Re: [PATCH v2] iommu: iommu_group_claim_dma_owner() must always assign a domain

[Jason Gunthorpe via iommu <iommu@lists.linux-foundation.org>]

On Mon, May 09, 2022 at 10:59:11AM +0100, Robin Murphy wrote:

> IOMMU_DOMAIN_DMA_FQ now effectively takes over the original
> semantics of IOMMU_DOMAIN_DMA as the one that depends on
> driver-specific functionality.

If I grasp the FQ stuff right, it seems that this only requires the
driver to implement ops->flush_iotlb_all()? I don't see anything
obvious in any driver that is specifically _FQ related?

If yes, it makes me wonder why I see drivers implementing
ops->flush_iotlb_all() but not supporting the _FQ domain during alloc?

Further, if yes, wouldn't it make sense to just trigger FQ based on
domain->op->flush_iotlb_all being set?

It seems like there is some confusion here, because I see the sysfs
default domain store path just does this:

	/* We can bring up a flush queue without tearing down the domain */
	if (type == IOMMU_DOMAIN_DMA_FQ && prev_dom->type == IOMMU_DOMAIN_DMA) {
		ret = iommu_dma_init_fq(prev_dom);
		if (!ret)
			prev_dom->type = IOMMU_DOMAIN_DMA_FQ;
		goto out;
	}

Which will allow a driver that rejected creating DMA_FQ during alloc
to end up with DMA_FQ anyhow???

> FWIW, mtk-iommu doesn't really have any need to reject
> IOMMU_DOMAIN_UNMANAGED, they just don't have any relevant client drivers
> that want it;

Ok..

> however arm-smmu definitely does need to continue rejecting
> IOMMU_DOMAIN_DMA when it can't rely on the DT code ensuring the
> correct probe ordering with the old DT binding (otherwise client
> drivers are liable to get broken DMA ops).

I saw this code and wondered what it does?

smmu alloc_domain returns NULL, which if I read everything right
causes NULL to become the default_domain.

But then what happens? This driver has no detach_dev so after, say
VFIO does stuff, how do we get back into whatever boot-time state NULL
represents?

Is it OK because dev_iommu_priv_get() in arm_smmu_attach_dev() will
always fail if legacy? If that is the case then why allow allocating
any domain at all?

It feels like this really wants a 'IOMMU_DOMAIN_PLATFORM_DMA_OPS' set
as the default_domain meaning that when that domain is assigned, the
platform's DMA ops are handling the iommu? If I get it right..

Anyhow, thanks, this sort of helps confirm my feeling that the domain
types are a little crufty..

Thanks,
Jason
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu