Wayland: missing acquire point on wp_linux_drm_syncobj_surface_v1 commit with NVIDIA EGL driver (KWin disconnects with protocol error)

[johnathonfox]

Summary

On Wayland with the proprietary NVIDIA driver + KWin compositor, the
first surfman commit triggers a wp_linux_drm_syncobj_surface_v1
protocol error the instant a WindowRenderingContext-backed Servo
WebView is installed — "explicit sync is used, but no acquire
point is set". KWin disconnects; Gdk exits with Error 71 (Protocol error) within ~50ms of the renderer becoming active. The embedder
process dies with exit code 1.

Environment

• Compositor: KWin (KDE Plasma 6) on Wayland, session
  wayland-0
• GPU / driver: NVIDIA GeForce RTX 5070 Ti, proprietary driver
  595.58.03
• OS: CachyOS (Arch-derivative), Linux 7.0.0
• surfman: 0.11.0 (from crates.io, transitive via
  servo = "0.1.0")
• Servo: 0.1.0 (released 2026-04-13)
• rust: 1.94.0 stable

Minimal reproduction shape

In a Tauri 2 app, create a plain tauri::window::Window (no
webview) and hand its raw-window-handle to
servo::WindowRenderingContext::new, then build a Servo instance +
WebView with that context and call load(data:text/html;…) on
the WebView. No other surfaces share the wl_surface; this isn't
a dual-context conflict.

Relevant log excerpt with WAYLAND_DEBUG=client:

[...] -> xdg_toplevel#46.set_title("Capytain Reader (Servo)")
INFO  paint::painter: Running on NVIDIA GeForce RTX 5070 Ti/PCIe/SSE2 with OpenGL version 3.2.0 NVIDIA 595.58.03
INFO  capytain-desktop: Servo renderer installed
[Display Queue] wl_display#1.error(wp_linux_drm_syncobj_surface_v1#48, 4,
    "explicit sync is used, but no acquire point is set")
Gdk-Message: Error 71 (Protocol error) dispatching to Wayland display.

KWin's interpretation: the client (NVIDIA EGL-Wayland driver, via
surfman's Wayland backend) subscribed to
wp_linux_drm_syncobj_surface_v1 for the surface but committed a
buffer without setting the acquire timeline point that the protocol
requires. KWin's protocol validator rejects the commit as a fatal
protocol error.

surfman 0.11.0 itself does not reference the wp_linux_drm_syncobj
strings — grep is clean — so the protocol subscription appears to
be happening inside NVIDIA's EGL-Wayland library when it sees KWin
advertising the interface. From the surfman side, that means the
first EGLSwapBuffers (or whatever path triggers the initial
commit) is missing the per-surface acquire-point wiring that
NVIDIA's driver expects to participate in.

Workarounds confirmed in testing

Workaround	Result
MESA_LOADER_DRIVER_OVERRIDE=llvmpipe LIBGL_ALWAYS_SOFTWARE=1 __EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json	Bypasses the error. Forces Mesa llvmpipe software GL instead of the NVIDIA EGL-Wayland path; no protocol error, app continues to run. Confirms the bug is NVIDIA-EGL-specific, not generic-Wayland.
GDK_BACKEND=x11 (force XWayland)	Gdk falls back to XWayland but parent.display_handle() on the Tauri plain window then returns HandleNotAvailable, so the Servo path silently disables and the app runs without the reader. Unrelated to the sync issue but worth noting.
KWIN_DRM_NO_EXPLICIT_SYNC=1	No effect. That env affects KWin's own DRM scanout sync, not whether it advertises wp_linux_drm_syncobj_surface_v1 to clients.
LIBGL_ALWAYS_SOFTWARE=1 alone	NVIDIA's EGL-Wayland driver ignores it.
Replacing the Tauri WebviewWindow parent with a plain tauri::window::Window (no webview)	Same error. Rules out "two GL contexts fighting over one wl_surface" — the error reproduces with Servo as the only subscriber.

Why filing against surfman

The protocol object comes from NVIDIA's closed-source EGL-Wayland
layer, which participates in the protocol automatically when it
sees KWin advertising it. But the observable fix has to live
somewhere the embedder can reach — either in surfman's Wayland
backend's first-commit path (setting an acquire point before
calling EGLSwapBuffers) or in an embedder-facing opt-out knob.
Filing here since surfman is the Servo-side Wayland owner; happy
to re-file against servo/servo or upstream NVIDIA if preferred.

Impact on downstream

For Capytain (a pure-Rust email client embedding Servo for the
reader pane), this is the single blocker between compile-validated
Servo integration on Linux and runtime-validated end-to-end
rendering on the host hardware class that's currently most common
in Linux dev boxes (NVIDIA + Wayland + KWin or similar). The
SoftwareRenderingContext path works fine as a headless corpus
harness but isn't viable as the user-facing reader pane because
it loses native text selection, scrolling, and the
prefers-color-scheme media query path.

Happy to run deeper diagnostics or test patches. The full Day 2
investigation including the WAYLAND_DEBUG protocol dump is linked
in the summary above.

[jdm]

Do you have any information about what the acquire point means in practice? There isn't a lot of information here to go on.

[jdm]

I wonder if #361 will help here?

[mrobinson]

@johnathonfox Do you mind running your application with WAYLAND_DEBUG=1 until your application crashes and attaching the log here?