[Nouveau] Blank console but X11 works on MCP79 - old regression since 3.8

[Ilia Mirkin]

On Fri, Nov 17, 2017 at 2:37 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> On Fri, Nov 17, 2017 at 2:25 PM, Ondrej Zary <linux at rainbow-software.org> wrote:
>> On Friday 17 November 2017 18:41:17 Ilia Mirkin wrote:
>>> On Fri, Nov 17, 2017 at 12:33 PM, Ondrej Zary
>>>
>>> <linux at rainbow-software.org> wrote:
>>> > @@ -483,8 +483,8 @@
>>> >  nouveau 0000:02:00.0: disp:    0860: 00000000 -> 00000500
>>> >  nouveau 0000:02:00.0: disp:    0864: 00000000
>>> >  nouveau 0000:02:00.0: disp:    0868: 00000000 -> 04000500
>>> > -nouveau 0000:02:00.0: disp:    086c: 00000000 -> 00100500
>>> > -nouveau 0000:02:00.0: disp:    0870: 0000e900 -> 00001e00
>>> > +nouveau 0000:02:00.0: disp:    086c: 00000000 -> 00100a00
>>> > +nouveau 0000:02:00.0: disp:    0870: 0000e900 -> 0000e800
>>> >  nouveau 0000:02:00.0: disp:    0874: 00000000 -> ffff0000
>>> >  nouveau 0000:02:00.0: disp:    0878: 00000000
>>> >  nouveau 0000:02:00.0: disp:    0880: 05000000
>>> >
>>> > Looks like it's using 8bpp (0x1e00) in 32MB case but 16bpp (0xe800) in
>>> > 64MB case. Why?
>>> >
>>> > I get blank screen even with 64MB with video=1280x1024-8 kernel
>>> > parameter. Console works with video=1280x1024-16 even with 32MB stolen
>>> > memory.
>>> >
>>> > Conclusions: 8-bit support is broken and bpp reduction is weird.
>>>
>>> OK, well that makes a *ton* of sense (8bpp being broken).
>>>
>>> I think the idea of bpp reduction is that when you're on your shiny
>>> new Riva TNT with 16MB of VRAM, you don't want to go crazy allocating
>>> all that to a pinned fbcon - almost half of that would go to a single
>>> 32bpp 1600x1200 buffer, more for 1920x1200. You want to be able to
>>> have at least a few fb-sized buffers for backbuffer rendering, etc.
>>>
>>> The specific limits could probably use tweaking - I think they only
>>> consider VRAM size, not the fb size.
>>>
>>> I guess 8bpp worked prior to the change you bisected though, so we
>>> should figure out what we did wrong in the new code.
>>
>> Yes, booted 3.7 (last working kernel) and it's running in 8bpp.
>
> By the way, instead of booting $kernel, you can use modetest from
> libdrm/tests. Not sure if it supports C8 though =/
>
> I think the issue is this:
>
> -               OUT_RING(evo, nv_crtc->lut.depth == 8 ?
> -                               NV50_EVO_CRTC_CLUT_MODE_OFF :
> -                               NV50_EVO_CRTC_CLUT_MODE_ON);
>
> Whereas now we always set 0xC0000000 (aka "ON").

In case I was being unclear, I'm talking about

https://github.com/skeggsb/nouveau/blob/master/drm/nouveau/nv50_display.c#L1808

and surrounding items. Looks like lut_clr sets 0x40000000 which was
previously not used. Not sure what the difference between that and
0x00000000 is. This is what we have in rnndb for it:

https://github.com/envytools/envytools/blob/master/rnndb/display/nv_evo.xml#L408

So bit 30 is mode, set is "high res", unset is "low res". So really
what we want is 0x80000000 which leaves the LUT enabled but uses the
low-res mode?

All this could use some playing-around with.

  -ilia