On Fri, Sep 09, 2022 at 05:00:16AM +0000, Tian, Kevin wrote:
> > I have started this effort by combining this list and the one from
> > the side thread:
> >
> > @@ -266,6 +266,13 @@ struct iommu_ops {
> > /**
> > * struct iommu_domain_ops - domain specific operations
> > * @attach_dev: attach an iommu domain to a device
> > + * Rules of its return errno:
> > + * ENOMEM - Out of memory
> > + * EINVAL - Device and domain are incompatible
> > + * EBUSY - Device is attached to a domain and cannot be changed
>
> With this definition then probably @attach_dev should not return -EBUSY
> at all given it's already checked in the start of __iommu_attach_group():
I think the EBUSY would be only for non-conforming drivers. The API
semantic is you can always attach a new domain and replace an existing
domain.
So things like AMD's "can't do anything but idenitity on RID when
PASID enabled" would be -EBUSY.
Seems right that it should be rare though.
> > + * ENODEV - Device or domain is messed up: device is not mapped
> > + * to an IOMMU, no domain can attach, and etc.
>
> if domain is messed up then should return -EINVAL given using another domain
> might just work. IMHO here -ENODEV should only cover device specific problems
> preventing this device from being attached to by any domain.
Agree
> > + * <others> - Same behavior as ENODEV, use is discouraged
>
> didn't get the "Same behavior" part. Does it suggest all other errnos should
> be converted to ENODEV?
It says all other errnos should be treated as ENODEV by the caller but
forwarded to userspace for further detail.
> btw what about -ENOSPC? It's sane to allocate some resource in the attach
> path while the resource might be not available, e.g.:
Seems resaonable that it is similar to ENOMEM
> As discussed in a side thread a note might be added to exempt calling
> kAPI outside of the iommu driver.
Sadly, not really.. The driver is responsible to santize this if it is
relevant. It is the main downside of this approach.
Jason
