On Thu, Oct 31, 2024 at 09:56:37AM -0700, Nicolin Chen wrote:
> On Thu, Oct 31, 2024 at 10:29:41AM -0300, Jason Gunthorpe wrote:
> > On Wed, Oct 30, 2024 at 02:35:27PM -0700, Nicolin Chen wrote:
> > > +void iommufd_vdevice_destroy(struct iommufd_object *obj)
> > > +{
> > > + struct iommufd_vdevice *vdev =
> > > + container_of(obj, struct iommufd_vdevice, obj);
> > > + struct iommufd_viommu *viommu = vdev->viommu;
> > > +
> > > + /* xa_cmpxchg is okay to fail if alloc returned -EEXIST previously */
> > > + xa_cmpxchg(&viommu->vdevs, vdev->id, vdev, NULL, GFP_KERNEL);
> >
> > There are crazy races that would cause this not to work. Another
> > thread could have successfully destroyed whatever caused EEXIST and
> > the successfully registered this same vdev to the same id. Then this
> > will wrongly erase the other threads entry.
> >
> > It would be better to skip the erase directly if the EEXIST unwind is
> > being taken.
>
> Hmm, is the "another thread" an alloc() or a destroy()?
I was thinking both
> It doesn't seem to me that there could be another destroy() on the
> same object since this current destroy() is the abort to an
> unfinalized object. And it doesn't seem that another alloc() will
> get the same vdev ptr since every vdev allocation in the alloc()
> will be different?
Ah so you are saying that since the vdev 'old' is local to this thread
it can't possibly by aliased by another?
I was worried the id could be aliased, but yes, that seems right that
the vdev cmpxchg would reject that.
So lets leave it
Jason
