU process shows no access
to dGPU nodes or sysfs during the capture.
- --disable-accelerated-video-decode: no change.
- --disable-features=WaylandLinuxDrmSyncobj: no change (9106 kwin_wayland
hits).
--- WHAT MAKES IT DISAPPEAR ---
- Same browser via Xwayland (--ozone-platform=x11): 0 resumes in 15 s.
- Same browser with --disable-gpu (shm buffers instead of dmabuf): no flipping.
So the trigger seems to be KWin handling dmabuf-based buffers from a native
Wayland client, with KWin issuing ioctls on its dGPU fd for each frame
(every amdgpu ioctl takes a runtime-PM reference).
--- WORKAROUND ---
Running Chromium/Electron apps through Xwayland.
--
You are receiving this mail because:
You are watching all bug changes.
