Nouveau - Mar 2021 - [PATCH 2/3] drm/nouveau/kms/nv50-: Report max cursor size to userspace

On Wed, Mar 3, 2021 at 7:41 AM Alex Riesen <alexander.riesen at
cetitec.com> wrote:
>
> Ilia Mirkin, Sat, Feb 27, 2021 22:26:57 +0100:
> > Can you try Alex's patch to modetest and confirm that you see
issues
> > with modetest? If so, can you (and maybe Alex as well) try an older
> > kernel (I'm on 5.6) and see if modetest behaves well there. [The
patch
> > in question was to expose 256x256 as the 'preferred' size, but
support
> > for the larger cursors has been around for a while.] Alex - if you
> > have time, same question to you.
>
> Sorry that it took so long. I retestet with the same kernel as Uwe did
> (5.4.101) and my results are the same: modetest with 64 and 128 cursors
looks
> good, 256 broken. Didn't test with X this time (this being my main
office
> machine), but can do later if required.
Thanks for confirming! No need to test X - that will work fine, since
we were reporting the smaller cursors on that kernel.

So I think we have definitive evidence that at least all Kepler
doesn't do 256x256 (and it's not just some bug in the display macro
rework that got introduced). And unless I'm blind and don't see the
corruption, Pascal seems fine. Question remains for Maxwell1/2 GPUs.
Unless someone has immediate plans to test on those, I'd recommend
bumping the minimum for 256x256 being reported to Pascal, and doing
128x128 for Kepler / Maxwell.

Simon, Lyude -- thoughts?

Cheers,

  -ilia

Am 03.03.21 um 14:12 schrieb Ilia Mirkin:
> On Wed, Mar 3, 2021 at 7:41 AM Alex Riesen <alexander.riesen at
cetitec.com> wrote:
>>
>> Ilia Mirkin, Sat, Feb 27, 2021 22:26:57 +0100:
>>> Can you try Alex's patch to modetest and confirm that you see
issues
>>> with modetest? If so, can you (and maybe Alex as well) try an older
>>> kernel (I'm on 5.6) and see if modetest behaves well there.
[The patch
>>> in question was to expose 256x256 as the 'preferred' size,
but support
>>> for the larger cursors has been around for a while.] Alex - if you
>>> have time, same question to you.
>>
>> Sorry that it took so long. I retestet with the same kernel as Uwe did
>> (5.4.101) and my results are the same: modetest with 64 and 128 cursors
looks
>> good, 256 broken. Didn't test with X this time (this being my main
office
>> machine), but can do later if required.
> 
> Thanks for confirming! No need to test X - that will work fine, since
> we were reporting the smaller cursors on that kernel.
> 
> So I think we have definitive evidence that at least all Kepler
> doesn't do 256x256 (and it's not just some bug in the display macro
> rework that got introduced). And unless I'm blind and don't see the
> corruption, Pascal seems fine. Question remains for Maxwell1/2 GPUs.
> Unless someone has immediate plans to test on those, I'd recommend
> bumping the minimum for 256x256 being reported to Pascal, and doing
> 128x128 for Kepler / Maxwell.
The only older model I have access to would be a 6600GT (NV43). Don't know
if this would have any
significance?

	Uwe

> Simon, Lyude -- thoughts?
> 
> Cheers,
> 
>   -ilia
>

On Wed, 2021-03-03 at 08:12 -0500, Ilia Mirkin wrote:
> On Wed, Mar 3, 2021 at 7:41 AM Alex Riesen <alexander.riesen at
cetitec.com>
> wrote:
> > 
> > Ilia Mirkin, Sat, Feb 27, 2021 22:26:57 +0100:
> > > Can you try Alex's patch to modetest and confirm that you see
issues
> > > with modetest? If so, can you (and maybe Alex as well) try an
older
> > > kernel (I'm on 5.6) and see if modetest behaves well there.
[The patch
> > > in question was to expose 256x256 as the 'preferred'
size, but support
> > > for the larger cursors has been around for a while.] Alex - if
you
> > > have time, same question to you.
> > 
> > Sorry that it took so long. I retestet with the same kernel as Uwe did
> > (5.4.101) and my results are the same: modetest with 64 and 128
cursors
> > looks
> > good, 256 broken. Didn't test with X this time (this being my main
office
> > machine), but can do later if required.
> 
> Thanks for confirming! No need to test X - that will work fine, since
> we were reporting the smaller cursors on that kernel.
> 
> So I think we have definitive evidence that at least all Kepler
> doesn't do 256x256 (and it's not just some bug in the display macro
> rework that got introduced). And unless I'm blind and don't see the
> corruption, Pascal seems fine. Question remains for Maxwell1/2 GPUs.
> Unless someone has immediate plans to test on those, I'd recommend
> bumping the minimum for 256x256 being reported to Pascal, and doing
> 128x128 for Kepler / Maxwell.
I'm fairly sure this will actually work fine on maxwell, I think Kepler is
the
only one I'm aware of so far that has issues with this. I've got one of
each gen
from kepler 1 up to ampere, so I'll try to find some today to test the cards
I've got and see where the support cutoff for this is and then submit a
patch
for the time being.

It's also really strange to me that this would be present in the display
classes
yet not be supported on the GPU, since there definitely aren't any
capabilities
for this. I'll also send an email to Andy and see if he might know
what's going
on here.
> 
> Simon, Lyude -- thoughts?
> 
> Cheers,
> 
> ? -ilia
> 
-- 
Sincerely,
   Lyude Paul (she/her)
   Software Engineer at Red Hat
   
Note: I deal with a lot of emails and have a lot of bugs on my plate. If
you've
asked me a question, are waiting for a review/merge on a patch, etc. and I
haven't responded in a while, please feel free to send me another email to
check
on my status. I don't bite!