> From: Jason Gunthorpe <jgg@xxxxxxxxxx> > Sent: Saturday, January 28, 2023 7:54 AM > > On Fri, Jan 27, 2023 at 09:58:46PM +0000, Robin Murphy wrote: > > > Please just add IOMMU_CAP_IOMMUFD to represent whatever the > nebulous > > requirements of IOMMUFD actually are (frankly it's no less informative > than > > calling domains "broken"), handle that in the drivers you care about > > and > > I don't want to tie this to iommufd, that isn't the point. > > We clearly have drivers that don't implement the iommu kernel API > properly, because their only use is between the iommu driver and some > other same-SOC driver. > > As a user of the iommu API iommufd and VFIO want to avoid these > drivers. > > We have that acknowledgment already via the IOMMU_DOMAIN_DMA stuff > protecting the dma_iommu.c from those same drivers. > > So, how about this below instead. Imagine it is followed by something along > the lines of my other sketch: > > https://lore.kernel.org/linux-iommu/Y4%2FLsZKmR3iWFphU@xxxxxxxxxx/ > > And we completely delete IOMMU_DOMAIN_DMA/_FQ and get dma- > iommu.c > mostly out of driver code. > > iommufd/vfio would refuse to work with drivers that don't indicate > they support dma_iommu.c, that is good enough. this is a good idea. Just one doubt. Robin mentioned that sprd-iommu might have a broken unmanaged domains: " I'd also question sprd-iommu, which hardly has a generally-useful domain size, and has only just recently gained the ability to unmap anything successfully." Want to understand why that restriction is not a problem for DMA API. what Jason proposed here assumes that existing driver support for DMA API can be safely applied to vfio/iommufd. But above looks warning certain exception? > diff --git a/drivers/iommu/intel/iommu.c b/drivers/iommu/intel/iommu.c > index 59df7e42fd533c..bb34d3f641f17b 100644 > --- a/drivers/iommu/intel/iommu.c > +++ b/drivers/iommu/intel/iommu.c > @@ -4165,8 +4165,6 @@ static struct iommu_domain > *intel_iommu_domain_alloc(unsigned type) > switch (type) { > case IOMMU_DOMAIN_BLOCKED: > return &blocking_domain; > - case IOMMU_DOMAIN_DMA: > - case IOMMU_DOMAIN_DMA_FQ: > case IOMMU_DOMAIN_UNMANAGED: > dmar_domain = alloc_domain(type); > if (!dmar_domain) { > @@ -4761,6 +4759,7 @@ static void intel_iommu_remove_dev_pasid(struct > device *dev, ioasid_t pasid) > } > > const struct iommu_ops intel_iommu_ops = { > + .use_dma_iommu = true, missed: + .allow_dma_fq = true,
