Messages in this thread | | | From | "Tian, Kevin" <> | Subject | RE: [PATCH v2 3/3] iommufd/device: Change iommufd_hw_pagetable_has_group to device centric | Date | Wed, 1 Feb 2023 03:07:32 +0000 |
| |
> From: Nicolin Chen <nicolinc@nvidia.com> > Sent: Monday, January 30, 2023 6:22 PM > > On Sun, Jan 29, 2023 at 02:38:55AM -0800, Nicolin Chen wrote: > > > > > @@ -385,10 +372,8 @@ void iommufd_device_detach(struct > > > > iommufd_device *idev) > > > > struct iommufd_hw_pagetable *hwpt = idev->hwpt; > > > > > > > > mutex_lock(&hwpt->ioas->mutex); > > > > - mutex_lock(&hwpt->devices_lock); > > > > refcount_dec(hwpt->devices_users); > > > > - list_del(&idev->devices_item); > > > > - if (!iommufd_hw_pagetable_has_group(hwpt, idev->group)) { > > > > + if (iommufd_hw_pagetable_has_device(hwpt, idev->dev)) { > > > > if (refcount_read(hwpt->devices_users) == 1) { > > > > iopt_table_remove_domain(&hwpt->ioas->iopt, > > > > hwpt->domain); > > > > @@ -397,7 +382,6 @@ void iommufd_device_detach(struct > iommufd_device > > > > *idev) > > > > iommu_detach_group(hwpt->domain, idev->group); > > > > } > > > > > > emmm how do we track last device detach in a group? Here the first > > > device detach already leads to group detach... > > > > Oh no. That's a bug. Thanks for catching it. > > > > We need an additional refcount somewhere to track the number of > > attached devices in the iommu_group. > > Wondering if we can let iommu_attach/detach_device handle this: >
that is the desired way to fully remove group awareness in iommufd.
but iirc there were some concerns on changing their semantics. But I don't remember the detail now. Jason might know. also +Baolu/Robin.
otherwise as long as the group attach/detach continues to be used then identifying last device in the group always needs some hack within iommufd itself.
| |