Re: [PATCH rc 1/3] iommufd/selftest: Do not try to destroy an access once it is attached

[Jason Gunthorpe <jgg@nvidia.com>]

On Tue, Jul 25, 2023 at 02:45:24PM -0700, Nicolin Chen wrote:
> On Tue, Jul 25, 2023 at 04:05:48PM -0300, Jason Gunthorpe wrote:
> > The access must be detached first.
> >
> > To make the cleanup simpler copy the fdno to userspace before creating the
> > access in the first place. Then there is no need to unwind after
> > iommufd_access_attach.
> >
> > Fixes: 54b47585db66 ("iommufd: Create access in vfio_iommufd_emulated_bind()")
> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
> 
> Hmm, I was expecting that the iopt_remove_access() call in the
> iommufd_access_destroy_object() could "detach" the access.

Hmm, it does yes

> If calling iopt_remove_access() isn't enough, it means that we'd
> need the full routine from the iommufd_access_detach() in cdev
> series, i.e. we are missing the unmap part?

At least that part is not missing the unmap because the close_device()
immediately before is supposed to be an implicit unmap for the driver.
 
> In that case, though this patch can fix the issue in selftest,
> yet does the emulated pathway potentially have the same issue?

It seems not, and this patch is more a cosmetic change than any
functional issue, but given that the vfio flows rely on this working
we should keep it here to test it.

So, lets drop this patch.

Jason