Environment
- GPU: NVIDIA GeForce RTX 4070 Laptop GPU
- Driver: 595.91.07 (
nvidia-driver-595-open, Ubuntu 24.04 “noble” packages on Zorin OS 18.1); previous driver 595.84 worked fine - Kernel: 7.0.0-34-generic — same bug reproduced on 7.0.0-31-generic, so it is kernel-independent
- Session: X11 (GNOME 46 / Zorin), single laptop display
- WebKitGTK: 2.52.6-0ubuntu0.24.04.1 (repro uses
MiniBrowserfromlibwebkit2gtk-4.1-dev)
Summary
After the unattended upgrade from 595.84 to 595.91.07, any WebKitGTK page that uses requestAnimationFrame sends WebKitWebProcess to ~100% CPU of one core, forever. The spin starts with the first rAF subscription and persists after the page animation ends. Heavier WebKitGTK applications never complete their first paint and show a blank white window; static pages (e.g. yelp) are unaffected; Chromium and Firefox on the same machine are unaffected.
Steps to reproduce
- Save the following as
repro.html— a bare 600-framerequestAnimationFramecounter, no canvas, no WebGL, no network:
<!DOCTYPE html>
<html>
<head><meta charset="utf-8"><title>rAF spin repro</title></head>
<body style="background:#fff; font:14px sans-serif; padding:2em">
<p id="p">animating…</p>
<script>
let n = 0;
const p = document.getElementById('p');
(function frame() {
p.textContent = 'frames: ' + n;
if (++n < 600) requestAnimationFrame(frame);
else p.textContent = 'done after 600 frames — CPU should be idle now';
})();
</script>
</body>
</html>
- Run it in WebKitGTK’s MiniBrowser:
DISPLAY=:1 /usr/lib/x86_64-linux-gnu/webkit2gtk-4.1/MiniBrowser file:///path/to/repro.html
- Watch
WebKitWebProcessintop: it pegs one core at ~99–100%. The 600-frame loop ends in ~10 s at 60 fps; measured 99.7% CPU a full 18 s after launch, and it never recovers until the process is killed.
Evidence that this is driver-side
- The very same application binary (byte-identical, sha256-verified) worked on 595.84 and broke immediately after the 595.91.07 upgrade (2026-09-25). The failure is identical on two different kernel versions.
- The spin reproduces under Xvfb too (software X server, no NVIDIA X driver in play) — note the NVIDIA EGL vendor libraries are still enumerated by glvnd in that scenario.
- Forcing the Mesa EGL vendor changes the failure mode to an immediate, clean abort instead of a spin:
__EGL_VENDOR_LIBRARY_FILENAMES=/usr/lib/x86_64-linux-gnu/libEGL_mesa.so.0→ WebProcess dies at startup withCould not create default EGL display: EGL_BAD_PARAMETER. Aborting... - Process characterization: the spin is on the WebProcess main thread, pure userspace (strace summary shows quiet futex/poll, no syscall storm), no crash, no Xid errors in dmesg.
Workarounds tried — none effective
WEBKIT_DISABLE_DMABUF_RENDERER=1, WEBKIT_DISABLE_COMPOSITING_MODE=1, WEBKIT_FORCE_VBLANK_TIMER=1, WEBKIT_DISABLE_SANDBOX_THIS_IS_DANGEROUS=1, LIBGL_ALWAYS_SOFTWARE=1, GDK_GL=disable, __GL_SYNC_TO_VBLANK=0/1, vblank_mode=0, clean $HOME (no caches). Recent WebKitGTK releases removed the non-accelerated compositing path, so the classic toggles no longer provide an escape hatch.
Ask
A fix in an upcoming 595.x release would be much appreciated — right now every WebKitGTK-based app with animations (including Epiphany/GNOME Web) is unusable on this machine. Happy to run tests, provide an nvidia-bug-report log, or more traces with instructions.
nvidia-bug-report.log.gz (386.9 KB)