Etnaviv issues on i.MX8M-Mini

[Guido Günther]

On Wed, Feb 26, 2020 at 04:54:35PM +0100, Lucas Stach wrote:
> On Mi, 2020-02-26 at 15:31 +0000, Schrempf Frieder wrote:
> > On 25.02.20 09:13, Frieder Schrempf wrote:
> > > Hi Lucas,
> > > 
> > > On 24.02.20 12:08, Lucas Stach wrote:
> > > > On Mo, 2020-02-24 at 10:53 +0000, Schrempf Frieder wrote:
> > > > > Hi Lucas,
> > > > > 
> > > > > On 24.02.20 11:37, Lucas Stach wrote:
> > > > > > Hi Frieder,
> > > > > > 
> > > > > > On Mo, 2020-02-24 at 10:28 +0000, Schrempf Frieder wrote:
> > > > > > > On 20.02.20 19:58, Chris Healy wrote:
> > > > > > > > For the jerkey transitions, can you determine if this is a symptom of
> > > > > > > > a low framerate or dropped frames or something else?
> > > > > > > > 
> > > > > > > > Perhaps you can start your app with
> > > > > > > > "GALLIUM_HUD=fps,cpu,draw-calls,frametime".  This may give some 
> > > > > > > > clues.
> > > > > > > 
> > > > > > > The framerate seems ok. I get something between 50 and 70 FPS.
> > > > > > > 
> > > > > > > I have a Qt demo app with a menu and an animated 'ball' that moves
> > > > > > > across the screen. When the menu is visible, the ball movement is 
> > > > > > > really
> > > > > > > jerky (ball seems to 'jump back and forth' instead of moving 
> > > > > > > linearly).
> > > > > > > 
> > > > > > > As soon as I hide the menu and show the animation fullscreen, the
> > > > > > > movements are perfectly smooth.
> > > > > > > 
> > > > > > > Running the same app with software rendering, everything looks 
> > > > > > > good, too.
> > > > > > > 
> > > > > > > No idea what that means, though. I probably need to look at the 
> > > > > > > code of
> > > > > > > the app and do some more experiments to get a better idea of what 
> > > > > > > might
> > > > > > > cause the distortion.
> > > > > > > 
> > > > > > > Unless some of the graphics experts here already have an idea of what
> > > > > > > can cause and/or how to debug such an issue!?
> > > > > > 
> > > > > > Which driver is used for the display side? It seems like the display
> > > > > > side doesn't properly handle the dma fences used to synchronize scanout
> > > > > > and rendering.
> > > > > 
> > > > > I ported/picked the drivers for the LCDIF and DSI controllers from
> > > > > development branch of the 5.4-based vendor kernel [1] to our own
> > > > > v5.4-based kernel [2]. So it is quite probable, that something could be
> > > > > wrong here.
> > > > 
> > > > Please just use DRM_MXSFB for the display side, instead of the
> > > > downstream driver.
> > > 
> > > Hm, good idea. I somehow forgot about the fact, that there is an 
> > > upstream driver for the LCDIF controller. On first try I couldn't get it 
> > > to run on the i.MX8MM, but I suspect that's due to some reset, 
> > > power-domain or clock setup, that is missing upstream. I will see if I 
> > > can get any further with this.
> > 
> > So I had a closer look and while the DRM_MXSFB looks ok on its own, I 
> > have some problem with the rest of the i.MX8MM display subsystem.
> > 
> > The vendor stack, that I'm currently using integrates into the imx-drm 
> > master/core driver [1] that binds all the components of the display 
> > subsystem, such as the LCDIF driver and the integrated SEC_DSIM DSI bridge.
> > 
> > And because of my lack of DRM skills, I have no idea how to get the 
> > DRM_MXSFB driver to bind to the imx-drm core, instead of running 
> > separately and connecting directly to some panel as it is done for 
> > i.MX23/28 and i.MX6SX/UL.
> 
> It's a separate hardware and it's a pretty major design issue of the
> downstream driver that it integrates into imx-drm. You don't want this
> with the upstream driver.
> 
> Maybe Guido (CCed) can give you some clues, as apparently he is using
> the mainline eLCDIF driver + some patches to drive the DSI display path
> on i.MX8MQ. A lot of this will probably be transferable to the i.MX8MM
> display path.

Newer mxsfb supports attaching a bridge so if you make your DSI host
controller driver a DSI bridge mxsfb can drive it:

     https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/mxsfb/mxsfb_drv.c#n268

this should be similar to what was done for the imx8mq here (imx8mm
users a different ip core though):

     https://source.puri.sm/guido.gunther/linux-imx8/commits/forward-upstream/next-20200217/mxsfb+nwl/v9-wip

There's also some additional mxsfb patches by Robert floating around
which aren't mainline yet which the above branch also has.

Which reminds me that i need to prepare and send out a v9.
Cheers,
 -- Guido