[FengHeting]

Summary

• EDID replies are validated before parsing, host scanout geometry outside 0..8192 is rejected, and the initial modeset tries the EDID preferred mode, the remaining EDID modes, the host scanout modes, the bootloader geometry and the configured default, in that order. The plane atomic check rejects modes and layouts the 2D scanout cannot scale or address, and a failed initial modeset now fails KMS init instead of leaving a half-configured pipeline.
• Closing the last DRM client clears the client cursor and flips the console framebuffer back instead of disabling the host scanout. The console framebuffer is pinned so rmfb cannot free it, the flush hooks detach when KMS is torn down, and a console flush or a dirty update from the committed primary takes the scanout back if a client flipped its own buffer. Flip error paths report which step failed.
• Control and cursor waits only use the wait queue when the scheduler is running, interrupts are enabled and the completion IRQ is available; otherwise they keep polling to the deadline. Each log reason also gets its own ratelimit state, fixing the five same-named static definitions that made virtgpu_vq.c fail to build with CONFIG_VIRTIO_GPU=y.

Dependencies

Stacked on the drm core PR: its three commits appear in this branch until that PR merges, after which this PR reduces to the three virtio commits.

Testing

Builds with CONFIG_VIRTIO=y, CONFIG_VIRTIO_PCI=y and CONFIG_VIRTIO_GPU=y.