Fullscreen windows are not unredirected on NVIDIA X11 even with unredirect-fullscreen-windows enabled, preventing G-SYNC/VRR #847

Activity

1. itz-me-zappex commented on Aug 31, 2026

   

   itz-me-zappex

   on Aug 31, 2026

   More actions

   Here is a function that figures out whether window should be unredirected or not.
   

   muffin/src/compositor/meta-surface-actor-x11.c

   Lines 254 to 265 in 0b7a6b9

   	gboolean
   	meta_surface_actor_x11_should_unredirect (MetaSurfaceActorX11 *self)
   	{
   	if (!meta_surface_actor_x11_is_opaque (META_SURFACE_ACTOR (self)))
   	return FALSE;
   	if (!self->does_full_damage &&
   	!meta_window_is_override_redirect (self->window))
   	return FALSE;
   	return TRUE;
   	}

   As you can see, there is a filter that prevents unredirect under some circumstances, and there was said that dumb redirect/unredirect is not what devs want. #715 (comment)
2. khameleonium commented on Sep 27, 2026

   

   khameleonium

   on Sep 27, 2026

   More actions

   Same thing here on Intel integrated graphics with a native Linux game, same Mint 22.3 / Muffin 6.6.3 / Cinnamon 6.6.9. So this doesn't look NVIDIA- or Wine-specific. On Intel it shows up as a clear FPS loss in every game. Below are my measurements, plus two related findings from the same investigation.

   Setup: Intel UHD (CometLake-U GT2, i915, modesetting DDX + glamor), Mesa 25.2.8, Xorg 21.1.12, i5-10210U laptop, kernel 7.0. Single active monitor (HDMI 1920x1080@60, laptop panel off). unredirect-fullscreen-windows = true. Only stock @cinnamon.org applets, and all desklets were removed for the last runs.

   1. Fullscreen games are never unredirected → 10–15 FPS lost

   Test game: Half-Life 2, native 32-bit hl2_linux (OpenGL/SDL, not Steam, no Wine), true fullscreen 1920x1080, in-game VSync off. The game window looks like a perfect unredirect candidate:

   _NET_WM_STATE(ATOM) = _NET_WM_STATE_FULLSCREEN, _NET_WM_STATE_FOCUSED
   _NET_WM_BYPASS_COMPOSITOR(CARDINAL) = 1
     Absolute upper-left X:  0   Y: 0   Width: 1920   Height: 1080   Depth: 24
     Override Redirect State: no
   

   Same machine, same game, same scene, 30 s averages (CPU = % of one core):

   	Cinnamon X11	GNOME 46 X11 (same Xorg)	KDE Plasma Wayland	Cinnamon X11, no desklets
   In-game FPS	45–55	60 (capped)	60	45–55
   iGPU clock / idle (RC6)	1100 MHz (max) / 0 %	1100 / 0 %	1100 / 1 %	1100 / 0 %
   Package temp	67 °C	66 °C	68 °C	72 °C
   hl2_linux	62.5 %	66.0 %	67.5 %	64.4 %
   Compositor	cinnamon 18.6 %	gnome-shell 0.1 %	kwin_wayland 6.7 %	cinnamon 9.5 %
   Xorg / Xwayland	4.9 %	4.8 %	2.5 %	5.1 %

   No thermal or power throttling, and the GPU is fully loaded in every session. The only real difference is the compositor. On the same Xorg, gnome-shell goes idle during the fullscreen game (as expected after unredirecting). cinnamon keeps working, and the game loses ~10–15 FPS. It's not just HL2: all my Steam games (native and Proton) run fine in GNOME and KDE (both X11 and Wayland sessions) and noticeably worse in Cinnamon.

   A/B toggle of the setting while the game is focused (3 s settle, then 10 s sample; samples discarded if focus was lost):

   unredirect=true   cinnamon 10%  Xorg 6%
   unredirect=false  cinnamon 9%   Xorg 5%
   unredirect=true   cinnamon 10%  Xorg 6%
   unredirect=false  cinnamon 9%   Xorg 5%
   unredirect=true   cinnamon 9%   Xorg 5%
   unredirect=false  cinnamon 10%  Xorg 7%
   

   The setting makes no measurable difference: ~2 FPS in-game, and identical compositor load. Unlike the Wine case in #701, this is a native SDL window without the 1x1 child window problem, so it looks like the same "never unredirects" behaviour as in this issue.

   2. A desklet refreshing every 3 s froze all games for 0.5–1 s every 3 s

   Before the measurements above, every game in Cinnamon X11 froze completely (image stuck for 0.5–1 s, then smooth again) exactly every 3 seconds. This affected windowed and fullscreen games, native and Proton, even very light ones like Blobby Volley. Other desktops on the same machine had no such freezes.

   The cause was the third-party desklet commandResult@ZimiZones with a refresh interval of 3 s (delay = 3.0). On each refresh it runs 5 small commands through Util.spawn_async and redraws its text:

   date +"%A %B %d, %I:%M %p"                      3 ms
   hostname -I                                     2 ms
   sensors coretemp-isa-0000 / nvme-pci-2e00       5–7 ms
   upower -i $(upower -e | grep BAT)              19 ms
   uptime -p                                       2 ms
   

   The commands themselves are fast and run asynchronously. Still, removing the desklet completely fixed the freezes, while enabling unredirect-fullscreen-windows did not help at all.

   I'm reporting this here because it points to the same root cause. Game frames on X11 keep going through Cinnamon's compositor even in fullscreen, so any stall in Cinnamon's main loop (an xlet update, GJS work) freezes the game's image. The desklet may not be written ideally, but a desktop widget shouldn't be able to freeze a fullscreen game like this. With working unredirection, this wouldn't happen.

   3. Side note: the Wayland session gives clients only software rendering

   We know "Cinnamon on Wayland" is experimental, but this seemed worth mentioning. On this machine, every client in that session gets llvmpipe. Muffin itself uses the GPU, but Xwayland has no DRI3 and Wayland clients can't find the render node. Games are a slideshow there.

   $ glxinfo -B                    → llvmpipe (LLVM 20.1.2), Accelerated: no; "screen 0 does not appear to be DRI3 capable"
   $ eglinfo -B                    → GBM: Mesa Intel UHD (CML GT2);  Wayland / X11 platforms: llvmpipe
   $ wayland-info                  → zwp_linux_dmabuf_v1 version 3 only, no wl_drm
   EGL_LOG_LEVEL=debug             → "failed to get driver name for fd -1", "MESA-LOADER: failed to retrieve device information"
   open /dev/dri fds               → cinnamon: card1; Xwayland: none; game: none
   

   My guess is that Mesa 25.x needs linux-dmabuf v4 feedback (or wl_drm) to pick the device. I can open a separate issue for this if you prefer.

   ---

   I can run any additional tests or debug builds. If there's a way to log Muffin's unredirect decision, I'd be happy to capture it.

   The investigation (measurement scripts, xprop and A/B tests, finding the desklet) was done with the help of Claude Code (Anthropic's AI assistant). I verified the in-game behaviour myself.

   👍React with 👍2Reacted by Nicolás Capdet and Net Bielecki