[PATCH v2 05/11] ARM: dma-mapping: Bypass the legacy path for core-owned DMA API domains

[Mikko Perttunen]

On Tuesday, October 20, 2026 1:55 AM Robin Murphy wrote:
> On 16/10/2026 7:46 am, Mikko Perttunen wrote:
> > ARM's arch_setup_dma_ops / arm_setup_iommu_dma_ops creates a custom
> > IOMMU domain for DMA API usage, and continues to do so when the IOMMU
> > driver doesn't explicitly request a DMA default domain type.
> > 
> > When, however, that DMA default domain is requested and the core owns
> > the domain, we need to skip our own domain management. Add checks to
> > arm_setup/teardown_iommu_dma_ops for these cases. In such a case IOVAs
> > can be allocated from anywhere in the domain's geometry, so also add a
> > check to ensure the device's dma-ranges cover it. Presently, this is the
> > case for all platforms except some OMAP platforms with a non-zero
> > dma-ranges base.
> 
> It's probably worth adding some cleverness to arm_iommu_detach_device() 
> as well - it might work out benign to attempt to detach from a real 
> core-owned default domain, but we'll now want to set dma_ops and 
> to_dma_iommu_mapping according to the default domain itself (since we'll 
> now have the possibility of switching between an explicit unmanaged 
> dma_iommu_mapping and the implicit default one.)

I was thinking of just disallowing the explicit arm_iommu_* API for 
devices that get a default domain and hoping that any device that opts 
in to the "modern way" is capable of getting rid of the custom handling 
:)

I'll add a check to arm_iommu_attach_device to disallow it if there's
a core owned domain unless you have other thoughts.

Thanks,
Mikko

> 
> Thanks,
> Robin.
> 
> > Signed-off-by: Mikko Perttunen <mperttunen at nvidia.com>
> > ---
> >   arch/arm/mm/dma-mapping.c | 25 ++++++++++++++++++++++++-
> >   1 file changed, 24 insertions(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> > index f2d3e8ec3487..d6315baeddb9 100644
> > --- a/arch/arm/mm/dma-mapping.c
> > +++ b/arch/arm/mm/dma-mapping.c
> > @@ -1761,13 +1761,33 @@ EXPORT_SYMBOL_GPL(arm_iommu_detach_device);
> >   
> >   static void arm_setup_iommu_dma_ops(struct device *dev)
> >   {
> > -	struct dma_iommu_mapping *mapping;
> > +	struct dma_iommu_mapping *mapping = to_dma_iommu_mapping(dev);
> >   	u64 dma_base = 0, size = 1ULL << 32;
> >   
> >   	if (dev->dma_range_map) {
> >   		dma_base = dma_range_map_min(dev->dma_range_map);
> >   		size = dma_range_map_max(dev->dma_range_map) - dma_base;
> >   	}
> > +
> > +	/*
> > +	 * An existing cookie means the core owns the domain. Only set the
> > +	 * per-binding DMA ops. IOVAs are allocated from the whole aperture,
> > +	 * so the device's dma-ranges must cover it.
> > +	 */
> > +	if (mapping) {
> > +		const struct iommu_domain_geometry *geo =
> > +			&mapping->domain->geometry;
> > +
> > +		if (dma_base > geo->aperture_start ||
> > +		    dma_base + size < geo->aperture_end) {
> > +			dev_err(dev, "dma-ranges do not cover the IOMMU aperture\n");
> > +			return;
> > +		}
> > +
> > +		set_dma_ops(dev, &iommu_ops);
> > +		return;
> > +	}
> > +
> >   	mapping = arm_iommu_create_mapping(dev, dma_base, size);
> >   	if (IS_ERR(mapping)) {
> >   		pr_warn("Failed to create %llu-byte IOMMU mapping for device %s\n",
> > @@ -1792,6 +1812,9 @@ static void arm_teardown_iommu_dma_ops(struct device *dev)
> >   	if (!mapping)
> >   		return;
> >   
> > +	if (!mapping->owns_domain)
> > +		return;
> > +
> >   	arm_iommu_detach_device(dev);
> >   	arm_iommu_release_mapping(mapping);
> >   }
> > 
> 
>