Add DMA-BUF support to screencopy

[emersion]

This would perform a GPU copy instead of a CPU one. screencopy already uses wl_buffer so it shouldn't be too hard.

• Indicate which device to allocate buffers on
• Indicate the modifier
• (Future issue) How does that play with explicit sync?

[emersion]

Implementation detail: on the compositor side, either of these could be used:

• https://www.khronos.org/registry/OpenGL/extensions/ANGLE/ANGLE_framebuffer_blit.txt (https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/3149)
• https://www.khronos.org/registry/OpenGL/extensions/NV/NV_framebuffer_blit.txt

[ammen99]

Out of curiosity, why would we want this instead of the export-dmabuf approach?

[ascent12]

I consider export-dmabuf to be borderline broken, and has some pretty bad race conditions. While copy-to-dmabuf isn't perfectly efficient, it'll still be pretty good, and I think it's much more preferable.

[emersion]

I disagree. Renderer v6 should allow us to mitigate race conditions in export-dmabuf.

[ascent12]

By throwing out any allocations once they've been exported?

[emersion]

No, by adding a "release" request and allowing the client to keep a reasonable number of DMA-BUFs locked.