[RFC] Host1x/TegraDRM UAPI

Mikko Perttunen cyndis at kapsi.fi
Thu Jun 25 09:17:19 UTC 2020


On 6/24/20 11:55 PM, Dmitry Osipenko wrote:
> 23.06.2020 15:09, Mikko Perttunen пишет:
> ...
>> * Previously present GEM IOCTLs (GEM_CREATE, GEM_MMAP) are not present.
>> Not sure if they are still needed.
> 
> Hello Mikko!
> 
> A quick comment for the starter. Switching away from the Tegra-specific
> GEM IOCTLs to the dma-buf heaps is a such radical change! But I like it!
> 
> Previously I was curious about whether we could have multiple CMA
> regions (one shared/reusable and other private, for example) for the
> Tegra DRM driver and at a quick glance the dma-buf heaps could be a nice
> solution that avoids re-inventing a driver-specific things for that.
> 
> I'm instantly foreseeing these types of heaps:
> 
> 1. Large reusable CMA heap.
> 2. Smaller non-reusable common CMA that could be used when allocation
> from a reusable CMA fails. Or vice versa, when we want to reduce the
> chance to block for a long time on allocation, for example.
> 3. Sparse (system) memory heap.
> 
> It's the first time I'm looking at the dma-buf heaps and it sounds like
> a very nice idea to utilize them!
> 
> The https://lkml.org/lkml/2019/11/18/787 says that the system-contiguous
> and carveout heaps we not added because they were not needed, but they
> will be needed for the Tegra20 drivers and for the case where IOMMU is
> disabled. It shouldn't be difficult to add these types of heaps.
> 
> I'll continue to examine the dma-buf heaps in a more details.
> 

Sure, let's go with this for now. We can anyway add GEM IOCTLs later if 
they are needed, without any compatibility issues.

Mikko


More information about the dri-devel mailing list