gfx: survive, and configure, a screen with no window manager

[jcelerier]

Opening a gfx output window on a board with no X and no Wayland killed score outright. setup_x11() falls back to eglfs there, and eglfs gives a screen to whichever window asks first: the second one is not an error, it is qFatal("EGLFS: OpenGL windows cannot be mixed with others.") inside QEglFSWindow::create(), so the editor being up meant the output took the process down.

Measured on an rk3588 with two HDMI outputs, Qt 6.12, one process per case:

what is opened	eglfs
one GL window	ok
one widget window	ok
several widget windows	ok (composited)
widget, then a GL window	qFatal
GL window, then a widget	ok
two GL windows, one screen	qFatal
widget @0 + GL @1, setScreen only	qFatal
widget @0 + GL @1, geometry set	ok
GL @0 + GL @1, geometry set	ok

Two findings, neither in the documentation and both confirmed against the Qt 6.12 sources:

1. A screen backs exactly one native surface. Raster windows after the first are fine — QOpenGLCompositor draws them onto it — which is why the order matters.
2. setScreen() does not place a window; its geometry does. The platform makes every window fullscreen and rederives the screen from where it landed. setGeometry(screen->geometry()) before show() is what works.

So ScreenNode and MultiWindowNode ask for a screen nobody holds, place by geometry, and decline to open with a clear message when every screen is taken, instead of aborting.

Configuring the displays

On an appliance there is no xrandr and no display panel: what the platform reads at startup is the only say anyone gets. score::gfx::DisplayConfig holds the outputs and what was asked of them, reads the real ones from /sys/class/drm — which answers before there is a QGuiApplication and while another process holds the display — and produces what each platform understands. It lives in score_lib_base because it runs from main() before plug-ins load.

The platforms are not equivalent and are not pretended to be. eglfs takes connector names, modes, formats, layout and cloning through its JSON. vkkhrdisplay has three integers, addresses displays by index, and cannot be told about layout at all, so it gets its own settings rather than a guessed translation. displayCapabilities() is what the UI asks, so the page offers less where less is possible, and says plainly on a desktop that nothing here applies to it.

requiresRestart is a capability, not an assumption: Windows and macOS can change displays while running, and a change that can be applied live can be undone. Neither backend is implemented yet, which is why perOutputConfiguration stays false there — the page must not offer what nothing behind it will do.

Getting back to an editor

A machine can end up showing output and nothing else, by configuration or because a display setting was wrong. The platform never takes a screen back, so the editor cannot be hidden and revealed — it has to not exist. editorUi decides whether one is built, and Ctrl+Alt+Shift+E writes the setting back and relaunches. On vkkhrdisplay it also sets platformOverride, since that platform creates a window for a widget and then draws nothing into it.

resolvePlatform refuses an override unless score was choosing an embedded platform anyway, so an appliance's file cannot hijack a desktop session.

Tests

test_unit_screen_placement and test_unit_display_config: screen selection, sysfs enumeration against a fixture, KMS JSON generation, save/load round-trip, the capability split, and the platform-override rules. Every commit builds on its own, checked by building each in turn.

[coveralls]

coverage: 15.294% (+0.07%) from 15.222% — fix/eglfs-screen-placement into master