Re: [PATCH 1/2] iommu: Pass domain to remove_dev_pasid() op

[Yi Liu <yi.l.liu@intel.com>]

On 2024/3/27 21:02, Jason Gunthorpe wrote:
> On Wed, Mar 27, 2024 at 05:54:32AM -0700, Yi Liu wrote:
>> Existing remove_dev_pasid() callbacks of the underlying iommu drivers get
>> the attached domain from the group->pasid_array. However, the domains
>> stored in group->pasid_array are not always correct. For example, the
>> set_dev_pasid() path updates group->pasid_array first and then invoke
>> remove_dev_pasid() callback when error happened. The remove_dev_pasid()
>> callback would get the updated domain. This is not correct for the
>> devices that are still attached with an old domain or just no attached
>> domain.
>>
>> To avoid the above problem, passing the attached domain to the
>> remove_dev_pasid() callback is more reliable.
> 
> I've relaized we have the same issue with set_dev_pasid, there is no
> way for the driver to get the old domain since the xarray was updated
> before calling set_dev_pasid. This is unlike the RID path. Meaning
> drivers can't implement PASID replace.
> 
> So we need another patch to pass the old domain into set_dev_pasid
> too...

yes. you've given this suggestion in below. :) I've already in this way.
Will send it out together with other iommufd pasid attach/replace/detach
patches.

https://lore.kernel.org/linux-iommu/20240318165247.GD5825@nvidia.com/

> 
> This looks fine
> 
> Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

thanks.

-- 
Regards,
Yi Liu