Re: [PATCH 3/3] drm/mediatek: Use drm_gem_cma_object instead of mtk_drm_gem_obj

From: Daniel Vetter
Date: Mon Oct 29 2018 - 05:16:27 EST

On Mon, Oct 29, 2018 at 11:11:16AM +0800, CK Hu wrote:
> Hi,Daniel:
> On Fri, 2018-10-26 at 12:21 +0200, Daniel Vetter wrote:
> > On Fri, Oct 26, 2018 at 03:22:03PM +0800, CK Hu wrote:
> > > After adding dma_dev in struct drm_device and
> > > drm_gem_cma_dumb_create_no_kmap(), drm_gem_cma_object could replace
> > > mtk_drm_gem_obj, so use drm_gem_cma_object instead of mtk_drm_gem_obj to
> > > reduce redundant code.
> > >
> > > Signed-off-by: CK Hu <>
> >
> > A few questions/thoughts:
> >
> > - Why do you need both drm_device->dev and drm_device->dma_dev? Can't you
> > just register the drm_device with the right struct device?
> >
> In [1], mmsys is the drm driver and ovl0 and ovl1 is the sub device
> which has dma function.
> In this drm, there are two crtc and each one is comprised of many
> component.
> This is an example of mt8173:
> crtc0: ovl0, color0, aal, od, rdma0, ufoe, dsi0
> crtc1: ovl1, color1, gamma, rdma1, dpi0
> In the device node of ovl0 and ovl1, there is a 'iommus' parameter in
> it, so use dma_alloc_xxx() and dma_map_xxx() with that device would get
> iova rather than pa. I don't think it's a good idea to register ovl0 or
> ovl1 as drm device because each one is just a component in a pipeline.
> mmsys controls the clock and routing of multi-media system which include
> this drm system, so it's better to register mmsys as drm device. Maybe
> we could move 'iommus' parameter from ovl device to mmsys device, so the
> dma device changes from ovl device to mmsys device. I'm not sure this
> would be a good choice, how do you think?
> [1]

Ah ok. But if you have 2 blocks that make up the overall drm device, why
don't you need to switch at runtime between them? I.e. buffer allocated
for crtc0 needs to be dma-mapped to crtc0, buffer allocated to crtc1 needs
to be dma-mapped on crtc1?

And if they're both the exact same iommu, then imo it would make indeed
sense to move the iommu attribute up. Since your current code cant'
actually handle truly separate dma-mappings. And neither can your patch
series here handled separate iommu for crtc0 and crtc1.

> > - You don't use drm_gem_prime_import_dev, so prime import isn't using the
> > right device either.
> Yes, you are right. I'm not familiar with whore drm core, so I start to
> modify what Mediatek drm use. But this function still works for the drm
> device that itself is dma device. If one day there is a drm device which
> itself is not a dma device and need this function, send a patch to
> modify this function and test it with that drm device. If you want me to
> modify all in advance, I'm ok but need others to test it because
> Mediatek drm driver does not use them.

I meant to say that mediatek should use drm_gem_prime_import_dev, but
currently isn't using that. And your patch series here doesn't fix that
either. So there's more bugs left in this area.

> > - exynos seems to have the same or at least similar issue, stronger case
> > for your patches if you can solve both.
> I'm still Mediatek's employee. If I modify other company's driver and it
> is not a MUST-BE for Mediatek, Mediatek may think I give contribution to
> other company. So I've better not to modify exynos driver.

This isn't how upstream works :-)

> > - I'd start out with using struct drm_gem_cma_object in mtk (similar to
> > what vc4 does), and then reusing as much as possible of the existing
> > helpers. And then looking later on what's still left (like the support
> > for leaving out the virtual mapping).
> I'm not clear what vc4 does. It looks like that you want me to redefine
> mtk_drm_gem_obj based on drm_gem_cma_object. So it would be like
> struct mtk_drm_gem_obj {
> struct drm_gem_cma_object base;
> void *cookie;
> unsigned long dma_attrs;
> };
> I could try to modify as this and see what have left.

Yup, that's my suggestion. Then we can look at what mtk can use unchanged
from the core helpers. And what would need to change and so better
evaluate whether it makes sense to do that.

I still think just moving the iommu is probably best.
Daniel Vetter
Software Engineer, Intel Corporation