Re: [PATCH v4 3/7] iommu: Support allocation of global PASIDs outside SVA

[Baolu Lu]

On 4/19/23 7:04 AM, Jacob Pan wrote:
> On Tue, 18 Apr 2023 10:06:12 +0800, Baolu Lu<baolu.lu@xxxxxxxxxxxxxxx>
> wrote:
>
> > On 4/18/23 12:46 AM, Jacob Pan wrote:
> > > On Wed, 12 Apr 2023 09:37:48 +0800, Baolu Lu<baolu.lu@xxxxxxxxxxxxxxx>
> > > wrote:
> > >    
> > > > On 4/11/23 4:02 PM, Tian, Kevin wrote:
> > > > > > From: Jacob Pan<jacob.jun.pan@xxxxxxxxxxxxxxx>
> > > > > > Sent: Saturday, April 8, 2023 2:06 AM
> > > > > > @@ -28,8 +26,8 @@ static int iommu_sva_alloc_pasid(struct mm_struct
> > > > > > *mm, ioasid_t min, ioasid_t ma
> > > > > >     		goto out;
> > > > > >     	}
> > > > > >
> > > > > > -	ret = ida_alloc_range(&iommu_global_pasid_ida, min, max,
> > > > > > GFP_KERNEL);
> > > > > > -	if (ret < min)
> > > > > > +	ret = iommu_alloc_global_pasid(min, max);
> > > > > I wonder whether this can take a device pointer so
> > > > > dev->iommu->max_pasids is enforced inside the alloc function.
> > > > Agreed. Instead of using the open code, it looks better to have a
> > > > helper like dev_iommu_max_pasids().
> > > yes, probably export dev_iommu_get_max_pasids(dev)?
> > >
> > > But if I understood Kevin correctly, he's also suggesting that the
> > > interface should be changed to iommu_alloc_global_pasid(dev), my
> > > concern is that how do we use this function to reserve RID_PASID which
> > > is not specific to a device?
> > Probably we can introduce a counterpart dev->iommu->min_pasids, so that
> > there's no need to reserve the RID_PASID. At present, we can set it to 1
> > in the core as ARM/AMD/Intel all treat PASID 0 as a special pasid.
> >
> > In the future, if VT-d supports using arbitrary number as RID_PASID for
> > any specific device, we can call iommu_alloc_global_pasid() for that
> > device.
> >
> > The device drivers don't know and don't need to know the range of viable
> > PASIDs, so the @min, @max parameters seem to be unreasonable.
> Sure, that is reasonable. Another question is whether global PASID
> allocation is always for a single device, if not I prefer to keep the
> current iommu_alloc_global_pasid() and add a wrapper
> iommu_alloc_global_pasid_dev(dev) to extract the @min, @max. OK?

No problem from the code perspective. But we only need one API.

We can now add the kAPI that we really need. In this series, the idxd
driver wants to allocate a global PASID for its kernel dma with pasid
purpose. So, iommu_alloc_global_pasid_dev() seems to be sufficient.

If, in the future, we will have a need to provide global pasid
allocation other than device drivers, we can easily add the variants.

Best regards,
baolu