[Intel-gfx] RFC frontbuffer write tracking

[Chris Wilson]

Having thrown around a few ideas for how to do PSR and FBC write
tracking on the frontbuffer, including the creation of a new SCANOUT
domain, the implementation that I've settled on is to detect writes to
either the GTT domain (hmm, a CPU write flush should also set the fb as
dirty) and then a deferred SCANOUT flush. The deferred flush then resets
the write domain and kicks the GTT mapping so that any further dumb
writes (i.e. dumb kms buffers or fbcon) cause a new frontbuffer
invalidation. This patch series also introduces framebuffer parameters
(which maybe should be properties) to allow userspace to opt out of the
dumb mechanism and elect to call dirtyfb itself when it requires the
scanout to be flushed.

I've been playing around with GFDT on SNB and IVB as a means for
prototyping the delayed updates and framebuffer parameters.
-Chris