Re: [RFC] iommufd: Use accurate dev_id in the PRI forwarding path

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

On 2025/3/6 11:10, Tian, Kevin wrote:
>> From: Jason Gunthorpe <jgg@nvidia.com>
>> Sent: Thursday, March 6, 2025 1:19 AM
>>
>> On Wed, Mar 05, 2025 at 05:03:56AM -0800, Yi Liu wrote:
>>> Hi Jason, Kevin, Baolu,
>>>
>>> This is more a query, the patch here does not really work as there
>>> is locking issue. So I'd like to hear from you if any good idea.
>>>
>>> Detail as below:
>>>
>>> Existing code has a problem when the PRI happens on a device that has
>>> alias. The PRI reporting path uses the idev stored in the handle. While
>>> this idev is the first idev that attachs to the domain.  If the PRI
>>> happens on devices other than the first attached device, then the idev
>>> stored in handle does not match.
>>
>> Do you actually care about this? Can we say that devices that have
>> aliases do not allow PRI?

It would be simpler if we can enforce it.

>> I'm confused about how that can even work if the RID is erased in the
>> fabric, how do we deliver the PRI response to the correct entity?
>>

If the RID is detached, kernel should response any pending PRIs. If user
sends a response, it won't suit any pending PRI then dropped.

> 
> Another tricky thing is when some devices in the group support PRI
> but others don't...
> 
> I'm inclined to agree that it's simpler to disable PRI on non-singleton
> group, than trying to fix and eanble it.

yes. This also tricky.

-- 
Regards,
Yi Liu