On Wed, Jun 25, 2025 at 09:37:18AM -0700, Nicolin Chen wrote:
> On Wed, Jun 25, 2025 at 03:38:19AM +0000, Tian, Kevin wrote:
> > > From: Nicolin Chen <nicol...@nvidia.com>
> > > Sent: Saturday, June 14, 2025 3:15 PM
> > >
> > > +int iommufd_access_notify_unmap(struct io_pagetable *iopt, unsigned long
> > > iova,
> > > + unsigned long length)
> > > {
> > > struct iommufd_ioas *ioas =
> > > container_of(iopt, struct iommufd_ioas, iopt);
> > > struct iommufd_access *access;
> > > unsigned long index;
> > > + int ret = 0;
> > >
> > > xa_lock(&ioas->iopt.access_list);
> > > xa_for_each(&ioas->iopt.access_list, index, access) {
> > > + if (!access->ops || !access->ops->unmap) {
> > > + ret = -EBUSY;
> > > + goto unlock;
> > > + }
> >
> > then accesses before this one have been notified to unpin the area
> > while accesses afterwards are left unnotified.
> >
> > in the end the unmap fails but with some side-effect incurred.
> >
> > I'm not sure whether this intermediate state may lead to any undesired
> > effect later. Just raise it in case you or Jason already thought about it.
>
> That's a good point. When an access blocks the unmap, there is no
> unmap happening so no point in notifying devices for ops->unmap.
>
> And, when the function is re-entered, there could be a duplicated
> ops->unmap call for those devices that are already notified once?
>
> So, if we play safe, there can be a standalone xa_for_each to dig
> for !access->ops->unmap. And it could be a bit cleaner to add an
> iommufd_access_has_internal_use() to be called under those rwsems.
Correct: it has to be under the xa_lock. So, this pre-check needs
to be done in this function.
Thanks
Nicolin
