Re: [PATCH 4/4] iommufd: Simplify the lifecycle of a hwpt

[Jason Gunthorpe <jgg@nvidia.com>]

On Tue, Feb 14, 2023 at 06:14:07AM +0000, Tian, Kevin wrote:
> > From: Jason Gunthorpe <jgg@nvidia.com>
> > Sent: Tuesday, February 14, 2023 2:03 AM
> > 
> > The HWPT is connected to the ioas when it is finished being created and
> > should remain connected until it is destroyed.
> > 
> > Instead of trying to disconnect the HWPT manually during
> > iommufd_device_detach() simply use the object reference counting
> > mechanism
> > and free it when the refcount reaches 0.
> > 
> > This means the ioas->hwpt_list can have an item with a 0 refcount during
> > races, a prior patch made sure this is safe.
> > 
> > Add a new API iommufd_hw_pagetable_finalize() to formalize this contract.
> 
> Can you add a few words how you expect this to be called in manually
> created hwpt case? the new API requires it to be called after
> iopt_table_add_domain() which is part of iommufd_device_do_attach()
> now. Do you intend to move iopt_table_add_domain() into hwpt
> creation path so this finalize API is also called there or leave it as today
> which means again need to track first device in the group to finalize?

It is all still pretty messy, let me try again.

Jason