Re: [REGRESSION] Invalid gather when using Tegra210 media engines

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Feb 04, 2025 at 12:26:46PM +0900, Mikko Perttunen wrote:
> On 2/4/25 4:13 AM, Jason Gunthorpe wrote:
> > On Mon, Feb 03, 2025 at 06:49:07PM +0000, Robin Murphy wrote:
> > > I'd hope the historical reasons for not supporting IOMMU_DOMAIN_DMA in
> > > tegra-smmu no longer apply, given that all the default domain stuff has now
> > > been integrated into host1x for the newer arm-smmu based Tegras.
> > 
> > Indeed I do see appropriate looking calls to the normal DMA API, and
> > the other mapping path is conditionalized by !host->domain.
> > 
> > So, why didn't it work for Diogo? Even in identity mode the DMA API
> > will return correct DMA addresses and the !host->domain path will skip
> > mapping them.
> > 
> > Poking around I wonder if there is some assumption that if other parts
> > of the stack, maybe the DRM driver, are using the special domain than
> > everyone is? It seems to blindly pass around IOVA without really
> > checking who is consuming it.
> 
> I'm not sure where that would be, but it's certainly possible given that
> this combination of code paths wouldn't have been tested.

I saw some weird stuff.. Like tegra_bo_pin() does:

	/*
	 * If we've manually mapped the buffer object through the IOMMU, make sure to return the
	 * existing IOVA address of our mapping.
	 */
	if (!obj->mm) {
	} else {
		map->phys = obj->iova;
		map->chunks = 1;

And obj->iova isn't obviously linked to a dma map on dev.. The comment
reads like it is making the assumption that there is only one iommu
domain shared by everyone (without checking dev is part of that
scheme)

> > Christian was telling me DMABUF had some drivers that made the
> > (incorrect!) assumption they were all sharing translation.
> > 
> > It does seem like a nice project for someone who has the hardware to
> > rip out all of this custom domain stuff and just have the iommu layer
> > setup a shared dma-iommu domain.
> 
> This has been a long-standing project. The issue is that some boot chains
> set up the display expecting identity mappings,

The smmu driver can match on compatible strings and leave some devices
in identity mode, if that helps. Other drivers are doing this to work
around various issues.

> See https://lore.kernel.org/linux-iommu/20220512190052.1152377-5-thierry.reding@gmail.com/

But using RMR seems like a better solution?

We could perhaps also figure out some way to leave the translation in
identity until the DRM driver completes the reset, then the DRM driver
could activate the DMA API translation manually.

Jason