https://bugs.kde.org/show_bug.cgi?id=521668

[email protected] changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[email protected]

--- Comment #2 from [email protected] ---
Also hitting the similar issue on Plasma 6.7.0 (Arch Linux, kernel
7.0.12-arch1-1).
In my case the dominant symptom is persistent flicker on the DisplayLink
outputs rather than a hard crash, which I think points at the same multi-GPU
copy path.

Setup:
- Primary render GPU: Intel (xe), driving one native DisplayPort monitor
(2560x1600@144) -- this output is never affected.
- Two external DVI monitors via a DisplayLink dock = evdi 1.14.16 virtual
devices (secondary GPU). The evdi outputs run ABGR8888.
- kwin_wayland 6.7.0, Wayland session.

Symptoms: persistent flicker on the two DisplayLink outputs, plus a single
kwin_wayland SIGSEGV inside libgallium (Mesa) at session startup that
self-recovered.
Downgrading the whole Plasma stack to 6.6.5 (keeping KF 6.27 + Qt 6.11.1) fully
fixes both the flicker and the crash.

Env-var test matrix (with the usual DisplayLink workarounds already set:
KWIN_DRM_DISABLE_TRIPLE_BUFFERING=1, KWIN_DRM_USE_MODIFIERS=0,
KWIN_DRM_NO_DIRECT_SCANOUT=1, KWIN_FORCE_SW_CURSOR=1, and KWIN_DRM_DEVICES
pinned Intel-first):
- KWIN_DRM_FORCE_GL_FINISH_MGPU_COPY=1: flicker disappears completely, BUT the
per-frame glFinish stalls the shared render loop and drags the ENTIRE desktop
-- including the native 144Hz Intel monitor -- down to ~30fps.
- KWIN_DRM_NO_AMS=1 (and plain atomic without the glFinish): smooth/full
refresh, but the flicker returns.
- Dropping the DisplayLink outputs from 75Hz to 60Hz: no effect on the flicker.

The fact that only a full pipeline drain (glFinish) immediately before the
cross-GPU copy eliminates the tearing suggests the new GPU-based multi-GPU copy
path from !9236 ("backends/drm: don't use a software renderer for multi GPU
copies", commit 0d6b757c, bug 520361) is letting the secondary/DisplayLink GPU
scan out before the copy has completed -- i.e. a missing fence/sync between the
copy and the secondary-GPU presentation. That change fixed the 520361 crash but
looks like it traded it for a frame-sync regression on the DisplayLink scanout.

-- 
You are receiving this mail because:
You are watching all bug changes.

Reply via email to