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

[Baolu Lu <baolu.lu@linux.intel.com>]

On 3/7/25 03:06, Jason Gunthorpe wrote:
> On Thu, Mar 06, 2025 at 03:09:34PM +0800, Yi Liu wrote:
>> +bool iommu_group_is_singular(struct iommu_group *group)
>> +{
>> +	bool res;
>> +
>> +	mutex_lock(&group->mutex);
>> +	res = list_count_nodes(&group->devices) == 1;
>> +	mutex_unlock(&group->mutex);
>> +
>> +	return res;
>> +}
>> +EXPORT_SYMBOL_NS_GPL(iommu_group_is_singular, "IOMMUFD_INTERNAL");
> I'm not so excited about something racy like this..
> 
> I still think this is a bit wrong, we only care about aliases, not
> multi device groups.
> 
> Aliases are created here:
> 
> 	if (pci_for_each_dma_alias(pdev, get_pci_alias_or_group, &data))
> 		return data.group;
> 
> And maybe in a few other places..
> 
> Can you tag the group that it has aliases in it directly during group
> creation so there are no races?
> 
> I've long had a desire to more clearly separate the two cases of
> groups:
> 
>   1) The HW cannot tell the different struct device sources on a per
>      DMA basis
>   2) There is a security intertwinement between the devices but they
>      can all run separately in the IOMMU HW
> 
> Maybe it is not so hard?

This is a long-standing issue. If I recall correctly, Joerg once
proposed the concept of iommu subgroup.

Thanks,
baolu