Re: [PATCH v8 04/12] iommufd/device: Only add reserved_iova in non-pasid path

[Nicolin Chen <nicolinc@nvidia.com>]

On Thu, Feb 27, 2025 at 09:50:29AM +0800, Yi Liu wrote:
> On 2025/2/27 08:05, Nicolin Chen wrote:
> > On Wed, Feb 26, 2025 at 03:40:24AM -0800, Yi Liu wrote:
> > > @@ -620,7 +622,7 @@ iommufd_device_do_replace(struct iommufd_device *idev, ioasid_t pasid,
> > >   		goto err_unresv;
> > >   	old_hwpt_paging = find_hwpt_paging(old_hwpt);
> > > -	if (old_hwpt_paging &&
> > > +	if (old_hwpt_paging && pasid == IOMMU_NO_PASID &&
> > >   	    (!hwpt_paging || hwpt_paging->ioas != old_hwpt_paging->ioas))
> > >   		iommufd_group_remove_reserved_iova(igroup, old_hwpt_paging);
> > 
> > I am a bit unclear about the use case with pasid and replace. If
> > we replace the old_hwpt_paging with pasid != IOMMU_NO_PASID, what
> > would happen those reserved iova on the old_hwpt_paging?
> > 
> > Or does that mean a device can multi-attach to one hwpt_paging via
> > different pasids so that only down the IOMMU_NO_PASID path would
> > it remove the reserved iova?
> 
> yes, and only the IOMMU_NO_PASID path adds it. So only this path removes
> it when the old_hwpt_paging is replaced or detached.

I see. This is replacing one of the domain in the per-device PASID
table, and only the master domain (PASID=0) owns the reserved_iova:

PASID=0 -> domainA -> attach_resv/remove_resv
PASID=1 -> domainB
PASID=2 -> domainC

?

Thanks
Nic