https://bugs.kde.org/show_bug.cgi?id=522547

mimara <[email protected]> changed:

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

--- Comment #8 from mimara <[email protected]> ---
Arch Linux, Wayland, single 4K output (LG C2 over HDMI):

  plasma-workspace   6.7.4-3
  kwin               6.7.4-7
  kwindowsystem      6.29.0-1
  plasma-integration 6.7.4-5

The thread so far covers 6.7.2/6.7.3 and kwindowsystem 6.27.0-6.28.0, so this
is a data point that it survives into 6.29.0. Same string, plasmashell rather
than LibreOffice:

  kwin_wayland_wrapper[1493]: error in client communication (pid 1627)
  plasmashell[1627]: ext_background_effect_surface_v1#233: error 0: tried to
set blur region on destroyed surface
  plasmashell[1627]: The Wayland connection experienced a fatal error: Protocol
error
  systemd[1371]: plasma-plasmashell.service: Main process exited, code=exited,
status=255/EXCEPTION

systemd restarts plasmashell about a second later, so what I actually see is
the panels blink. Annoying rather than fatal, though it does take whatever
panel popup I had open down with it.

Worth flagging for anyone trying to reproduce it: because this is a protocol
error the client exits 255 with no signal, so there is no coredump. It is
invisible to any coredump-based search and to DrKonqi. I lost some time on that
before I worked out why nothing was being captured.

Frequency here is 8 occurrences over 14 days, and it is bursty rather than
steady:

  Aug 21: 08:40, 12:55, 16:35, 17:47
  Aug 25: 07:30
  Aug 31: 08:47, 11:51
  Sep  3: 18:35

Four in one day, then six days of nothing. That matters if you are trying to
catch it, because a couple of quiet days means nothing at all. I ran a
WAYLAND_DEBUG=client capture on plasmashell for 2.5 days and it recorded zero
crashes, which given those gaps was not surprising in hindsight.

One thing I want to be careful about, since it is easy to report the wrong
thing here. That capture did record a popup teardown where plasmashell destroys
the wl_surface 12us before the ext_background_effect_surface_v1 bound to it:

  -> wl_surface#230.destroy()
  -> ext_background_effect_surface_v1#227.destroy()

I was ready to call that the bug, but that teardown did not crash - it is a
trace of the normal path, and both objects do get destroyed. So it is not
evidence of the fault, and I mention it only so nobody chases destroy ordering
on my account. The stale-object analysis in comments 2 and 5 looks like the
real shape of it: the effect object outliving its surface and then being used.

The last time it went I had the Plasma notification history panel open.
Notification popups are short-lived blurred surfaces, so that is at least the
right class of object, but the last notification render was 5 minutes 13
seconds before the crash and there is nothing notification-related in the final
two minutes of the log, so I would not read much into it. I cannot check it
against the earlier seven because those journals have rotated out.

Happy to run something specific if it would help - I have the machine
reproducing this every few days and I can arm a capture on request, though as
above it needs to run for a week or so to have a decent chance of catching one.

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

Reply via email to