Re: [PATCH v2 02/10] iommu: Introduce a new iommu_group_replace_domain() API

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Feb 15, 2023 at 06:10:47AM +0000, Tian, Kevin wrote:
> > From: Nicolin Chen <nicolinc@nvidia.com>
> > Sent: Wednesday, February 8, 2023 5:18 AM
> > 
> > +int iommu_group_replace_domain(struct iommu_group *group,
> > +			       struct iommu_domain *new_domain)
> > +{
> > +	int ret;
> > +
> > +	if (!new_domain)
> > +		return -EINVAL;
> > +
> > +	mutex_lock(&group->mutex);
> > +	ret = __iommu_group_set_domain(group, new_domain);
> > +	if (ret)
> > +		__iommu_group_set_domain(group, group->domain);
> 
> Just realize the error unwind is a nop given below:
> 
> __iommu_group_set_domain()
> {
> 	if (group->domain == new_domain)
> 		return 0;
> 
> 	...
> 
> There was an attempt [1] to fix error unwind in iommu_attach_group(), by
> temporarily set group->domain to NULL before calling set_domain().
> 
> Jason, I wonder why this recovering cannot be done in
> __iommu_group_set_domain() directly, e.g.:
> 
> 	ret = __iommu_group_for_each_dev(group, new_domain,
> 					  iommu_group_do_attach_device);
> 	if (ret) {
> 		__iommu_group_for_each_dev(group, group->domain,
> 					  iommu_group_do_attach_device);
> 		return ret;
> 	}
> 	group->domain = new_domain;

We talked about this already, some times this is not the correct
recovery case, eg if we are going to a blocking domain we need to drop
all references to the prior domain, not put them back.

Failures are WARN_ON events not error recovery.

Jason