kern/55828: Further problems on Ironlake GPUs, even on UXA

[jbryn%students.wcpss.net@localhost]

>Number:         55828
>Category:       kern
>Synopsis:       Further problems on Ironlake GPUs, even on UXA
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    kern-bug-people
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Fri Nov 27 06:50:00 +0000 2020
>Originator:     Jackson Bryn
>Release:        NetBSD-current 9.99.75
>Organization:
N/A
>Environment:
NetBSD E5510-NetBSD.dell.foo 9.99.75 (GENERIC) #0: Mon Nov 16 14:12:46 UTC 2020  mkrepro%mkrepro.NetBSD.org@localhost:/usr/src/sys/arch/amd64/compile/GENERIC amd64
>Description:
Not sure why the previous form did not get sent, but here we go:
Investigation needs to be done beyond a mere UXA fallback workaround. For various Ironlake boards I have, UXA works for basic 2D acceleration such as dragging windows, but complex canvases stutter, and 3D does not work.

Tested on a Dell Latitude E5510 and an HP EliteBook 8440p. These bugs do not happen if I instead use DRM via biosboot into VESA, disabling KMS.
>How-To-Repeat:
2D issues: build Firefox with native Xorg and MesaLib. I get slow canvas rendering at 30-40fps. With biosboot/VESA, I get a constant 60fps instead.

3D issues: Run "glxgears." The output fails to render past the first two frames. Drag it around, and the window should flicker, go crazy, freeze then crash with this message:
"i965: failed to submit batchbuffer"
On biosboot/VESA, the program in contrast works fine.

With firefox + native Xorg/MesaLib, WebGL fails to initialize on hardware, instead resorting to llvmpipe.
>Fix: