RE: [PATCH v3 09/17] iommufd: Add IOMMU_HWPT_INVALIDATE

[Liu, Yi L]

> From: Nicolin Chen <nicolinc@xxxxxxxxxx>
> Sent: Thursday, August 3, 2023 10:17 AM
> 
> On Mon, Jul 31, 2023 at 10:19:09AM -0300, Jason Gunthorpe wrote:
> > On Mon, Jul 31, 2023 at 10:07:32AM +0000, Liu, Yi L wrote:
> > > > > +		goto out_put_hwpt;
> > > > > +	}
> > > > > +
> > > > > +	/*
> > > > > +	 * Copy the needed fields before reusing the ucmd buffer, this
> > > > > +	 * avoids memory allocation in this path.
> > > > > +	 */
> > > > > +	user_ptr = cmd->data_uptr;
> > > > > +	user_data_len = cmd->data_len;
> > > >
> > > > Uhh, who checks that klen < the temporary stack struct?
> > >
> > > Take vtd as an example. The invalidate structure is struct
> iommu_hwpt_vtd_s1_invalidate[1].
> > > The klen is sizeof(struct iommu_hwpt_vtd_s1_invalidate)[2].
> iommu_hwpt_vtd_s1_invalidate
> > > is also placed in the temporary stack struct (actually it is a union)[1]. So the klen
> should
> > > be <= temporary stack.
> >
> > Ohh, I think I would add a few comments noting that the invalidate
> > structs need to be added to that union. Easy to miss.
> 
> Added here:
> 
> -        * Copy the needed fields before reusing the ucmd buffer, this
> -        * avoids memory allocation in this path.
> +        * Copy the needed fields before reusing the ucmd buffer, this avoids
> +        * memory allocation in this path. Again, user invalidate data struct
> +        * must be added to the union ucmd_buffer.
> 
> > > It's not so explicit though. Perhaps worth to have a check like below in this patch?
> > >
> > > if (unlikely(klen > sizeof(union ucmd_buffer)))
> > > 	return -EINVAL;
> >
> > Yes, stick this in the domain allocate path with a WARN_ON. The driver
> > is broken to allocate a domain with an invalid size.
> 
> And here too with a WARN_ON_ONCE.
> 
> +       /*
> +        * Either the driver is broken by having an invalid size, or the user
> +        * invalidate data struct used by the driver is missing in the union.
> +        */
> +       if (WARN_ON_ONCE(hwpt->domain->ops->cache_invalidate_user &&
> +                        (!hwpt->domain->ops->cache_invalidate_user_data_len ||
> +                         hwpt->domain->ops->cache_invalidate_user_data_len >
> +                         sizeof(union ucmd_buffer)))) {
> +               rc = -EINVAL;
> +               goto out_abort;
> +
> +       }
> 
> Though I am making this cache_invalidate_user optional here, I
> wonder if there actually could be a case that a user-managed
> domain doesn't need a cache_invalidate_user op...

If user-managed domain is the stage-1 domain in nested, then seems not
possible as cache invalidate is a must. But I think this logic is fine as not
all the domains allocated by the user is user-managed. It may be kernel
managed like the s2 domains.

Regards,
Yi Liu