Re: [PATCH v8 13/23] intel_iommu_accel: Bind/unbind guest page table to host

[Nicolin Chen <nicolinc@nvidia.com>]

On Fri, Dec 12, 2025 at 02:12:09AM +0000, Duan, Zhenzhong wrote:
> 
> 
> >-----Original Message-----
> >>> +static void vtd_destroy_old_fs_hwpt(HostIOMMUDeviceIOMMUFD
> >*idev,
> >>> +                                    VTDAddressSpace *vtd_as)
> >>> +{
> >>> +    if (!vtd_as->fs_hwpt) {
> >>> +        return;
> >>> +    }
> >>> +    iommufd_backend_free_id(idev->iommufd, vtd_as->fs_hwpt);
> >>> +    vtd_as->fs_hwpt = 0;
> >>is it a valid assumption a valid ID cannot be null? Is it documented
> >>somewhere?
> >
> >I didn't find it in uAPI doc, but it's hard coded in kernel:
> >
> >xa_init_flags(&ictx->objects, XA_FLAGS_ALLOC1 | XA_FLAGS_ACCOUNT);
> >
> >Hi @jasowang@redhat.com, @nicolinc@nvidia.com do you have a guideline
> >on this?
> >Could I take zero id reserved?
> 
> Sorry, typo error, I meant to @Jason Gunthorpe

It is not documented in uAPI but there is some implication that
ID=0 is restricted from being an actual object ID:
 297 /**
 298  * enum iommufd_option - ioctl(IOMMU_OPTION_RLIMIT_MODE) and
 299  *                       ioctl(IOMMU_OPTION_HUGE_PAGES)
 300  * @IOMMU_OPTION_RLIMIT_MODE:
 301  *    Change how RLIMIT_MEMLOCK accounting works. The caller must have privilege
 302  *    to invoke this. Value 0 (default) is user based accounting, 1 uses process
 303  *    based accounting. Global option, object_id must be 0

So, I think that it'd be easier for user space to follow, if we
add a line of statement:
-----------------------------------------------------------------
diff --git a/Documentation/userspace-api/iommufd.rst b/Documentation/userspace-api/iommufd.rst
index f1c4d21e5c5e..c250a9eafdd5 100644
--- a/Documentation/userspace-api/iommufd.rst
+++ b/Documentation/userspace-api/iommufd.rst
@@ -135,6 +135,8 @@ Following IOMMUFD objects are exposed to userspace:
   HW QUEUE, the VMM must request a pair of mmap info (offset/length) and pass in
   exactly to an mmap syscall via its offset and length arguments.
 
+Each object must be allocated via uAPI and returned with a non-zero ID.
+
 All user-visible objects are destroyed via the IOMMU_DESTROY uAPI.
 
 The diagrams below show relationships between user-visible objects and kernel
diff --git a/drivers/iommu/iommufd/main.c b/drivers/iommu/iommufd/main.c
index ce775fbbae94..4410f0561034 100644
--- a/drivers/iommu/iommufd/main.c
+++ b/drivers/iommu/iommufd/main.c
@@ -311,6 +311,7 @@ static int iommufd_fops_open(struct inode *inode, struct file *filp)
 	}
 
 	init_rwsem(&ictx->ioas_creation_lock);
+	/* ID=0 is reserved for global use in uAPI */
 	xa_init_flags(&ictx->objects, XA_FLAGS_ALLOC1 | XA_FLAGS_ACCOUNT);
 	xa_init(&ictx->groups);
 	ictx->file = filp;
-----------------------------------------------------------------

Jason?

Thanks
Nicolin