Re: [PATCH v3 08/17] iommu: Make iommu_group_do_dma_first_attach() work with owned groups

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Apr 06, 2023 at 05:37:57PM +0100, Robin Murphy wrote:

> I don't have a particularly strong opinion on changing this behaviour, but I
> guess what's nagging me is that we could do a better job of calling out that
> it *is* a behaviour change, and its implications.

This patch got dropped with your other suggestion.

Instead this distinction is handled inside __iommu_device_set_domain():


	if (dev->iommu->attach_deferred) {
		if (new_domain == group->default_domain)
			return 0;
		dev->iommu->attach_deferred = 0;
	}

Relying on the owned means group->domain != default_domain thinking we
used elsewhere.

> slightly better if the device binds to the same owner, in that it could rely
> use the current domain straight off without the risk of being in ACS limbo,
> but also slightly worse in that if the attach fails for any reason, it will
> now trash aliased RIDs of devices that *were* working OK.

The functional issue is that iommufd uses iommu_attach_group() - so
conceivably it is possible to hot plug in a new group member, add a
vfio driver to it and then hot add it to an existing iommufd. In this
case it is required that the core code set the iommu_domain properly
since iommufd won't call iommu_attach_group() again.

On the alias RID side, I would say a driver should be checking if the
acutal RID is already set and not disturb anything..

> the device could bind to any *other* driver with driver_managed_dma and have
> a much better guarantee of having access to the other owner's
> domain,

Right, this is what the 'owner' argument to
iommu_device_claim_dma_owner() is supposed to allow.

I added the above note about iommufd to the patch "iommu: Fix
iommu_probe_device() to attach the right domain"

Jason