Public bug reported:
[ Impact ]
In mutter 50.1 every Wayland cursor surface a client creates is leaked,
together with its
`MetaCursorWayland` sprite. Each leaked surface also stays connected to the
cursor
renderer's `cursor-painted` signal, so every cursor repaint on every output
walks all of
them on the compositor thread. The cost grows with client churn: browsers,
Electron apps,
Playwright/automation, anything that sets a texture cursor. The desktop slowly
degrades to
roughly half rate on all monitors, including a stuttering pointer, until the
user logs out.
Measured on an affected 26.04.1 desktop (3× 4K, 2 days 8 hours of uptime):
20,388
handlers on `cursor-painted`, all the same callback (`on_cursor_painted`);
~380,000
`g_closure_invoke` calls per second; gnome-shell main thread at 42 % vs ~15 %
when healthy;
non-primary CRTCs at 50–66 ms per frame. Restarting the client apps freed none
of it.
Independently confirmed by a second user on an AMD iGPU laptop (Ubuntu 26.04),
so it is not
NVIDIA-specific; logging out clears it. 50.1-0ubuntu2.5 (resolute-proposed,
2026-10-08) does not
include the fix.
Cause: 95d44e93 ("wayland: Keep ref on surface at MetaWaylandCursor") created a
reference
cycle, MetaWaylandSurface → role (MetaWaylandCursorSurface) → MetaCursorWayland
→ strong
ref back to the surface. Upstream fixed it on gnome-50 with 918da17b
("cursor/wayland:
Don't hold a ref on the surface", MR !5097), first released in 50.3.
50.1-0ubuntu2.4 does
not include it.
[ Test Plan ]
A headless reproducer (nothing is shown on the real desktop) is at
https://github.com/nickallevato/mutter-cursor-leak-fix (`test/`). It runs
`mutter --headless` on a private D-Bus session, moves a virtual pointer with the
RemoteDesktop API, and starts N short-lived GTK4 clients that each set a
texture cursor
and exit. An LD_PRELOAD shim counts live `MetaWaylandCursorSurface` /
`MetaCursorWayland` objects with weak refs.
Results with 30 clients:
- 50.1-0ubuntu2.2: 30 of 30 cursor surfaces still alive after all clients exited
- 50.1-0ubuntu2.4: 30 of 30 still alive
- 50.1-0ubuntu2.4 + 918da17b: 0 of 30 (never more than 1 alive at a time)
Manual check on a real session: use texture-cursor apps (e.g. a browser)
heavily for a
day. With the fix, gnome-shell CPU stays at baseline and the cursor stays
smooth; without
it, both degrade.
[ Where problems could occur ]
The patch turns the sprite's surface reference into a weak pointer. Code that
used the
surface through the sprite after the surface was destroyed now sees NULL.
Upstream guards
the two places that do this (`prepare_at` and `get_buffer`). A missed caller
would show up
as a gnome-shell crash on cursor changes while a client is exiting. The
reproducer covers
exactly that case, a client exiting while its cursor is displayed, 30 times per
run, with
no crash or critical. The change is 4 hunks in one file, is upstream in
50.3–50.5, and
applies cleanly on top of 50.1-0ubuntu2.4.
[ Other Info ]
Investigated, backported and tested with the help of an AI coding agent (Claude
Code);
findings and numbers were checked on the affected machine. Workaround until
fixed: log out
and back in.
** Affects: mutter (Ubuntu)
Importance: Undecided
Status: New
** Tags: regression-release resolute wayland
** Attachment added: "patch for the fix"
https://bugs.launchpad.net/bugs/2171024/+attachment/6009561/+files/0001-cursor-wayland-dont-hold-a-ref-on-the-surface.patch
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2171024
Title:
Leaked Wayland cursor surfaces make GNOME Shell progressively slower
over uptime (fixed upstream in 50.3 by 918da17b)
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/mutter/+bug/2171024/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
